www.allitebooks.

com
3D Integration in VLSI
Circuits

www.allitebooks.com
 Devices, Circuits, and Systems
Series Editor
Krzysztof Iniewski

3D Integration in VLSI Circuits

Implementation Technologies and Applications
Katsuyuki Sakuma
Low Power Semiconductor Devices and Processes for Emerging
Applications in Communications, Computing, and Sensing
Sumeet Walia and Krzysztof Iniewski
Biomaterials and Immune Response
Complications, Mechanisms and Immunomodulation
Nihal Engin Vrana
Low Power Circuits for Emerging Applications in
Communications, Computing, and Sensing
Krzysztof Iniewski and Fei Yuan
High-Speed and Lower Power Technologies
Electronics and Photonics
Jung Han Choi and Krzysztof Iniewski
X-Ray Diffraction Imaging
Technology and Applications
Joel Greenberg and Krzysztof Iniewski
Compressed Sensing for Engineers
Angshul Majumdar
IoT and Low-Power Wireless: Circuits, Architectures,
and Techniques
Christopher Siu and Krzysztof Iniewski
Sensors for Diagnostics and Monitoring
Kevin Yallup and Laura Basiricò
Energy Efficient Computing: Devices, Circuits, and Systems
Santosh K. Kurinec and Sumeet Walia
Radio Frequency Integrated Circuit Design
Sebastian Magierowski
For more information about this series, please visit:
https://www.crcpress.com/Devices-Circuits-and-Systems/
book-series/CRCDEVCIRSYS

www.allitebooks.com
3D Integration in VLSI
Circuits
Implementation Technologies
and Applications

Edited by
Katsuyuki Sakuma
Managing Editor
Krzysztof Iniewski

www.allitebooks.com
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742

© 2018 by Taylor & Francis Group, LLC

CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Printed on acid-free paper

International Standard Book Number-13: 978-1-1387-1039-9 (Hardback)

This book contains information obtained from authentic and highly regarded sources. Reasonable efforts
have been made to publish reliable data and information, but the author and publisher cannot assume
responsibility for the validity of all materials or the consequences of their use. The authors and publishers
have attempted to trace the copyright holders of all material reproduced in this publication and apologize
to copyright holders if permission to publish in this form has not been obtained. If any copyright material
has not been acknowledged please write and let us know so we may rectify in any future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter
invented, including photocopying, microfilming, and recording, or in any information storage or retrieval
system, without written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access www.copyright.
com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood
Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that provides licenses and
registration for a variety of users. For organizations that have been granted a photocopy license by the
CCC, a separate system of payment has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are
used only for identification and explanation without intent to infringe.

Library of Congress Cataloging‑in‑Publication Data

Names: Sakuma, Katsuyuki, author.

Title: 3D integration in VLSI circuits : implementation technologies and
applications / [edited by] Katsuyuki Sakuma.
Description: Boca Raton, FL : CRC Press/Taylor & Francis Group, 2018. |
Series: Devices, circuits, & systems | Includes bibliographical references
and index.
Identifiers: LCCN 2018010530| ISBN 9781138710399 (hardback : acid-free paper)
| ISBN 9781315200699 (ebook)
Subjects: LCSH: Three-dimensional integrated circuits. | Integrated
circuits--Very large scale integration.
Classification: LCC TK7874.893 .A16 2018 | DDC 621.39/5--dc23
LC record available at https://lccn.loc.gov/2018010530

Visit the Taylor & Francis Web site at

http://www.taylorandfrancis.com

and the CRC Press Web site at

http://www.crcpress.com

www.allitebooks.com
Contents

Preface .................................................................................................................... vii

Series Editor ............................................................................................................xi
Editor..................................................................................................................... xiii
Contributors ...........................................................................................................xv

1. Three-Dimensional Integration: Technology and Design ..................... 1

P. Franzon

2. Three-Dimensional System-in-Package for Application-Specific

Integrated Circuit and Three-Dimensional Dynamic
Random-Access Memory Integration ....................................................... 15
Li Li

3. A New Class of High-Capacity, Resource-Rich

Field-Programmable Gate Arrays Enabled by
Three- Dimensional Integration Chip-Stacked Silicon
Interconnect Technology ............................................................................. 41
Suresh Ramalingam, Henley Liu, Myongseob Kim, Boon Ang,
Woon-Seong Kwon, Tom Lee, Susan Wu, Jonathan Chang, Ephrem Wu,
Xin Wu, and Liam Madden

4. Challenges in 3D Integration ..................................................................... 71

M. Koyanagi, T. Fukushima, and T. Tanaka

5. Wafer-Level Three-Dimensional Integration Using Bumpless

Interconnects and Ultrathinning ..............................................................85
Takayuki Ohba

6. Three-Dimensional Integration Stacking Technologies

for High-Volume Manufacturing by Use of Wafer-Level
Oxide-Bonding Integration ...................................................................... 117
Spyridon Skordas, Katsuyuki Sakuma, Kevin Winstel, and
Chandrasekharan Kothandaraman

v
vi Contents

7. Toward Three-Dimensional High Density ........................................... 145

S. Cheramy, A. Jouve, C. Fenouillet-Beranger, P. Vivet, and
L. Di Cioccio

8. Novel Platforms and Applications Using Three-Dimensional

and Heterogeneous Integration Technologies ...................................... 185
Kuan-Neng Chen, Ting-Yang Yu, Yu-Chen Hu, and Cheng-Hsien Lu

Index ..................................................................................................................... 211

Preface

More and more products are using three-dimensional (3D) integration

technology nowadays. Through-silicon vias (TSVs) have been used in high
bandwidth memory (HBM) modules and will become mainstream for
other high-end products such as graphics processing units (GPU) and high-
performance computing (HPC), with applications in databases, security,
computational biology, molecular dynamics, deep learning, and automotive.
There is no doubt that 3D integration is gaining a significant attention as a
promising means to improve performance as it can provide higher intercon-
nect speeds, greater bandwidth, increased functionality, higher capacity, and
lower power dissipation.
Currently, the term 3D integration includes a wide variety of different
integration methods, such as 2.5-dimensional (2.5D) interposer-based inte-
gration, 3D-integrated circuits (3DICs), 3D systems-in-package (SiP), 3D het-
erogeneous integration, and monolithic 3D ICs. The goal of this book is to
provide readers with an understanding of the latest challenges and issues
in 3D integration. TSVs are not the only technology element needed for 3D
integration. There are numerous other key enabling technologies required
for 3D integration and the speed of the development in this emerging field
is very rapid. To provide readers with state-of-the-art information on 3D
integration research and technology developments, each chapter has been
contributed by some of the world’s leading scientists and experts from aca-
demia, research institutes, and industry from around the globe.
Chapter 1 by Prof. Franzon from North Carolina State University
(NCSU) provides a brief review of 2.5D and 3D technology options,
including interposers and TSV-stacking technologies. As illustrated by
successful commercial 3D products and experimental results of complemen-
tary metal–oxide–semiconductor (CMOS) stacks with a copper thermo-
compression-bonded interface, Chapter 1 discusses the reasons why 3D
integration is superior in terms of power efficiency, performance enhance-
ment, and cost reduction.
Chapter 2 by Dr. Li from Cisco Systems describes up-to-date manufactur-
ing technologies for 2.5D and 3D SiP for integrating application-specific inte-
grated circuits (ASICs) with multiple 3D dynamic random-access memory
(DRAM) stacks. In 2.5D integration, both ASIC chip and 3D DRAM stacks
are packaged in a planar format on an interposer. This type of integration
has emerged as another killer application for 3D integration. This chapter
also examines large-size high-density silicon/organic interposers, micro-
bump interconnects, warpage behaviors, packaging assembly processes for
ASIC and HBM integration, and board-level reliability tests, and character-
ization of 3D-integrated packages.

vii
viii Preface

Chapter 3 presents architecture, design, and technology implementations

for 3D field-programmable gate array (FPGA) integration and was written
by Dr. Ramalingam et al. from Xilinx, one of the major players in this area.
Multiple FPGA dies are placed side by side and interconnected on a large
silicon interposer. The design simulation methodology, reliability assess-
ment, and future challenges are discussed. On these topics, the authors
provide an industry perspective based on volume production of the largest
3D-integrated FPGA as of today, which contains 4.4  million logic cells,
600 thousand microbumps, and 19 billion transistors in a 55 mm package.
Chapter 4 by Prof. Koyanagi et al. from Tohoku University, Japan, covers
various unique 3D system-on-chips technologies, such as 3D integration
using self-assembly and electrostatic bonding, TSV formation based on
directed self-assembly with nanocomposites, and hybrid bonding technol-
ogy using Cu nano-pillar. Technologies for a 3D-integrated CMOS image
sensor module for a driver assistance system are presented, and future 3D
integration challenges are also discussed.
Chapters 5 and 6 focus on fabrication approaches for 300 mm wafer-level
3D integration without microbumps. For high-volume manufacturing,
TSVs are formed after a wafer-bonding process. The leakage current and
electrical resistance of TSVs, Ion/Ioff characteristics for field-effect transistor
(FET) devices, and the characteristics of wafer-bonding technologies are also
discussed. Chapter 5 by Prof. Ohba from the Tokyo Institute of Technology,
Japan, deals with the 3D integrations of permanently adhesive-bonded ultra-
thin wafers. In Chapter 6, Dr. Skordas et al. from IBM discussed the 3D inte-
gration technology based on low-temperature oxide-bonding for integrating
high-performance POWER7™ 45  nm silicon-on-insulator (SOI)-embedded
DRAM.
Chapter 7 by Ms. Cheramy et al. from CEA-Leti, France, provides the
principles, process integration, and detailed overviews of both monolithic
3D ICs (CoolCube™) and Cu/SiO2 hybrid bonding technologies. Monolithic
3D ICs enable the stacking of multiple transistor layers in the third dimension,
with a vertical interconnect pitch in the range of a few tens of nanometers,
and the bottom layer can be any CMOS type, be it bulk planar FET, FinFET,
or fully-depleted silicon-on-insulator (FDSOI). Potentially, it will be a key
technology driver for the next generation of 3D integration. The Cu/SiO2
hybrid bonding technology enables wafer-to-wafer and die-to-wafer connec-
tivity with a vertical interconnect pitch in the range of a few micrometers. This
chapter also addresses the issue of thermal dissipation in 3D integration.
Chapter 8 by Prof. Chen et al. from National Chiao Tung University, Taiwan,
offers examples of novel platforms and application demonstrations, includ-
ing terahertz (THz) optical components, piezoresistive pressure sensors, and
flexible neural sensing biosensors using 3D integration technologies. This
illustrates that 3D integration technology can be used in a wide variety of
applications and that it will open a new era of electronics and sensors that
cannot be achieved with conventional 2D microelectronics.
Preface ix

I would like to sincerely thank all of the authors for their hard work and
commitment. Without their contributions, it would not have been possible to
provide an up-to-date review of these innovative technologies and the chal-
lenges in 3D integration. It is my hope that this book will provide readers
with a timely and comprehensive view of current 3D integration technology.

Katsuyuki Sakuma
Yorktown Heights, New York
Series Editor

Krzysztof (Kris) Iniewski is managing R&D at

Redlen Technologies Inc., a startup company
in Vancouver, Canada. Redlen’s revolutionary
production process for advanced semiconduc-
tor materials enables a new generation of more
accurate, alldigital, radiation-based imaging
solutions. Kris is also a founder of ET CMOS
Inc. (http:// www.etcmos.com), an organization
of high-tech events covering communications,
microsystems, optoelectronics, and sensors. In
his career, Dr. Iniewski held numerous faculty
and management positions at University of Toronto (Toronto, Canada),
University of Alberta (Edmonton, Canada), Simon Fraser University (SFU,
Burnaby, Canada), and PMC-Sierra Inc (Vancouver, Canada). He has pub-
lished more than 100 research papers in international journals and con-
ferences. He holds 18 international patents granted in the United States,
Canada, France, Germany, and Japan. He is a frequently invited speaker and
has consulted for multiple organizations internationally. He has written and
edited several books for CRC Press (Taylor & Francis Group), Cambridge
University Press, IEEE Press, Wiley, McGraw-Hill, Artech House, and
Springer. His personal goal is to contribute to healthy living and sustain-
ability through innovative engineering solutions. In his leisurely time, Kris
can be found hiking, sailing, skiing, or biking in beautiful British Columbia.
He can be reached at [email protected].

xi
Editor

Katsuyuki Sakuma is a research staff member

at the IBM T.J. Watson Research Center. He has
over 19  years of experience of researching
3D  integration technologies and performing
various semiconductor packaging research and
development projects. His research interests
include 3D integration technologies, bonding
technologies, advanced packaging, and bio-
medical sensors.
He has published more than 85 peer-reviewed
journal papers and conference proceeding
papers, including three book chapters in the
semiconductor and electronic packaging area.
He also holds over 35 issued or pending U.S. and international patents.
He  has been recognized with the IBM Eleventh Invention Achievement
Award in 2017 and an Outstanding Technical Achievement Award (OTAA)
in 2015 for his contribution and leadership in the area of 3D integration
technology development. He was also given the 2018 Exceptional Technical
Achievement Award from the IEEE Electronics Packaging Society, and
the 2017 Alumni Achievement Award from his Alma Mater, the School of
Engineering at Tohoku University, for his contribution to 3D chip stack tech-
nology development in the electronics packaging industry. He was a core-
cipient of the IEEE Components, Packaging, and Manufacturing Technology
(CPMT) Japan Society Best Presentation Award in 2012, and the IMAPS
“Best of Track” Outstanding Paper Award in 2015.
Dr. Sakuma received his B.E. and M. Eng. degrees from Tohoku University
and the Ph.D. degree from Waseda University, Japan. He is currently serving as
an associate editor for IEEE Transactions on CPMT. He served as an associate
editor of the Institute of Electronics, Information and Communication
Engineers (IEICE, Japan) from 2003 until 2005. He has served as commit-
tee member of the IEEE ECTC Interconnections subcommittee since 2012,
for the IEEE International Conference on 3D System Integration (IEEE 3DIC)
since  2016, and for the IEEE International Reliability Physics Symposium
(IEEE IRPS) since 2017. He has been a senior member of IEEE since 2012.

xiii
Contributors

Boon Ang Yu-Chen Hu

Xilinx, Inc. Department of Electronics
San Jose, California Engineering
National Chiao Tung University
Jonathan Chang Hsinchu City, Taiwan
Xilinx, Inc.
San Jose, California
A. Jouve
CEA-Leti
Kuan-Neng Chen
Grenoble, France
Department of Electronics
Engineering
National Chiao Tung University Myongseob Kim
Hsinchu City, Taiwan Xilinx, Inc.
San Jose, California
S. Cheramy
CEA-Leti
Chandrasekharan Kothandaraman
Grenoble, France
IBM T.J. Watson Research Center
Yorktown Heights
L. Di Cioccio
Albany, New York
CEA-Leti
Grenoble, France
M. Koyanagi
C. Fenouillet-Beranger New Industry Creation Hatchery
CEA-Leti Center (NICHe)
Grenoble, France Tohoku University
Sendai, Japan
P. Franzon
North Carolina State University
Woon-Seong Kwon
Raleigh, North Carolina
Google LLC
Mountain View, California
T. Fukushima
Department of Mechanical Systems
Engineering Tom Lee
Tohoku University Xilinx, Inc.
Sendai, Japan San Jose, California

xv
xvi Contributors

Li Li Spyridon Skordas
Cisco Systems, Inc. IBM Research Division
San Jose, California Albany, New York

Henley Liu T. Tanaka

Xilinx, Inc. Department of Biomedical
San Jose, California Engineering
Tohoku University
Cheng-Hsien Lu Sendai, Japan
Department of Electronics
Engineering P. Vivet
National Chiao Tung University CEA-Leti
Hsinchu City, Taiwan Grenoble, France

Kevin Winstel
Liam Madden IBM Research Division
Xilinx, Inc. Albany, New York
San Jose, California
Ephrem Wu
Takayuki Ohba Xilinx, Inc.
Laboratory for Future San Jose, California
Interdisciplinary Research of
Science and Technology Susan Wu
Tokyo Institute of Technology Xilinx, Inc.
Tokyo, Japan San Jose, California

Xin Wu
Suresh Ramalingam Xilinx, Inc.
Xilinx, Inc. San Jose, California
San Jose, California
Ting-Yang Yu
Katsuyuki Sakuma Department of Electronics
IBM T.J. Watson Research Center Engineering
Yorktown Heights National Chiao Tung University
Albany, New York Hsinchu City, Taiwan
1
Three-Dimensional Integration:
Technology and Design

P. Franzon

CONTENTS
1.1 Introduction .................................................................................................... 1
1.2 Three-Dimensional Integrated Circuit Technology Set ...........................2
1.3 Three-Dimensional Drivers..........................................................................5
1.4 Miniaturization .............................................................................................. 5
1.5 Cost Reduction ............................................................................................... 6
1.6 Heterogeneous Integration ...........................................................................7
1.7 Performance Enhancement ..........................................................................8
1.8 Power Efficiency ........................................................................................... 10
1.9 Conclusion .................................................................................................... 12
References............................................................................................................... 12

1.1 Introduction
3D and 2.5D integration technologies permit substantial improvement in
form factor, power, performance, functionality, and sometimes even cost.
Though not providing a direct replacement for Moore’s law, 3D technologies
can permit a generation or more of exponential scaling in power per unit of
performance and other factors.
This chapter is structured as follows. First there is a review of 3D technolo-
gies, followed by a general discussion of commercial 3D success stories and
technology drivers. Included in that we review 3D logic projects conducted
at North Carolina State University, United States, before closing out on het-
erogeneous integration.

1
2 3D Integration in VLSI Circuits

1.2 Three-Dimensional Integrated Circuit Technology Set

There are a variety of technologies that contribute to the 3D technology set.
The purpose of this chapter is to review these technologies and not to pro-
vide details of materials and manufacturing options. We will review inter-
posers, 3D stacking technologies and monolithic 3D technologies.
Figure 1.1 shows an interposer, often referred as a multi-chip module in
the past. Interposers get its name from their functionality of being placed
between the chip and the package. There are two standard ways to make
interposers. One approach is to use multiple thin film layers of spun-on
dielectrics and metal patterning. While permitting thick metals, it does not
permit high density. High density can be achieved by leveraging a legacy
back-end-of-line (BEOL) process from semiconductor chip manufacturing.
A BEOL process permits multiple levels of planarized wiring, typically 4–6
layers.
A defining technology in 3D integration is the through-silicon via (TSV). The
TSV goes through the silicon substrate, connecting the front and back sides
of the structure. It is created by etching a near vertical hole, lining it with a
dielectric before filling it with metal. The wafer is then thinned to expose
the metal. Typical dimensions for interposer structures are 100 μm of thick-
ness for the overall structure and 1–10 μm width and space for the metals.
TSVs are typically on at least a 100 μm pitch. The microbumps are typically
around a 40 μm pitch, whereas the bumps to the package must be on conven-
tional scales at 150+ μm pitch.
Figure 1.2 provides an illustration for some of the key three-dimensional
integrated circuit (3DIC) stacking technologies. This shows a three-chip
stack, two of which incorporate TSVs. The top two chips are joined with a
face-to-face technology (F2F). Here face refers to the top side of the chip—
where the transistors and interconnect are. Several methods are used to cre-
ate high-density F2F connections. Microsolder bumps can be used but today

Microbump Oxide
Wiring layers
Through-silicon vias (TSVs)
Bumps Silicon

Interposer

FIGURE 1.1
Interposer (2.5D) technology.
3D Integration: Technology and Design 3

Microbumps or direct bond

Transistor/wiring layers
Through-silicon vias (TSVs)
Microbumps to package or
interposer

FIGURE 1.2
3D stacking technologies.

they are limited to 40 μm pitch (with potential for 25 μm). Copper–copper
thermo-compression can be used down to sub-5 μm pitches. Alternatively,
hybrid bonding can be used. In hybrid bonding the top surface, with metal
plugs in it, is planarized and then bonded to another such surface. An exam-
ple is the Ziptronix data-based individualization (DBI) process [1]. This pitch
can be built down to 1 μm pitch but 6–8 μm are more typical. Hybrid bond-
ing is used to make many cell phone cameras today (Please see Section 1.6
on heterogeneous integration). With such high interconnect densities, many
interesting architectures can be explored, as will be explained later.
The bottom two chips in Figure 1.2 are joined using a face-to-back (F2B)
arrangement in which the bottom (back) of one die is joined to the top ( face) of
another. TSVs are needed to provide the connection to the joining backside.
As the sidewalls of the TSV are not entirely vertical, the TSV pitch is limited
to approximately the thickness of the wafer, typically 10–25 μm. Thus a F2B
connection provides a lot of less density than a F2F connection.
TSVs can be inserted before, during, or after complete wafer processing.
These are referred to, respectively, as via early, middle, and last. Figure 1.3
shows a process via middle. Wafers are partially completed, say to metal 1.
A vertical side-wall via is etched part way through the wafer. This via is
created using the Bosch process [2]. In the Bosch process, deep reactive-ion
etching (DRIE) is alternated with a deposition step multiple times to cre-
ate a near vertical wall via. Today the wall steepness is typically 10:1. Thus
the via depth has to be less than 10× the diameter of the opening at the
top. The via sidewalls are passivated, typically with an oxide, and then filled
with a metal, typically tungsten or copper. The chip metal stack (BEOL) is
then completed. The wafers are then flipped and thinned, exposing the TSV
metal. The exposed metal can then be used directly for a joining process, or a
bump structure is added before joining with another wafer or chip to create
a 3D stack.
4 3D Integration in VLSI Circuits

Transistor/early
wiring layers
Bulk silicon

Via creation Sidewall

passivation

Metal fill Completion of Completion of

BEOL processing BEOL processing

Microbumping

Stacking

FIGURE 1.3
Through-silicon vias (TSVs) fabrication steps.

Note in Figure 1.2 that the top chip has not been thinned. Usually a 3D
chip stack is assembled with the chip in wafer form. Two wafers are attached
with the second left unthinned. This wafer stack can be attached to another
(thick) wafer for further handling. In most cases, one wafer is left in a thick
format so that the wafer stack can be easily handled. Thin wafer handling is
possible but increases the cost.
Most 3DICs are assembled in wafer format. Again, the driver is cost.
Wafer-to-wafer attachment processing costs less per die and has higher
yields. Although chip to wafer attachment is possible it is not widely used.
An intriguing complement to 3D chip stacking technology is monolithic
3D in which there is only one silicon substrate (and thus no TSVs). The most
commercially successful one of these technologies is 3D NAND Flash in
which the string of NAND transistors in a nonvolatile flash device is fabri-
cated vertically. This approach brings substantial density improvements and
cost savings over conventional NAND flash technology. However, to date it
has not been commercially applied to any other logic or memory structure.
Another monolithic (-like) 3D technology set involves techniques in which
silicon-on-insulator (SOI) wafers are the starting wafer. This is shown in
Figure 1.4. In this approach, fabricated wafers are joined face to face using an
oxide–oxide bond. As the transistors are built on the top of an oxide layer, a
silicon-selective back etch can be used to remove the silicon part of the SOI
substrate while not affecting the transistors and interconnect layers. Simple
3D Integration: Technology and Design 5

BEOL oxide
Epi
Metal
transistors Buried oxide
Bulk silicon

4. Repeat

1. Oxide–oxide bond 2. Silicon etch 3. Via formation

FIGURE 1.4
3DIC chip stack using SOI wafers.

through-oxide vias can then be used to create vertical connections between

what were previously separate chips. An example of this process can be
found in Reference 3. If the first two chips in the stack are fabricated without
interconnect, then one gets two directly connectable transistor layers in what
would be considered a monolithic 3D technology. This is monolithic in the
sense there is only one bulk wafer left at the end of processing. However,
many would not consider it to be truly monolithic as multiple separate
wafers have to be created first.

1.3 Three-Dimensional Drivers

In general, the decision to employ 3D technologies is driven by (1) the need
for miniaturization, (2) the desire for cost reduction, (3) heterogeneous inte-
gration, (4) the need for performance enhancement, and/or (5) the need for
improvement in performance/power. Often several of these go hand in hand.

1.4 Miniaturization
An early application of TSVs was providing the I/O connections cell phone
camera front-side imaging sensor [4,5]. The goal was not to leverage 3D chip
stacks—these were single die—but to reduce the overall sensor height, at
least when compared with conventional packaging approaches.
6 3D Integration in VLSI Circuits

FIGURE 1.5
Miniaturized sensor as a two-chip stack.

3D chip stacking can be used to make such sensors with low integrated
volume. Though fabricated using wire bonding, Chen et  al. demonstrated
an integrated power-harvesting data-collecting sensor with the photovoltaic
power-harvesting chip mounted on top of the logic and RF chips [6]. This
maximizes the photovoltaic power-harvesting area while minimizing the
volume. TSVs and bonding technologies would permit further volume reduc-
tion. Lentiro [7] described a two-chip stack aimed at simulating a particle of
meat for the purposes of calibrating a new food processing system. One chip
is a radio-frequency identification (RFID) power harvester and communica-
tions chip, the second is the temperature data logger. It is a two-chip stack
with F2F connections and TSV-enabled I/O. It is integrated with a small bat-
tery for data collection purposes only as the RFID cannot be employed in
the actual processing pipes. The two-chip stack permits smaller imitation
food particles than otherwise would be the case and is shown in Figure 1.5.
The RFID coil can be seen. The chip includes capacitors for temporary power
storage.

1.5 Cost Reduction

There have been several examples of 3DIC products with a focus on
cost reduction, though often they achieve performance enhancement
simultaneously.
Many cell phone cameras are today made as a two-chip stack. One chip is a
backside-illuminated pixel array that does not include interconnect layers or
even complete complementary metal–oxide–semiconductor (CMOS) transis-
tors. The second chip is a complete CMOS chip on which the analog-to-digital
converters (ADC) are built and interconnected for all other functionalities
required of an image sensor. A high-density connection process such as
hybrid bonding is used to create the needed face-to-face connections on a
3D Integration: Technology and Design 7

few micron pitch. Examples are reported in References 8 and 9. This is also
an example of heterogeneous integration, as the two chips in the stack go
through different manufacturing processes.
Another example of cost reduction is found in building high-end field-
programmable gate arrays (FPGAs). To a first approximation, the cost of
a large CMOS chip goes up with the square of the area. This is because
the probability of a defect occurring on the chip and thus killing the chip
goes up with the chip area, whereas the cost of making the chip in the first
place also goes up with the area. Thus it is worth considering partitioning a
large chip into a set of smaller ones, if the cost of integration and the addi-
tional test are less than the savings accrued to increase CMOS yield. Xilinx,
California, United States, investigated this concept for large FPGAs and is
now selling FPGA modules containing 2–4 CMOS FPGA chips, tightly inte-
grated on an interposer. Details are not available but they claim an overall
cost savings [10].
A third example is that of mixing technology nodes. In general, Moore’s
law tells us that a digital logic gate costs less to make in a more advanced
technology due to the reduced area for that gate in that node. However, in
contrast, many analog and analog-like functions such as ADC and high-
speed serial deserializer I/Os (SerDes) do not benefit in such a fashion. The
reason is that the analog behavior of a transistor has higher variation for
smaller transistors than for larger ones. Thus for many analog functions that
rely on well-matched behaviors of different transistors in the circuit, no ben-
efit is accrued from building smaller transistors. More simply put, analog
circuit blocks do not shrink in dimensions with the use of more advanced
technologies. Thus the cost of these functions in a more advanced process
node can actually be higher, than in the old node, as the old node costs less to
make per unit of area. Again this is an example of heterogeneous integration,
the heterogeneity being that of mixing technology notes.
Although Wu [10] also explored this concept generically, Erdmann et al.
[11] have explored this concretely for a mixed ADC/FPGA design. Their
design consisted of two 28 nm FPGA logic dies, integrated with two 65 nm
ADC array dies on an interposer. Thus two sets of cost benefits are accrued,
first the yield-related savings from splitting the logic die into two and the
fabrication cost savings of keeping the ADCs in an older technology.

1.6 Heterogeneous Integration

In Section 1.5, two examples of heterogeneous integration were given in
which wafers from different silicon process lines were integrated using 3DIC
technologies. There is also considerable interest in integrating wafers with
different underlying technologies, not just silicon.
8 3D Integration in VLSI Circuits

Another interesting use of 3D technologies has been to build nonvisible

light sensors, sometimes using a nonsilicon technology for the sensing layer.
Examples include IR imagers, X-ray imagers [12], and other images for high-
energy physics investigations [13].
The example is that of mixing III-V and silicon technologies. This is best
exemplified by The Defense Advanced Research Projects Agency Diverse
Accessible Heterogeneous Integration (DARPA DAHI) program in which
GaN and InP chips are integrated on top of the CMOS chips through micro-
bumps and other technologies [14]. More specifically, CMOS can be used
for most of the transistors in a circuit, whereas GaN high-electron-mobility
transistors (HEMTs) can be used for its high-power capability, and InP het-
erojunction bipolar transistors (HBTs) can be used for their very high speed.
An example of the latter is an ADC. In an ADC, only a few transistors gen-
erally determine the sampling rate. Thus with the DAHI technology, these
few transistors can be built in a high speed, but expensive and low yielding,
InP chiplet, whereas the rest of the ADC is built in cheaper and more robust
CMOS.

1.7 Performance Enhancement

A major goal of using 3D technologies is to increase the performance in some
key aspects. For example, a two-layer image sensor might enable bigger pix-
els in the image layer and thus greater sensitivity. Alternatively, integrating
a high-performance low device count InP chiplet onto a sophisticated CMOS
die leads to performance improvement, often in conjunction with the high
I/O count that the 3D technologies permit.
However, many 3D products and research objects are driven purely by
performance, specifically the ability to integrate many silicon parts, with a
lot of accumulated area into one 3D chip stack.
3D memories do exactly this and are positioned to be the next large-
volume application of 3DIC technologies. To date dynamic random-access
memory (DRAM) has relied on one-signal-per-pin signaling using low cost,
low pin count, and single-chip plastic packaging. As a result, DRAM has
continued to lag logic in terms of bandwidth potential and power efficiency.
Furthermore, the I/O speed of one-signal-per-pin signaling schemes are
unlikely to scale a lot beyond what can be achieved today in double data
rate (DDR4) (up to 3.2 Gbps per pin) and graphics double data rate (GDDR6)
(8 Gbps). Beyond these data rates, two-pins-per-signal differential signaling
is needed. Furthermore, the I/O power consumption, measured as pJ/bit, is
relatively high, even for the low-power DDR (LPDDR) standards (intended
for mobile applications).
3D Integration: Technology and Design 9

TABLE 1.1
Comparison of 2D and 3D Memories
Power Efficiency I/O Efficiency DQ
Technology Capacity BW (GB/s) (W) (mW/GB/s) (mW/Gb/s) Count

DDR4-2667 4 GB 21.34 6.6 309 6.5–39 32

LPDDR4 4 GB Up to 42 5.46 130 2.3 32
HMC 4 GB 128 GBps 11.08 86.5 10.8 8 Serdes
lanes
HBM 16 GB 256 GBps 48 1024
Wide I/O 8–32 GB 51.2 GBpsa 42b 256
DiRAM4 64 GB 8 Tbps 4096
a Wide I/O2, Wide I/O1 was half of this.
b Wide I/O1, Wide I/O2 should be lower.

There have been several proposed 3D memory technologies, which are

summarized together with 2D memories in Table 1.1. Note that in this table,
both B (Byte) and b (bits) are used. Also note that 1 mW/Gb/s is equivalent
to 1 pJ/bit.
The Hybrid Memory Cube (HMC) is a joint Intel-Micron standard that cen-
ters on a 3D-stacked part including a logic layer and multiple DRAM layers
organized as independent vertical slices. This 3D chip stack is then provided
as a packaged part, so the customer does not have to deal with any 3DIC or
2.5D packaging issues. At the time of writing this chapter, Micron offered
2 GB and 4 GB parts with a maximum memory bandwidth of up to 160 GBps
(The 128 GBps part is used in Table 1.2). The data I/O is organized as an eight
high-speed serial channels or lanes. HMC is mainly aimed at computing
applications. However, it does not appear to be widely adopted.
The high bandwidth memory (HBM) is a JEDEC (i.e., industry) stan-
dard that is not intended to be packaged on its own but to be further inte-
grated using an interposer or 3D chip stack. Its high pin count, configuring
8 × 128-bit wide interfaces, prevents it from being packaged easily. The inter-
faces run at 2 Gbps per pin and the pins are placed on a dense 48 × 55 μm
grid. It is fabricated as a stack of multibank memory die, connected to a logic

TABLE 1.2
Improvements in 3D Design over 2D Using Logic Cell Partitioning
Total Wire Length Fmax Total Power Power/
(% Change) (% Change) (% Change) MHz

Radar PE −21.0% +22.6% −12.9% −38%

Advanced encryption −8% +15.3% −2.6% −18%
standard (AES)
Multiple-input and +21.6% +17.1% −5.1% −23%
multiple-output (MIMO)
10 3D Integration in VLSI Circuits

die through a TSV array, the TSV arrays running through the chip centers.
Each chip is F2B mounted to the chip beneath it. The eight channels are oper-
ated independently. Details for a first-generation HBM (operating at 3.8 pJ/bit
power level at 128 GB/s) can be found in Reference 15. The use of HBM in
graphics module products has been announced by Nvidia and AMD.
Wide I/O is also a JEDEC-supported standard, aimed largely at low-power
mobile processors. Although intended to be mounted on top of the logic
die in a true 3D stack, side by integration on an interposer is also possible.
Wide I/O is a DRAM-only stack—there is no logic layer. Instead the DRAM
stack is exposed through a TSV-based interface and the memory controller
is designed separately on the CPU/logic die that is customer designed. To
date the thermal challenges of mounting a DRAM on an already hot mobile
processor logic die have been insurmountable, especially as it is desired to
operate the DRAM at a lower temperature than logic (85°C for DRAM vs.
105°C for logic) to control leakage and refresh time. This has been a barrier
to adoption.
The Tezzaron DiRAM4 is a proprietary memory still in development. It
has 4096 data I/O organized across 64 ports. It is intended only for 3D and
interposer integration. It has a unique organization in that the logic layer
is not only used for controller and I/O functions but also houses the global
sense amplifiers and addresses decoders that in other 3D memories are
on the DRAM layers. This permits faster operation for these circuits. The
DiRAM4 has potential for a very high bandwidth (up to 8  Tbps) and fast
random cycles (15 ns) [16].

1.8 Power Efficiency

In theory, taking a 2D chip and turning it into a stack of two chips should
reduce the power of the chip by about 35%. The reason is quite simple. In
most digital chips, about half (50%) the power is consumed in driving wires.
When stacked, the two chips in the stack should be half of the area of the
original chip, and thus the wires should be √2 shorter. As the power needed
to drive a wire is roughly proportional to its length, each wire driver should
see a power reduction of around 70%, that is, the chip as a whole should see
a power reduction of around 35%.
However, this is difficult to achieve as short wires are unlikely to be routed
between the two chips. On the other hand, reduced wire loads also means
the chip can operate faster. Thus one would still expect to see a substantial
improvement in the power/performance ratio over the 2D equivalent.
In one project at North Carolina State University, we demon-
strated this using a two-chip CMOS stack-attached F2F with a copper
thermo-compression-bonded interface at a 6.3  μm pitch. We put together
3D Integration: Technology and Design 11

Logic only

Logic, clocks,
flip-flops Die photo

FIGURE 1.6
Two-chip stack.

a CAD flow to do this that could reuse 2D CAD tools, especially place and
route tools. To make that feasible, all flip-flops are kept in one tier so that 3D
clock distribution was not required. The radar PE was implemented in the
Tezzaron bulk CMOS 3D process [17] (Figure 1.6). The results are summa-
rized in Table 1.2. On average, performance per unit of power was increased
by 22% due to the decreases in wire length achieved through this parti-
tioning approach. The radar processor had an improvement in performance
per unit of power of 21%. The other designs achieved 18% and 35%. The
achieved improvement was roughly equivalent to one generation of Moore’s
law scaling.
In a different project at North Carolina State University, we took a very differ-
ent approach to improving performance/power using 3D technologies. A stack
of two different CPUs are integrated vertically using a vertical thread transfer
bus that permits fast compute load migration from the high-performance CPU
to and from the low-power CPU when an energy advantage is found [18]. In this
design, the high-performance CPU can issue two instructions per cycle, whereas
the low-power CPU is a single-issue CPU. The transfer is managed using a low-
latency, self-testing multisynchronous bus [19]. The bus can transfer the state
of the CPU in one clock cycle by using a wide interface and by exploiting a
high-density copper–copper direct bond process. The caches are switched at
the same time, removing the need for a cold cache restart.
Simulation with Specmark workloads shows a 25% improvement in the
power/performance ratio compared with executing the sample workload
solely in the high-performance processor. In contrast, if the workload was
executed solely in the single-issue (low-power) CPU, there would be 28%
total energy savings, compared with keeping the workload in the high-
performance CPU but at the expense of a 39% reduction in performance. If
the workload was allowed to switch every 10,000 cycles, there would be 27%
total energy savings but at the expense of only a 7% reduction in performance
that is, a 25% improvement in power per unit of performance is achieved. The
die photograph is shown in Figure 1.7.
12 3D Integration in VLSI Circuits

FIGURE 1.7
3D heterogeneous processor die.

1.9 Conclusion
3D and 2.5D technologies considerably open the design space for semicon-
ductor technologies. Dimensions for exploration include miniaturization,
cost reduction, achieving new modalities via heterogeneous integration, per-
formance improvement, and improvement in performance/power.

References
1. P. Enquist, G. Fountain, C. Petteway, A. Hollingsworth, and H. Grady, Low cost
of ownership scalable copper direct bond interconnect 3D IC technology for
three dimensional integrated circuit applications, IEEE International Conference
on 3D System Integration, 3DIC 2009, San Francisco, CA, 2009, pp. 1–6.
2. U.S. Patents 5,501,893 and 6,531,068.
3. J.A. Burns, B.F. Aull, C.K. Chen, C. Chang-Lee, C.L. Keast, J.M. Knecht,
V. Suntharalingam, K. Warner, P.W. Wyatt, and D. Yost, A wafer-scale 3-D cir-
cuit integration technology, IEEE Transactions on Electron Devices, 52(10): 2507–
2516, 2006.
4. http://image-sensors-world.blogspot.com/2008/09/toshiba-tsv-reverse-
engineered.html
5. http://www.semicontaiwan.org/en/sites/semicontaiwan.org/files/docs/
4._mkt__jerome__yole.pdf
6. G. Chen, M. Fojtik, D. Kim, D. Fick, J. Park, M. Seok, M.-T. Chen, Z. Foo,
D. Sylvester, and D. Blaauw, Millimeter-scale nearly perpetual sensor system
with stacked battery and solar cells, 2010 IEEE International Solid-State Circuits
Conference-(ISSCC), San Francisco, CA, 2010, pp. 288–289.
7. A. Lentiro, Low-density, ultralow-power and smart radio frequency telemetry
sensor, PhD Dissertation, NCSU, 2013.
3D Integration: Technology and Design 13

8. P. Enquist, 3D integration applications for low temperature direct bond tech-

nology, 2014 4th IEEE International Workshop on Low Temperature Bonding for 3D
Integration (LTB-3D), Tokyo, Japan, 2014, p. 8.
9. http://www.sony.net/SonyInfo/News/Press/201201/12-009E/index.html,
http://www.reuters.com/article/2014/03/25/us-sony-sensors-idUSBREA
2O0PQ20140325
10. X. Wu, 3D-IC technologies and 3D FPGA, 3D Systems Integration Conference
(3DIC), 2015 International, Sendai, Japan, 2015, pp. KN1.1–KN1.4.
11. C. Erdmann et al., A heterogeneous 3D-IC consisting of two 28 nm FPGA die
and 32 reconfigurable high-performance data converters, IEEE Journal of Solid-
State Circuits, 50(1): 258–269, 2015.
12. G.W. Deptuch et  al., Fully 3-D integrated pixel detectors for X-rays, IEEE
Transactions on Electron Devices, 63(1): 205–214, 2016.
13. http://meroli.web.cern.ch/meroli/DesignMonoliticDetectorIC.html
14. S. Raman, C.L. Dohrman, and T.H. Chang, The DARPA diverse accessible het-
erogeneous integration (DAHI) program: Convergence of compound semi-
conductor devices and silicon-enabled architectures, 2012 IEEE International
Symposium on Radio-Frequency Integration Technology (RFIT), Singapore, 2012,
pp. 1–6.
15. D.U. Lee et  al., A 1.2  V 8  Gb 8-channel 128  GB/s high-bandwidth memory
(HBM) stacked DRAM with effective I/O test circuits, IEEE Journal of Solid-State
Circuits, 50(1): 191–203, 2015.
16. www.tezzaron.com and Evolving 2.5D and 3D integration, RTI 3D ASIP,
December 2013.
17. T. Thorolfsson, S. Lipa, and P.D. Franzon, A 10.35  mW/GFLOP stacked SAR
DSP unit using fine-grain partitioned 3D integration, Proceedings in Custom
Integrated Circuits Conference 2012, pp. 1–4.
18. E. Rotenberg, B.H. Dwiel, E. Forbes, Z. Zhang, R. Widialaksono, R. Basu Roy
Chowdhury, N. Tshibangu, S. Lipa, W.R. Davis, and P.D. Franzon, Rationale for
a 3D heterogeneous multi-core processor, 2013 IEEE 31st International Conference
on Computer Design (ICCD), pp. 154, 168, October 6–9, 2013.
19. Z. Zhang, B. Noia, K. Chakraparthy, and P. Franzon, Face to face bus design
with built-in self-test in 3DICs, Proceedings IEEE 3D Integration Conference.
2
Three-Dimensional System-in-Package
for Application-Specific Integrated
Circuit and Three-Dimensional Dynamic
Random-Access Memory Integration

Li Li

CONTENTS
2.1 Three-Dimensional SiP Introduction ........................................................ 15
2.2 Enabling Technologies for 3D SiP ............................................................. 17
2.2.1 Three-Dimensional Stackable Memory ........................................ 17
2.2.2 High-Density Interposer ................................................................. 19
2.2.2.1 Silicon Interposer .............................................................. 19
2.2.2.2 Organic Interposer ............................................................ 21
2.2.3 Microbump Interconnect ................................................................ 23
2.3 3D SiP for Application-Specific Integrated Circuits and High
Bandwidth Memory Integration ................................................................ 25
2.3.1 Organic Interposer Design ............................................................. 26
2.3.2 Simulation and Results ................................................................... 27
2.4 Three-Dimensional SiP Assembly............................................................. 29
2.5 Test and Characterization ........................................................................... 32
2.6 Reliability Challenge ................................................................................... 33
2.7 Summary....................................................................................................... 37
References............................................................................................................... 38

2.1 Three-Dimensional SiP Introduction

The bandwidth for high-performance networking switches and rout-
ers increases two to ten times for every new generation. This in turn
drives the bandwidth requirements for the application-specific integrated
circuits (ASICs) and their external memory devices designed for the

15
16 3D Integration in VLSI Circuits

high-performance network systems. To meet these bandwidth require-

ments, the ASIC packaging technology has to keep up with the integrated
circuit (IC) technology scaling. Recently several 2.5-dimensional (2.5D) and
3-dimensional (3D) IC integration or packaging technology platforms have
been developed to address the gap seen between the slowdown of Moore’s
law scaling and the ever-increasing system integration requirements. In the
remainder of this chapter, the terms 2.5D and 3D IC integration and 2.5D
and 3D packaging are used interchangeably.
The early success of FCAMP that stands for Flip Chip and Memory Package
provided a much-needed solution to the bandwidth challenges between the
ASIC and memory devices designed for a high-end networking application
[1]. The FCAMP, sometimes also referred to as a system-in-package (SiP), con-
tains an ASIC die that is attached to an organic substrate using the flip chip
technology. The memory devices are packaged, tested at-speed and burned-
in before putting on the SiP substrate. This process flow is compatible to the
now industry-standard flip chip assembly process and has achieved high
manufacturing yield. The other innovation that made the FCAMP a success
is the availability of large size (>52.5 mm), high density, thin-core or core-
less, and high-performance organic substrates that are based on the build-
up, micro-via technologies [2]. The bandwidth improvement through the
FCAMP design has had a profound impact on high-performance network-
ing switches and routers, similar to computing applications.
Leveraging 3D IC integration, a concept of 3D ASIC and memory integra-
tion with a silicon interposer and through-silicon via (TSV) was proposed
and a feasibility study was conducted [3]. 3D ASIC and memory heteroge-
neous integration can be considered as an extension of the FCAMP that was
developed about a decade ago. It has the advantages of reducing power con-
sumption, improving the bandwidth between the ASIC and memory, and
modularizing system hardware designs. Critical components for enabling
3D ASIC and memory integration include large-size silicon interposer,
3D-stacked memory with TSV, and microbump or micropillar interconnects.
The 3D ASIC and memory heterogeneous integration not only optimizes
the package’s size but also imposes certain thermal limitations, as heat gen-
erated from the bottom logic chip has to pass through the entire die stack to
get dissipated. To overcome the challenges associated with 3D IC integration;
especially in the areas of generating TSVs in the active devices, thermal man-
agement and packaging; a series of multichip packaging technologies using
a silicon interposer or a fine-pitch organic interposer have been developed
[4]. This type of IC integration is nicknamed 2.5D IC integration as the chips
are still packaged in a planar format on the interposer as seen in previous
generations of multichip module (MCM) technologies. Furthermore, a 3D
SiP combining 2.5D and 3D IC integration has been developed for integrat-
ing one or more ASIC chips with one or more 3D-stacked memory die stacks.
Both the ASIC chip and 3D memory die stacks reside on one side of the inter-
poser [5]. A schematic of the 3D SiP is shown in Figure 2.1.
3D System in Package 17

H BM

HBM
ASIC H BM

Inter H BM
pos er
Pack
age subs
trate

FIGURE 2.1
A schematic of a 3D SiP with one ASIC die and four HBM die stacks.

Here HBM stands for high bandwidth memory, which is a 3D-stacked

dynamic random-access memory (DRAM) device developed by major
DRAM suppliers in recent years.
The interposers used in 2.5D or 3D SiP can provide much higher wiring
density than the conventional organic build-up substrate. Together with
the microbump interconnect technology, the 3D ASIC and memory SiP can
greatly increase the width and density of the interface between the ASIC
and the memory devices, and hence the bandwidth. The 3D SiP for ASIC and
memory integration has the following advantages:

• High package thermal performance similar to the single flip-chip

module
• Higher interconnect wiring density than the conventional organic
build-up substrate
• Compatible to the flip-chip assembly process
• Lower CTE mismatch between the silicon chips and the interposer
• Lower stresses on the low-k dielectrics and interconnects of the sili-
con chips by leveraging the Cu-filled interposer as a stress buffer

2.2 Enabling Technologies for 3D SiP

2.2.1 Three-Dimensional Stackable Memory
In the past several years, memory suppliers started the development of low-
power, high-bandwidth, 3D-stackable memory devices. Industry standards
for the 3D-stackable memory devices have also been developed. An early
18 3D Integration in VLSI Circuits

example of 3D IC integration is the application of 3D-stacked (3DS) DRAM

with TSV [6]. Published in December 2013, the addendum to JESD79-3 defines
the 3DS DDR3 SDRAM specification, including features, functionalities, AC
and DC characteristics, packages, and ball/signal assignments. Since then,
3DS DDR3 and 3DS DDR4 have been developed and are now in mass pro-
duction to meet the ever-increasing demand in memory module density for
server, high-performance computing (HPC) and networking applications. In
parallel to the development of 3DS DDR3 and 3DS DDR4, Wide I/O mobile
DRAM was developed leveraging the same 3D IC stacking technology. The
Wide I/O single date rate (SDR), which is 512-bit wide, running at 200 MHz
can provide a total bandwidth of 12.8  GB/s and consumes only half the
power of a low-power double data rate (LPDDR2) on the per bandwidth base
[7]. The second-generation Wide I/O DRAM, Wide I/O 2, provides four times
the memory bandwidth (up to 68 GB/s) of the previous version but at lower
power consumption [8]. From a packaging perspective, the Wide I/O DRAM
is designed specifically to stack on top of a system-on-chip (SoC) logic device
such as an application processor (AP) die using TSVs and microbumps or
micropillars to minimize electrical interference and the overall package
form factor that is much needed for mobile applications.
Using a similar 3D heterogeneous integration approach, a consortium led
by Micron Technology developed the Hybrid Memory Cube (HMC) [9]. The
HMC consists of a 3D stack of DRAM dies on top of a logic memory interface/
controller die. The DRAM dies within the stack are each of 1 GB in DRAM
density and are divided into 16 partitions. Each partition has two banks
(also called arrays), and its own data and control TSVs. HMC is designed
to emphasize massive amounts of bandwidth at higher power consumption
and cost than Wide I/O 2.
For the 3D ASIC and memory SiP designs, these 3D-stackable memory
devices have advantages that include the following:

• Memory and ASIC (logic) device can each be built in their own spe-
cific processes.
• Further exploit the process/cost differences between the logic and
memory devices.
• Very high data rate (bandwidth) with low latency and low-power
per bandwidth.
• Wide interface enabled by very wide interdie interconnect.
• Low parasitic enabled by short, direct interconnect.

Recently, this SiP approach has been extended to include the HBM DRAM
devices based on 3D IC integration with TSV and micropillar intercon-
nects [5].
HBM is a new class of DRAM developed by major DRAM suppliers
leveraging wide I/O and TSV technologies [10]. It is targeted for graphics
3D System in Package 19

processor units (GPU), HPC, server, networking, and client applications

where peak bandwidth, bandwidth per watt, and capacity per area are
valued metrics. The high bandwidth is delivered using a 1024-bit wide
memory interface that is divided into eight independent channels. To
accommodate the 1024-bit interface between HBM and the host compute
die, micropillars with a pitch of 55 microns are placed at the bottom of the
HBM DRAM die stack.

2.2.2 High-Density Interposer

2.2.2.1 Silicon Interposer
The introduction of 3D IC integration also brought new requirements and
sometimes disruptions to the existing microelectronics manufacturing sup-
ply chain. One example is the manufacturing and supply of silicon interposers
(Si-IP). Currently there are three commonly used interposer-manufacturing
process flows. The first can be referred to as the Foundry Process flow. The sili-
con interposer is fabricated completely by the wafer foundry and sometimes
the service even includes packaging the 3D IC subassembly to the package
substrate [11].
The second can be referred to as the middle-end-of-line (MEOL) process flow.
MEOL is often used by the outsourced assembly and test (OSAT) suppli-
ers for processing TSV wafers with active circuits. It is usually started at
the wafer foundry for via generation and filling, and fabrication of front-
side metal wiring layers. Then the full thickness wafers are delivered to the
OSAT for further downstream processes that include wafer thinning, TSV
reveal, backside metal layer generation, passivation, and bumping. The third
process can be called as the substrate process flow. In this case, the silicon
interposer is fabricated and supplied by the traditional packaging substrate
suppliers. A comparison of these silicon interposer manufacturing and sup-
ply chain flows is given schematically in Figure 2.2.

Foundry process

TSV and Front side Wafer Back side Assembly

Foundry Debond
FEOL μBump thinning bump /ship

MEOL process at OSAT

TSV and Front side Wafer Back side

Foundry Debond Assembly
FEOL μBump thinning bump

Substrate process
Subs TSV and Front side Wafer Back side
Debond Ship
supplier FEOL μBump thinning bump

FIGURE 2.2
A comparison of silicon interposer manufacturing and supply chain flows.
20 3D Integration in VLSI Circuits

For both the foundry and MEOL processes, TSVs are generated using the
deep reactive-ion etching (DRIE) process [12]. The front-side interconnects
or wiring layers are made with Cu damascene techniques. For the backside
interconnection, MEOL usually uses redistribution layer (RDL) process [13].
For true 3D wafers with active circuits and TSVs, MEOL may be preferred
but for passive interposers, an alternative and cost-effective way may exist.
This alternative, substrate process flow is the focus of this study.
The 3D SiP in Figure 2.1 can be supported with a silicon interposer with
TSV. The TSVs that are typically 10–25  μm in diameter are formed by the
DRIE process. The walls of the TSV are lined with the SiOx dielectric. Then,
a diffusion barrier and a copper seed layer are introduced. The via holes
are filled with copper through the electrochemical deposition process. The
chemical–mechanical planarization (CMP) process is used to remove the
copper overburden.
Recently, manufacturing of cost- and performance-effective, large-size
silicon interposer has been investigated [14]. The existing supply chain
and infrastructure of high-performance flip-chip packaging substrates are
leveraged. There are several advantages in this approach. One is mini-
mal disruption to the existing supply chain. The silicon interposer is con-
sidered as a packaging material rather than another piece of silicon chip.
Secondly, large-size silicon interposers can be manufactured with a line
width and line spacing in the range of a few micrometers. This type of sili-
con interposer (Si-IP) is often referred to as coarse-pitch silicon interposer
to distinguish itself from the ones made by wafer foundries. Table  2.1
shows a comparison between the coarse-pitch silicon interposer to the
fine-pitch silicon interposer.
For coarse-pitch silicon interposers, the semi-additive process (SAP) is
used to fabricate Cu wiring on either side of the interposer. The SAP method

TABLE 2.1
Comparison of Coarse-Pitch and Fine-Pitch Silicon Interposers
Features Coarse-Pitch Si-IP Fine-Pitch Si-IP

Cu wiring (dielectric) SAP (polyimide) Damascene (oxide)

Microbump material Cu/Ni/SnAg, etc. Cu/Ni/SnAg, etc.
Minimum microbump size/pitch (μm) 30/50 20/40
Minimum front-side wiring 3/3/3 0.5/0.5/1.0
Line width/space/thickness (μm)
RDL via size (μm) 20 1.0
TSV size/pitch/depth (μm) 60/150/200 10/50/100
Minimum back-side wiring 10/10/3 10/10/1.0
Line width/space/thickness (μm)
Back-side pad and bump size (μm) Ni/Au Ni/Au
100/150 100/150
3D System in Package 21

FIGURE 2.3
A top view of a 35 × 35 mm silicon interposer attached to a package substrate.

has fewer process steps and uses conventional equipment that is also used
for fine-pitch printed wiring board fabrication [13]. On the other hand, fine-
pitch silicon interposer relies on the damascene technique for Cu wiring fab-
rication that requires both chemical–mechanical polishing (CMP) and dry
etching processes. As it involves fewer process steps and uses conventional
equipment for fabrication, coarse-pitch silicon interposers will be less costly
when compared to fine-pitch silicon interposers [13].
Figure 2.3 shows the top view of a 35 × 35 mm silicon interposer fabricated
with the SAP method.
The front side of the interposer has two metal wiring layers, whereas
the backside has one wiring layer. The interposer shown is attached on a
50 × 50 mm HiTCE ceramic substrate. Major fabrication steps used are shown
schematically in Figure 2.4.
The size of the silicon interposers from the leading foundries is currently
limited to 26  ×  32  mm, which is the reticle size used in the lithographic
wafer processing. This size limitation can be a disadvantage for ASIC and
memory integration as the die sizes for high-performance ASICs can be as
large as 25 × 25 mm and the ASIC chips require multiple external memory
devices as illustrated in the FCAMP case. Use of reticle stitching to increase
the silicon interposer size is under development; however, it has its own
limitations as well.

2.2.2.2 Organic Interposer

Another approach to overcome the size limitation of the silicon interposer
technology and for better ASIC and memory integration is through an
22 3D Integration in VLSI Circuits

(1)

(2)

(3)

(4)

(5)

FIGURE 2.4
Major fabrication steps used to fabricate the coarse-pitch silicon interposer. (1) Silicon wafer;
(2) TSV formation, thermal oxidation silicon thickness: 200 μm, TSV: diameter 60 μm/pitch
150  μm; (3) TSV filling, planarization; (4) multilayer wiring (double-sided), RDL—
semi-additive process, insulator—photosensitive resin, top side—two-layer, and bottom
side—one-layer; and (5) double-sided bumping, Cu/Ni/SnAg bump (electroplating), and
diameter—30 μm.

organic interposer technology. A new 3D SiP based on an organic interposer

is being developed [5]. In this study, application of a fine-line, fine-pitch,
large-size organic interposer for integrating ASIC and HBM is investigated.
There are several advantages of organic interposers. One is less disruption
to the existing supply chain while leveraging the infrastructure developed
over the years for manufacturing high-performance flip-chip packaging
substrates. The organic interposer is considered as a packaging material
rather than another piece of silicon chip for the 3D SiP assembly. Second,
large-size organic interposers can be manufactured with a panel process and
can have line width and line spacing in the range of a few micrometers mak-
ing them suitable for integrating large die-size ASIC or other logic devices
with multiple HBM DRAM stacks. Table 2.2 shows a comparison between
the organic interposer and a typical fine-pitch silicon interposer.
To accommodate a total of 1024 wires and micropillars for the HBM inter-
face, a few more routing layers are needed for the organic interposer due to
its coarse line width and spacing compared to the silicon interposer. Major
manufacturing steps for making organic interposers are the same as that for
the organic build-up substrates. These include the following:
3D System in Package 23

TABLE 2.2
A Comparison of Organic Interposer and Silicon Interposer
Features Organic Interposer Silicon Interposer

Cu wiring (dielectric) SAP pattern plating (organic) Damascene (oxide)

Minimum front-side pad size/pitch (μm) 30/55 20/40
Minimum front-side wiring 6/6/10 0.5/0.5/1.0
Line width/space/thickness (μm)
Number of routing layers (typical) 10 4
Wiring layer via size (μm) 20 1.0
PTH or TSV size/pitch/depth (μm) 57/150/200 10/50/100
Back-side pad size/pitch (μm) 100/150 100/150
Minimum back-side wiring 6/6/10 10/10/1.0
Line width/space/thickness (μm)

• Plated-through hole (PTH) generation and filling for the core layer
• Circuitization of the core layer
• Building Cu-wiring layers on two sides of the core layer with the
micro-via and build-up processes

Pattern plating as part of the SAP is used to fabricate Cu wiring and micro-
vias for all the build-up layers. The pattern plating method has been used
extensively in manufacturing high-density, build-up organic packaging
substrates. On the other hand, fine-pitch silicon interposer relies on the
damascene technique for Cu-wiring fabrication that requires both chemical–
mechanical polishing (CMP) and dry etching processes. As it involves fewer
process steps and uses a panel format for fabrication, organic interposers
will be less costly when compared to silicon interposers and can offer much
larger size interposer for high-performance ASIC and HBM integration.
Figure 2.5 shows the top view of a 38 × 30 mm organic interposer fabricated
with the build-up and pattern-plating processes.
As shown in the above-mentioned figure, four micropillar footprint pat-
terns are included for attaching the HBM DRAM die-stacks. A close-up view
on the pads for micropillars is shown in Figure 2.6.

2.2.3 Microbump Interconnect

To accommodate the higher wiring density enabled by silicon or organic
interposers and to join 3D-stackable memory devices, the pitch and size of
conventional controlled collapsed chip connection (C4) solder bumps have to
be reduced greatly. Based on the same electroplating process developed for
C4 solder bumps, a microbump interconnect technology is developed and
applied to a 3D SiP test vehicle [3,5]. A 3D view of the electroplated micro-
bumps is shown in Reference 3. The bump shear test is used to characterize
24 3D Integration in VLSI Circuits

TL TR

BL BR

FIGURE 2.5
A top view of a 38 × 30 mm organic interposer manufactured.

TL (× 100) TR (× 100)

(× 500)
(A)

(B)

BL (× 100) BR (× 100)

FIGURE 2.6
A close-up view of a 38 × 30 mm organic interposer on the pads for micro-pillars.
3D System in Package 25

the mechanical integrity of the microbumps after the plating process and
after aging at 150°C for 72 hours. The results for the bump shear test before
aging are also shown in Reference 3. The average shear force is 3.47 g/bump.
After aging, the average shear force decreased to 2.1 g/bump and the growth
of intermetallic compound (IMC) Cu6Sn5 is observed [3]. Another approach
for the solder bump pitch reduction and density increase is through a cop-
per (Cu) pillar approach. The Cu pillar technology was first introduced to
advanced logic devices in 2006 and has since been developed especially for
flip chip applications with a bump pitch below 100 μm. Copper pillars are
less prone to electromigration that makes the technology a good choice for
applications with reduced bump pitches, sizes, and increased current densi-
ties. The Cu pillar interconnection used for 2.5D and 3D IC integration is also
referred to as micropillar interconnection. For example, the micropillar array
used to connect the current generation of HBM DRAM to interposers has a
pitch of 55 μm [15].

2.3 3D SiP for Application-Specific Integrated Circuits

and High Bandwidth Memory Integration
A 3D SiP is designed and manufactured that includes a large-size organic
interposer with five Cu wiring layers on each side of the core layer. The organic
interposer has a size of 38 × 30 × 0.4 mm. A high-performance ASIC die mea-
sured in 19.1 × 24 × 0.75 mm is attached on top of the organic interposer along
with four HBM DRAM die stacks. The HBM DRAM die stack with a size of
5.5 × 7.7 × 0.48 mm includes one base buffer die and four DRAM-core dice,
which are interconnected with TSVs and fine-pitch micropillars. The HBM
stacks are placed equally in quantity along the two sides of the ASIC die. For
assessing thermomechanical reliability, two of the HBM stacks are replaced
with their daisy-chain versions.
The ASIC and the HBM 3D IC subassembly is then assembled on an
organic packaging substrate with the conventional C4 solder bumps.
Communications between the ASIC die and the HBM DRAM stacks are
made through the Cu wiring and micro-via interconnections of the organic
interposer. A top view of the 3D SiP designed is shown schematically
in Figures 2.7 and 2.8 provides a schematic cross-sectional view of the
assembly.
Here HBM-M represents the mechanical or daisy-chain version of the
HBM DRAM stack.
The 3D SiP was designed and built to provide functions and capabilities
to test the signal and power integrity of the signal connections, of the HBM
memory channels, from the ASIC die to the two HBM DRAM die stacks.
26 3D Integration in VLSI Circuits

50 mm

38 mm

HBM-M HBM

30 mm

50 mm
ASIC

HBM HBM-M

Organic interposer

Build-up substrate

FIGURE 2.7
A schematic top view of the 3D SiP designed.

Heat spreader Organic interposer

Cu micropillar C4 bumps

HBM ASIC HBM

1-2-1 Build-up substrate

FIGURE 2.8
A schematic cross-sectional view of the 3D SiP designed.

A  memory controller and PHY for the HBM were designed and imple-
mented for the host ASIC on the organic interposer. To shorten the develop-
ment cycle, the ASIC design including the IP for the HBM memory controller
and PHY was first implemented through a field-programmable gate arrays
(FPGA) device.

2.3.1 Organic Interposer Design

The 38  ×  30  mm organic interposer has 12 layers: 5 top routing layers,
2 layers around the core, and 5 bottom routing layers. This configuration is
also known as a 5-2-5 layer stack-up. The 50 × 50 mm package substrate has
4 layers: 1 top routing layer, 2 layers around the core, and 1 bottom routing
layer. This configuration is known as a 1-2-1 layer stack-up.
Although the package substrate uses regular dielectric material used for
standard build-up substrates, the organic interposer is designed and fabricated
3D System in Package 27

with a low-loss dielectric material (0.005@10 GHz loss tangent) and has a coef-
ficient of thermal expansion (CTE) closely matched to the CTE of the dielec-
tric material used in the packaging substrate. The substrate has an 800  μm
thick core to help minimizing the warpage effects during reflow, whereas the
organic interposer has a 200 μm thick core. This new, low-loss dielectric mate-
rial used for the organic interposer allows ultrafine line spacing (line width/
spacing = 6 μm/6 μm), low transmission loss, and high insulation reliability.

2.3.2 Simulation and Results

Time-domain and frequency-domain simulations have been performed to
validate the 3D SiP design and to check if the performance specification of
the HBM channels is met. The frequency-domain simulations have been
done with the Cadence Sigrity PowerSI, whereas for the time-domain simu-
lations, HSPICE was used.
The results for the frequency-domain analysis are shown in Figures 2.9
through 2.11, respectively.
Very good margins, based on standard specifications, for the 0–1 GHz oper-
ating frequency bandwidth are achieved. The insertion loss is −0.5 dB@1 GHz
(with −1 dB@1 GHz standard specification), the return loss is approximately
−23 dB@1 GHz (with −10 dB@1 GHz standard specification), and the cross-
talk is approximately −45 dB@1 GHz (with −30 dB@1 GHz standard specifi-
cation). The results in these figures also show that in fact good performances
are valid for a very high-frequency bandwidth, up to 20 GHz.

S Amplitude (dB) HBM - Insertion loss (S12)

0

−0.5

−1

−1.5

−2

−2.5

−3

−3.5

−4

−4.5

−5
0 1 2 4 6 8 10 12 14 16 18 20
Frequency (GHz)

FIGURE 2.9
Insertion loss of the HBM channel simulated.
28 3D Integration in VLSI Circuits

S Amplitude (dB) HBM - Return loss (S11)

0

−10
−15
−20

−30

−40

−50

−60

−70

−80

0 0.5 1 2 4 6 8 10 12 14 16 18 20
Frequency (GHz)

FIGURE 2.10
Return loss of the HBM channel simulated.

S Amplitude (dB) HBM - X-talk

0

−20

−30
−40

−60

−80

−100

−120
0 0.5 1 2 4 6 8 10 12 14 16 18 20
Frequency (GHz)

FIGURE 2.11
Loss due to cross-talk for the HBM channel simulated.
3D System in Package 29

Read mode
PRBS31
stimulus HBM Probe
ASIC
IBIS Package
IBIS
model model
model

Write mode
PRBS31
Probe stimulus
HBM ASIC
Package
IBIS IBIS
model
model model

FIGURE 2.12
A schematic for setting up the time-domain analysis.

For the time-domain analysis, multiple corners have been simulated and
only the worst-case results are reported here. From the package design file,
an HSPICE model was extracted and used for the simulations, along with
I/O buffers for the ASIC and HBM memories. The setup used in the HSPICE
simulations is shown schematically in Figure 2.12.
A random PRBS 31 (pseudo-random bit sequence) is generated at the input
and comparing the resulted sequence at the output with and without any
aggressors (one signal was considered victim, and three aggressors on each
side) is used to analyze the effect of cross-talk. The results are a comparison
for the victim, in the situation with and without the aggressors, so that we
can see the output data waveforms when the memory has low activity com-
pared to the situation when the memory has high activity. As such, in READ
MODE (see block diagram in Figure 2.12), the worst-case skew due to cross-
talk is ~46 ps, whereas in WRITE MODE, the worst-case skew is ~42 ps.
Eye diagrams have been computed as well for the two modes. In READ
MODE, with cross-talk from aggressors, the eye has a height of ~900  mV
and a width of ~800 ps, and the introduced jitter is about 200 ps. In WRITE
MODE, with cross-talk from aggressors, the eye has a height of ~1 V and a
width of ~825 ps, whereas the introduced jitter is about 175 ps. All the results
mentioned here are with cross-talk and in the worst-case possible. As it can
be expected, the results without the cross-talk, in the worst-case, are slightly
better. Good confirmation to the frequency-domain results is achieved.

2.4 Three-Dimensional SiP Assembly

To develop the optimal chip joining process for attaching HBM die stacks
with micropillars and an overall packaging assembly process flow for
the final 3D SiP module, shadow moiré method was used to analyze the
30 3D Integration in VLSI Circuits

16.0

14.0
HBM_1 HBM_2
12.0
Warpage (micron)

10.0

8.0

6.0

4.0

2.0

0.0
25
100
120
150
183
200
220
240
260
280
260
240
220
200
183
150
120
100
25
Temperature (°C)

FIGURE 2.13
Warpages measured for the HBM die stack from room temperature to 280°C.

thermomechanical behavior of each constituent component of the 3D SiP.

Figure 2.13 shows the warpage measurement results for the HBM die stacks
over the temperature range from room temperature to 280°C.
As expected, the warpage of the HBM die stack was small (<8 μm) and did
not change much with temperature.
Figure 2.14 shows the warpage measurement results for the organic inter-
poser from room temperature to 260°C.

140.0

120.0

100.0
Warpage (micron)

80.0

60.0
Interposer_1 Interposer_2
40.0

20.0

0.0
25
75
90
100
110
120
130
150
183
200
220
240
260
240
220
200
183
150
130
120
110
100
90
75
25

Temperature (°C)

FIGURE 2.14
Warpages measured for the organic interposer from room temperature to 260°C.
3D System in Package 31

It can be seen that the warpage of the organic interposer was relatively
small (about 100 μm) and stayed in that range when the temperature was
increased up to 260°C.
Based on the thermal deformation analysis for each components of the
3D SiP, a series of design-of-experiments (DOE) were planned and con-
ducted to overcome the challenges in developing the suitable assembly
process for the ASIC and HBM 3D SiP. As the warpages of the silicon dice
and organic interposer were small and did not change with tempera-
ture, the ASIC die and HBM DRAM die stacks were assembled onto the
interposer to form the ASIC and HBM subassembly first. Underfill encap-
sulation was used to protect the joints made of micropillars and regu-
lar bumps. A top view and a bottom view of the 3D SiP subassembly are
shown in Reference 5.
The structure of the 3D SiP subassembly was then analyzed using X-ray,
acoustic, and optical microscopy to ensure good micropillar solder joints are
formed. Pictures of the cross-sections for the micropillar joints for attaching
HBM die stacks can be found in Reference 5.
In the final assembly process step, the ASIC and HBM subassembly was
attached to the package substrate using the conventional C4 solder bump
interconnection. The C4 bumps were then encapsulated using the underfill
material. A top view of the finished 3D SiP (without the lid) is shown in
Figure 2.15 and a cross-sectional view of the finished 3D SiP (with the lid) is
shown in Figure 2.16.

FIGURE 2.15
A top view of the finished 3D SiP.
32 3D Integration in VLSI Circuits

ASIC (FPGA)
HBM ASIC

Micro bumps

Interposer

C4 bumps

Substrate

BGA

HV spot mag WD 2 mm
10.00 kV 4.0 100 × 21.2 mm C3SiP module

FIGURE 2.16
A cross-sectional view of the finished 3D SiP.

2.5 Test and Characterization

Electrical performance of the ASIC-HBM memory interface and functional-
ities of the HBM DRAM stacks as well as post-assembly open and short per-
formance were verified and tested using an application evaluation board
designed and fabricated.
The application board with a socket to host the 3D SiP modules was
included as part of the test platform. The application board along with a 3D
SiP placed in the test socket is shown in Figure 2.17.
The application board measures at 9.5  ×  7 inches and has 16 layers: 7
ground layers, 3 power layers and 4 signal layers, and 2 high-speed layers,
on the top and bottom of the printed circuit board (PCB). Although the top
and bottom layers use Nelco N4000-13 EPSI dielectric material, the other lay-
ers use FR4 as dielectric.
Among the communication interfaces present on this application board,
an I2C interface was included to provide the communication path from a
Windows PC to the device and to output the results of the tests. The I2C inter-
face was selected due to reuse of the memory controller built-in tests that
also included an I2C interface. On the application board, test points and a
connector are also included for testing the daisy-chain connections designed
into the organic interposer, JTAG connection for configuring the ASIC, HBM
direct access connector, and power distribution and clocking necessary for
the operation of the ASIC and the HBM dies. On the PC, a Perl script-driven
interface allows for the configuring, running, and analyzing of several
3D System in Package 33

FIGURE 2.17
A top view of the application board with a 3D SiP module in the test socket.

different built-in memory tests within the ASIC memory controller design.
The Perl script calls a Windows program and the necessary drivers that con-
trol a USB to I2C protocol conversion board that sits between the Application
Board’s I2C interface and the PC’s USB port. Through the ASIC’s built-in tests
and via the I2C/Perl script platform, the electrical interface between PHY
and memory is tested for general connectivity, cross-talk between lines, and
simultaneously switching data lines. Defective cells in the memory are also
detected through the tests.

2.6 Reliability Challenge

Compared with traditional 2D IC packaging, the emerging 2.5D and 3D IC
integration involves several new elements in design, manufacturing, and
supply chain processes. The traditional boundaries between the wafer fab
and an OSAT supplier are blurred. These new elements include

• Thin die.
• TSV.
• Chip-to-chip (C2C) or chip-to-wafer (C2W) joining and interconnect.
• Micropillar or microbump for C2C or C2W connection.
• Die/interposer backside (BS) RDLs.

A summary of the new elements introduced with the 2.5D and 3D IC integra-
tion technologies and their implications on reliability of the final products is
included in Table 2.3.
34 3D Integration in VLSI Circuits

TABLE 2.3
New Elements Introduced by 2.5D and 3D IC Integration
Element Failure Mechanism

Thin die • Warpage-induced stress effect on device performance

• Interconnect/Si delamination-cracking and misalignment
• Deterioration of device characteristics by Cu diffused
from the back side
TSV • Si cracks and interconnect cracks/delamination
• TSV-stress effects on device performance
• Electromigration (EM)/stress migration (SM)
C2C connection • Stress generated by chip-stacking effects device
performance
• Delamination/cracking in the interconnect
BS–RDL • Stress-induced cracking/delamination
• Cu contamination on the back surface of the thinned
wafer/stress-enhanced diffusion
Microbumps • Cracking, EM/SM, and electrical resistance
Large volume of copper • Cu contamination on the back surface of the thinned
wafer/stress-enhanced diffusion
Increased power density and • Device performance
large thermal gradients • Thermal runaway
Removal of I/O structure • Electrostatic discharge (ESD) damage

The JEDEC publication, JEP158 titled “3D Chip Stack with Through-Silicon
Vias (TSVS): Identifying, Evaluating and Understanding Reliability Interactions”
[16] was developed about a decade ago when industry started to develop
2.5D and 3D IC integration for commercial applications. It was intended as
a guideline to describe the extension of the standard tests to 3D IC com-
ponents. The main element of this extension is the addition of appropriate
test structures to evaluate the reliability of the TSVs and other new features
introduced in the fabrication of 3D IC products. JEP158 together with JEDEC
standard JESD47, “Stress-Test-Driven Qualification of Integrated Circuits” [17]
are used as a starting point for technology qualification to assess the intrin-
sic reliability of TSV and micropillar interconnection that are new elements
introduced by 2.5D and 3D IC integration.
Besides component-level reliability evaluation leveraging JEP158 and
JESD 47, system-level and board-level validations are also needed. The fol-
lowing is an example of board-level reliability evaluation of a 3D IC compo-
nent for networking applications. A test board was designed to mimic the
actual system board. Four large body-size, high-performance Cisco ASIC
packages were also included for the reliability test board assembly along
with the two 3D IC components. Figure 2.18 shows a top view of the fully
assembled test board assembly.
3D System in Package 35

3D IC

3D IC

FIGURE 2.18
A top view of the reliability test board assembly for the ASIC and 3D IC components.

The 3D IC component consists of five dice vertically stacked and intercon-

nected with TSV and microbumps. Over 3000 test features including chains
going through all layers of the die stack and package substrate are designed
into the 3D IC components. Two stress conditions were used. To assess the
reliability of micropillars further, an enhanced stress condition was applied.
It included 1000 hours of thermal aging at 125°C and followed by 6000 cycles
of the temperature cycling with a range of 0°C–100°C. The stress conditions
and final results are summarized in Table 2.4.
For the 3D SiP with ASIC and HBM components, a reliability test board
was also designed, fabricated, and assembled [18]. A top view of the finished
test board assembly designed for further board-level reliability evaluation is
shown in Figure 2.19.

TABLE 2.4
Board-Level Reliability Evaluation of a 3D IC Package
Test Condition Sample Size Results

Aging 1000 hours at 125°C 25 Zero failure

Aging plus 1000 hours at 125°C; 25 Zero failure
temperature cycling 6000 cycles 0°C–100°C
36 3D Integration in VLSI Circuits

FIGURE 2.19
A top view of the reliability test board assembly for the 3D SiP with ASIC and HBM DRAM.

Board-level temperature cycling test was conducted using an ESPEC

TCC-150 environmental testing chamber, and real-time continuity resistance
monitoring was done using an ESPEC AMR data logger. Daisy chains that
cover the micropillar connection from the HBM stacks to the organic inter-
poser, the regular solder bump connection from the organic interposer to
the SiP substrate, and finally the ball grid array (BGA) connections to the test
board are used as the test features for the reliability evaluation.
The reliability testing board measures at 10 × 8 inches and has 16 layers: 6
ground layers, 2 power layers and 6 signal layers, plus 2 layers on the top and
bottom surfaces of the PCB. A front view showing the wiring and setup of
the test boards inside the chamber is given in Figure 2.20.

1 2 3

5
4 6

FIGURE 2.20
A front view showing the wiring and setup of test boards inside the environmental chamber.
3D System in Package 37

A 0°C–100°C accelerated temperature cycling test was used for the board-
level reliability test. The temperature profile inside the chamber was adjusted
to minimize temperature variation from board to board. A measured tem-
perature profile is given in Reference 18. The duration for the temperature
cycling was 6000 cycles. Except for a few very early fails on some chains in
the sub 100 cycle range, all the other daisy chains or test features monitored
passed 6000 cycles with no failures.

2.7 Summary
To meet the requirements of next-generation Information and Communication
Technology (ICT) systems, several 2.5D and 3D IC integration or packag-
ing technology platforms have been developed. 3D-stackable memory,
large-size silicon/organic interposers, and microbump interconnection
are key-enabling technologies for realizing 2.5D and 3D IC integration.
Leveraging the existing supply chain and infrastructure of high-performance
flip-chip packaging substrates, a 3D SiP is designed and manufactured that
includes a large-size organic interposer with a total of 12 Cu-wiring layers,
an ASIC die, and 4 HBM DRAM stacks.

• The large-size organic interposer is manufactured with the SAP

method established for the micro-via and build-up substrates, so it
can be cost- and performance-effective for the applications such as
integrating high-performance ASIC and HBM DRAM stacks.
• Thermal deformation of each constituent component of the 3D SiP
was analyzed using real-time, metrological techniques such as
shadow moiré to optimize material selection, assembly process, and
reliability.
• A series of DOE and structural analyses were conducted to dem-
onstrate that the 3D SiP module can be built with micropillar inter-
connects, a unique C2C joining, and the final packaging assembly
processes developed.
• Electrical performance of the ASIC-HBM memory interface and
functionalities of the ASIC and HBM DRAM stacks as well as the
open and short of critical interconnects were verified and tested using
an application evaluation board fabricated.

Understanding the new elements introduced by 2.5D and 3D IC integration

and assessing their impact on reliability of the 2.5D and 3D devices are key
to successfully introducing these new packaging technologies to server,
HPC, and networking applications.
38 3D Integration in VLSI Circuits

New elements introduced by 2.5D and 3D IC integration are reviewed and

their impact on product reliability is discussed.

• The stress-test-driven methodology outlined by JEDEC standard

JESD 47 and JEP157 is a good start point for evaluating reliability
of 2.5D and 3D IC components and for assessing the individual and
collective effects of the new elements introduced by 2.5D and 3D IC
integration.
• Board-level or system-level characterization with purposely
designed test vehicles and actual 3D IC components is used to vali-
date reliability of 2.5D and 3D IC packages. Examples with one 3D IC
component and one 3D SiP with ASIC and HBM DRAM were given
as case studies.

A lot of progress has been made in the last decade through technology devel-
opment and process improvement. Several 2.5D and 3D SiP-packaging plat-
forms have emerged and appear to be ready for production.

References
1. Priest, J., M. Ahmad, L. Li, J. Xue, and M. Brillhart, Design optimization of a high
performance FCAMP package for manufacturing and reliability, Proceedings
55th Electronic Components and Technology Conference, Orlando, FL, May 2005,
pp. 1497–1501.
2. Priest, J. and L. Li, Challenges in substrate design, assembly, and reliability of
SiP package for a high end networking application, 39th International Symposium
on Microelectronics, IMAPS, San Diego, CA, October 2006, pp. 8–12.
3. Li, L., S. Peng, J. Xue et al., Addressing bandwidth challenges in next genera-
tion high performance network systems with 3D IC integration, Proceedings of
the 62nd Electronic Components and Technology Conference (ECTC), San Diego, CA,
May 2012, pp. 1040–1046.
4. S. Lakka, Xilinx SSI technology, Hot Chips: A Symposium on High Performance
Chips, Palo Alto, CA, August 2012.
5. Li, L., P. Chia, P. Ton, et al., 3D SiP with organic interposer for ASIC and mem-
ory integration, Proceedings of the 66th Electronic Components and Technology
Conference (ECTC), Las Vegas, NV, June 2016, pp. 1445–1450.
6. JEDEC Standard, Addendum No. 3 to JESD79-3: 3D Stacked SDRAM,
JESD79-3-3, December 2013.
7. JEDEC Standard, Wide I/O single data rate (wide I/O SDR), JESD229, December
2011.
8. JEDEC Standard, Wide I/O 2, JESD229-2, August 2014.
9. Hybrid Memory Cube. Retrieved from http://hybridmemorycube.org/
specification-v2-download-form/
3D System in Package 39

10. JEDEC Standard, High bandwidth memory (HBM) DRAM, JESD235A,

November 2015.
11. TSMC. CoWoS® Services. Retrieved from http://www.tsmc.com/english/
dedicatedFoundry/services/cowos.htm
12. Yoon, S. W., D. J. Na, W. K. Choi et al., 2.5D/3D TSV processes development and
assembly/packaging technology, Proceedings of the 11th Electronics Packaging
Technology Conference (EPTC), Singapore, December 2011, pp. 336–340.
13. Sunohara, M., A. Shiraishi, Y. Taguchi, K. Murayama, M. Higashi, and
M.  Shimizu, Development of silicon module with TSVs and global wiring
(L/S = 0.8/0.8 μm), Proceedings of the 59th Electronic Components and Technology
Conference (ECTC), San Diego, CA, May 2009, pp. 25–31.
14. Li, L., M. Higashi, A. Takano, J. Xue, and G. Ikari, Cost and performance effec-
tive silicon interposer and vertical interconnect for 3D ASIC and memory inte-
gration, Proceedings of the 64th Electronic Components and Technology Conference
(ECTC), Orlando, FL, May 2014, pp. 1366–1371.
15. JEDEC Product Outline, MO-316, HBM micropillar grid array package (MPGA),
October 2015.
16. JEDEC Publication, JEP158 3D chip stack with trough-silicon vias (TSVS):
Identifying, evaluating and understanding reliability interactions, November
2009.
17. JEDEC Standard, JESD47, Stress-test-driven qualification of integrated circuits,
2015.
18. Li, L., P. Ton, M. Nagar, and P. Chia, Reliability challenges in 2.5D and 3D
IC integration, Proceedings of the 67th Electronic Components and Technology
Conference (ECTC), Orlando, FL, May 2017, pp. 1504–1509.
3
A New Class of High-Capacity,
Resource-Rich Field-Programmable Gate
Arrays Enabled by Three- Dimensional
Integration Chip-Stacked Silicon
Interconnect Technology

Suresh Ramalingam, Henley Liu, Myongseob Kim,

Boon Ang, Woon-Seong Kwon, Tom Lee, Susan Wu,
Jonathan Chang, Ephrem Wu, Xin Wu, and Liam Madden

CONTENTS
3.1 Introduction ..................................................................................................42
3.2 Architecture, Design, and Product Enablement......................................43
3.2.1 Key Challenges.................................................................................44
3.2.1.1 Limited Connectivity and Bandwidth ...........................44
3.2.1.2 Excessive Latency ..............................................................44
3.2.1.3 Power Penalty ....................................................................44
3.2.2 Xilinx-Stacked Silicon Interconnect Technology......................... 45
3.2.2.1 Creating Field-Programmable Gate Array Die
Slices with Microbumps for Stacked Silicon
Integration .......................................................................... 46
3.2.2.2 Silicon Interposer with Through-Silicon Vias .............. 47
3.2.2.3 Three-Dimensional Integration Chip Analysis
Methodology Enablement with Simulation
Program with Integrated Circuit Emphasis
Simulators .......................................................................... 48
3.2.2.4 Silicon Interposer Signal Integrity.................................. 50
3.2.3 Three-Dimensional Integration Chip Resource-Rich
Field-Programmable Gate Arrays Product Offerings .................54
3.3 Stacked Silicon Technology Development and Package Reliability ..... 55
3.3.1 Key-Enabling Technologies ............................................................ 55
3.3.1.1 Silicon-Grinding Quality Optimization ........................ 57
3.3.1.2 Wafer Edge and Bevel-Cleaning Optimization ............ 57
3.3.1.3 Silicon-Footing Improvement.......................................... 58

41
42 3D Integration in VLSI Circuits

3.3.2
Three-Dimensional Integration Chip Development Test
Vehicles .............................................................................................. 59
3.3.2.1 28 nm Test Vehicle-Driven Process and Reliability
Improvements .................................................................... 60
3.3.2.2 Improvements to 20 nm Test Vehicle ............................. 61
3.3.2.3 Board-Level Reliability Test Vehicle and Results .........63
3.3.3 Three-Dimensional Integration Chip Reliability
Look-Ahead Assessments ...............................................................65
3.3.3.1 Continuous Process Yield Improvements ..................... 66
3.4 Potential Three-Dimensional Integration Chip Future Challenges ..... 67
Acknowledgment .................................................................................................. 68
References............................................................................................................... 68

3.1 Introduction
As the role of the field-programmable gate arrays (FPGA) becomes more sig-
nificant in larger and more complex system designs, it demands higher logic
capacity and more on-chip resources and functionalities. Until 40 nm node,
FPGAs have depended predominantly on Moore’s law scaling to respond to
this need, delivering nearly twice the logic capacity with each new process
generation. However, keeping pace with today’s high-end market demands
requires more than Moore’s law can provide.
To respond to these requirements, Xilinx first introduced a three-
dimensional integration chip (3D-IC) stacked silicon interconnect technol-
ogy (SSIT) for building FPGAs that offer bandwidth and capacity that are
not realized by traditional Moore’s law scaling (Figure 3.1). SSIT uses silicon
interposers with microbumps (μBumps) and through-silicon vias (TSV) to
integrate multiple FPGA die slices in a single package.
Starting from 2006 (Figure 3.2), Xilinx worked with research consortiums,
equipment vendors, foundry, and outsourced assembly and test (OSAT) part-
ners to develop TSVs, μBump, bonding, and 3D-IC integration technology.
Xilinx’s first 3D-IC SSIT FPGA, Virtex®-7 2000T, introduced in late 2011
mainly for emulation applications has been well adopted by the industry.
3D-IC devices since have expanded into networking, data center cloud
computing, and high-performance computing (HPC) areas in volume pro-
duction and account for more than 30% of the 20 nm/16 nm products offered.
The largest SSIT FPGA as of today is UltraScale XCVU440 with 4.4 million
logic cells, 600 k μBumps, 19 billion transistors in a 55 mm package, and has
been shipping in volume production since early 2016.
This chapter describes Xilinx’s 3D-IC development experiences gained
by launching process learning vehicles, developing design simulation
methodology, pursuing early-stage reliability assessment, and optimizing
A New Class of High-Capacity, Resource-Rich FPGA 43

Largest FPGA logic cells counts

5.0E+06
4.4E+06
4.0E+06 3D-IC implemented

3.0E+06

2.0E+06
2.0E+06

1.0E+06 7.6E+05
3.3E+05
2.7E+04 4.3E+04 1.0E+05 2.0E+05
0.0E+00
1998 2000 2002 2004 2006 2009 2011 2014

250 nm 180 nm 130 nm 90 nm 65 nm 40 nm 28 nm 20/16 nm

Year/Technology nodes

FIGURE 3.1
FPGA capacity doubling as 28 nm generation enabled by 3D-IC.

Heterogeneous
stacked silicon
Stacked 90 nm
90 nm 65 nm 28 nm interconnect
silicon Process
interconnect integration
Test Test Test Design tools technology
vehicle vehicle vehicle available
development and modular
completed completed completed
started development

2006 2007 2008 2009 2010 2011 2012

Design
Initial enablement World’s first 3D-
Reliability Design
reliability and supply stacked silicon
assessment validation
look-ahead chain interconnect
validation device

FIGURE 3.2
FPGA 3D-IC history—from concept to world’s first programmable products on 28 nm.

the process baseline. All the development efforts paved the foundation
for a robust production line. At the end, the potential future 3D-IC tech-
nology challenges are outlined.

3.2 Architecture, Design, and Product Enablement

Let us start with a summary of the challenges faced implementing large
designs with multiple monolithic FPGAs and how SSIT addresses them [1].
44 3D Integration in VLSI Circuits

3.2.1 Key Challenges

3.2.1.1 Limited Connectivity and Bandwidth
System-on-chip (SoC) designs comprise millions of logic gates connected by
complex networks of wires in the form of multiple buses, complicated clock dis-
tribution networks, and multitudes of control signals. Successfully partitioning
a SoC design across multiple FPGAs requires an abundance of I/Os to imple-
ment the nets spanning the gap between FPGAs. With SoC designs including
buses as wide as 1024 bits, even when targeting the highest available pin-count
FPGA packages, engineers must use data buffering and other design optimiza-
tions that are less efficient for implementing the thousands of one-to-one con-
nections needed for high-performance buses and other critical paths.
Packaging technology is one of the key factors to this I/O limitation.
The most advanced packages currently offer approximately 1200 ~ 1500 I/O
pins, far short of the total number of I/Os required.
At the die level, I/O technology presents another limitation because I/O
resources do not scale at the same pace as interconnect logic resources with
each new process node. When compared to transistors used to build the
programmable logic resources in the heart of the FPGA, the I/O transistors
are much larger to deliver the currents and withstand the voltages required
for chip-to-chip I/O standards. Thus, increasing large number of standard
I/Os on a die is not a viable solution to provide the connections for combining
multiple FPGA die.

3.2.1.2 Excessive Latency

Increased latency is another challenge with the multiple FPGA approach.
Standard device I/Os impose pin-to-pin delays that degrade the overall cir-
cuit performance for designs that span multiple FPGAs. Moreover, using
time-domain multiplexing (TDM) on standard I/Os to increase the virtual
pin count by running multiple signals on each I/O imposes even greater
latencies that can slow I/O speeds further down by a factor of 4x–32x or more.

3.2.1.3 Power Penalty

Connecting multiple FPGAs via standard I/Os approaches also results
in  higher power consumption. When used to drive hundreds of package-
to-package connections across printed circuit board (PCB) traces, standard
I/O pins carry a heavy-power penalty compared to connecting logic nets on
a monolithic die.
Similarly, multichip module (MCM) technology offers potential form-
factor reduction benefits for integrating multiple FPGA die in a single pack-
age. The MCM approach, however, still suffers from the similar restrictions
of limited I/O count and undesirable latency and power consumption
characteristics.
A New Class of High-Capacity, Resource-Rich FPGA 45

3.2.2 Xilinx-Stacked Silicon Interconnect Technology

To overcome these limitations and roadblocks, Xilinx developed a new
approach that enables high-bandwidth connectivity between multiple dies
by providing a much greater number of interdie connections. It also imposes
much lower latency and consumes dramatically lower power than the mul-
tiple FPGA via standard I/Os approach, while enabling the integration of
massive quantities of interconnect logic and on-chip resources within a sin-
gle package.
Xilinx arrived at such a solution by applying several proven technologies
in an innovative way. By combining TSV and μBumps with its application-
specific modular block (ASMBL™) architecture, Xilinx is building a new class
of FPGAs that delivers the capacity, performance, capabilities, and power
characteristics required to address the programmable imperative. Figure 3.3
shows the top view of the die stack-up with four FPGA die slices, silicon
interposer, and package substrate. SSIT technology combines enhanced
FPGA die slices and a passive silicon interposer creates a die stack that
implements tens of thousands of die-to-die connections to provide ultrahigh
interdie interconnect bandwidth with far lower power consumption and one
fifth of the latency of standard I/Os.
Being originally developed for use in a variety of die-stacking design
methodologies, silicon interposers provide modular design flexibility and
high-performance integration suitable for a wide range of applications. The
silicon interposer acts as a sort of microcircuit board in silicon on which
multiple dies are set side by side and interconnected. Compared to organic or
ceramic substrates, silicon interposers offer far finer interconnect geometries
(approximately 20X denser wire pitch) to provide device-scale interconnect
hierarchy that enables more than 10,000 die-to-die connections.

® 7
ex -
Virt 0T FPGA Dies
200

SI interposer

Package

FIGURE 3.3
Stacked silicon interconnect package top view.
46 3D Integration in VLSI Circuits

3.2.2.1 Creating Field-Programmable Gate Array Die Slices

with Microbumps for Stacked Silicon Integration
The foundation of Xilinx SSIT technology is the company’s ASMBL archi-
tecture, a modular structure comprising Xilinx FPGA building blocks in
the form of tiles that implement key functionality such as configurable logic
blocks (CLBs), block RAM, digital signal processing (DSP) slices, SelectIO™
interfaces, and serial transceivers. Each type of tile is organized into blocks
in columns and then combines the columns to create an FPGA. By vary-
ing the height and arrangement of columns, assortment of FPGAs can be
created with different amounts and mixes of logic, memory, DSP, and I/O
resources (Figure 3.4). The FPGA contains additional blocks for generat-
ing clock signals and for programming the static random access memory
(SRAM) cells with the bitstream data that configure the device to implement
the end user’s desired functionality. The architecture can be extended for
heterogeneous die integration as well [2,3].
Starting with the basic ASMBL architectural construct, three key modi-
fications that enable stacked silicon integration (Figure 3.5) have been
introduced. First, each die slice receives its own clocking and configura-
tion circuitry. Then the routing architecture is modified to enable direct
connections to routing resources within the FPGA’s logic array, bypassing
the traditional parallel and serial I/O circuits. Finally, each slice under-
goes additional processing steps to fabricate microbumps that attach the
die to the silicon interposer. All of these enable connections in far greater

FIGURE 3.4
FPGA tiles built with ASMBL™ architecture.

FIGURE 3.5
FPGA die or SLR (super logic region) optimized for stacked silicon integration.
A New Class of High-Capacity, Resource-Rich FPGA 47

numbers, with much lower latency, and much less-power consumption

than is possible using traditional I/Os (100x the die-to-die connectivity
bandwidth per watt versus standard I/Os).

3.2.2.2 Silicon Interposer with Through-Silicon Vias

The passive silicon interposer interconnects the FPGA die. It is built on a
low-risk, high-yielding 65 nm process and provides four layers of metalliza-
tion (3 Cu and 1 Al) for building the tens of thousands of traces that connect
the logic regions of multiple FPGA die (Figure 3.6).
Figure 3.7 illustrates the concept of the assembled die stack. It contains a
stack-up of four FPGA die mounted side by side on a passive silicon interposer

FIGURE 3.6
Passive silicon interposer with die-to-die interconnect wiring.

ASMBL Silicon
optimized interposer
FPGA slice glue

High-density
Additional stitching
slices side- • >10K connections
by-side • ∼1ns latency

FIGURE 3.7
Transparent view of the assembled die stack and die-to-die interconnects using silicon
interposer.
48 3D Integration in VLSI Circuits

SLR0 SLR1 SLR2 SLR3 μBump

C4 Silicon interposer TSV

Package substrate

BGA

FIGURE 3.8
Cross-section view of FPGA-silicon interposer integration on a package substrate.

(bottom view). The interposer is shown as transparent to enable a view of the

FPGA die slices connected by traces on the silicon interposer (not to scale).
The TSVs combined with controlled-collapse chip connection (C4)
solder bumps enable mounting the FPGA/interposer stack-up on a
high-performance package substrate using flip-chip assembly techniques
(Figure 3.8). The coarse-pitch TSVs provide the connections between the
package and the FPGA for the parallel and serial I/O, power/ground,
clocking, configuration signals, and so on.

3.2.2.3 Three-Dimensional Integration Chip Analysis

Methodology Enablement with Simulation Program
with Integrated Circuit Emphasis Simulators
As with any other new technology, 3D-IC also brings new challenges to tra-
ditional ad hoc design simulation tools because of the integration of multiple
dice and technologies. In order to take full advantage of 3D-IC integration,
Xilinx initiated and codeveloped with Synopsys a new simulation program
with integrated circuit emphasis (SPICE) simulation methodology for 3D-IC
module in 2012.
Although traditional circuit simulation tools work well on single-die simu-
lations, it is challenging to handle multidie simulations. One challenge is
in the common naming convention, such as global node names, subcircuit
(subckt) names, and model names. Global parameters confliction brings
another challenge. Extra effort is needed to modify the netlists and technol-
ogy models, making it a very time-consuming and error-prone process. In
addition, this workaround cannot handle different technologies with varied
process parameters such as tnom, scale, and geoshrink not to mention for
third-part IPs where the netlists are encrypted, it becomes impossible to do
any netlist modifications.
In contrast, the new 3D-IC features taking on a bottom-up approach pro-
vide a complete integrated simulation solution that works for any combina-
tion of integrated technologies.
By introducing module block syntax for each die, the approach enables
multidice chip-level simulations under the existing SPICE simulation envi-
ronment. Besides providing the simulation capabilities for 3D-IC designs,
A New Class of High-Capacity, Resource-Rich FPGA 49

FIGURE 3.9
Module-based netlist for 3D-IC integration chip with two dice from different technologies.

the major advantage of the new approach over the traditional one is its capa-
bility to maintain libraries and netlist files to help reduce memory usage and
simulation errors.
An example netlist in Figure 3.9 below illustrates the simple 3D IC integra-
tion chip with two dice of different technologies and temperatures as “xic1”
and “xic2.”
Each design team is expected to exercise the full design verification flow
for each IC module before the 3D-IC integration. As each integrated IC mod-
ule is regarded as a completed product, retaining the IC module netlist integ-
rity becomes a primary requirement. New HSPICE netlist features are then
introduced to target the multiple-die and multitechnology integration for
both heterogeneous and homogeneous designs [4].
New 3D-IC-specific configurations and analysis requirements are intro-
duced as well, such as multidie multitechnology integration, verification with
multiple operational IC temperature domain, and exponentially increasing
corner simulations for multidie integration with different simulation corners
on each IC module.
For both prelayout and postlayout design analysis and verification, the
new SPICE simulators offer 3D-IC-specific methodology for functional veri-
fication, timing verification, and power analysis.
The flow chart in Figure 3.10 provides the steps for 3D-IC integration simu-
lations using Synopsys’ 3D-IC simulation features. The new 3D-IC simula-
tion features and the methodology offer an efficient simulation solution for
timing closure of multidie and multitechnology integration under the con-
ventional SPICE simulation environment.
50 3D Integration in VLSI Circuits

Start with the fully verified netlists

from each of the design blocks
Create one module
for each design block

In each module, define a top-level subckt

Create a 3D IC top-level simulation deck

to connect the modules defined above

Add connectivity and control statements

Run HSPICE/CustomSim
simulations

FIGURE 3.10
3D-IC SPICE simulation flow.

Silicon interposer signal integrity also requires careful consideration both

from transmission and signal loss perspective and there are a few Xilinx
publications on this topic [2,5,6]. In Section 3.2.2.4, we will examine homo-
geneous and heterogeneous signal integrity analysis partitioning used suc-
cessfully by Xilinx to deliver 4 × 100 GB/s line card solution. It is to be noted
that this case study used a low-temperature cofired ceramic (LTCC) substrate
though actual product implementation was done using a low-loss organic
substrate.

3.2.2.4 Silicon Interposer Signal Integrity

Although the silicon interposer is only hundred microns thick, failure
to consider high-frequency effects of the TSVs will degrade the rise/fall
time, increase cross-talk, increase noise injection, and cause significant
performance degradation of the signal transmitted through the high-
speed channel. In addition, die-to-die signals that pass through the micro-
bumps and laterally through the fine-pitch metallization also need careful
consideration.
This section summarizes signal integrity studies for two types of off-die
signals in 3D designs: (1) Type I and (2) Type II signals. A Type I signal con-
nects one or more active ICs and a package pin. A Type II signal supports
interdie communication only and is not connected to any package pins. For
clarity, Type I signals are electrically connected to TSVs in the silicon inter-
poser, whereas Type II signals are not.
A New Class of High-Capacity, Resource-Rich FPGA 51

GTZ SLR0 SLR1

μBump
c
a b Silicon interposer TSV
C4

FIGURE 3.11
Signal Type I (vertical to package pin, as “a”) and Type II (interdie, as “b” and “c”).

Signal “a” in Figure 3.11 is a Type I signal. Typically, multiple microbumps

are connected to one C4 bump to ensure adequate electromigration margin.
Signal “b” and “c” are Type II signals from one top die to another top die.
Sections 3.2.2.4.1 and 3.2.2.4.2 discuss signal integrity related to the GTZ
where GTZ is a stand-alone 28  Gbps SerDes die and super logic regions
(SLRs) are 2 FPGA dies as described previously.

3.2.2.4.1 Type I Signals

The GTZ-IC receives eight differential serial streams at up to 28.05  GB/s
each, deserializes them with serial-in-parallel-out (SIPO) blocks, and sends
the parallel data and necessary clocks to the SLR. Similarly, when in output,
it receives parallel data from the SLR, serializes the traffic with parallel-in-
serial-out (PISO) blocks, and outputs the resulting differential serial streams.
The GTZ Type I signals are in the analog supply domain, which has its
own power and ground metal planes. The parallel Type II signals are in the
digital supply domain, sharing the digital and ground planes with the SLR.
Analog power and ground decoupling capacitors are distributed on the die,
on the interposer, and on the package substrate to ensure optimal signal
integrity at 28 GB/s.

3.2.2.4.1.1 Silicon Interposer TSV Modeling and Optimization A full 3D elec-

tromagnetic (EM) field solver is used to model the silicon interposer TSV
accurately over a wide frequency range. A broadband S-parameter model is
generated with an upper frequency limit of 50 GHz. The interposer consists
of TSVs, four metal layers for die-to-die connections, microbumps for die-to-
interposer connections and C4 bumps for interposer-to-package connections.
All components are incorporated into the model to predict actual interposer
performance in the system. A silicon interposer test vehicle was fabricated
for measurement and verification (Figure 3.12). Various TSV test structures
were characterized across a wide frequency range. The TSV is approximately
100 μm in height. De-embedding microprobe effects is critical for accurate
frequency-domain vector network analyzer (VNA) measurements. As a first
step, a microprobe calibration is performed to move the reference plane to
the end of microprobe. Depending on the metallization of the probe tip and
the surface of the pad, a residual, uncompensated resistive impedance in the
order of 10 ~ 100 mΩ could still be present. The flexing or overtravel of the
52 3D Integration in VLSI Circuits

200

Conductance, S
Capacitance, fF

2e−03

100
1e−03

0 10 G 20 G 30 G 0 10 G 20 G 30 G
(a) Frequency, GHz (b) Frequency, GHz

FIGURE 3.12
Simulation versus measurement comparison for (a) effective capacitance and (b) effective
conductance.

microprobe can also result in an uncompensated, residual inductance of as

much as 100 pH. To overcome these parasitic, a 2-port VNA was used.
The 2-port VNA eliminates the residual contact resistance and inductance.
However, mutual magnetic coupling between probes still exists if probes are
located face to face in close proximity. To improve high-frequency measure-
ment accuracy, 90° orthogonal probing is performed to cancel the magnetic
field coupling from the two probes. These result in improved measurement
especially for higher frequencies (>5 GHz).
As the focus is on multigigabit operation within the interposer, the model
must support and have a good agreement for frequencies up to tens of giga-
hertz. The measured data are compared with simulation results. Figure 3.12
shows the comparative results of effective capacitances which were extracted
from the open test structures. Two different silicon substrates are fabricated
to compare the substrate resistivity effects. One test interposer sample has
a 10 Ω-cm silicon resistivity, representative of a typical commercial process
and the other test interposer sample has a 20 Ω-cm silicon resistivity.
In Figure 3.12, the dotted traces represent the measured data and the solid
traces represent the simulated data of effective capacitance and conductance
values over frequency from the 10 Ω-cm and 20 Ω-cm silicon resistivity sub-
strates, respectively. Good agreement is observed between the measured and
simulated results. The effective capacitance decreases as a function of fre-
quency (slow-wave model to quasi-TEM mode). This effect can be modeled
as resistance and capacitance in the silicon. The measured effective capaci-
tance from the 10 Ω-cm silicon resistivity substrate was about 205.2 fF at low
frequency and 19.4 fF at high frequency, whereas from the 20 Ω-cm silicon
resistivity substrate, the measured effective capacitance was about 115.8 fF at
low frequency and 12.4 fF at high frequency. Both low- and high-frequency
capacitance are important electrical parameters because they directly affect
data eye opening, power consumption, and delay.
A second through test structure is fabricated to evaluate insertion for serially
connected TSVs. S21 insertion loss is measured and compared with simulation
as shown in Figure 3.13. The measured insertion loss at the Nyquist frequency
A New Class of High-Capacity, Resource-Rich FPGA 53

0 db
Insertion loss, db

−1.0 db

Measured 20 ohm-cm silicon interposer

Simulated 20 ohm-cm silicon interposer
Measured 10 ohm-cm silicon interposer
Simulated 10 ohm-cm silicon interposer
−2.0 db
0 10 G 20 G 30 G
Frequency, GHz

FIGURE 3.13
Simulation versus measurement comparison for insertion loss.

(14 GHz) is −0.414 dB from the 20 Ω-cm silicon substrate and −0.822 dB from
10  Ω-cm silicon substrate, respectively. The high-resistivity silicon substrate
obviously provides lower loss over the entire frequency range.
Based on these measurements and simulation results, we conclude that a
20 Ω-cm high-resistivity silicon substrate is preferred for very high-speed sig-
naling applications. Higher than 20 Ω-cm resistivity silicon substrates are also
available in the industry if there is a need to further reduce TSV insertion loss.

3.2.2.4.2 Type II Signals

3.2.2.4.2.1 Silicon Interposer Signal Routing The silicon interposer lateral
routing consists of four layers of metal: one layer for redistribution, one for
ground reference, and two for signal routing. The ground reference layer
separates the two signal-routing layers. In each signal-routing layer, every
signal wire runs parallel to a grounded side shield.
Figure 3.14 illustrates the effect of these grounded side shields on a victim
wire. Eleven interposer wires of 4.5  mm in a single metal layer on top of
a ground reference layer are simulated. The victim wire is at the center of
the eleven-wire bus. The interposer wires are modeled with 3D RLC ele-
ments. The simulation consists of 10 aggressors transitioning with a range of
delay values relative to the victim. Figure 3.14a and b illustrates the victim
waveform without and with side shields inserted into the eleven-wire bus to
eliminate overshoots and undershoots, respectively.

3.2.2.4.2.2 Interdie Timing Verification Interdie routing requires rigorous

timing analysis to verify adequate setup and hold time.
There are over 4000 Type II signals between the GTZ and its neighboring
SLR. Figure 3.15 illustrates the wire length distribution of these Type II sig-
nals. All but five signals are routed on traces no more than 6 mm in length,
54 3D Integration in VLSI Circuits

Voltage, V

Odd mode Odd mode

without side with side
shields shields

(a) Time, nS (b) Time, nS

FIGURE 3.14
Even-mode coupling (a) Even mode without side shields and (b) Even mode with side shields.

Wire length histogram

1000 3000 5000 7000 9000

Wire length (μm)

FIGURE 3.15
Type II signal wire length distribution.

whereas 85% of the traces are less than 3.75 mm in length. Clocking between
the GTZ and its neighboring SLR is system synchronous. Careful balancing
of clock networks on these two pieces of silicon across process, voltage, and
temperature is necessary. Static timing analysis (STA) verifies interdie tim-
ing efficiently. However, during XC7VH580T development, STA tools did not
accept RLC interconnect models. As a result, interdie propagation delay and
transition time values were calibrated between SPICE simulations with 3D
RLC interconnect models and STA with RC interconnect models. The layout
ensures that Type II signaling in the XC7VH580T is not in transmission-line
mode and STA with RC-based interconnects suffices.

3.2.3 Three-Dimensional Integration Chip Resource-Rich

Field-Programmable Gate Arrays Product Offerings
SSIT provides multi-terabit-per-second die-to-die bandwidth through
more than 10,000 device-scale connections—enough for the most complex
A New Class of High-Capacity, Resource-Rich FPGA 55

multidie designs. Xilinx first used this technology to create Virtex®-7 (28 nm)

FPGA products in 2011, which offered unprecedented capabilities including
up to 2 million logic cells; 65 Mb of block RAM; 2375 GMACS of DSP perfor-
mance (4750 GMACS for symmetric filters); 1200 SelectIO pins supporting
1.6 GB/s low-voltage differential signaling (LVDS) parallel interfaces; and 72
serial transceivers delivering 1886 GB/s aggregate bidirectional bandwidth.
Besides homogeneous partitioning, Xilinx also qualified and productized
heterogeneous 3D integration via silicon interposer of third-party 28 Gbps
SerDes dies in 28  nm generation [2,7]. XC7VH580T and XC7VH870T are
two heterogeneous 3D-IC integration products introduced in 2013. Other
examples include internally designed analog-to-digital converter/digital-to-
analog converter (DAC/ADC) heterogeneously integrated on an interposer
with 2 FPGA SLRs, which was demonstrated at ISSCC 2014 [3].
Since then, Xilinx has expanded the 3D-IC technology envelope further
with Ultrascale (20  nm) and Ultrascale+ (16  nm fin field-effect transistor
[FINFET]) products to support emulation, 200G/400G and datacenter applica-
tions. New development such as larger-than-reticle interposer (X-interposer)
with lithography-stitched metal lines and 600,000 μBumps interconnections
[8–10] have been qualified and are in volume production. Homogeneously
partitioned lithography-stitched devices in production include XCVU440,
XCVU190, XCVU11P, and XCVU13P. High bandwidth memory (HBM)-
integrated products have also been announced [8].

3.3 Stacked Silicon Technology Development

and Package Reliability
To meet the high 3D-IC FPGA demand, the process line in the fab needs to
maintain high yield and reliability all the time. 3D-IC FPGA assembly yield
improvement was discussed by B. Banijamali et al. [7,9], R. Chaware et al. [11]
and reliability was discussed by L. Lin et al. [12], respectively. New process
modules are also needed to enable silicon interposer TSV reveal, the so-called
middle-end-of-line (MEOL) steps. Section 3.3.1 describes key-enabling tech-
nologies and how a well-defined test vehicle played a significant role in
Xilinx’s 3D-IC technology development and reliability evaluation.

3.3.1 Key-Enabling Technologies

One possible silicon interposer MEOL scheme (soft reveal) is illustrated
in Figure 3.16. Other approaches such as flat reveal are also commonly
employed. The main difference is whether the TSV is protected by a capping
layer or not during silicon grinding. Nevertheless, these various process
schemes have several common process steps—microbump, carrier mount,
56 3D Integration in VLSI Circuits

0. i-wafer w TSV 3. Silicon thinning 6. CMP contact open

Si Grind/Polish CMP polish contact
AP (775 μm → 115 μm) (No corrosion/pit)

65 nm 4 L

TSV
(105 μm depth)

1. Microbump 4. Si recess etch 7. C4 bumping

Si recess C4 bumping (90 μm BH)
Micropad 45 μm
(115 μm → 100 μm) >180 μm

2. Carrier mount 5. P-capping 8. Carrier demount

Handling wafer Carrier demount/clean
Passivation (SiN) capping
Temporary glue

Dicing tape

Frame

FIGURE 3.16
MEOL soft reveal process flow scheme: 0—i-wafer with TSV, 1—Microbump, 2—Carrier
mount, 3—Silicon thinning, 4—Si recess etch, 5—P-capping, 6—CMP contact open, 7—C4
bumping, and 8—Carrier demount.

silicon thinning, Si etch, backside passivation, chemical–mechanical polish-

ing (CMP) contact open, C4 bumping, and carrier demount.
The μBump formation requires significant interface optimization for tight
resistance control. As the critical dimensions (CD) are much smaller than
regular C4 Cu-pillar bumps (for e.g., 25 μm vs. 50 μm), the sensitivity of resis-
tance and reliability to process variations and defect is higher.
Many factors can cause considerable yield loss in foundry TSV manufacturing
and OSAT MEOL processes, including the different definitions of critical killer
defects from existing fab/OSAT integration, glue residue/contamination at bevel
and edge areas due to plasma inhomogeneity at wafer edge, wafer bow due to
film stress, nonuniformity in etch profiles and post-CMP film thickness, and
plasma-induced or handling-induced damage at thin wafer edge. Considerable
efforts have been made to enhance line yield and interposer die yield in various
aspects of particle control, wafer-edge engineering, process uniformity improve-
ment, surface treatment/modification, thin wafer handling damage, and so on.
From a process integration perspective, statistical process control (SPC), which
is extremely useful in process control and optimization, is also applied for yield
improvement in MEOL module. The following are several case studies illustrat-
ing yield improvement efforts and substantial enhancement outcome [13].
A New Class of High-Capacity, Resource-Rich FPGA 57

3.3.1.1 Silicon-Grinding Quality Optimization

During the MEOL process, the front side of the TSV interposer wafer is
temporarily glued to a carrier wafer. Bonding glue total thickness variation
(TTV), edge bead amount, glue coverage for trimmed bevel, and bonding void
are optimized to improve line yield and accommodate subsequent process
integration. TSV interposer wafers are ground with successively finer grade
abrasives (Z1/Z2/Z3) to thin down until a certain amount of silicon is left on
TSV copper bottom. Post-grind surface property of silicon is very critical for
particle control during subsequent process steps. For example, hydrophobic
Si surface tends to repel water droplets that leave organic material. When
the water droplet is removed from the hydrophobic surface during wafer-
drying process, the watermarks remained and organic residues are formed.
This watermark defect is invisible right after grinding step and appears after
reactive ion etching (RIE) step. This is a very crucial defect mode for subse-
quent RIE etch-back and final known good die (KGD) electrical testing. It is
necessary to remove the watermark and contaminants from the post-grind
Si surface. Figure 3.17 shows that both Si CMP optimization and post-grind
surface clean could result in defect-free surface. The control of the wettabil-
ity of post-grind surface and post-CMP cleaning method plays an important
role on the removal of particles and watermarks from Si surface.

3.3.1.2 Wafer Edge and Bevel-Cleaning Optimization

During the MEOL manufacturing process, defects, carrier wafer damages,
and contamination are likely to occur at wafer edge caused by temporary
glue particulate, dry etching, film deposition, and wafer handling. These
edge defects and particles are easily transferred to polished silicon surface
during subsequent process steps, for example, RIE, chemical vapor deposi-
tion (CVD), or wafer transport steps. Figure 3.18 depicts the representative
defect images after RIE etch-back step. Foreign material (a) could come from

Method CMP 1 CMP 2 CMP 2

+CLN1 +CLN1 +CLN2 (POR)
Surface Hydrophobic Hydrophilic Hydrophilic
contact angle
Defect map

Defects Watermark, Watermark No defect

chemical (100% yield)
residue

FIGURE 3.17
Defect improvement before and after postgrind surface optimization.
58 3D Integration in VLSI Circuits

(a) (b) (c)

FIGURE 3.18
Defect types (a) foreign material, (b) Sulfur hexafluoride (SF6) by-product, and (c) pre-existing
particle masking defect.

either edge glue particulate or environment random defects. Etch by-product

(b) is the in-process particle. Hard mask defect (c) can originate from pre-
etch defects that act as Si hard mask. Wafer edge or bevel area is a primary
source of yield lowering defects. Therefore the cleanliness at wafer edge (top
and bevel) is crucial for yield enhancement.
A special wafer edge bevel cleaning (wet + dry) technique can be effective
in reducing defect source from wafer edge during RIE & CVD steps. In order
to see the impact of bevel clean in single RIE etch step, different bevel clean-
ing methods prior to RIE etch-back step were evaluated. Figure 3.19 shows
defect reduction and yield improvement due to bevel clean. The minor fall-
on particles shown in Figure 3.19c could be scrubbed away in subsequent
wafer cleaning step.

3.3.1.3 Silicon-Footing Improvement

Plasma etch has localized etch rate difference between dissimilar mate-
rials (Si and SiO2) interface due to etch radical charging on oxide surface
that causes either Si footing or notching. It becomes much worse with over-
recessed TSV tip height (or pillar height), which is more likely to occur at
wafer edge. Figure 3.20a displays the worst-case electrical failure case due to
Si footing defect particularly at wafer edge.
In order to enhance TSV tip height uniformity after RIE step, gradi-
ent Si etch rate across the wafer is applied to compensate for the thickness

No bevel clean Bevel wet clean Bevel wet + dry

clean (POR)

(a) (b) (c)

FIGURE 3.19
Die yield improvement comparing (a) without bevel clean, (b) bevel wet clean, and (c) bevel
wet + dry clean.
A New Class of High-Capacity, Resource-Rich FPGA 59

Bevel angle Pt coating center Edge

found near TSV
Bump
5,505 nm 5,560 nm

1 1 1 Silicon
22 1 1 1 1 21 TSV
1 1 1 1 1 1 1 1 21
1 1 1 1 1 1 1 1 1 1 21 Cu (b)
1 1 1 1 1 1 1 1 1 1 BS PASV 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 25 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
Oxide
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
Si 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 11
(a) (c) 1 1

FIGURE 3.20
(a) Interposer wafer edge failures caused by nonoptimized etching revealed, (b) center and
edge revealed TSVs after Si gradient etch implementation, and (c) no failures after etch
optimization.

nonuniformity on incoming wafers due to carrier bond and lapping steps.

The focused ion beam (FIB)/SEM images in Figure 3.20b show revealed TSV
tip with optimized RIE etch-back. It proves that TSV tip height variation
across wafer (center to edge) is uniform and oxide liner loss is little on top
of exposed TSV tip. With aforementioned gradient Si etching technique and
localized charge reduction effort, the TSV tip height uniformity and silicon
footing size are substantially improved and final KGD yield is almost closed
to ~100% as shown in Figure 3.20c.

3.3.2 Three-Dimensional Integration Chip Development Test Vehicles

Test chips or test vehicles (TVs) are important to understand technology,
design, and/or their interactions before a new product launch. 3D-IC is an
integration of new silicon, packaging, and 3D-IC technologies together, thus
test chips play especially an important role. Xilinx works closely with foundry
by running multiple test chips in almost all technologies (Figure 3.21).

Year 20XX 20XX+1

Quarter Q1 Q2 Q3 Q4 Q1 Q2 Q3 Q4
Month 1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12
SPICE V0.01 V0.1 V0.5 V0.5

DRM/DRC V0.01 V0.1 V0.5 V0.5

TV (MPW) TV1 TO TV2 TO TV3/3D-IC TO

FIGURE 3.21
Milestone of PDK delivery and test vehicle development for new technology node and 3D-IC.
60 3D Integration in VLSI Circuits

3.3.2.1 28 nm Test Vehicle-Driven Process and Reliability Improvements

This section discusses the 28 nm 3D-IC TV. The top die of the 28 nm 3D-IC
test vehicle (TV) was populated by 1.8 V I/O blocks with long passive scan
chains. Each of these long passive scan chains in a block has ~2.7 K of micro-
bumps. Each test chip unit contains ~110  K microbumps. Each 3D-IC test
chip consists of four top dies on an interposer (Figure 3.22).
Constrained by resources in this TV, the top die could not be probed at
wafer level to check the yield. The yield was validated at chip level and
found it was relatively straightforward to differentiate the yield scan chain
versus open/short scan chain electrically. However, the physical failure
analysis (FA) feedback loop takes longer time to identify the failure site
from the long chain. The typical 3D-IC FA flow used in Xilinx is illustrated
(Figure 3.23).
After e-test at chip level, x-ray and confocal scanning acoustic micros-
copy (C-SAM) detection are the frequently used constructive inspection
approaches to validate the test results, in particular failing or delamination
locations. In Figure 3.24, one of the typical FA case results is shown.
Detecting microbump interconnect robustness is a main focus in Xilinx’s
28 nm 3D-IC TV as the high microbump count (~200,000 in largest device)
requires joining yield to be extremely high. The TV helped to detect either
microbump process systematic (defect, particle/residue, chip warpage) or
excursion issues during the development period. If looking back and weigh-
ing the significance of the 3D-IC TV, one would recognize it as a very suc-
cessful 3D-IC TV that paved the foundation of microbump process baseline.

Bali top die

Microbump
Interposer
C4 bump
Interposer

Substrate

BGA

∼27.1 mm
Bali top
dies
Bali top die
Interposer
∼7.5 mm
∼32.6 mm

∼24.7 mm

FIGURE 3.22
A schematic of Xilinx 28 nm 3D-IC TV (code named “Bali”).
A New Class of High-Capacity, Resource-Rich FPGA 61

Electrical test

X-ray inspection

C-SAM
(Confocal scanning acoustic microscopy)

TDR
(Time domain reflectometry)

OBIRCH
(Optical beam-induced resistance charge)

Delayer/X-section

FIGURE 3.23
A typical 3D-IC physical FA working flow.

BGA

X-ray image TDR CSAM image OBIRCH image

FIGURE 3.24
One typical FA case using X-ray and C-SAM inspection approaches.

In fact the TV helped with choosing chip-on-wafer (CoW) as the preferred

assembly integration platform (Figure 3.25). In Figure 3.26, typical micro-
bump open/short cases are demonstrated.

3.3.2.2 Improvements to 20 nm Test Vehicle

In Xilinx’s 20 nm 3D-IC TV, to meet the product features (microbump scal-
ing from pitch 45–40  μm with 600,000  microbumps and X-interposer that
has a size larger than a regular lithography reticle size), long passive scan
chain was replaced by defect-monitoring vehicle (DMV). Xilinx has been
using DMV for many generations due to its rapid defect detection capability
and accuracy. DMV is large and sensitive enough to catch PPM-level open
and short defects. The DMV was redesigned to fit into 3D-IC development
for X-interposer with 40 μm pitch microbump defect detection. Figure 3.27
illustrates the floor plan of X-interposer with 3 slices of DMV (each slice is
23.25 × 14.56 mm) that contains a total of ~375 K microbumps.
62 3D Integration in VLSI Circuits

CoWoS CoW-Last
∼ 100 μm FPGA FPGA FPGA FPGA

Hold
Maximum die-level warpage (μm)

UF UF

Full thickness wafer

Cavity-handling wafer

CoS
FPGA FPGA
UF1
UF

UF2 ∼ −150 μm

Organic substrate
Concave warpage

0 50 100 150 200 250 300

Temperature (°C)

FIGURE 3.25
Warpage behavior during microbump joining for different integration schemes.

FIGURE 3.26
Demonstration of one typical microbump open and short FA case, respectively.

As expected, DMV’s failing location identification capability and efficiency

assisted to identify open/short microbump defect. In particular, the DMV
can have wafer-level test capability implemented so the KGD can be achieved
before 3D-IC integration on X-interposer. This 20 nm 3D-IC TV also helped to
identify the μBump density variation versus yield sensitivity, which led to a
specific design design for manufacturability (DFM) establishment.
Board-level reliability (BLR) of package is another important aspect to con-
sider from a system-use perspective. Introducing new technology elements
such as μBump, multidie, X-interposers, and large package sizes requires a
careful study of failure modes to ensure a good understanding of product
performance in cycling, shock, bend, and other tests.
Xilinx, working with our lead customer inputs, defined a special vehicle
with hooks that go well beyond traditional ball grid array (BGA) daisy-chain
monitoring to include C4, TSV, µBumps, and die edge crack monitoring. The
focus was on indoor Telecom use conditions that are stringent due to use of
A New Class of High-Capacity, Resource-Rich FPGA 63

Top die
die seal ring

Interposer 14.56 mm
DMV TD

Interposer
die seal ring
DMV TD
45.3 mm

23.25 mm

DMV TD

24.7 mm

FIGURE 3.27
20 nm 3D-IC TV with 3 DMV top dies on X-interposer containing ~375 K μBumps.

very thick, high-layer count boards (high stress on BGA), and long life time
expectation. In Section 3.3.2.3, we briefly discuss the vehicle and results seen
in board level 0°C–100°C thermal cycling and Shock and Bend.

3.3.2.3 Board-Level Reliability Test Vehicle and Results

A specially designed test vehicle was developed to conduct BLR test.
Traditionally, per IPC-9710, BGA solder balls are monitored through daisy
chains in both package peripheral and under die perimeter areas. For the
X-interposer package, daisy chains were added to monitor C4 bumps, TSV,
and microbumps. To achieve this purpose, top die was redesigned, inter-
poser and substrate from the original component package, as shown in
Figure 3.28, which illustrates a daisy chain that includes BGA ball, C4 bump,
TSV, and μBump.
The package was assembled on a BLR board and tested with and without
heat sink. The board is 140 × 383 mm2 and 3.42 mm thick. It has 28 metal
layers with high-speed, low-loss dielectric material. Board design, assembly,
and test are based on IPC-9710. The BLR test was done using TC1 Condition
from 0°C to 100°C with 10 minutes of ramp up and cool down along with
10 minutes of dwell time. One complete cycle takes ~40 minutes.
64 3D Integration in VLSI Circuits

Top die AP layer

Interposer FPGA Die FPGA Die FPGA Die
AP layer μBump
TSV
C4 bumps
Substrate
BGA balls
Board

FIGURE 3.28
20 nm 3D-IC board-level reliability (BLR) test vehicle.

First failure without heat sink occurred at 8619 cycles and characteristic
life (63.2% failure) was reached at 10,792 cycles. Dye and pry FA on the first
failed unit indicated the failure was caused by solder ball crack at pack-
age corner ball, as shown in Figure 3.29. All failed units have first failure
at corner solder balls and not on C4 bump or μBump. This agreed with
simulation that the X-interposer package has similar BLR characteristic as
a large die flip chip package where corner balls fail early and should be

(a)

Failing ball

(b)

FIGURE 3.29
(a) Passing sample, no corner ball crack (with six corner depopulated pads). (b) Failed sample
with corner ball crack (with six corner depopulated pads).
A New Class of High-Capacity, Resource-Rich FPGA 65

considered critical. Including heat sink, onset of first failure was sooner
at 5092 cycles and characteristic life was reached at 7565 cycles. However
failure mode remained the same, that is, corner BGA balls failed and there
was no impact to elements within the package such as μBump, C4, FPGA
die ELK, interposer, TSV, and substrate.
The package was also assembled on a Shock and Bend board and tested
without heat sink. The board is 185 × 185 mm 2 and 3.2 mm thick. It has 16
metal layers with high-speed, low-loss dielectric material. Board design,
assembly, and test are based on JESD22-B110A for Shock and JESD22-B113
for Bend. The failure criteria for Shock are 10% increase in daisy-chain
net resistance and for Bend test is 20% net resistance increase. Results
showed Shock test passed both 100G of Condition C and an extended
test of 125G; for Bend, the global strain to failure ranged from 1775 to
2359 microstrains.

3.3.3 Three-Dimensional Integration Chip Reliability

Look-Ahead Assessments
Reliability is always a significant and inseparable part of technology devel-
opment. Therefore Xilinx has been constantly running iterations of reliabil-
ity assessments to discover and to improve any hidden defect, technology
weakness, or testing/screening coverage issues. These reliability assess-
ments are especially important during 3D-IC development.
The main 3D-IC reliability assessment (component level), which include tem-
perature cycling (TC, −55°C ~125°C), high-temperature storage (HTS, 150°C),
and unbiased highly accelerated test condition-A (uHAST, 130°C/85%RH)
with MSL-4 preconditioning (30°C/60%RH/96  hr + 245°C Reflow×3), was
applied before these reliability tests.
Figure 3.30a illustrates a 28  nm μBump HTS failure at 1000  hours SEM
cross section, significant Sn sidewall wetting, and voiding. The 28 nm μBump

(a) (b)

FIGURE 3.30
HTS 150°C 1000 hours failing μBump shown in (a) and an improved one shown in (b) (post HTS
150°C 4000 hours of stress).
66 3D Integration in VLSI Circuits

process was then further improved from metallurgical perspective to enhance

the robustness. With the improved μBump, HTS 4000 hours passed without
electrical failure (although minor void was observed [Figure 3.30b]).

3.3.3.1 Continuous Process Yield Improvements

KGD is a critical issue for all multidie integration. Xilinx has built-in test
capability, thus its 3D-IC top dies are fully tested at wafer level, similar to
their monolithic dies done normally. However it is more challenge to do
built-in test in passive silicon interposer KGD. A new approach connects
the power and ground network in the interposer die into a giant open-short
test network in scribe area enabling almost KGD test capability. After dic-
ing, these open-short test network return to their normal power and ground
function (Figure 3.31).
In addition to design for testability (DFT), redundancy and self-repairing
schemes have also been developed [14]. Due to these yield improve-
ment efforts, both 28 and 20 nm largest FPGA (each having ~200,000 and
~600,000 μBumps, respectively) yielded well.
Figure 3.32 illustrated another example of C4 joining yield improvement.
Dummy dies were included in chip-on-wafer (CoW) integration, which
results in ~10%–15% C4 joining yield gain in a 28 nm heterogeneous 3D-IC
product (XC7VH870T). The C4 bridging (caused by warpage of the CoW) at
corners of the interposer was then eliminated.
Three generations of Xilinx’s FPGA products (28, 20, and 16 nm) are now
shipping in production volumes with their defect density approaching
monolithic level. Comprehensive reliability tests well beyond JEDEC have
been conducted to understand new failure modes and solutions. All of these
are due to an effective test vehicle and development strategy.

Die sawing

FIGURE 3.31
Plan view of interposer die highlighted with power-ground lines.
A New Class of High-Capacity, Resource-Rich FPGA 67

Dummy Dummy

Dummy Dummy

(a) (b)

FIGURE 3.32
~10%–15% C4 bridging at four corners of the XC7VH870T was detected (a). The C4 bridging was
eliminated after dummy dies included in CoW integration (b).

3.4 Potential Three-Dimensional Integration

Chip Future Challenges
3D-IC will be getting more tractions and acceptance from the industry
in future. One reason is because 3D-IC technology is reaching maturity,
whereas other driving forces and advanced silicon technology continuous
scaling trend are becoming more challenging and more expensive.
The challenges expected and which the industry should resolve soon to
penetrate high-volume mainstream adoption are given in the following:

1. The cost-down tendency and margin of the 3D-IC: It will become a key
factor to promote 3D-IC adoption by new markets other than emula-
tion, hyperscale data center, and networking.
2. The 3D-IC capacity and/or supply chain expansion to keep pace with the
adoption demand: Today’s limited capacity is expected to be con-
sumed by one or two major fabless design houses. More capacity
investment is pursued by foundry and OSATs but the pace needs to
accelerate.
3. The need for an innovative/effective 3D-IC development learning vehicle:
How would it be possible to define a test vehicle as close as to the
future 3D-IC product in the development period of a new technol-
ogy node?
68 3D Integration in VLSI Circuits

For vertical stacking integration the challenges are more pronounced to iden-
tify the technology, design/architecture, and power/thermal space. In the
future with advanced silicon technology scaling only becoming more chal-
lenging not only from technical but an economic standpoint, it is expected
that 3D vertical stacking would gain more interest and momentum in the
industry.

Acknowledgment
The authors would like to express our appreciation for the great support
from Xilinx’s architecture, design, reliability, and FA teams and our partners
TSMC’s 3D-IC development team, TSMC BTSD team, and SPIL CRD Team.
Particular thanks go to our management team for their full commitment and
support right throughout the program.

References
1. K. Saban, Xilinx stacked silicon interconnect technology delivers breakthrough
FPGA capacity, bandwidth, and power efficiency, Xilinx White Paper: Virtex-7
FPGAs, WP380 (Initial v1.0) October 2011 and (updated v1.2), December 11,
2012.
2. L. Madden, E. Wu, N. Kim, B. Banijamali, K. Abugharbieh, S. Ramalingam, and
X. Wu, Advancing high performance heterogeneous integration through die
stacking, in Proceedings of the IEEE 38th European Solid-State Circuits Conference
(ESSCIRC), Bordeaux, France, 2012, pp. 18–24.
3. C. Erdmann, D. Lowney, A. Lynam, A. Keady, J. McGrath, E. Cullen,
D. Breathnach et al., A heterogeneous 3D-IC consisting of two 28 nm FPGA die
and 32 reconfigurable high-performance data converters, IEEE Journal of Solid-
State Circuits, 50(1): 258–269, 2015.
4. S. Wu, J. Wei, and H. Lam, A new SPICE simulation approach for 3D IC integra-
tion, SNUG 2013.
5. N. Kim, D. Wu, D.W. Kim, A. Rahman, and P. Wu, Interposer design optimiza-
tion for high frequency signal transmission in passive and active interposer
using through silicon via (TSV), in Proceedings of the 2011 IEEE 61st Electronic
Components and Technology Conference (ECTC), Lake Buena Vista, FL, 2011,
pp. 1160–1167.
6. N. Kim, D. Wu, J. Carrel, J.H. Kim, and P. Wu, Channel design methodology for
28Gb/s SerDes FPGA applications with stacked silicon interconnect technol-
ogy, in Proceedings of the 2012 IEEE 62nd Electronic Components and Technology
Conference (ECTC), San Diego, CA, 2012, pp. 1786–1793.
A New Class of High-Capacity, Resource-Rich FPGA 69

7. B. Banijamali, S. Ramalingam, K. Nagarajan, and R. Chaware, Advanced reli-

ability study of TSV interposers and interconnects for the 28nm technology
FPGA, in Proceedings of the 2011 IEEE 61st Electronic Components and Technology
Conference (ECTC), Lake Buena Vista, FL, 2011, pp. 285–290.
8. Xilinx, Inc., All programmable 3D ICs, http://www.xilinx.com/products/
silicon-devices/3dic.html.
9. B. Banijamali, H. Liu, S. Ramalingam, I. Barber, T. Lee, J. Chang, M. Kim, and
L. Yip, Reliability evaluation of an extreme TSV interposer and interconnects
for the 20  nm technology CoWoSTM IC-package, in Proceedings of the IEEE
Electronic Components and Technology Conference (ECTC), San Diego, CA, May
26–29, 2015, pp. 276–280.
10. G. Hariharan, L. Yip, R. Chaware, I. Singh, M. Shen, K. Ng, and A. Xu, Reliability
evaluations on 3D IC Package beyond JEDEC, in Proceedings of the IEEE 67th
Electronic Components and Technology Conference (ECTC), Lake Buena Vista, FL,
2017, pp. 1517–1522.
11. R. Chaware, G. Hariharan, J. Lin, I. Singh, G. O’Rourke, K. Ng, and S. Pai,
Assembly challenges in developing 3D IC package with ultra high yield and
high reliability, in Proceedings of the IEEE Electronic Components and Technology
Conference (ECTC), San Diego, CA, May 26–29, 2015, pp. 1147–1451.
12. L. Lin, Reliability characterization of chip-on-wafer-on-substrate (CoWoS) 3D
IC integration technology, in Proceedings of the IEEE Electronic Components and
Technology Conference (ECTC), Las Vegas, NV, May 28–31, 2013, pp. 366–371.
13. W.S. Kwon, M. Kim, J. Chang, S. Ramalingam, L. Madden, G. Tsai, S.  Tseng,
J.Y. Lai, T. Lu, and S. Chiu, Enabling a manufacturable 3D technologies and
ecosystem using 28  nm FPGA with stack silicon interconnect technology, in
Proceedings of the 46th International Symposium on Microelectronics (IMAPS),
Orlando, FL, 2013, pp. 217–222.
14. R.C. Camarota, J. Wong, H. Liu, and P. McGuire, Applying a redundancy
scheme to address post-assembly yield loss in 3D FPGAs, in Digest of Technical
Papers, 2014 Symposium on VLSI Technology, Hawaii, HI, 2014, pp. 150–151.
4
Challenges in 3D Integration

M. Koyanagi, T. Fukushima, and T. Tanaka

CONTENTS
4.1 Introduction .................................................................................................. 71
4.2 Past Challenges in Three-Dimensional Integration ............................... 72
4.3 Challenges in Three-Dimensional System Integration .......................... 73
4.4 Challenges in Three-Dimensional Heterogeneous Integration ............ 76
4.5 Challenges toward Future Three-Dimensional Integration .................. 79
4.6 Summary....................................................................................................... 81
References............................................................................................................... 81

4.1 Introduction
The 3D integration technology using through-silicon via (TSV) has
significantly progressed for these years as represented by 3D-stacked
dynamic random-access memory (DRAM) such as hybrid memory
cube (HMC) and high bandwidth memory (HBM) [1,2]. The 3D-stacked
structure is also employed in a complementary metal–oxide–semiconductor
(CMOS) image sensor (CIS) [3]. In addition to these 3D-stacked DRAM
and 3D-stacked image sensor, heterogeneous 3D integration technology
has increasingly attracted much attention as it is indispensable for future
Internet of Things (IoT). In a heterogeneous 3D integration technology,
different kinds of chips such as microelectromechanical systems (MEMS),
sensor, photonic device chip, and spintronic device are stacked on CMOS
chips. Low-power consumption, small form factor, and multifunctionality
are required for embedded devices in IoT. Heterogeneous 3D integration
can provide these embedded devices with low-power consumption, small
form factor, and multifunctionality. We have developed new heterogeneous
3D integration and system integration technologies using self-assembly

71
72 3D Integration in VLSI Circuits

to achieve multichip-to-wafer stacking  with high throughput and high

precision. In this chapter, these new heterogeneous 3D integration and
system integration technologies for IoT are described.

4.2 Past Challenges in Three-Dimensional Integration

The first 3D large-scale integration (LSI) test chip having three device
layers was reported by Akasaka and Nishimura [4]. In this 3D LSI test
chip,  metal–oxide–semiconductor field-effect transistors (MOSFETs) are
fabricated in poly-Si films deposited on an LSI wafer. The poly-Si film
is recrystallized by laser annealing before fabricating the MOSFETs to
improve the electrical characteristics of the poly-Si film. A 3D LSI fabri-
cation process using a thinned-wafer transfer method was proposed by
Hayashi et  al.  [5]. In this 3D LSI process, an LSI wafer is glued to a sup-
porting material and then the LSI wafer is thinned from the back surface
by mechanical grinding and polishing to a thickness of approximately
0.5  μm. This thinned LSI wafer is directly bonded to another LSI wafer
and then the supporting material is removed from the thinned LSI wafer.
To avoid the transfer of extremely thinned wafer, we proposed a 3D LSI
with vertical interconnection called a TSV as shown in Figure 4.1a in 1989
[6–9]. We have developed four kinds of TSVs so far as shown in Figure 4.1b.
We fabricated various 3D prototype chips using poly-Si TSV, W-TSV, and
Cu-TSV. Our first 3D prototype chip is a 3D-stacked image sensor with
poly-Si TSV fabricated in 1999 as shown in Figure 4.2 [10]. Furthermore we
fabricated a 3D-stacked memory in 2000, 3D-stacked artificial retina chip
in 2001, and 3D-stacked processor in 2002 using poly-Si TSV [11–13].

0.7 μm 12 μm
Supporting
Cu material
2.5 um
Glue layer
30 μm

W Metal wiring
18 μm
18 μm

Metal microbump TSV Poly-Si

MOSFET MOSFET SiO2 Active

7 μm
region

Si MOSFET Si Si 3 μm Cu
SiO2
Poly-Si TSV W-TSV Cu-TSV Cu-TSV
MOSFET (Via first) (Via middle) (Via middle) (Via last)
Bonding layer Si substrate
(1995) (2001) (2006) (2007)
(a) (b)

FIGURE 4.1
Cross-sectional structure of 3D LSI with TSVs (a) and cross-sectional view of TSVs (b).
Challenges in 3D Integration 73

First layer Second layer Third layer

(Photosensor (Register (ADC and
circuit) circuit) ALU circuit)

Image sensor array

Pipeline operation

Parallel processing
P
C

AMP
M
D
A

and ADC
10 11
00 01 02
SHIFT
register
cu ut
it
cir up

ALU and output

LC
O

circuit
Configuration of Array of processing units
one processing unit to form one frame

Configuration of block paralled

3D-stacked image sensor Quartz class
Si interposer
Stacked chips

Photo of 3D-stacked image sensor

FIGURE 4.2
Photomicrographs of fabricated 3D-stacked image sensor test chip with poly-Si TSVs.

4.3 Challenges in Three-Dimensional System Integration

We can create new intelligent system modules for not only high-end infor-
mation system application but also mobile and consumer applications by
employing the 3D integration technology [14,15]. A typical example of these
intelligent system modules is a 3D-stacked image sensor system module for
advanced driver assistance systems (ADAS) as shown in Figure 4.3 where
two 3D-stacked image sensors and 3D-stacked multicore processors are inte-
grated on a silicon interposer [16,17].
A prototype 3D-stacked CMOS image sensor with the resolution of quarter
video graphics array (QVGA) (320 × 240 pixels) and the frame rate of 10,000
frames/s was fabricated by the 3D integration technology with the backside-
via [18]. A configuration of a prototype 3D-stacked CMOS image sensor
with a block-parallel architecture that contains a number of image signal-
processing elements is shown in Figure 4.4 [19–21]. The 3D-stacked CMOS
image sensor consists of four layers of image sensor, correlated double sam-
pling (CDS) with programmable gain amplifier (PGA) array, analog-to-digital
converter (ADC) array, and interface circuit array. The CMOS image sen-
sor (CIS) layer and CDS layer were designed with a standard 0.18 μm CMOS
image sensor and mixed signal technologies, respectively. We have employed a
block-parallel architecture for ADC to convert a large amount of analog data
from image sensors to digital data with high speed. This block-parallel ADC
architecture employs a 9-bit time-interleaved successive approximation (SAR)
74 3D Integration in VLSI Circuits

r
so
sen Image sensor chip
age Analog chip
d im ADC chip
ke Interface chip
st ac
3 D-

Si interposer
3D-stacked processor

FIGURE 4.3
Configuration of 3D-stacked image sensor system module for advanced driver assistance sys-
tems (ADAS).

Frame rate: 10,000 frames/s

(High-speed digital output) Pixel
circuit

No.00
Image sensor Image sensor (CIS) Pixel No.01

Correlated double Correlated Pixel No.n

double CIS
sampling (CDS) sampling (CDS) TSV TSV
A/D converter A/D converter (ADC)
(ADC)
Amp
Si interposer CDS
Si TSV
for testing interposer
(a) SAR
ADC
TSV
(b)

FIGURE 4.4
Configuration of one image signal processing element in 3D-stacked image sensor. (a) Structure
of stacked image sensor and (b) configuration of one circuit block.

ADC with digital noise  cancellation  circuit. The  ADC was designed with a
standard 90-nm 1-Poly 9-Metal CMOS technology [20]. The three-dimensional
structure  of  designed 3D-stacked image sensor is depicted in Figure 4.5a.
Figure 4.5b shows X-ray CT scan image and SEM cross-sectional view of fabri-
cated 3D-stacked image sensor. It is clearly seen in the X-ray CT scan image that
four layers with many TSVs are vertically stacked. The die size of each layer is
5 × 5 mm2 and each layer has approximately two thousands of TSVs. The thick-
ness of Si substrate and the diameter of TSV are approximately 50 and 5 μm,
respectively. We confirmed the successful operation of fabricated prototype
3D-stacked image sensor.
A prototype 3D-stacked dependable multicore processor was fabricated by
the 3D integration technology with the backside-via [22–23]. The conceptual
Challenges in 3D Integration 75

Cu TSV M1

Metal pad Oxide Oxide

5 μm

4th layer TSV

rd M1
3 layer TSV

nd
Cu TSV

50 μm
2 layer TSV
st Cu TSV
1 layer TSV
5 μm

Si Sub.
Metal μ-bump

(a) (b)

FIGURE 4.5
Three-dimensional display of designed 3D-stacked image sensor by computer graphics (a)
and X-ray CT scan image and SEM cross-sectional view of fabricated 3D-stacked image
sensor (b).

Vertical bus using TSVs

OSTC PBB Core VBB SM Mem

Ctrl
Memory Tier 3
layers
OSTC PBB Core VBB SM Mem
Ctrl
Tier 2

Processor OSTC PBB Core VBB SM Mem

Ctrl System
layers Tier 1
Test bus bus

memory
External
OSTC PBB Core VBB SM Mem
Self-test/ Tier 0 Ctrl
Self-repai Test access port
layer (3D TAP) System bus Memory bus
(a) (b)

FIGURE 4.6
Conceptual structure (a) and functional circuit block diagram (b) of 3D-stacked multicore pro-
cessor with self-test and self-repair function.

structure of this 3D-stacked multicore processor is illustrated in Figure 4.6.

A circuit block diagram and a die photo of core processor fabricated by a
standard 90-nm 1-Poly 9-Metal CMOS technology are shown in Figure 4.7a.
The die size is 5 × 5mm2. Approximately two thousands of TSVs per each
core processor layer were assigned to form the system bus, memory bus, and
test bus including TSVs in I/O circuits. This core processor chip exhibited the
performance of 350 Mips (Dhrystone 2.1) at a clock frequency of 200 MHz.
We fabricated a four-layer stacked multicore processor and a four-layer stacked
cache memory. Figure 4.7b shows the X-ray CT scan image and SEM cross-
sectional view of fabricated 3D-stacked multicore processor. It is clearly
seen in the figure that the stacked structure with many TSVs is successfully
formed. We confirmed using internal built-in-self-test (BIST) circuits that
the fabricated prototype 3D-stacked multicore processor exhibited excellent
characteristics [24].
76 3D Integration in VLSI Circuits

Tier 1 TSV

Multilevel
TSVs for Cu wiring
3D-stacked
memory Tier 0 8 μm 50 μm μ-bump

3D-shared TSVs for I/O

General- memory buffer, power
controller
purpose ground
processor core Local Tier 3
5.0 mm

system bus TSVs for

vertical
Vertical
system bus
Tier 2
system bus
bridge
PLL Tier 1 Cu-TSV
Peripherals
TAP Cu-RDL/
Input/Output buffer, power, ground Tier 0 CuSn μ-bump
5.0 mm
(a) (b)

FIGURE 4.7
A circuit block diagram and a die photo of core processor (a) and X-ray CT scan image and SEM
cross-sectional view of fabricated 3D-stacked multicore processor (b).

4.4 Challenges in Three-Dimensional

Heterogeneous Integration
We challenge to achieve a new heterogeneous 3D LSI for IoT application
called a superchip as shown in Figure 4.8 by stacking various kinds of chips
with different sizes, different devices, and different materials. We have
developed a new 3D integration technology as shown in Figure 4.9 to fab-
ricate such a new heterogeneous 3D LSI. We call this new 3D integration
technology a reconfigured wafer-to-wafer (R-W2W) 3D integration technol-
ogy [25–27]. A  reconfigured wafer is fabricated by bonding many known
good dies (KGDs) onto a carrier wafer and then these reconfigured wafers
are bonded in this 3D integration technology. We have developed new
self-assembly and electrostatic temporary bonding methods to bond many

MEMS chip
3D super chip Sensor chip
CMOS RF-IC
MMIC
Power IC
Control IC
Logic LSI
Flash memory
DRAM
SRAM
Microprocessor

FIGURE 4.8
An example of heterogeneous 3D LSI.
Challenges in 3D Integration 77

Known good Third layer chip Reconfigured wafer

die (KGD)
Failure die

Second layer chip

First layer chip Target

interposer
wafer

LSI wafer with/without TSV Carrier wafer 3D super chip

FIGURE 4.9
Production procedure for heterogeneous 3D LSIs in reconfigured wafer-to-wafer (R-W2W) 3D
integration technology.

Hydrophilic

Hydrophobic
Dielectric Hydrophobic
layer KGD
Water
SAE carrier

Bipolar electrode

Chip Chip Chip Chip Chip

Water droplet
Wafer

FIGURE 4.10
Concept of self-assembly and electrostatic temporary bonding and a photo of 8-inch wafer
with hydrophilic areas and hydrophobic areas after simultaneously supplying liquid droplets
with four sizes on hydrophilic areas.

known good dies (KGDs) onto a carrier wafer with high alignment accuracy
and high throughput. The surface tension of liquid droplet is utilized in the
self-assembly to simultaneously align many dies as shown in Figure  4.10.
Hydrophilic areas and hydrophobic areas are formed on the surface of
wafer or chip. We have succeeded in simultaneously aligning five hundreds
of chips with the average alignment accuracy of 0.5 μm within 0.1 second.
78 3D Integration in VLSI Circuits

Cu sidewall
interconnection
MEMS chip

LSI chip

Substrate

FIGURE 4.11
A photo of heterogeneous 3D LSI test chip where a pressure sensor MEMS chip is integrated
on CMOS chip, which is stacked on an active interposer wafer.

So far we have fabricated several heterogeneous 3D LSI test chips by stacking

MEMS chip, spin memory chip, and photonic device chip on CMOS chips
using self-assembly [28–30]. Figure 4.11 shows a photo of heterogeneous 3D
LSI test chip where a pressure sensor MEMS chip is integrated on CMOS
chip [28]. The process flow to fabricate 3D LSIs by the R-W2W integration
technology is described in Figure 4.12 [31]. Many chips (KGDs) with metal
microbumps and nonconductive films (NCFs) are simultaneously face-down
bonded onto the carrier wafer after self-assembly. Then these chips are simul-
taneously thinned from the backside, and metal microbumps are formed on

1 6
NCF
μ-bump
TSV

2 7 Top chip φ 8 μm
KGD
TSV
Hydrophobic
3 CF
NCF 45 μm thick
Water 8 chip
SiO2
4
Die stacking Wafer
9
5

FIGURE 4.12
Fabrication process flow to fabricate 3D LSIs by a reconfigured wafer-to-wafer (R-W2W) inte-
gration technology. 1—Lamination, 2—dicing, 3—first-layer die release, 4—self-assembly
(alignment), 5—die-bonding, 6—Wafer thinning/TSV and μ-bump formation, 7—second-
layer die release, 8—self-assembly and die-bonding, and 9—completed 3D stacking.
Challenges in 3D Integration 79

the backside surface of chips after thinning. A number of 3D-stacked LSIs

can be simultaneously fabricated by repeating this sequence. Cross-sectional
photomicrograph of 3D-stacked test chip fabricated according to such a 3D
integration technology is also shown in Figure 4.12.

4.5 Challenges toward Future Three-Dimensional Integration

We are challenging to achieve a future high-performance and low-power
3D LSI with high-density TSV and microjoint of more than one million.
The pitches of TSV and microjoint will be reduced down to less than
5 μm to achieve such a 3D LSI. The size of microjoint and the diameter of
TSV should be less than 2 and 1  μm, respectively. We have proposed to
employ W-TSV for a via-middle and Ni-TSV for a via-last to achieve such a
fine TSV. We have succeeded to completely fill W into the deep silicon hole
with the diameter of 0.5 μm and the aspect ratio of more than 30 by atomic
layer deposition (ALD) [32]. We also filled Ni into the deep silicon hole by
electroless plating [33]. A metal barrier layer and a seed layer for electro-
plating are not necessary in W-TSV and Ni-TSV. However, we still need
Cu-TSV with larger diameter for power delivery. Therefore, W/Cu or Ni/Cu
hybrid TSVs will be employed in a future 3D LSI [34]. To further increase the
TSV density in a future 3D LSI, the diameter of TSV for signal propagation
should be less than 0.1 μm. A completely new method will be needed to fab-
ricate such a TSV with extremely small diameter. Then we propose a novel
method to fabricate a fine TSV with the diameter of less than 0.1 μm using
a directed self-assembly (DSA) of diblock copolymer. Figure 4.13 shows the
new concept of TSV formation based on advanced DSA with nanocompos-
ites consisting of diblock copolymers and nanosized metal particles [35].
TSVs are formed by the phase separation of block copolymers with simple
heating at below 300°C. Nanosized metal particles are included into copo-
lymer with cylindrical structure after the phase separation of copolymers.
To confirm the capability to fabricate such a fine TSV using DSA, we filled
diblock copolymer with metal nanoparticles into a deep Si trench with the
diameter of 3 μm and the depth of 10 μm and confirmed that nano-ordered
lamella and cylindrical structures were formed in the diblock copolymer,
PS-b-PMMA with the molecular ratio of 1:1 and 2:1, respectively, after the
phase separation at 280°C. The PS-b-PMMA with higher molecular weight
showed a larger width of the periodic structure of more than 50 nm.
The microjoint with the small size of less than 2  μm is indispensable to
fabricate a fine TSV with the diameter of less than 1 μm. A hybrid bonding
may be useful to bond two wafers with many microjoint with the small size
of less than 2 μm [36]. However it is difficult to employ a conventional hybrid
bonding in chip-to-chip bonding and chip-to-wafer bonding. Then we have
80 3D Integration in VLSI Circuits

Si substrate
Metal
Insulator
Supporting substrate
(a)
Copolymer

10 μm 2.5 μm

Si substrate
Si
Oxide liner (DSA guide) TEOS liner
(b)

Diblock copolymer TEOS liner

including metal (SiO2)
nanodots with PS PMMA
low melting point

(c) 200 nm Si

Metal filament Cross-sectional structure of diblock

(TSV) copolymer inside deep hole

Copolymer A Copolymer B
(d)

FIGURE 4.13
New concept of TSV formation based on advanced directed self-assembly (DSA) with nano-
composites consisting of diblock copolymers and nanosized metal particles. (a) Thinning of
Si substrate, (b) formation of Si deep hole (RIE), (c) deposition of diblock copolymer including
metal nanodots with low melting point, and (d) nanophase separation.

proposed a new hybrid bonding method for chip-to-chip bonding and chip-
to-wafer bonding in which a novel inorganic anisotropic conductive film
(i-ACF) comprising ultrahigh density of Cu nanopillar (CNP) and alumina
matrix is used. SEM images of i-ACF film with the top view and the cross-
sectional view are shown in Figure 4.14. The i-ACF film is formed by anodic
oxidation of aluminum film and Cu electroplating [36]. The diameter and
pitch of Cu nano-pillars are 60 and 100  nm, respectively. In order to eval-
uate the electrical characteristics of Cu–Cu joining bonded through the
i-ACF film, we fabricated a test chip with a huge number of Cu electrodes of
4.3 million. The size and pitch of Cu electrodes are 3 and 6 μm, respectively.
The cross-sectional image of fabricated test chips bonded using i-ACF film
is also shown in Figure 4.14. Test chip and interposer chip are electrically
connected by ultrahigh-density CNPs. We confirmed that all of 4.3 million
Cu joints with the electrode size of 3 μm were completely connected through
ultrahigh-density CNPs [37–38]. The joining resistance was approximately
30 mΩ for each pair of joining.
Challenges in 3D Integration 81

CNP Alumina

100
nm

(Top view of i-ACF) Upper chip

CNP Alumina

20 μm i-ACF

Lower wafer Cu electrode

(Cross-sectional view CNP: Cu nanopillar (ϕ 60 nm)
of i-ACF)
(a) (b)

FIGURE 4.14
SEM images of i-ACF film with the top view and the cross-sectional view (a) and cross-sectional
image of fabricated test chip bonded using i-ACF film (b).

4.6 Summary
We have developed various 3D integration technologies using TSV and
metal microbump for a long time. In addition, we have also developed vari-
ous 3D system-on-chips and heterogeneous 3D LSIs using self-assembly and
electrostatic bonding. Furthermore, we are challenging to develop new tech-
nologies such as DSA TSV and CNP hybrid bonding for future 3D LSIs with
high-density TSVs and microjoints.

References
1. J. Jeddeloh and B. Keeth, Hybrid memory cube new DRAM architecture
increases density and performance, Digest of Technical Papers, Symposium on VLSI
Technology, pp. 87–88 (2012).
2. D. U. Lee et  al., A 1.2V 8Gb 8-channel 128GB/s high-bandwidth memory
(HBM) stacked DRAM with effective microbump I/O test methods using
29nm process and TSV, IEEE International Solid-State Circuits Conference (ISSCC),
pp. 432–433 (2014).
3. S. Sukegawa et al., A 1/4-inch 8Mpixel back-illuminated stacked CMOS image
sensor, IEEE International Solid-State Circuits Conference (ISSCC), pp.  484–485
(2013).
4. Y. Akasaka and T. Nishimura, Concept and basic technologies for 3-D IC struc-
ture, IEEE International Electron Devices Meeting (IEDM), pp. 488–491 (1986).
82 3D Integration in VLSI Circuits

5. Y. Hayashi et  al., A new three dimensional IC fabrication technology,

stacking thin film dual-CMOS layers, IEEE International Electron Devices
Meeting (IEDM), pp. 657–660 (1991).
6. M. Koyanagi, Roadblocks in achieving three-dimensional LSI, Proceedings of
the 8th Symposium on Future Electron Devices, pp. 50–60 (1989).
7. T. Matsumoto et al., Three-dimensional integration technology based on wafer
bonding technique using micro-bumps, International Conference on Solid State
Devices and Materials (SSDM), pp. 1073–1074 (1995).
8. T. Matsumoto et al., New three-dimensional wafer bonding technology using
the adhesive injection method, Japanese Journal of Applied Physics, 1(3B): 1217–1221
(1998).
9. M. Koyanagi et al., Future system-on-silicon LSI chips, IEEE Micro, 18(4): 17–22
(1998).
10. H. Kurino et al., Intelligent image sensor chip with three dimensional struc-
ture, IEEE International Electron Devices Meeting (IEDM), pp. 879–882 (1999).
11. K. W. Lee et  al., Three-dimensional shared memory fabricated using wafer
stacking technology, IEEE International Electron Devices Meeting (IEDM),
pp. 165–168 (2000).
12. M. Koyanagi et  al., Neuromorphic vision chip fabricated using three-
dimensional integration technology, IEEE International Solid-State Circuits
Conference (ISSCC), pp. 270–271 (2001).
13. T. Ono et al., Three-dimensional processor system fabricated by wafer stacking
technology, IEEE COOL Chips, pp. 186–193 (2002).
14. M. Koyanagi et al., Three-dimensional integration technology based on wafer
bonding with vertical buried interconnections, IEEE Transactions on Electron
Devices, 53(11): 2799–2808 (2006).
15. M. Koyanagi et al., High-density through silicon vias for 3-D LSIs, Proceedings
of the IEEE, 97(1): 49–59 (2009).
16. M. Koyanagi, Heterogeneous 3D integration -Technology enabler toward future
super-chip, IEEE International Electron Devices Meeting (IEDM), pp. 8–15 (2013).
17. M. Koyanagi, 3D system module with stacked image sensors, stacked mem-
ories and stacked processors on a Si interposer, IEEE International Solid State
Circuits Conference (ISSCC) 3D-Forum (2014).
18. K.-W. Lee et  al., Characterization of chip-level hetero-integration technology
for high-speed, highly parallel 3D-stacked image processing system, IEEE
International Electron Devices Meeting (IEDM), pp. 785–788 (2012).
19. K. Kiyoyama et al., A block-parallel SAR ADC for CMOS image sensor with 3-D
stacked structure, International Conference on Solid State Devices and Materials
(SSDM), pp. 1055–1056 (2011).
20. K. Kiyoyama et al., Very low area ADC for 3-D stacked CMOS image processing
system, IEEE International 3D System Integration Conference (3DIC), p. 5.1 (2012).
21. K. Kiyoyama et al., A block-parallel ADC with digital noise cancelling for 3-D
stacked CMOS image sensor, IEEE International 3D System Integration Conference
(3DIC) (2013).
22. T. Fukushima et al., New heterogeneous multi-chip module integration tech-
nology using self-assembly method, IEEE International Electron Devices Meeting
(IEDM), pp. 499–502 (2008).
Challenges in 3D Integration 83

23. T. Fukushima et  al., New chip-to-wafer 3D integration technology using hybrid
self-assembly and electrostatic temporary bonding, IEEE International Electron
Devices Meeting (IEDM), pp. 789–792 (2012).
24. H. Hashimoto et  al., Highly efficient TSV repair technology for resilient 3-D
stacked multicore processor system, IEEE International 3D System Integration
Conference (3DIC) (2013).
25. T. Fukushima et al., New three-dimensional integration technology based on
reconfigured wafer-on-wafer bonding technique, IEEE International Electron
Devices Meeting (IEDM), pp. 985–988 (2007).
26. T. Fukushima et al., Three-dimensional integration technology based on recon-
figured wafer-to-wafer and multichip-to-wafer stacking using self-assembly
method, IEEE International Electron Devices Meeting (IEDM), pp. 349–352 (2009).
27. T. Fukushima et al., Self-assembly technology for reconfigured wafer-to-wafer
3D integration, IEEE Electronics Components and Technology Conference (ECTC),
pp. 1050–1053 (2010).
28. K.-W. Lee et al., A cavity chip interconnection technology for thick MEMS chip
integration in MEMS-LSI multichip module, Journal of Microelectromechanical
Systems, 19(6): 1284–1291 (2010).
29. K.-W. Lee et  al., Three-dimensional hybrid integration technology of CMOS,
MEMS, and photonics circuits for optoelectronic heterogeneous integrated sys-
tems, IEEE Transactions on Electron Devices, 58(3): 748–757 (2011).
30. T. Tanaka et al., Ultrafast parallel reconfiguration of 3D-stacked reconfig-
urable spin logic chip with On-chip SPRAM (spin-transfer torque RAM),
Symposia on VLSI Technology and Circuits (VLSI2012), pp. 169–170 (2012).
31. Y. Ito et  al., Development of highly-reliable microbump bonding technology
using self-assembly of NCF-covered KGDs and multi-layer 3D stacking chal-
lenges, IEEE Electronics Components and Technology Conference (ECTC), pp. 336–
341 (2015).
32. H. Kikuchi et al., Tungsten through-silicon via technology for three-dimensional
LSIs, Japanese Journal of Applied Physics, 47(4): 2801–2806 (2008).
33. K.-W. Lee et al., Effects of electro-less Ni layer as barrier/seed layers for high
reliable and low cost Cu TSV, International 3D System Integration Conference
(3DIC) (2014).
34. M. Murugesan et al., High density 3D LSI technology using W/Cu hybrid TSVs,
IEEE International Electron Devices Meeting (IEDM), pp. 139–142 (2011).
35. T. Fukushima et al., New concept of TSV formation methodology using directed
self-assembly (DSA), IEEE International 3D System Integration Conference (3DIC)
(2016).
36. K. Yamashita et al., Copper-filled anodized aluminum oxide -a potential material
for low temperature bonding for 3D packaging, ICEP- IAAC, pp. 571–574 (2015).
37. K.-W. Lee et al., Novel reconfigured wafer-to-wafer (W2W) hybrid bonding tech-
nology using ultra-high density nano-Cu filaments for exascale 2.5D/3D inte-
gration, IEEE International Electron Devices Meeting (IEDM), pp. 185–188 (2015).
38. K.-W. Lee et  al., Novel W2W/C2W hybrid bonding technology with high
stacking yield using ultra- fine size, ultra-high density Cu nano-pillar (CNP)
for exascale 2.5D/3D integration, IEEE Electronics Components and Technical
Conference (ECTC), pp. 350–355 (2016).
5
Wafer-Level Three-Dimensional
Integration Using Bumpless
Interconnects and Ultrathinning

Takayuki Ohba

CONTENTS
5.1 Introduction .................................................................................................. 86
5.2 Co-Engineering by 3D and 2D ................................................................... 87
5.2.1 Delay of Three-Dimensional Integration Technology ................ 87
5.2.2 Economic and Technical Issues for Lithography ........................ 87
5.2.3 Co-Engineering Using Three Dimensional for Next
Generation of Manufacturing ........................................................ 88
5.3 Bumpless Interconnecting and Wafer-Level Three-Dimensional
Integration .....................................................................................................90
5.3.1 Overview of Bumpless Interconnects ........................................... 90
5.3.2 Thickness of Wafer and Ultrathinning......................................... 91
5.4 Details of Wafer-Level Three-Dimensional Process ............................... 92
5.4.1 Thinning Module............................................................................. 92
5.4.2 Stacking Module .............................................................................. 93
5.4.3 Through-Silicon Via Module.......................................................... 94
5.4.4 Packaging Module (Singulation/Packaging Module) ................ 97
5.5 Device Characteristics after Ultrathinning .............................................. 98
5.5.1 Retention Time Change of Dynamic Random- Access
Memory after Thinning .................................................................. 98
5.5.2 Ultrathinning and Estimation of Critical Thickness .................. 99
5.5.3 Cu-Diffusion Phenomenon at Ground Surface ......................... 101
5.6 Characteristics of Low-Aspect Ratio Through-Silicon Via
Interconnects .............................................................................................. 103
5.6.1 Step Coverage and Cu Diffusion ................................................. 103
5.6.2 Stresses in Cu Through-Silicon Vias ........................................... 104
5.6.3 Electrical Characteristics .............................................................. 106

85
86 3D Integration in VLSI Circuits

5.7 Thermal Resistance of Bumpless Interconnects and Thin Wafers........ 106

5.8 Concurrent Manufacturing Using WOW ............................................... 110
5.8.1 Enhancement of Memory Capacity ............................................. 110
5.8.2 Yield in Wafer Stacking................................................................. 111
5.8.3 Benefits of Co-Engineering........................................................... 112
5.9 Summary and Conclusions ...................................................................... 114
References............................................................................................................. 114

5.1 Introduction
Integrated circuits (ICs) based on planar technology started in the 1960s and
have led to today’s huge semiconductor industry and will be a key technology
in realizing a global infrastructure for the Internet-of-Things (IoT) and the
Internet-of-Everything (IoE) [1]. Twenty years after the advent of ICs, the con-
cept of three-dimensional integration (3DI) was proposed. The beginning of
3DI was the so-called transistor-based 3DI in the front-end-of-line (FEOL)
for stacking complementary metal–oxide–semiconductor (CMOS) devices
(a monolithic-stacked structure of n-MOSFETs [metal–oxide–semiconductor
field-effect transistor] and p-MOSFETs) to fabricate high-density ICs. In 2000s,
packaging-based 3DI, such as chip-on-board (COB) and chip-on-chip (COC)
with wire bonding, was developed to fabricate high-performance electron-
ics. A typical product is a system-in-package (SiP) consisting of a stack of 4–5
chips for mobile applications. Due to these two different approaches, that is,
transistor- or packaging-based 3DI, some technical misunderstandings often
occurred when people heard the term 3DI technology. For instance, current
interconnecting ICs in back-end-of-line (BEOL) for microprocessing units
(MPUs) or graphics processing units (GPUs) were realized by 12–14 level Cu/
low-k multilevel interconnects. This is obviously a 3D structure and comes
from the requirements for high device performance and density. Packaging-
based 3DI consists of a chip-based stack after singulation of the wafer, and
many fabrication methods have been introduced leading to complication
and confusion in attempts to classify 3DI technologies. In any case, attempts
to apply the early generation of 3DI were delayed. This is because conven-
tional miniaturization, involving scaling in planar and/or 2D processes, had
succeeded in reaching the submicron level, enabled by inexpensive ICs with
high density and high capacity. Wafer enlargement, such as from 8 to 12 inch
in diameter, also helped to reduce chip costs.
Interest in packaging-based 3DI technology using wafer-level process-
ing has been increasing again. This is driven by the physical and eco-
nomic limits of conventional scaling, which is no longer a main stream
for the increasing demands for device performance, system form factor,
and total manufacturing cost. This chapter discusses a BEOL-compatible
Wafer-Level 3D Integration 87

wafer-level 3DI approach, including state-of-the-art interconnecting tech-

nologies such as bumpless vertical interconnects and ultrathinning of
300 mm wafers down to micrometer thickness [2–4].

5.2 Co-Engineering by 3D and 2D

5.2.1 Delay of Three-Dimensional Integration Technology
The two-dimensional (2D) semiconductor industry and its growth are
supported by mature scaling technologies following Moore’s empirical law [5]:
a doubling transistor density every 18 months by scaling. To discuss 3D and
to compare chip-based packaging, utilizing wafers for IC manufacturing was
a significant difference and it made it easy to increase production volumes.
Thus, scaling and wafer processing are the key factors for improving perfor-
mance, shrinking chip size, and reducing costs, simultaneously. Although
chip-based manufacturing has an advantage in the case of small production
volumes and short cycle-time from design to production, most chip-based
processes, including tools and materials, were not compatible to that of wafer
based and need to adjust to the specifications for volume production. Those
facility trends were another reason for the delay in adaption of 3DI processes.

5.2.2 Economic and Technical Issues for Lithography

Two-dimensional scaling will soon face a severe economic crisis due to the
expensive lithography processes and facilities required. Reducing costs
requires the adoption of advanced lithography technologies, which, together
with peripheral support facilities, account for one-third to one-fourth of the
total cost of a manufacturing line. In short, although useful for reducing chip
size, scaling is extremely burdensome in terms of capital investment. Large-
scale investments have so far been made considering the technologies that
will be available two to three generations ahead, for example, 10 nm technol-
ogy should be applied to nodes of <5 nm, which will also face physical limi-
tations. This is a business scenario based on the empirical rule that profits
are made several generations after investments, for reasons involving the
tradeoffs between product sales and facility depreciation.
According to this empirical rule, an investment in 15 nm technology needs
to be made in consideration of its applicability to 10, 5, and 3  nm technolo-
gies. However, the price of extreme ultraviolet (EUV; λ = 13.5 nm) lithography
machines is >100 million USD, which is more than twice that of existing ArF
immersion lithography machines, and their current throughput is around one-
tenth or less. When converted into the processing capacity of a current large-
scale fabrication facility (e.g., 50,000 incoming wafers per month), based on this
system performance, an investment of approximately 2 billion USD will be
88 3D Integration in VLSI Circuits

required for EUV technology. Assuming that the past lifelong sales for each
generation are approximately 10 times the corresponding business investment,
the corresponding market size necessary for this investment is more than
20 billion USD. Based on the 300 billion USD, total worldwide semiconductor
market, this expected market size for one product and one manufacturer is not
realistic. In short, this is the limit of 2D scaling in light of the economics of the
industry, and it is difficult to find a scenario of victory at present.

5.2.3 Co-Engineering Using Three Dimensional for Next

Generation of Manufacturing
3D stacking, in combination with improved I/O signals of 2D, is anticipated
to overcome these scaling problems [6–8]. In the case of globally intercon-
necting large chips with high density, the signal traffic latency increases with
an increase in the physical length (L2D) from block-to-block as well as logic-
to-memory due to wire resistance–parasitic capacity (RC) delay, as shown in
Figure 5.1. This latency will be reduced by 3D-stacked structure when the
physical length (L3D) through one chip to the chip below becomes shorter
than L2D. The shorter interconnects (L3D  <<  L2D) are also advantageous for
power reduction, noise reduction, and clock cycle propagation across the
entire chip. Therefore, how to shrink the vertical length of 3D interconnects
and how to increase their density will be key factors in developing future
technology roadmaps, instead of the features of 2D scaling technology.
Vertical interconnects without bump electrodes between chips, such as
Cu BEOL processes, provide the lowest RC characteristics because of the

L = Physical length of electrical wiring

L2D 2D layout
PCB Power = N ( f · C · V 2DD + Ioff · VDD)
interconnect
Wire bonding PTr ≈ lnW/MHz/Tr
BGA
RDL
Bandwidth ∼ S/L2
Cu/Low-k RC delay ∼ L2/S
interconnect
Wiring capacitance ∼ L H/d
Large foot-print

Block-to-block length 1 mm ∼ 1 cm → 10 ∼ 100 μm@3DI

L3D
3D layout
RDL
Cu/Low-k Power ∼1/100 = L3D/L2D
interconnect >100 GBps bandwidth
TSV Stable power distribution
interconnect Large integration
Low total cost by WOW
Die-scale

FIGURE 5.1
A comparison of wiring length and layout for 2D and 3D chip sets. Miniaturizing the layout
using 3DI provides low-power consumption, higher bandwidth, and higher integration.
Wafer-Level 3D Integration 89

Current Hyper system

TSV
Microbump
RDL line
4 × DRAM C4-bump CPU/GPU/SoC

∼500 μm
Controller logic PHY PHY

Interposer

PKG substrate

(a)
Bumpless and ultra thin WOW/COW technology User requirements
Form factor, energy
efficiency, and cost
TSV CPU/GPU/SoC
8 × DRAM
Controller logic <200 μm
PHY PHY
interposer

PKG substrate

(b)

FIGURE 5.2
A comparison of (a) bump and (b) bumpless interconnects using TSVs for 3D memory stack
structures, assuming memory stacks containing four memories + controller (five stacks) and
eight memories + controller (nine stacks). Bumpless interconnects used for the ultrathin WOW
process can be formed with higher density (narrower pitch) compared with TSVs and bumps
due to the limitations of bump size and pitch.

RC reduction by no bump structure. Figure 5.2 shows a comparison of

bump and bumpless interconnects using through-silicon vias (TSVs) for 3D
memory–memory-stacked structures. 3D memory is implemented by plac-
ing memories and MPUs side-by-side on an interposer. In the case of bump-
less and ultrathin processes, the vertical length of interconnects from die to
die is 1/10th compared with bumps, assuming a bump height of 30 μm and
a die thickness of 70 μm. Even for stack containing two times more memory
chips, the total height of the 3D memory will be a half of the height of the
conventional approach.
Wafer-level 3DI for such high-capacity memories has productivity benefits
compared with other 3D methods such as COC and chip-on-wafer (COW).
Chip-based processes need a pick-and-place process (~2 s/chip) after wafer
singulation and thus suffer from low throughput. When the cycle-time
in pick-and-place is reduced, misalignment of chip placement typically
increases by more than 5 μm, which limits the density of TSVs.
A stack containing three 300  mm wafers provides a total silicon surface
larger than that of a single 450 mm wafer. This is important in considering
90 3D Integration in VLSI Circuits

whether wafer enlargement should be continued, taking into account the

cost benefits of increasing wafer sizes to 450 mm compared with utilizing
300  mm wafers. Moreover, retaining the standard 300  mm wafer size for
stacking ensures compatibility with existing manufacturing facilities in
front-end processing and helps utilize the mature process technology that
has been developed for wafer processing.

5.3 Bumpless Interconnecting and Wafer- L evel

Three-Dimensional Integration
5.3.1 Overview of Bumpless Interconnects
Bumpless interconnects using TSVs are a second-generation alternative to the
use of TSVs with microbumps. The bumpless interconnects process involves
a thinning-first process before bonding wafers, followed by a via-last process,
meaning that interconnects are formed after bonding the wafers, as shown
in Figure 5.3. This method is compatible with the BEOL, similar to multi-
level metallization in which dielectric deposition using thinned wafers and

Thinning module
• Wafer-edge trimming
Support glass or Silicon • Temporary adhesive
• Exposure
coating
Al or Cu Pad Device wafer 1 • Development
• Bonding onto support
glass 5∼10 μm diameter

• Backside grinding • RDL etching

• Post thinning • TSV etching
• Surface conditioning • Cleaning
5∼10 μm thickness

Bonding module • Permanent adhesive • Dielectrics deposition

coating • RIE bottom etching
• Bonding onto base
wafer
• Alignment of top and On-chip TSV Off-chip TSV
bottom wafers • TiN/Ti barrier deposition
• Cu seed deposition
Device wafer 1 • Deboning support • Cu ECD
Device wafer 0 glass • CMP
• Cleaning

Bumpless interconnects module Packaging module

• PR coating for • Backside grinding
lithography of RDL • Dicing/singulation
and TSVs • BGA
x2 device capacity

FIGURE 5.3
Bumpless interconnects using ultrathinning, TSVs, and wafer-on-wafer (WOW) process flow.
Vertical interconnects for TSVs are formed after wafer bonding from the front side. Additional
wafers can be stacked on top without any limitation on the number of wafers. These modules
can also be applied to chip-on-wafer (COW) integration. On-chip and off-chip TSV, respec-
tively, represent bumpless interconnects formed in the device area and the area around
devices, including gap fill materials in COW.
Wafer-Level 3D Integration 91

Al  and/or Cu metallization is replaced with bumpless interconnects using

TSVs. After wafer thinning and bonding, via-hole etching is carried out on a
silicon substrate from the front surface of the wafer having the device layer
and a dielectric layer for the multilevel interconnects. In principle, any num-
ber of thinned 300  mm wafers can be stacked to fabricate a large-capacity
memory device. In the case of a heterogeneous stack using different chip sizes
such as memory, logic, and passive device, COW can be used. For instance,
after preparing a redistribution line (RDL) on an MPU wafer, the RDL pad lay-
out is adjusted to match the dynamic random-access memory (DRAM) I/Os,
and the DRAM chip is picked up and placed on the MPU wafer, so that the
chip and wafer are interconnected by bumpless TSVs. For a 3D stack memory,
a side-by-side array with an MPU is possible, where a 3D memory and MPU
are stacked in a face-to-face manner onto an interposing 300 mm wafer. By
applying a wafer-level packaging process for the front side, the interposing
wafer is thinned down to <5 μm, and then TSV and RDL processes are car-
ried out on the thinned silicon wafer, as shown in Figure 5.2b.
The development of WOW has proceeded through four modules, classified
along the process flow. The modules include (1) a thinning module for thin-
ning the wafer substrates in which devices are implemented, (2) a stacking
module for bonding and stacking the wafers, (3) a TSV interconnects mod-
ule for forming Cu interconnects embedded in upper and lower wafers with
TSVs, and (4) a packaging module for singulating the stacked wafers. Dual-
damascene interconnects form (RDL) and also serve as a counter electrode
for the subsequent stacked wafer.

5.3.2 Thickness of Wafer and Ultrathinning

The thickness of the wafer is a critical dimension for the aspect ratio of TSVs
because the aspect ratio is determined by the diameter and the wafer thick-
ness. As a thinned wafer has low rigidity caused by bowing and/or warpage
due to a mismatch in the coefficients of thermal expansion (CTE) between sil-
icon and dielectrics, there is a thickness limit of ~50 μm at which standalone
thinned wafers can be safely handled. In our process, a wafer was bonded
to a support substrate (glass or silicon) using a temporary bonding material,
and thinning was carried out from the back side. The thinned wafer was
bonded to another wafer by back-to-face bonding using a permanent bond-
ing material, and then the support substrate was removed. Thus, there is no
need for high rigidity even with a thin wafer.
In our recent work, the typical thickness of thinned wafers ranged from 5
to 10 μm. When the thicknesses of the device layers in a DRAM and an MPU
are assumed to be approximately 5 and 10 μm, respectively, the aspect ratio
(depth-to-diameter ratio) of a TSV is only 2–4 for a TSV diameter ranging
from 5 to 10  μm, whereas conventional TSVs for a 50  μm wafer thickness
have aspect ratios ranging from 5 to 10. With decreasing aspect ratio, the pro-
cessing time in the TSV processes, such as etching, thin film deposition, and
92 3D Integration in VLSI Circuits

Aspect ratio = 5
Aspect ratio = 1 100 μm thick Si
20 μm thick Si Si3
BEOL
BEOL
MN Si4
BEOL Cu Cu
525 MPa
Si3
BEOL 225 MPa
Si2
MX

BEOL
Si1 Si2
MX

BEOL
Compressive Tensile
−600 −400 −200 0 200 400 600 900 (MPa)
(a) (b)

FIGURE 5.4
Stress simulation using the finite element method (FEM) for Si thicknesses of 100 (b) and 20 μm
(a) after stacking three wafers and forming a TSV with a diameter of 30 μm. A 10 μm Cu/low-k
BEOL layer is formed on every wafer surface, and thus the depths of the TSVs are 110 and
30 μm, respectively.

metal filling, decreases to about 1/5 at most, and the step coverage signifi-
cantly improves. A reduced TSV length is also considerably advantageous
for data transfer and power distribution with high-energy efficiency.
In addition, with the use of small TSVs, stress induced by a mismatch in
the CTE between Cu and Si decreases with decreasing aspect ratio of the
TSV, as shown in Figure 5.4 [9]. Stress at the center of a Cu plug in a 100 μm
thick wafer is 525 MPa and decreases to 225 MPa in a 20 μm wafer. The small
aspect ratio provided by an ultrathin wafer also has the advantage of reduc-
ing stresses generated in the silicon itself, in the bottom and top Cu-TSV, and
in interface regions having different CTEs.

5.4 Details of Wafer-Level Three-Dimensional Process

5.4.1 Thinning Module
A wafer is bonded to a support substrate in advance with a temporary adhe-
sive (thermoplastic resin). Thinning is performed by grinding (Back Grind,
or BG) within several micrometers of the target thickness, followed by pol-
ishing until the final thickness is achieved. The thinned wafer is permanently
bonded (thermosetting resin) to the device surface of another wafer, and then
the support substrate is removed. In the case of face-to-face WOW employing
bumps between wafers, the wafer thickness is limited by its poor rigidity and
the critical thickness is 50–100 μm. By using bumpless and back-to-face WOW
Wafer-Level 3D Integration 93

processes, the wafer thickness is limited mainly by the mechanical precision

of the grinding. With optimization of mechanical parallelism during grind-
ing, a thickness precision of 2 μm in a 300 mm-diameter DRAM wafer has
been achieved [10]. A 300 mm wafer on which 32 nm-node SDRAM devices
(2GB DDR3) are fabricated can be thinned down to 4 μm with a total thick-
ness variation (TTV) of around 1 μm, which is just 0.5% of the initial thick-
ness [11]. With a wafer thickness of 4 μm, visible light starts to pass through
the wafer. Remarkably, no degradation of retention time before and after
thinning was observed, even in a Si wafer with a thickness of 4 μm.

5.4.2 Stacking Module

An organic adhesive such as benzo-cyclo-butene (BCB) polymer [12–14] and/
or an alternative thermosetting resin with a thickness of approximately
<5  μm was used to permanently bond the wafer. BCB adhesives start to
polymerize with increasing temperature and are solidified at temperatures
of 200°C–250°C. For WOW, the wafers are aligned just before being perma-
nently bonded. To ensure appropriate alignment, infrared light passing
through the silicon substrate is used. Wafers to be bonded to one another
in WOW are originally thin and are therefore highly transmissive. With a
low-temperature bonding process and an optimized curing duration and
viscosity, the average misalignment between wafers can be made as small
as several micrometers as shown in Figure 5.5. This is predominantly due
to the bonding temperature, that is, reducing the bonding temperature pro-
vides smaller misalignment. On the other hand, because any gaseous solvent

Permanent Outgassing Postcure

Precure
adhesive 140°C/ Bonding 250°C/ Measurement
RT∼230°C
coating 3 min 30 min

1E+11
Misalignment <1 μm
1E+10
1E+09
Viscosity (MPa.sec)

1E+08
1E+07
>10 μm
1E+06
1E+05
1E+04
80 μm
1E+03
1E+02
0 50 100 150 200 250 300
Temperature (°C)

FIGURE 5.5
Relationship of misalignment of wafer bonding and viscosity of permanent adhesive material
as a function of temperature.
94 3D Integration in VLSI Circuits

escaping from the adhesive after the bonding process would form a cavity
(void) in the adhesive layer, measures should be taken to prevent this, such
as preheating after applying the adhesive or performing the bonding process
under a reduced pressure.

5.4.3 Through-Silicon Via Module

For bumpless interconnects, including RDLs, the Damascene method is
employed to simplify the processes. The Damascene wiring method also
enables fabrication of high-density interconnects as well as Cu intercon-
nects in the BEOL, as shown in Figure 5.6. For TSV processing, dry etching
through the dielectric (device layer), Si, and adhesive layer is carried out.
TSVs with a small aspect ratio, for example <3, have the advantage of short-
ening the process time for both etching and metal filling because those are
proportional to the volume of the TSV, as shown in Figure 5.7. Assuming
that the etching rate follows the mass transport limit reaction, the etching
times, t and t1, at different TSV diameters, D and D1, and depths, d and d1,
follow t1/t  =  (D1/D)2  ×  (d1/d), that is, t1/t  =  0.1 at D  =  D1, d  =  50  μm, and
d1  =  5  μm, which suggests 1/10th of the etching time for the same TSV
diameter and 1/10th of the depth. Reducing process time is the key for low-
cost manufacturing. After TSV etching, a SiN or SiO2 film is deposited by
plasma-enhanced chemical vapor deposition (PECVD) to provide electrical
insulation from the Si substrate. The barrier dielectrics at the bottom of the

M2 RDL trench
SiN
PR TSV hole
Si #2
Si #2 Si #2
Adhesive
M1 RDL
Si #1 Si #1 Si #1
(a) (b) (c)

ECD-copper
PECVD-SIN Ti liner

M2 RDL
Si #2 Si #2 Si #2 TSV
M1 RDL
Si #1 Si #1 Si #1
(d) (e) (f )

FIGURE 5.6
TSV formation and damascene Cu plug processes. After bonding of the thinned wafer to another
wafer surface, RDL (redistribution line) patterning, TSV etching, Cu plug formation, and pla-
narization by CMP are carried out. (a) Wafer level bonding, (b) TSV PR & Si DRIE, (c) PECVD,
(d) SiN self-align etching, (e) PVD metal barrier and Copper ECD, and (f) planarization.
Wafer-Level 3D Integration 95

Process time for TSV etching

(AR)2 d3
200 Process time =
180−200 NO! D2R
180 160−180 Assuming etching rate 10 μm/min (R) at 50 μm diameter (D) and
etching rate followed by mass-transport limited, where AR =
160 140−160
aspect ratio and d = depth, D = diameter, respectively.
Etching time (min)

120−140
140
100−120
120 80−100
NO!
100 60−80
40−60 100
80 90
20−40 NO! 80
60 70
0−20 60
40 50
40
20 30 Depth (μm)
Our target
0 20 Wf thickness
20 19 18 17 10
16 15 14 13 7.5
12 11 10 9
8 7 6 5
4 3 2 1 5
Aspect ratio (AR)

FIGURE 5.7
Silicon etching time versus aspect ratio of TSV and depth, assuming that the total etching vol-
ume follows the mass transport limit reaction. Etching time increases with increasing depth
and aspect ratio.

TSV is removed by bias sputtering, and Ti/TiN and Cu are deposited on

the barrier metal and the seed layer, respectively, by sputtering. For Cu plug
interconnects, electrochemical-plated Cu (ECP-Cu) is used. ECP-Cu pla-
narization is carried out by chemical–mechanical polishing (CMP), as shown
in Figure 5.8. In case of permanent adhesive formed on Cu pad after CMP,

As-deposited ECD-Cu After CMP

Top view X-ray CT

RDL TSV

50 μm

Chain structure
RDL

Si Si = 10 μm

TSV (ϕ = 10 um, Adhesive

10 μm Base Si wafer
h = 15 um)

FIGURE 5.8
SEM images of Cu TSVs after ECD-Cu deposition and CMP. For ECD-Cu, to reduce polishing
time in CMP, the overburden of Cu was controlled to be thin. Chain structure shows the cross-
section of Cu pad and damascene TSV for two stack wafers.
96 3D Integration in VLSI Circuits

50
Glass Void
BCB (2 μm) IR image of 2WOW Microscope image of 2WOW
40
BCB void area (%)

Si w/o M2 Cu metallization with M2 Cu metallization

Cu pad 200 μm × 200 μm
30
Cu pad Residual Cu
Void Cu PAD
20
IR

10 Si2 Si2
BCB BCB
Si1 Si1
0
0 100 200 300 400 500
(a) Dishing depth (nm) (b)

FIGURE 5.9
Occupied void area (BCB) formed on as a function of Cu pad as a function of dishing depth
of Cu pad (a). Top view images of second wafer surface after Cu CMP using IR and optical
microscope (b).

voids formation in adhesive occurs and depends on the dishing depth at

Cu pad as Figure 5.9a. This dishing phenomenon (microsurface roughness)
affects the next process in which Cu residue is removed by CMP in the sec-
ond wafer process (Figure 5.9b). This suggests that Cu dishing have to reduce
less than 200 nm in depth for the CMP process.
Figure 5.10 shows the via structure after etching and the leakage cur-
rent, compared with the Bosch and direct etching methods [15–16].

Bosch etching Direct etching

6.87 μm
SiO2 + Low-k
Device layer

Si SiON Cu Si SiON Cu
Ti/TiN
SiOC Ti/TiN
Si
Cu Si Cu
5.85 μm
SiOC
100 nm
Crack 100 nm
5.78 μm
−3 −3
10 10
400°C
10−5 10−5 5.45 μm
Leakage current (A)

Leakage current (A)

400°C
10−7 300°C 10−7 300°C
Si substrate

10−9 10−9
200°C
10 −11
10−11 200°C
w/o anneal
w/o anneal
10 −13
10 −13 5.39 μm
0 10 20 0 10 20
V (V) V (V) 10.0 um
(a) (b)

FIGURE 5.10
A comparison of leakage currents in two types of TSV samples made by Bosch etching and
direct etching (a). Cracks are observed in the side walls of the TSV in the Bosch-etched sample,
which had a rough interface called scalloping. The leakage current as a function of applied
voltage after annealing at temperatures up to 400°C was measured. With increasing tempera-
ture, the leakage current increased but was two orders of magnitude higher in Bosch etching.
By optimizing the scalloping shape, SEM images of TSV etched off through Cu/Low-k BEOL
layer and Si are shown (b).
Wafer-Level 3D Integration 97

As Bosch etching was conducted by cyclic isotropic-etching and deposi-

tion, microsteps called scalloping were formed at the sidewall of the via.
The scalloping causes cracks due to stress concentration in the dielec-
tric layer and poor step coverage for thin films deposited by chemical
vapor deposition (CVD) and physical vapor deposition (PVD). In contrast,
anisotropic direct etching resulted in a smooth surface profile along the
side wall. The leakage current in Bosch etching was one order of magni-
tude higher than that in direct etching. The leakage current was caused
by Cu diffusion at the side wall of the TSV that took place at a thinner part
of the dielectric containing cracks. Thus, direct etching of TSVs, in other
words etching that is capable of producing a smooth sidewall, is required
for TSV processing. In addition, low-aspect ratio Cu TSVs cause lower
stress concentration in thermal processing compared with high-aspect
ratio Cu TSVs. Low stress reduces Cu deformation and stress propagation
to the device regions.

5.4.4 Packaging Module (Singulation/Packaging Module)

After wafer stacking, the same procedure as in conventional packaging
(bumping, singulation by dicing, resin packaging) is followed. For dicing of
a seven-level wafer stack, the adhesive layer and silicon chips were found to
be free of defects or delamination, as shown in Figure 5.11. After the stacked
chips were packaged with epoxy resin, they were subjected to heat stress
testing at temperatures of −65°C to 150°C. Scanning acoustic tomography
(SAT) was adopted for internal observation, and after up to 100 repeated
heat stress tests, no delamination was found at the interfaces between the
molding compound and chips, nor at the chip stack interfaces, as shown in
Figure 5.12.

4
Thinned Si wafer 20 μm
3

Adhesive layer 5 μm Wafer 1

FIGURE 5.11
SEM cross-sectional image of seven-wafer stack after dicing. No crack and delamination of
bonding material were found.
98 3D Integration in VLSI Circuits

Thermal stress test (−65°C∼150°C, 10 min/10 min)

Pristine 1 cycle 10 cycle 100 cycle

Scanning acoustic tomography image

FIGURE 5.12
Thermal stress testing at temperatures of –65°C to 150°C for 100 cycles, observed by scanning
acoustic tomography (SAT).

5.5 Device Characteristics after Ultrathinning

5.5.1 Retention Time Change of Dynamic Random- Access
Memory after Thinning
The effect of wafer thinning on device performance was examined to evalu-
ate the thinning limits for 300  mm silicon wafers, as shown in Figure  5.13
[10,11,17]. The ferroelectric RAM (180  nm-node FRAM or FeRAM)
device wafer was thinned down to 9  μm. There is a capacitor formed of

Device FRAM (VLSI2010) SRAM (IEDM2009) DRAM (IEDM2014)

Node ∼180 nm 45 nm (Lg 35 nm) 40 nm (2 Gb)
Wafer thickness 9 μm 7 μm 4 μm
FRAM memory 35 nm SRAM device
Cu/L
ow-k
35-n interc er
lay
SEM mL o
g SR nnects
AM DR
AM
9-μm PZT
picture Si Si su
bstra m
Sili
con
te ∼7-μm 4μ

BCB
5 μm

99.99 NMOS PMOS 1.E+00

99.9 After 1.E−01
99
Cumulative probability (%)

thinning
Cumulative bit failure (a. u.)

1.E−02
95
Ion/Ioff

90 1.E−03
Electrical 80
70 1.E−04
50
property 30
20
1.E−05
1.E−06
10
Junction leakage

5 1.E−07
1 Before 1.E−08 Before thinning
−.1 thinning 1.E−09 After thinning
−.01 1.E−10
0.6 0.8 1 1.2 1.4 1 10 100 1000 10000
Switching charge (a.u.) Refresh time (a.u.)

FIGURE 5.13
Device characteristics before and after wafer thinning for ferroelectric random access mem-
ory (FeRAM), MPU (SRAM), and DRAM. In the FeRAM, the quantity of residual polarization
charge (Qsw) was evaluated; in the SRAM, the on- and off-current, and leakage current; and in
the DRAM, the retention time. There were no significant changes in these characteristics after
wafer thinning down to 9 and 4 μm.
Wafer-Level 3D Integration 99

lead-zirconate-titanate PbZrxTi1-xO3 (PZT) dielectric underneath the bit line

(M1). There was no significant shift of switching charge within the range of
the design specifications. In the case of a microprocessor (45 nm-node MPU)
device, thinning was carried out down to 7  μm. For both n-MOSFETs and
p-MOSFETs, the on-current (Ion), off-current (Ioff), and roll-off characteris-
tics of threshold voltage (Vth) were nearly the same as those before thinning.
In the case of DRAM devices, which are most sensitive to induced leakage
current due to defects, the wafer was thinned to a final thickness of 4 μm,
which is about 0.5% of the thickness of the bulk wafer (725 μm) and thinner
than the device layer. The TTV within the 300 mm wafer was low enough to
realize multistacking: A TTV of 1.02 μm was achieved at an average thick-
ness of 4 μm. The wafer was annealed during the bonding process at around
250°C for 60  minutes. No significant change with retention time (refresh
time) in the entire wafer before and after thinning was observed for the Si
thicknesses of 40, 20, 8, and 4 μm [10]. This suggests that the thinning process
developed in this study did not affect the junction leakage current, which
degrades the retention time more sensitively than other leakage phenomena
such as subthreshold leakage, capacitor dielectric leakage, and gate-induced
drain leakage (GIDL).

5.5.2 Ultrathinning and Estimation of Critical Thickness

Wafer thinning was carried out as follows: coarse grinding (#320 grit size)
down to ~50 μm, fine grinding (#2000 grit size) to <20 μm, and post-thinning
using CMP as an option. Figure 5.14 shows a schematic of the thickness

Wafer grinding
Inclination angle
2.5
Grinder wheel Initial angle
Si

Adhesive
2.0
Support glass
First step
Chuck table
TTV measurement using NCG
1.5
TTV (μm)

Non contact
gauge (NCG)

1.0

Adjusting chuck table angle

and final grind 0.5

0.0
(a) (b) Inclination angle (a.u.)

FIGURE 5.14
Improvement in total thickness variation (TTV) using so-called auto-TTV process employing NCG
(noncontact gauge) methods (a) and TTVs for 300 mm wafer with various inclination angles (b).
100 3D Integration in VLSI Circuits

variation control in the grinding process. The wafer thickness uniformity

after grinding was determined by the contact angle between the wheel and
wafer surface [18]. Usually, the wafer was very slightly bowed after bond-
ing due to elastic deformation of the temporary adhesive at the wafer edge,
and this shape was also reflected in the contact angle. By adjusting the con-
tact angle to follow the geometric shape of the wafer, the TTV decreased
to as low as 0.5  μm within the 300  mm wafer. With these thinning pro-
cesses, the thickness of the damaged layer, including point defects such
as vacancy-type defects, was decreased from micrometer level to several
nanometers when evaluated by transmission electron microscopy (TEM)
and positron annihilation spectroscopy analyses, as shown in Figure 5.15
[19,20]. To evaluate the critical thickness, thinning was carried out on a
300 mm DRAM wafer (DDR3 SDRAM) down to ~2 μm, and Cu was inten-
tionally formed at a density on the order of 1013  atoms/cm2 on the back
side after thinning. Then, the wafer was annealed at 250°C for 60 minutes.
Retention degradation and yield loss within the wafer were observed at
thicknesses less than 2.6  μm, as shown in Figure 5.16 [10]. According to
those results, in case of a DRAM, there was a critical thickness of the Si
wafer of around 2 μm.

0.553 0.555

No.5 : Coarse75→Fine50 No.5

No.9 : Coarse425→Fine50 0.550 No.9
0.548
No.10 : No.5→CMP1 No.10
No.12 : No.5→CMP5
S parameter
S parameter

0.545
0.543
0.540 200 nm No.5

0.538
0.535

5 nm No.10
0.533 0.530
0 5 10 15 20 25 30 0 50 100 150 200
(a) Positron energy (keV) (b) Depth (nm)

FIGURE 5.15
Doppler broadening spectrum of positron annihilation spectroscopy analyses for wafer
backside after thinning. (a) S parameter as a function of incident positron energy E for fine
grinding (Samples No. 5 and 9) and CMP (Samples No. 10 and 12). The number in terms
such as “Coarse75” represents 75 μm coarse grinding. (b) Depth profiles of S parameter for
fine grinding and CMP. The S parameter is normalized by that of Sample No. 12, which
is assumed as a reference. Thicknesses of the defect layer were ~200 and 3–4 nm for fine
grinding and CMP, respectively, observed by TEM analysis, which agrees with the depth
profiles.
Wafer-Level 3D Integration 101

1.2

Retention time (normalized)

1
w/o
0.8 Cu

0.6 w/ Cu
contamination
0.4

0.2

0
1 10 100
Si thickness (μm)

FIGURE 5.16
Retention time of DRAM as a function of Si thickness. Cu was formed at a density on
the order of 1013  atoms/cm 2 at the back side after thinning. Then wafers were annealed
at 250°C for 60  minutes. Normalized retention time was the time at 80% yield. Wafer
map represents Si thickness, where the average thickness and TTV were 2.66 and 1.6 μm,
respectively.

5.5.3 Cu-Diffusion Phenomenon at Ground Surface

A Cu-diffusion phenomenon was evaluated at the back side of the ground
Si surface [19]. Island-like Cu aggregations were observed from back surface,
as shown in Figure 5.17, in which Cu was detected with a high-angle annular
dark field (HAADF) scanning TEM and energy-dispersive X-ray spectros-
copy (EDX). The sample was annealed at 250°C for 60 minutes. In order to

HAADF
Device layer

Si

1.0 μm BF

EDX
Cu
200 nm

2 μm
1.0 μm Cu K

FIGURE 5.17
Cross-sectional TEM image of DRAM wafer after grinding and Cu contaminated. Cu aggregation
at the back side of the Si surface was observed by HAADF-STEM and EDX analyses.
102 3D Integration in VLSI Circuits

1.00E+21
w/o ANL
200°C
1.00E+20
300°C
800°C
Concentration (atoms/cm3)

1.00E+19 1000°C
200°C-GAUSS
1.00E+18 300°C-GAUSS
800°C-GAUSS
1.00E+17 1000°C-GAUSS

Simulation
1.00E+16

1.00E+15

1.00E+14
0.0 0.2 0.4 0.6 0.8 1.0
Depth (μm)

FIGURE 5.18
Backside SIMS depth profiles of Cu in Si after back-side grinding for various annealing tem-
peratures, and Gaussian distribution of Cu into Si (dotted line) calculated using the diffusion
kinetic model.

compare Cu diffusions, the annealing temperature was gradually increased

to 1000°C in 30 minutes under the vacuum during the grinding process, and
Cu was formed at the back side as well. The depth profile of the Cu was
measured by time-of-flight secondary ion mass spectrometry (TOF-SIMS),
as shown in Figure 5.18. The diffusion profile based on a diffusion kinetic
model, the dotted line in Figure 5.18, followed the equation:
 x2 
− 
Q  4 Dt 
C= e 
πD0

C and Q were Cu concentration in depth and Cu concentration at

Si surface, respectively, where the diffusion coefficient of Cu in Si,
D0 = 4.7 × 103 × exp(−0.43 eV/kT) cm2s−1 was used [21]. According to the
SIMS results, the diffusion depth of Cu was around 400 nm even at 800°C,
whereas Cu is predicted to diffuse by more than 1  μm from the diffusion
model. This suggests that the Cu was stable at the damaged layer at the back
side of the wafer and thought to be damaged layer having point defects that
act as gettering sites for Cu atoms. Recrystallization of silicon occurred at
a high temperature (>800°C). As those damaged layers consisted of atomic
level vacancy-type defects, these vacancy defects were substituted by Cu
atoms, which led to a stable state preventing further Cu diffusion.
Wafer-Level 3D Integration 103

5.6 Characteristics of Low-Aspect Ratio

Through-Silicon Via Interconnects
5.6.1 Step Coverage and Cu Diffusion
For high-reliability Cu metallization, formation of a conformal barrier
layer is necessary for preventing Cu diffusion into the Si substrate [15,22].
Figure 5.19 shows a cross-sectional TEM image of TSVs at the side wall of
a 10 μm diameter via. The aspect ratio was about 1.5. The dielectric barrier
layer of the SiON film deposited by the conventional PECVD process had a
thickness at the bottom of the via that was 30% of that at the top of the via.
A very thin barrier metal film composed of Ti/TiN deposited by the PVD
process has no film continuity because of the low coverage. Even with a low-
aspect ratio TSV, cracks occurred at the scallop-shaped microsteps due to
the Bosch process, whereas no cracks were observed in the smooth side wall
processed by anisotropic direct etching. These cracks cause leakage current.
Cu diffusion increases when the density of the SiON layer decreases. In the
case of a SiON layer with a relative density of 0.6 deposited at 100°C, Cu
easily reached the interface between the dielectric layer and the Si substrate.
Figure 5.20 shows the critical thickness (defined as the thickness that pre-
vents Cu diffusion within the dielectric layer) of dielectric films as a func-
tion of the relative density. Leakage current occurred at sidewall of TSV due
to Cu diffusion has been observed by the infrared-optical beam irradiation
(IR-OBIRCH) method [23]. The point at a relative density of 0.8 illustrates the
case for the cap layer of Cu interconnects used in Cu/low-k studies at a com-
paratively higher deposition temperature. For a relative density below 0.5, a
dielectric layer thicker than 1 μm would be needed to prevent Cu diffusion

After seed formation at R.T. After 400°C TC

Si SiON Cu Si Si SiON Cu
Cu Si

Ti/TiN SiOC
Ti/TiN

Crack
100 nm 100 nm
(a) (b)

FIGURE 5.19
TEM cross-sectional image of TSV after SiON/Ti/TiN/Cu formation. Scallop shape at the side
wall and crack propagation at the SiON layer were found for Bosch etching (a). No crack was
observed for direct etching, even after annealing (b).
104 3D Integration in VLSI Circuits

10000
Si(O)N/Si TSV 1015 atms/cm3
Cu 1010 atms/cm3
Thickness of Si(O)N (t: nm)

1000 after 400°C/10 hr aging

LT SiON (low dense) atms/cm3
1022 Cu
10 15 Si(O)N

t
100 Cu 1010
Diffused Cu
Si
LT SiON (high dense)
10
HT SiN

LT = Low-temperature PECVD (∼150°C)

HT = High-temperature PECVD (∼350°C)
1
0.2 0.4 0.6 0.8 1.0
Relative density of Si(O)N

FIGURE 5.20
Critical thickness for Cu barrier of dielectric layer as a function of film density. Density repre-
sents relative value with respect to that of bulk Si3N4 (3.44 g/cm3).

into the Si substrate. This suggests that the effective Cu area decreases with
an increase in the thickness of the dielectric, which limits the scalability of the
Cu TSV diameter, for example, a Cu TSV diameter of only 1 μm remains at a
3 μm thick dielectric.

5.6.2 Stresses in Cu Through-Silicon Vias

For the bumps attached to the device layer, delamination occurs at the
Cu/low-k interconnect, as shown in Figure 5.21a. This is due to excessive
stress induced during the process of bonding process the low-mechanical
stress Cu/Low-k interconnects. Furthermore, the bumps themselves tend to
fracture under the high-stress concentration due to a large CTE mismatch
between the bumps and underfill materials. As bumps are absent in bump-
less TSVs, no delamination is observed, as shown in Figure 5.21b. The stress
induced at the Cu/low-k interconnect layer was estimated by finite element
method (FEM) analysis using Chebyshev’s orthogonal function obtained
by the design of experiment (DOE), and the ratio of the thickness of the
adhesive layer (TAdhesive) to the Si (TSi) as a function of the aspect ratio of the
TSV is shown in Figure 5.22. These results suggest that the induced stress
increases when the thickness of the adhesive layer and aspect ratio increases.
Therefore, the use of small TSVs and a thin adhesive layer, in other words, a
no bump structure, provides a better solution for stress reduction.
Wafer-Level 3D Integration 105

Bump

PI
UBM
Cu/Low-k BEOL

Delamination

Wafer 1 20 μm

(a)

Bumpless
Cu TSV

Wafer 2

Wafer 1
Cu/Low-k BEOL 20 μm
(b)

FIGURE 5.21
Cross-sectional SEM images of (a) bump and (b) bumpless TSV formed on Cu/Low-k BEOL
interconnects.

0.5
Intel Leti 1.2 GPa (× 3) Si
(IITC2007) Tensile
BEOL
0.4
1300 Cu
1 GPa
TAdhesive/ TSi

0.3 Si
Samsung
(3D art.2007) σ
Elpida
0.2 (ECTC2007)
Bumpless TSV 800 MPa (× 2) BEOL
Fraunhofer
Leti (3D art.2004) Cu SiN Si
0.1 (Low temp.2007) −100
600 MPa (× 1.5) ASET (MPa) Induced stress
RTI (Low temp.2007) Adhesive
400 MPa (× 1) IMEC
(IEDM2006) RTI (Low temp.2007)
0.0
2 4 6 8 BEOL Cu/Low-k
Aspect ratio of TSV Si sub.

FIGURE 5.22
Stress at Cu/Low-k interconnects layer versus the ratio of thickness of adhesive layer to thick-
ness of Si as a function of aspect ratio of TSV.
106 3D Integration in VLSI Circuits

99.99
99.9 (a) w/o TSV
99 251k dense chain
BEOL: 0.4 μm dense chain
31k sparse chain
90
Probability (%)

50
w/o TSV (b) w/bumpless TSV
+ TSV: 10 μm Multi-TSV
10
w/ TSV
1
w/o TSV w/ TSV
.1
.01 0.21 mΩ/blocks
0 20 40 60 80 100 120
Resistance (kΩ)

FIGURE 5.23
Via chain resistance cumulative failure distribution of Cu BEOL interconnects with and with-
out Cu-TSVs. Schematic of test structure for electrical measurement: (a) 65-nm Cu intercon-
nects and (b) with multiple TSVs in two-wafer stack.

5.6.3 Electrical Characteristics

Figure 5.23 shows the cumulative failure distribution of the Cu-BEOL via
chain resistance and with multiple bumpless Cu TSVs [22]. The Cu-BEOL
via chains having a total number of vias equal to 31 × 103 (sparse chain) and
251 × 103 (dense chain) were compared. The diameter of the TSV and total
depth of the TSV, including the adhesive layer, are 5 and 15 μm, respectively.
The contact resistance of the Cu-TSV to the Cu-BEOL is estimated to be
0.21 mΩ. There are no open failures or significant changes in the resistance
distribution with the TSVs. The leakage current between TSVs blocks was as
low as 2.3 × 10−11  A at 4.0 V. Thermal stress testing of a structure in which
a Cu-TSV was contacted with a BEOL chain was carried out from −55°C to
125°C. In the case of a dense chain, the initial resistance of 104 kΩ was changed
to 100 kΩ and 101 kΩ after TSV formation and 1000 test cycles, respectively.
This suggests that there was no degradation after thermal stressing.

5.7 Thermal Resistance of Bumpless

Interconnects and Thin Wafers
Heat dissipation becomes more difficult in a 3DI structure due to the ther-
mal resistance of the stack of multiple devices and interconnect-dielectrics
layers. The total thermal resistance for a stack of multiple devices was
Wafer-Level 3D Integration 107

BEOL
Layer 8
Vertical interconnects

1st–
7th layer
Layer 0

(a)
TSV

Si
TSV

BEOL,
device layer

Adhesive, (c)
underfill

Microbump

(b)

FIGURE 5.24
Schematic diagrams of thermal resistance estimation for Si substrate having a BEOL layer and
TSV vertical interconnections with and without bumps: (a) nine-device stacked structure with
bumps, (b) details of TSV and bump with underfill, and (c) TSV (no bump) and ultrathin wafer
with organic bonding adhesive.

evaluated using bump and bumpless interconnects [24]. The thermal resis-
tance of a Si substrate having a BEOL layer and TSV vertical intercon-
nects with and without bumps was estimated, as illustrated in Figure 5.24.
There are polymer and composite materials, such as underfill in the case
of the bump interconnects and bonding adhesives in the case of bump-
less interconnects. The thermal resistance Rth of the vertical interconnects
and the total thermal resistance were calculated using the FEM and a ther-
mal network method, respectively. Assuming a stack of eight DRAMs and
one controller wafer, the thermal resistance was estimated by the follow-
ing sequence: estimating the effective thermal conductivity of each layer
and calculating the temperature rise using the thermal network method.
The thermal conductivities used were 148, 160.5, and 1.44 W/mK for Si, Si
with TSVs, and BEOL, respectively [25]. Figure 5.25 shows thermal resis-
tance as a function of normalized via occupancy. The thermal resistance
decreased as both the thicknesses of the underfill material and bonding
108 3D Integration in VLSI Circuits

100
Gap
= 50
μm
10 25 μ
m TSV
w/o Under-fill
Interconnection thermal

Bump Gap
Gap = 50 μm
resistance (Kcm2/W)

1
25 μm w/ Under-fill

Adh
0.1 es ive =
5 μm
2 μm
Bumpless TSV
0.01
TSV

0.001 Adhesive thickness

0.0001
0.001 0.01 0.1 1
Normalized occupancy of via area

FIGURE 5.25
Thermal resistance of interconnects as a function of normalized occupancy of via area, where
microbump and TSV (no bump) are compared.

adhesive decreased, and two orders of magnitude lower the bumpless

structure when compared with the conventional bump structure. This is
due to thickness reduction of the polymer layer and bumpless structure,
which has higher thermal resistance and the shortened interconnects. This
suggests that only 1% the total metal area of bumpless TSVs is required to
achieve a thermal resistance equivalent to that of the bump structure. The
total thermal resistances of a stack of eight wafers with and without bumps
are estimated as shown in Figure 5.26. The bump size and pitch were 25 μm
(thickness of underfill material) and 50  μm. For bumpless TSVs, the TSV
size, pitch, and thickness of the bonding adhesive were 10, 20, and 5 μm,
respectively. The total thermal resistances were 1.54 and 0.46 Kcm 2/W for
the bump and bumpless structures, respectively. According to these total
thermal resistance values, the temperature increase in the wafer stack was
estimated as shown in Figure 5.27, where the temperature indicates the rel-
ative increase in temperature from the bottom wafer. The temperature rises
with an increase in the number of stacked wafers and becomes approxi-
mately 20°C for the bump structure, which is about four times higher than
that of the bumpless structure. Therefore, if such a temperature increase is
allowed, a stack containing five-times more wafers will be possible with
bumpless TSV interconnects.
Wafer-Level 3D Integration 109

Microbump Bumpless
51.4 μm pitch 25 μm bump 512*16 TSV 5 μm gap

Thermal Thermal
Equivalent thermal Thickness Thickness
Components resistance resistance
conductivity (W/mK) (μm) (μm)
(Kcm2/W) (Kcm2/W)
Si 148 150 0.049 5 0.00034
1.44 15 0.104
BEOL
Top 3.99 15 0.038
Chip
Interconnection
2.54 20 0.079
microbump
Rth1
Interconnection
2.56 5 0.020
bumpless
DRAM
Si with TSV 160.5 50 0.003 5 0.00031
2–8 1.44 15 0.104
BEOL
Chip 3.99 15 0.038
7
layers Interconnection
2.54 20 0.079
microbump

Interconnection
Rth2–8 2.56 5 0.020
bumpless

Logic RthL Si with TSV 160.5 50 0.003 5 0.0003

Total thermal
Rth1+7*Rth2–8+RthL 1.54 0.46
resistance

FIGURE 5.26
Summary of total thermal resistance of TSV + microbump and TSV (no bump) structure.

25
Microbump Bumpless

20
Temperature rise (°C)

15

10

0
Base 8 7 6 5 4 3 2 1 0
Device layer

FIGURE 5.27
Temperature rise for each layer in stacked devices in bump and no-bump (bumpless) structures.
110 3D Integration in VLSI Circuits

5.8 Concurrent Manufacturing Using WOW

5.8.1 Enhancement of Memory Capacity
As the WOW process, thinning-first, back-to-face bonding, and then TSV-last,
allows thinning of silicon wafers down to 4 μm without any degradation
of the device characteristics, the total wafer thickness, including the device
layer and the adhesive layer, becomes 10–20 μm, which is 1/10th to 1/100th
of the thickness of conventional bump interconnects using TSVs. Even if
the number of stacked wafers is 100, assuming that the wafer thickness is
10 μm, the total thickness after stacking is only 1 mm. If a cubic space can
be utilized, 1000 device layers can be stacked. By following these multilevel
stacking processes, with a memory device such as a DRAM fabricated with
22 nm technology and having a memory density of 30 GB/cm2, when four,
eight, sixteen, and so on, of these devices are stacked, the total capacity of
the memory device can be increased to 120, 240, 480 GB, and so on, respec-
tively. In case of Flash nonvolatile memory, having 3D-transitor memory
cells, terabit-capacity memory can be realized by stacking only 4 or 8 wafers.
In contrast, to achieve equivalent capacity with a single wafer using extreme
scaling or 3D transistors, 1 nm processing technology or more than 128 tran-
sistor stacks will be required, as shown in Figure 5.28.

10000

gration
inte Terabit generation
1000 3D
it 1 nm
DRAM density (Gb/cm2)

y lim
og raph
, lith 3 nm
st
h co 5 nm
Hig
7 nm 4-stack/64 Gb
ing
Multistack

100 scal 10 nm 8-stack/64 Gb

2D 14 nm 16-stack/64 Gb
18 nm 32-stack/64 Gb
4-stack/128 Gb
22 nm 8-stack/128 Gb
Stacked dies / DRAM generation
16-stack/128 Gb
10 32 nm node 32-stack/128 Gb
4-stack/32 Gb
8-stack/256 Gb
A 100 Gb/cm2 memory capacity can be achieved by stacking 8-die and 4-die 16-stack/256 Gb
of 32 nm and 22 nm node, respectively, which is equivalent to the DRAM 32-stack/256 Gb
capacity at <14 nm Node. ITRS 2011 production
1
2008 2010 2012 2014 2016 2018 2020 2022 2024 2026 2028 2030
Production year (risk start)

FIGURE 5.28
Trend of DRAM density using 2D conventional scaling and 3D multistacking using existing
DRAM. DRAM capacity in the 3D case corresponds to the number of stacked dies, assuming
that redundancy is eliminated by cell blocks at each layer.
Wafer-Level 3D Integration 111

5.8.2 Yield in Wafer Stacking

In the case of die-to-die series connections, the total yield in wafer stack-
ing can be estimated by Yn, where Y and n are the nondefective die yield
of one wafer and the number of stacked wafers, respectively, as shown
in Figure 5.29 [26]. The total yield of stacked wafers without any design
improvements decreases with increasing number of stacked wafers, for
example, Y = 0.9 and four stacked wafers give a total yield of Y4 = 0.656.
By using multiple channel or memory mat for designing the architecture of
the memory layout, the defect area will be limited at the channel region. Thus,
the memory capacity within one memory die, excluding failed channels, is
maintained. In other words, when one memory die has multiple channels, the
loss of memory capacity would be reduced even at the same defect density.
As the bumpless TSVs and ultrathin WOW process enable interconnections
with high density at the minimum physical length, those channels are verti-
cally connected to the controller independently, as shown in Figure 5.30. For
instance, if one channel requires 64-bit I/Os, the total vertical interconnects
from 4 dies correspond to a total of 4096 bit I/Os. This is difficult to achieve
with small TSV area using conventional bump processes because the density
of bumps and TSV processes is limited to the bump size and pitch (~50 μm).
In the case of a conventional die-level stack, the yields of four good dies
and three good dies for the case where four wafers are stacked, based on the
combination probability at Y = 0.68, are 0.17 and 0.86, respectively [3,4,25].
With a memory density of 30 Gbit/cm2 (equivalent to 22 nm technology), the
memory capacity of four dies and three dies reaches 120 and 90 Gbit, respec-
tively. If the number of effective chips is 1700 per 300 mm wafer at a chip
size of ~0.4 cm2 and 8 Gbit/die, the expected number of chip sets per unit
memory capacity is 289/32 Gbit (4 stacked dies), 1462/24 Gbit (3 stacked dies),
and 1088/8  Gbit (single die). As the multichannel design supports further

0.8−1.0
1.0 0.6−0.8
0.4−0.6
n=4 0.8
0.2−0.4
m/4 0.6 0.0−0.2
n = 3 Y3D 1/4
0.4
2/4
n=2 0.2
3/4
0.0
n=1 0.9 0.8 0.7 0.6 4/4
Multi-TSVs 0.5 0.4
(a) (b) YS

FIGURE 5.29
Schematic of stack wafers of a combination in four-level stacked wafer (a), and yield for
four-level 3D wafer stack and a comparison of good-die combination at die size = 1.148 cm2
and defect density D0 = 0.2/cm2 (b).
112 3D Integration in VLSI Circuits

DRAM3 12
DRAM2 13
14
CH15 Controller
DRAM1
12
13
14
TSV0
DRAM0 12 13
14
CH15
12
Controller 13 CH15
CH/DRAM0
14
CH15 TSV1

CH/DRAM1
TSV2
CH/DRAM2
TSV3
y
(x, y, 3 CH/DRAM3
)
z (x, y, 2
)
(x, y, 1
)
x (x, y, 0
)

FIGURE 5.30
Schematic diagram of DRAM stack structure. Here, one memory die has 16 channels (CH0
to CH15) in total. Wafers following DRAM2, DRAM1, DRAM0, and controller are stacked to
DRAM3 (base wafer) using bumpless interconnects and a WOW process. Bumpless intercon-
nects are connected independently to the controller die from each channel of the DRAM layers.

yield improvements as described earlier, sufficiently high memory capacity

with high yield can be expected in the WOW memory stack.

5.8.3 Benefits of Co-Engineering

Achieving a capacity of over 24  Gbit per single chip would require tech-
nology two or three generations ahead, such as 10–14  nm technologies.
Defects, however, will become a bottleneck due to the difficulty in control-
ling defects. For instance, subnanometer roughness control of the gate width
along the gate line must be achieved after patterning. This is because, as
scaling proceeds, the so-called stealth defects (invisible defects) increase pro-
portionally, and the control of process variations approaches its limit. When
stealth defects become dominant, variations cannot be improved statisti-
cally and thus die yield deceases with scaling. Considering the technology
roadmap, the issues of scaling technology and technology for fabricating
three-dimensional structures are often discussed separately; for example,
scaling belongs to the front-end wafer process, whereas 3DI belongs to back-
end packaging. However, these two technologies are not always mutually
exclusive. Scaling would be relieved of the stringent requirements by
co-engineering using three-dimensional WOW technology combined with
mass-production technology. In other words, a sufficiently long learning
period would be ensured, and further cost reductions could be expected by
concentrating on the control of variations among generations and shortening
the process. As cost reduction requires the adoption of advanced lithography
technology, advanced lithography and peripheral support facilities account
for one-third to one-fourth of the total cost of a manufacturing line. In short,
although useful for reducing chip cost, scaling is very burdensome in terms
Wafer-Level 3D Integration 113

of capital investment, and the chip cost per unit area is currently saturated,
even after taking account of shrinking chip sizes. In combination with
three-dimensional stacking to overcome such problems, a roadmap toward
high-density integration backed up by production costs can be made. This is
because the capital investment for 3D wafer processes is not high compared
to that of lithography. Moreover, keeping the wafer shape as-is for stacking
ensures compatibility with manufacturing facilities in front-end process-
ing and helps utilize the mature process technologies developed for wafer
processing. If the processes up to three-dimensional stacking can be handled
as units in the manufacturing line, the throughput will be 1/100th of that in
stacking starting with chips. Therefore, future semiconductor manufactur-
ing is expected to advance with a roadmap in which the number of stacked
wafers, the wafer thickness, and the number of TSV interconnects serve as
indices, as shown in Figure 5.31.

105
Multistack 3D multistack × 32
vertical × 32 × 16
Feature size interconnects
10000 × 16 ×8
Feature size (nm), Total Si surface (cm2)

104
7000 ×8 450 mm × 4
3000 300 mm × 4

103 1000 450 mm

500 300 mm
350 Low cost
200 mm Low power
180
150 mm 130 High density
102 90
100 mm 65
45
32
50 mm 2D scaling 22
15
101 Wafer size 10
5
g-line i-line KrF ArF EUV
436 nm 365 nm 248 nm 193 nm 13.5 nm
100
1960 1970 1980 1990 2000 2010 2020 2030 2040
Production year (Risk start)

FIGURE 5.31
Trends of two-dimensional (2D) scaling and wafer size including total Si surface of wafer stack.
Conventional scaling will face difficulties such as physical limits and inability to minimize
costs, whereas 3D integration will become superior to scaling. By combining conventional two-
dimensional integration (2DI) with three-dimensional stacking to overcome such problems
associated with device scaling and increasing wafer size, it is possible to make a roadmap
toward high-density integration backed up by production costs. In volume production, 3D
wafer stacking (WOW) enables a lower cost than chip-on-chip (COC) and high-density inte-
gration, reaching Terabit level. Bumpless interconnects using TSVs and ultrathinning provide
high-density I/Os connecting top and bottom device layers and achieve a small form factor
1/10th that of bump structures.
114 3D Integration in VLSI Circuits

5.9 Summary and Conclusions

WOW technology and bumpless interconnects using TSVs for 3D stacking
in wafer form have been described. It was found that an optimized wafer-
thinning process for determining the stack thickness does not cause degra-
dation of the device characteristics in advanced commercial devices, even
with the smallest thickness of 4 μm. As bumpless interconnects using TSVs
can be connected directly to the upper and lower substrates by self-
alignment, the package thickness can be reduced by an amount equivalent
to the thickness of electrodes, such as bumps, which are not necessary when
bumpless interconnects are used in combination with wafer thinning. As the
design pitch of TSVs is determined by the bump size, high-density TSVs can
be formed in bumpless interconnects by following TSV-patterning processes.
At the same time, size reduction of the finished shape allows the wiring
between the upper and lower chips to be made shorter, which reduces the
total wiring impedance and makes it easier to ensure high bandwidth with
higher energy efficiency. Furthermore, by stacking wafers, high-density
integration and system block arrangements become more flexible, and the
design space is extended.

References
1. Cisco, 2013. http://www.cisco.com/c/dam/en_us/about/business-insights/
docs/ioe-value-at-stake-public-sector-analysis-faq.pdf, By comparison, the
“Internet of Things” (IoT) refers simply to the networked connection of physical
objects (doesn’t include the “people” and “process” components of IoE). IoT is a
single technology transition, while IoE comprises many technology transitions
(including IoT).
2. T. Ohba, N. Maeda, H. Kitada, K. Fujimoto, K. Suzuki, T. Nakamura, A. Kawai,
and K. Arai, Microelectron. Eng., 87: 485–490, 2010.
3. T. Ohba, Electrochem. Soc. Trans., 34 (1): 1011–1016, 2011.
4. T. Ohba, Y. S. Kim, Y. Mizushima, N. Maeda, K. Fujimoto, and S. Kodama, IEICE
Electron. Expr., 12 (7): 1–14, 2015.
5. G. Moore, Electron. Mag., 38 (8), April 19, 1965.
6. J. U. Knickerebocker, P. S. Andry, B. Dang, P. R. Horton, M. J. Interrante,
C. S. Patel et al., IBM J. Res. Dev., 50 (4/5): 553–567, 2006.
7. M. Koyanagi, T. Nakamura, Y. Yamada, H. Kikuchi, T. Fukushima, T. Tanaka,
and H. Kurino, IEEE Trans. Electron Devices, 53 (11): 2799–2808, 2006.
8. E. Beyne, P. D. Moor, W. Ruythooren, R. Labie, A. Jourdain, H. Tilmans,
D. S. Tezcan, P. Soussan, B. Swinnen, and R. Cartuyvels, IEEE IEDM Technical
Digest, pp. 495–498, 2008.
9. H. Kitada, N. Maeda, K. Fujimoto, K. Suzuki, A. Kawai, K. Arai, T. Suzuki,
T. Nakamura, and T. Ohba, IEEE IITC, pp. 107–109, 2009.
Wafer-Level 3D Integration 115

10. Y. S. Kim, S. Kodama, Y. Mizushima, T. Nakamura, N. Maeda, K. Fujimoto,

A. Kawai, K. Arai, and T. Ohba, IEEE IEDM Technical Digest, pp. 189–192, 2015.
11. Y. S. Kim, S. Kodama, Y. Mizushima, N. Maeda, H. Kitada, K. Fujimoto,
T. Nakamura et al., IEEE VLSI Symposium, pp. 22–23, 2014.
12. P. S. Foster, E. Ecker, E. Rutter Jr., and E. S. Moyer, U.S. Patent 5,882,836, 1999.
13. Y. H. So, D. M. Scheck, C. L. Murlick, G. S. Becker, and E. S. Moyer, World Patent
9631805, 1996.
14. Y. Kwona, A. Jindala, R. Augurb, J. Seokc, T. S. Calea, R. J. Gutmanna, and
Jian-Qiang, J. Electrochem. Soc., 155 (5): 280–286, 2008.
15. H. Kitada, N. Maeda, K. Fujimoto, Y. Mizushima, Y. Nakata, T. Nakamura, and
T. Ohba, Jpn. J. Appl. Phys., 50 (5): 05ED02, 2011.
16. D. Diehl, H. Kitada, N. Maeda, K. Fujimoto, S. Ramaswami, K. Sirajuddin,
R. Yalamanchili et al., Microelectronic Eng., 92: 3–8, 2011.
17. N. Maeda, Y. S. Kim, Y. Hikosaka, T. Eshita, H. Kitada, K. Fujimoto, Y. Mizushima
et al., IEEE VLSI Symposium, pp. 105–106, 2010.
18. Y. S. Kim, N. Maeda, H. Kitada, K. Fujimoto, S. Kodama, A. Kawai, K. Arai,
K. Suzuki, T. Nakamura, and T. Ohba, Microelectron. Eng., 107: 65–71, 2013.
19. Y. Mizushima, Y. S. Kim, T. Nakamura, R. Sugie, H. Hashimoto, A. Uedono, and
T. Ohba, J. Appl. Phys., 53: 05GE04, 2014.
20. A. Uedono, Y. Mizushima, Y. Kim, T. Nakamura, T. Ohba, N. Yoshihara,
N. Oshima, and R. Suzuki, J. Appl. Phys., 116: 134501-1–134501-5, 2014.
21. E. R. Weber, Appl. Phys., 30 (1): 1–22, 1983.
22. H. Kitada, N. Maeda, K. Fujimoto, Y. Mizushima, Y. Nakata, T. Nakamura,
W. Lee, Y.-S. Kwon, and T. Ohba, IEEE IITC, 2010.
23. Y. Mizushima, H. Kitada, K. Koshikawa, S. Suzuki, T. Nakamura, and T. Ohba,
J. Appl. Phys., 51: 05EE03, 2012.
24. H. Ryoson, K. Fujimoto, and T. Ohba, JIEP Proceedings of the International
Conference on Electronics Packaging (ICEP), pp. 522–525, 2017.
25. K. Matsumoto, S. Ibaraki, K. Sueoka, K. Sakuma, H. Kikuchi, Y Orii, F. Yamada,
K. Fujihara, J. Takamatsu, and K. Kondo, 29th IEEE SEMI-THERM Symposium,
2013.
26. T. Ohba, IEEE Proceedings of IITC, 12.1, 2010. doi:10.1109/IITC.2010.5510747.
6
Three-Dimensional Integration Stacking
Technologies for High-Volume
Manufacturing by Use of Wafer-Level
Oxide-Bonding Integration

Spyridon Skordas, Katsuyuki Sakuma, Kevin Winstel,

and Chandrasekharan Kothandaraman

CONTENTS
6.1 Introduction ................................................................................................ 118
6.1.1 Rationale for Three-Dimensional Integration ........................... 118
6.1.2 Rationale for Wafer-Bonding Technology .................................. 120
6.1.3 Description of Work ...................................................................... 120
6.2 Chip-Level Three-Dimensional Integration with 22 nm
Complementary Metal–Oxide–Semiconductor Technology ............... 121
6.3 Wafer-Bonding Technology for Three-Dimensional
Integration Stacking .................................................................................. 125
6.3.1 Metal–Metal Bonding.................................................................... 126
6.3.2 Hybrid Bonding ............................................................................. 127
6.3.3 Oxide Bonding ............................................................................... 128
6.4 Oxide-Bonding Technology for Embedded Dynamic
Random Access Memory Stacking .......................................................... 129
6.4.1 Oxide-Bonding Layer Preparation .............................................. 129
6.4.2 Wafer-Level Three-Dimensional Integration with
Oxide Bonding ............................................................................... 131
6.5 Key Electrical Metrics Performance Results for Embedded
Dynamic Random Access Memory Stacking ........................................ 135
6.6 Summary..................................................................................................... 140
Acknowledgments .............................................................................................. 141
References............................................................................................................. 141

117
118 3D Integration in VLSI Circuits

6.1 Introduction
It is a very well understood fact that device and interconnect (IC) scaling in
semiconductor technology applications is becoming increasingly difficult
and more expensive with each new technology node. The traditional model
of continuous electric field scaling that had been for decades the key to the
success and growth of the microelectronics industry [1], and which enabled
Moore’s law, has ceased to be applicable, at least not in a straightforward way.
However, the semiconductor industry is still expected to follow Moore’s law
in the economic sense and continues to deliver significant improvements in
device and chip performance to enable yet more applications, with each new
generation of technology, while still supporting concomitant decreases in cost
per function. The driving force for the last two to three decades is the inces-
sant implementation of new innovations to enable the continuation or even
acceleration of the business model based on Moore’s law. Many new materi-
als have been introduced along with new advanced processes and increas-
ingly complex integration schemes and device architectures. This has been
undoubtedly necessary as the traditional silicon-based materials and tradi-
tional complementary metal–oxide–semiconductor (CMOS) architectures
reached their physical limitations as the scaling trends demanded struc-
tures increasingly in the nanoscale. Furthermore, interconnect scaling has
also resulted in resistance–capacitance (RC) delays now becoming a major
contributing factor to limiting overall performance of systems, which were
usually significantly gated primarily by transistor performance in the past.
All these innovative approaches have driven collectively the tooling and fab
costs much higher and have increased the risks, cost, complexity, and the time
involved to deliver technologies through the pipeline, from early research
and development (R&D) to manufacturing ramp up, and to fully mature
product. To make matters worse, the performance gains with each technol-
ogy node are not as sizeable or straightforward as in the era of traditional
scaling, which makes the cost-benefit considerations for future investments
even more challenging. It is becoming increasingly apparent that revolution-
ary rather than evolutionary solutions are needed to sustain or even acceler-
ate again the desired rate of combined improvements in the performance and
the costs of the technology. This continued demonstration of technological
prowess is still the hallmark of the semiconductor industry and, combined
with the cost efficiencies, it is expected by both enterprise and individual cus-
tomers and has been an essential factor in the proliferation of semiconductor
technology applications across all areas of human life and enterprise.

6.1.1 Rationale for Three-Dimensional Integration

In keeping with this spirit of urgent and revolutionary innovation, one
of the most promising avenues toward the desired direction of increased
3D Integration Stacking Technologies 119

performance and cost-effective system scaling is three-dimensional integra-

tion (3DI) technology [2,3]. According to the 3DI technology paradigm, system
performance scaling can be achieved by the stacking of chips in a homoge-
neous or heterogeneous fashion that can enable performance and functional-
ity gains even beyond Moore’s law. Due to its use of the third dimension, 3DI
can realize advantages with regard to form factors and can achieve higher
device density per unit area compared to two-dimensional (2D) architec-
tures [4]. Furthermore, the stacking of chips can enable significant reduction
of length for long-range interconnect paths relative to traditional 2D designs,
which can in turn enhance performance and also can contribute to lower
power demands than the 2D designs [5]. This 3D stacking of chips can be
realized in a variety of ways, which include package-to-package stacking,
die-to-die stacking, die-to-wafer stacking, and wafer-to-wafer stacking. The
main advantage of the package-to-package and die-to-die stacking is that
they inherently require and enable the selection of known good die (KGD) for
the stacking, thus providing guaranteed higher yield versus other stacking
schemes [5,6]. However, the performance improvements compared to tradi-
tional 2D designs can be limited due to the nature of the bonding technol-
ogy and integration schemes. In addition, these involve a higher cost penalty
because typically each die needs to be individually handled during the built
of the stack. In these 3D integration schemes, the through-silicon vias (TSVs)
that are necessary for chip-to-chip interconnections in the package and die-
level stacks are typically of relatively large size. This is definitely a limiting
factor with regard to the overall bandwidth and performance of these sys-
tems, as they limit the chip-to-chip interconnection density. On the other
hand, 3D stacking by employing wafer-to-wafer bonding can allow for the
use of traditional silicon fab wafer-level processing, which in turn can enable
the scaling of TSVs by a factor of >20 [7,8]. This in turn enables much better
performance and higher bandwidth with lower associated manufacturing
cost, as the stacking of wafers can enable the massively parallel stacking of
chips and inherently would be more suitable and efficient for high-volume
manufacturing. However, wafer-stacking technology itself has its own chal-
lenges, primarily due to the compounded yield losses that can occur due to
any nonfunctional die in the wafers used as these stacks are manufactured.
It would therefore necessitate very high wafer-level yields or the implemen-
tation of efficient yield–loss mitigation strategies at the design level through
fault-tolerant circuit design and the integration in the design of repair circuit
architectures. These would enable the mitigation of potential yield losses
but would also increase the design complexity with respect to chip real
estate. Also, wafer-level integration places strict requirements for high pro-
cess yield for the wafer bonding and thinning modules. Overall, wafer-level
bonding is very promising for the realization of cost-effective, high-volume
manufacturing of high-performance 3DI stacks. With the ever-increasing
consumer demand for higher performance chips/stacks of smaller area size,
increased device density, and reasonable cost, which is driven primarily by
120 3D Integration in VLSI Circuits

the memory business and the backside image sensor business, the industry
is increasingly considering wafer-level stacked technology. It is noteworthy
that the scaling of TSVs depends on the aspect ratio of TSV diameter to the
thickness of thinned silicon wafer. It follows that the thinned silicon wafer
thickness needs to scale with the TSV diameter so as to achieve a fixed aspect
ratio as much as possible, and therefore all the related enabling processes
(bonding, thinning, etching, TSV liner deposition, and metal fill) must be
able to perform well to these specifications.

6.1.2 Rationale for Wafer-Bonding Technology

In the case of wafer stacking and thinning, standard wafer-level process-
ing promises more efficient handling and thinning versus the single-die
handling. Also, wafer-level processing allows for smaller TSVs and smaller
associated TSV keep-out areas, which are needed for ensuring that stress
around the TSVs will not affect device performance. This in turn results in
significantly higher vertical interconnection density, thus achieving much
higher bandwidth performance versus chip-level stacking. In addition,
by designing the TSVs so that they directly connect to a wiring level on
another wafer, the need of micropillars for the 3D interconnection can be
eliminated in wafer stacking. This can lead to remarkable improvement
compared to the limitations that are imposed by the pitch-scaling limita-
tions of the rather large structures typically used for chip-level bonding.
This provides another advantage when it comes to bandwidth potential
comparison between these two technologies. It can be surmised that with
respect to implementation in mainstream high-volume manufacturing, the
characteristics of wafer-bonding technology point to two applications pri-
marily. (1) The first is the massively parallel manufacturing of simple 3D
cores and (2) the second is the 3D scaling of commodity memory, as in
the case of memory cube type applications. In both cases, the prohibitively
increasing costs, mainly due to lithography, and the yield and reliability
issues associated with the scaling of 2D technology have made these two
applications prime candidates for implementing this paradigm shift. It is
also a great advantage that wafer-scale 3D stacking allows scaling by lever-
aging existing technology nodes by using chips of established technologies
to achieve better performance, thus reducing costs, risks, and time-to-
market potentially.

6.1.3 Description of Work

In this chapter, we describe two key technologies for chip stacking, (1) one
achieved by use of die-level stacking and (2) another based on wafer-level
stacking, with the main emphasis on the wafer-level stacking technology.
For the former, 3D die-level integration technology was developed using
large-die 22  nm CMOS technology where die assembly was achieved on
3D Integration Stacking Technologies 121

a laminate with seven layers of build-up circuitry on each core side. The
design features integrated TSVs, microbumps, back-end-of-line (BEOL) wir-
ing structures, and assembled controlled collapse chip connections (C4)
joints, with Cu TSVs integrated in the BEOL. Void-free TSVs were formed,
and the additional BEOL levels were fabricated after TSV processing and
subsequently planarized by chemical–mechanical planarization (CMP). The
types of bumps used to achieve interconnection were low solder volume
with Cu pillars for the top die to the bottom die, and high solder volume for
the bottom die to the laminate with the package exhibiting reliable assembly
process, high-quality connections, and very good thermal performance.
For the latter, wafer-bonding process and integration technology was
developed to achieve the stacking of high-performance POWER7™ cache
cores [9] that were built based on 45 nm silicon-on-insulator (SOI) technology
with embedded dynamic random access memory (EDRAM) and a total of 13
BEOL metal levels. For this work, copper TSVs of 5 μm in diameter were used
at a 13 μm pitch for the signal communication and the power delivery to the
stacked cache cores. The individual wafers, each featuring nine BEOL metal
levels, were bonded permanently to each other by using a low-temperature
oxide bonding. The topside wafers were thinned to about 13 μm using a com-
bination of mechanical grinding, chemical–mechanical polish, and dry etch-
ing. Subsequently the TSVs were formed by using conventional lithographic
alignment and definition techniques. Four additional metal levels were also
built post-bonding and TSV definition to complete the interconnection of the
chips in the two wafers and to enable testing. Electrical testing shows very
good performance and device stability.

6.2 Chip-Level Three-Dimensional Integration with 22 nm

Complementary Metal–Oxide–Semiconductor Technology
Realizing the advantages of the KGD benefits for chip stacking, a 3D chip-
level integration technology was demonstrated that involves large Si die
(>600  mm2) design, based on 22  nm CMOS technology with ultra low-k
(ULK) BEOL at two different microbump pitch sizes, 61 and 131 μm [10,11].
A cross-section high-magnification optical microscope image of the pack-
age in Figure 6.1a shows the 3D integration with a 22 nm CMOS die stack
using 61 μm pitch microbump on a laminate. Optical image of a bird-eye
view of the 3D package is shown in Figure 6.1b. Die assembly was executed
on a laminate with seven layers of build-up circuitry on each core side. This
is a challenging integration scheme considering the effects of warpage in a
TSV-featuring thin die, which makes the use of conventional furnace reflow
problematic. With laminate substrates also undergoing additional warping
during heating due to thermal coefficient of expansion (CTE) mismatch,
122 3D Integration in VLSI Circuits

22 nm CMOS die 22 nm CMOS die

Microbump
BEOL

TSV

C4 UF
Laminate

Laminate

(a)

22 nm CMOS top die

(>600 mm2)

22 nm CMOS
bottom die

Laminate

(b)

FIGURE 6.1
3D module with IBM’s 22 nm CMOS die (>600 mm2) joined face to face with microbumps, TSVs,
and C4s: (a) cross-section image and (b) optical image.

thin die warpage during 3D integration is increasingly challenging with an

increase in the die size and a decrease in the respective component thick-
ness [11,12].
These challenges, especially when it comes to maintaining thin die
and  laminate coplanarity, were addressed by using a new assembly
technology [11]. More specifically, the Cu TSVs were introduced in the BEOL
wafer manufacturing flow, with additional BEOL levels fabricated after
TSV integration, before die singulation. Two types of bumps were used to
achieve the desired interconnections during individual die stacking and
packaging. The cross-section image in Figure 6.1a shows the integrated TSV,
microbump, BEOL levels, and the assembled C4 joints to the laminate. As
mentioned earlier, Cu-TSV is introduced in the BEOL by using a bottom-up
3D Integration Stacking Technologies 123

Cu electroplating process for TSV filling to ensure void-free TSV formation,

without delamination of the BEOL dielectric stack. Additional BEOL levels
were subsequently fabricated and planarized appropriately by using CMP.
Two types of bumps were used for the interconnection. Microbump
(Cu  pillar/SnAg solder) was used for the top die to bottom die bonding,
whereas C4 was used for the bottom die to laminate bonding. The top solder
system consisted of Cu and SnAg and was joined to Ni/Au pads on the bot-
tom die side, whereas the bottom die side featured traditional SnAg bumps
with terminal Ni joined to SAC 305 (Sn-3%Ag–0.5%Cu) presolder on the lami-
nate side. The gap between the bottom die and the laminate is stable and no
microbump/C4 joint failures were observed.
In order to assess the integrity of the interconnections at various locations,
four-point measurements were used. More specifically, there were a total of
41 different locations of the module assembly in which four-point measure-
ments were executed [11]. This included locations such as die centers but
also potentially more challenging areas such as die corners. In Figure 6.2a
there is a graph featuring the cumulative percent distribution of electrical
resistance from these measurements. It should be noted that the resistance
values obtained include contributions from the microbumps, the TSVs, the
BEOL wiring, and the laminate wiring as well. There were two microbump
geometries that were used for the connection of the bottom and top die. One
was at a pitch of 61 μm and the other at a pitch of 131 μm, respectively. The
main difference between the two cases was that the bump diameter was dif-
ferent and, as expected, with a different height for the pillars. The measure-
ments suggest that the resistance of the smaller bumps at the smaller of the
two pitches is about three to four times higher compared to the resistance
for the larger bumps used at the longer pitch, which is not surprising based
on the larger size of the bumps in the latter case. Based on the tight distribu-
tion exhibited by the measured resistance values, we can conclude that the
assembly process is quite reliable, resulting in modules where the die-to-die
interconnections are of good quality.
In order to validate the promise of this chip-level 3D integration technology,
the large chip stacks with 131 μm pitch microbumps integrated with the lami-
nate underwent thermal cycling (TC), temperature humidity bias (THB), and
high-temperature storage (HTS) testing [10]. Table 6.1 summarizes the stress
conditions that were used for the evaluation of the 3D modules and the corre-
sponding test results. All stress times for these tests were performed accord-
ing to the specifications of the JEDEC standard. The number of lots sampled
and sample size per lot are also shown in Table 6.1. The test methodology
included comprehensive continuity and leakage testing of a variety of macros
that were specifically designed to assess the integrity of the upper-bump to
TSV to lower bump connection, the laminate wiring to these structures, and
the overall BEOL wiring level integrity in both chips in this package.
The thermal performance of the completed package is a key metric with
regard to thermal reliability assessment. The interface from the top of the
124 3D Integration in VLSI Circuits

99.9 99.9

99 99

90 90
Cumulative percent

70 70
50 50
30 61 μm pitch -1 61 μm pitch -2 30
61 μm pitch -3 61 μm pitch -4
10 61 μm pitch -5 61 μm pitch -6 10
61 μm pitch -7 61 μm pitch -8
131 μm pitch -1 131 μm pitch -2
1 131 μm pitch -3 131 μm pitch -4
1

0.1 0.1
(a) Resistance (Ω)

20%
Change in thermal resistance from time 0

Average Max Min

15%

10%
(% of center perf.)

5%

0%

−5%

−10%

−15%

−20%
0
166
412
1000
1500
0
166
412
1000
1500
0
166
412
1000
1500

0
166
412
1000
1500
0
166
412
1000
1500
0
166
412
1000
1500

Center Mid Edge Center Mid Edge

(b) Delta_Rint @ DTC Cycles Delta_Rd-d @ DTC Cycles

FIGURE 6.2
Reliability test results: (a) a cumulative distribution graph of electrical resistance including
microbumps, TSV, BEOL wiring, and the laminate wiring, and (b) thermal reliability of TIM1
(R int) and die-to-die (Rd-d) interfaces through 1500 cycles of DTC (−55/+125°C) accelerated
stressing.

stacked die to the package lid through a thermal interface material level-1
(TIM1), also known as Rint, is monitored by implementing 25 top-chip sensors
that are referenced to the external thermocouple attached to the lid. The sen-
sors are classified based on the location on the die (center, middle, and edge)
to track potential thermal performance degradation during deep thermal
cycling (DTC) (−55°C to +125°C) by a TIM tearing, shearing, or loss of adhe-
sion mechanism.
3D Integration Stacking Technologies 125

TABLE 6.1
Reliability Stress Conditions for Chip-Level 3D Modules and Test Results
Cell Stress Condition JEDEC Spec Requirements Quantity (Lots) Result
A TC-K 0/125°C A104 1000 cycles/0 Fail 31(4) Pass
B TC-G −40/125°C A104 850 cycles/0 Fail 13(1) Pass
C THB 85°C/85% RH/3.6V A101 1000 hours/0 Fail 30(2) Pass
D HTS 150°C A103 1000 hours/0 Fail 13(2) Pass

A new thermal interface in the interconnection level between the die is

unique to this stacked-chip structure. In this face-to-face interconnection
configuration, this interface consists primarily of two BEOL stacks featuring
bump and underfill material. Heat generated in the bottom die must transfer
through this interconnection interface into the top die before it can travel
through the Rint path to the exterior of the package. The key thermal resis-
tance parameter characterizing this interface is resistance die-to-die (Rd-d). The
Rd-d values are also classified as center, middle, and edge tracking groups,
similar to the sensors.
Both Rint and Rd-d are calculated generally as follows:

Rint (C/W ) =
( Ti (top die) − Tthermocouple ) (6.1)
power

Rd-d (C/W ) =
( Ti (bottom die) − Ti (top die)) (6.2)
power

Thermal reliability is defined by the change in Rd-d or Rd-d over the duration
of an accelerated stress test, such as DTC. In this case, a cumulative 1000
DTC was used with thermal readouts at the 166, 412, 1000, and 1500 cycle
points. As shown in Figure 6.2b, both Rint and Rd-d exhibited extremely stable
behavior throughout stressing. A maximum thermal degradation of <4°C
per 1000 W of power dissipation was observed at the end-of-stress 1500 cycle
readout.

6.3 Wafer-Bonding Technology for Three-Dimensional

Integration Stacking
As described earlier, wafer-scale bonding can be a key enabler with regard
to substantial benefits among which are interconnects (IC) with lower IC
delay, much higher data bandwidth due to tighter pitches, and better power
performance. From the manufacturing costs perspective, the much higher
126 3D Integration in VLSI Circuits

throughput is beneficial as multiple chip stacks can be formed in massively

parallel fashion and then these stacks can be singulated during the subse-
quent process steps. This becomes possible with relatively few additions
and modifications to established infrastructure in semiconductor fabs as
long as the most suitable 3DI wafer-bonding integration scheme is utilized.
We provide in what follows a brief review of main options for wafer bond-
ing that are currently considered and we discuss the rationale for selecting
oxide–oxide wafer bonding to achieve 3D integration of advanced CMOS for
this work.

6.3.1 Metal–Metal Bonding

In the case of metal–metal wafer bonding, metal pillars, metal studs, or
metal microbumps are created on the bonding surface side in each wafer
first. Subsequently, proper alignment and wafer bonding establish the direct
electrical interconnection between elements in the two wafers. There are
many approaches for forming these metallic features that are needed for the
metal-to-metal wafer bonding. Starting with the well-established method
of using C4 interconnects, which is typically used in chip packaging [13],
it quickly becomes obvious that it is not very suitable especially when it
comes to attempting multiple wafer stacking. This is primarily not only due
to the thermal budget constraints that it entails but also due to the large
size of C4s themselves, which would of course lead to very large IC pitches
and therefore would limit bandwidth severely [14]. One alternative method
that is also being considered is using gold/tin (Au/Sn) soldering technol-
ogy for the bonding. This involves microbump fabrication and is based on
the nature of the eutectic compositions of Au/Sn metallurgical systems [15].
This can definitely provide an improvement versus C4s as it achieves direct
electrical connection with low soldering temperature processes. It also pos-
sesses the added benefit of achieving self-aligning during the bonding pro-
cess. Furthermore, it exhibits very good wetting behavior and reasonably
adequate resistance against corrosion. Still, there can be concerns with the
mechanical robustness and the stability of the bonded systems and for this
reason, it typically requires underfill processes to enhance the robustness
[16]. This of course increases the cost and complexity, and when combined
with the fact that alignment performance, critical dimension (CD), and pitch
scaling of interconnections can be challenging, it limits the possibilities for
wide implementation for wafer bonding. As it is limited by the microbump
dimensions and with throughput also being rather limited, it does not appear
to be suitable for high-volume manufacturing.
Solid–liquid interdiffusion (SLID) is another metal-to-metal wafer bonding
method where a higher melting point metal, such as Cu, is combined with
a lower melting point metal, such as Sn. This method relies on the place-
ment of the lower melting point metal in between the higher melting point
metal stud structures [17,18]. A subsequent thermo-compression bonding
3D Integration Stacking Technologies 127

process ensures the melting and diffusion of the lower melting point metal
atoms into the higher melting point metal atoms, thus forming stable inter-
metallic states. However, this technique also is of rather limited scalability
with respect to multiwafer stacking due to the inadequate thermal stabil-
ity of the bonds compared to what is actually needed to handle the rigors
of downstream wafer-level processing. The direct metal-to-metal thermal
compression bonding is, on the other hand, more promising for wafer-scale
bonding compared to other metal-to-metal wafer-bonding techniques. In
this technique, the metallic structures, such as Cu microstuds and pillars,
are built such that they protrude out of the wafer surface, typically after a
recess process for the surrounding dielectric is done [19]. To prevent corro-
sion of these structures, it is essential to achieve clean metal surface before
the thermo-compression bonding. For this purpose, typically clean pro-
cesses are used to remove the surface oxides and any other impurities that
may be detrimental to the bonding quality over time. By using this method,
the advantage of direct electrical connection is retained, whereas high bond
energy is achieved for the metal–metal bond interfaces with stability that
can be enough to withstand downstream wafer-level processing. Having
said that, underfill is typically needed to guard against mechanical stability
and reliability risks, and this impacts the complexity and cost negatively. To
ensure tight pitches and CD, the bonding alignment overlay performance on
a wafer scale must be very accurate and more so in the case of smaller CD
interconnect features. Another challenge is derived from the thermal and
compression stress involved in the bonding process itself, which can affect
alignment performance, process yield, and is also typically of relatively low
throughput, which makes implementation for high-volume manufacturing
less probable.

6.3.2 Hybrid Bonding

In the case of hybrid-bonding wafer-bonding processes, the bonding of the
metallic interconnect pillar features on the wafer surfaces (as described in
metal-to-metal bonding earlier) is also accompanied by the bonding of the
surrounding dielectric surfaces, which can be polymer or more traditional
dielectric materials [20,21]. If a dielectric is used, the most common case is
the use of silicon oxide or silicon oxide-based interlayer dielectric (ILD). In
the case of polymeric adhesive materials, a common material used is benzo-
cyclobutene (BCB) adhesive. To achieve the simultaneous bonding of these
surfaces, specific surface clean and surface preparation and planarization
steps are combined, such as chemical mechanical polishing, plasma cleans,
and wet cleans for surface cleaning and surface activation. An essential
requirement for the successful hybrid bonding is achievement of low-wafer
topography with atomically smooth roughness of the bonding surfaces. It
is also quite common to use recessed metal structures with respect to the
surrounding dielectric. In hybrid wafer bonding, there is typically an initial
128 3D Integration in VLSI Circuits

room temperature-bonding step to preserve bonding alignment overlay,

which is followed by the main bonding process, the thermo-compression
bonding step. This allows the recessed metal features to expand and bond
with such opposing structures in the other wafer or in the case of polymer/
metal hybrid bonding; it ensures bonding of both polymer–polymer and
metal–metal surfaces.
The main advantage of this method is that direct electrical connection is
achieved with a high bond strength across the surface of the wafer, thus
allowing for robust bonding and therefore adequate resistance to down-
stream process stress, both mechanical and thermal. As an additional benefit,
underfill is not required, which is a key differentiation versus metal–metal
bonding techniques due to the increased mechanical stability achieved. The
main challenges with hybrid bonding are primarily with respect to the bond-
ing overlay and alignment performance, which must be very accurate and
more so for increasingly smaller CD features [20,21]. A key concern is also
with the reliability performance as Cu surfaces are not encased in impervi-
ous barriers during the simultaneous bonding of metal and insulator sur-
faces. Thermal and mechanical stresses inherent to the bonding process can
also be a concern for process yield in the case of polymer/metal hybrid bond-
ing. Overall, throughput is relatively low, especially if long thermal compres-
sion steps are involved to ensure strong bonding.

6.3.3 Oxide Bonding

Oxide wafer-to-wafer bonding requires the formation and preparation of
silicon oxide-bonding layers on the wafers to be bonded to each other. These
bonding layers are then typically activated and cleaned by using plasma and
aqueous solution cleaning treatment. This is immediately followed by the
wafers being aligned to each other, contact being initiated, and then releas-
ing the wafers to spread this initial bonding and to preserve the alignment.
This is subsequently followed by a brief post-bonding batch anneal process
that ensures full-strength covalent bonding between the wafers at the oxide-
bonding interface [7]. Wafer-scale integration using oxide-to-oxide bond-
ing promises some very desirable properties that make it a very interesting
choice as an upfront solution when it comes to high-volume manufacturing
and with low-reliability risks [7,22].
To begin with, the use of oxide-bonding layers provides flexibility in
terms of overcoming any topographic challenges on the surface of the
wafers as their preparation and planarization can be tailored accordingly to
accommodate incoming wafer topographies. The throughput of the oxide-
bonding process compared to thermo-compression bonding can be almost
an order of magnitude higher. Furthermore, the bonding alignment perfor-
mance is generally quite accurate and stable as the initial bonding, which
locks the alignment at room temperature. This technique also does not
entail the significant thermal and mechanical stresses that are inherent in
3D Integration Stacking Technologies 129

thermo-compression bonding. It should be noted that bumpless interconnect

formation that is enabled provides not only a significantly higher IC density
for signal transmission and higher power distribution but also more efficient
thermal dissipation. One key disadvantage is that it does not provide for
direct electrical connection, compared to the methods described earlier. But
this can actually become an advantage, as this means that risks with respect
to reliability (compared to the metal-to-metal or hybrid bonding) are mini-
mal if any, as the interconnections are defined later through oxide layers.
The one great advantage of oxide bonding is in the fact that it is much more
robust when it comes to enabling standard wafer-level processing steps and
with relatively high throughput. This is crucial when it comes to the pursuit
of tighter IC scaling of CD and pitch through TSV scaling, which can be
much smaller than the typical microbump-based schemes.

6.4 Oxide-Bonding Technology for Embedded

Dynamic Random Access Memory Stacking
6.4.1 Oxide-Bonding Layer Preparation
The key requirements for the reliable oxide-bonding technology are driven
by the necessity for good-quality oxide films and surfaces [7]. More specifi-
cally, the bonding films must be dense enough to allow for adequate bonding
site density, and they must be atomically smooth to avoid nanoscale voids
in the bonding interface, and they must neutralize any short- and medium-
range wafer surface topography enough to avoid the formation voids during
bonding. In addition, they must be very clean to avoid bonding defectivity
and voids around particles at the bonding interface. Bonding surface flat-
ness and the absence of short- and medium-range topography are crucial for
defect-free bonding.
Considering these requirements, a two-film oxide-bonding film stack solu-
tion was developed. Both the films in this stack were deposited by using
plasma-enhanced chemical vapor deposition (PECVD). This was first depos-
ited based on a tetraethoxysilane (TEOS) process and was optimized in
terms of its thickness and the subsequent planarization so as to overcome
any potential incoming wafer topography, as needed. There was also a spe-
cially designed annealing step that enhanced overall surface cleanliness that
enabled good adhesion properties with respect to the subsequently deposited
second layer. A CMP step polishing step was also used to achieve sufficient
flatness, followed by a cleaning step before proceeding with the deposition
of the second film. The second bonding layer was then deposited based on
silane chemistry to serve as the primary bonding layer interface. Also for
this film, special thermal annealing to improve cleanliness and a polishing
130 3D Integration in VLSI Circuits

step to ensure low-surface roughness were used, according to the demands

for defect-free oxide bonding. Representative results from stylus profilom-
etry measurements obtained from challenging areas on the wafer in terms of
topography, both before and after the planarizing film over a 2 mm by 5 mm
area, suggest that topography is drastically reduced from ~250  nm range
over the scanned area to ~5 nm, as it shown in Figure 6.3a [7]. Furthermore,
atomic force microscopy (AFM) was used to characterize the topmost bond-
ing film surface after preparation is completed. Such characterization shows

Incoming: ∼2.5 KA range topography Post treatment: ∼50 A range topography

3D profilometry—topography across (2 × 5) mm field (50 μm steps) before/after preparation

(a)

Measure Pan Data zoom 10.0 nm

0.0 1:Z output 10.0 μm

(b)

FIGURE 6.3
(a) 3D profilometry imaging results for a (2 × 5) mm field, in 10 nm steps, before and after the
bonding film preparation over a challenging area of incoming wafer topography showing that
topography is improved to ~5 nm over the scanned range. (b) Atomic force microscopy image
showing the atomically smooth surface of the prepared bonding film stack.
3D Integration Stacking Technologies 131

atomically smooth surface with root mean square (RMS) roughness values
between 0.2 and 0.4  nm, as demonstrated in Figure  6.3b. Thus, a smooth
bonding surface is achieved as required for the bonding process. Prior work
has shown the value of using the two-layer bonding stack system.
Although the first TEOS-based layer is thick enough to accommodate
incoming topography, the bond strength that it can deliver upon bonding
is not optimal. However, the use of silane-based low-temperature oxide as
the topmost bonding layer provides better performance. The presence of
defective regions in the silane-based oxide, where multiple Si-OH groups
are concentrated, can accommodate H2O that evolves later during the inter-
facial condensation reactions that take place during the bonding process and
this results in higher bonding energy [23]. This interpretation is confirmed
from the fact that prebaking of these films can increase their bonding energy,
as any already absorbed H2O is removed, leaving room for H2O molecules
generated later during the post-bonding anneal. The fact that the bonding
energy is typically dependent on the N2 plasma power for these films sug-
gests that the surface density of Si-OH is key, and therefore silane-based
films are more suitable than TEOS-based films.

6.4.2 Wafer-Level Three-Dimensional Integration with Oxide Bonding

The oxide bonding was executed by using a 300 mm wafer platform that was
equipped with integrated wet clean module, plasma chamber, wafer pre-
aligner, dual-microscope wafer aligner stage, anneal module, and vacuum
thermal compression module. The bonding process involved the following
steps: (1) loading the host and donor wafers, respectively, (2) activating
the  oxide-bonding layers by use of dual radio frequency (RF) nitrogen
plasma, (3) cleaning the top oxide surface on both wafers via ultrasonication
in aqueous solution, therefore terminating the activated hydrophilic surface
with silicon to hydroxyl bonds, (4) loading the wafers to the aligner stage
chucks with the two bonding surfaces facing each other, (5) aligning the host
(bottom) wafer and the donor (top) wafers by detecting alignment key marks
patterned on the wafers during the last metal-level fabrication through the
calibrated dual aligner microscopes, (6) bringing the two wafers in close
proximity with the bonding surface layers facing each other in the <10 μm
range, (7) initiating contact from the center of the top wafer through a small
piston pin assembly on the backside of the wafer, (8) releasing the donor
wafer from the top chuck to conclude the initial bonding, and (9) waiting
briefly for the van der Waals initial bonding to stabilize and preserve the
alignment before returning the now bonded wafer pair into the wafer cas-
sette for subsequent processing downstream.
The bonded wafer pairs were then batch annealed in a furnace at tem-
perature ranges compatible with BEOL-processing specifications, typi-
cally <350°C, for 2  hours to ensure the full-strength covalent bonding at
the bonding interface. Characterization of the bonding performance of this
132 3D Integration in VLSI Circuits

process has consistently shown void-free bonding as exhibited by scan-

ning acoustic microscopy (SAM). Equally important is the fact that the bond
strength, as characterized by the Maszara method [24], is typically in excess
of >2 J/m2, as it indicates that sufficient strength exists so as to withstand the
rigors of downstream processing, especially the wafer thinning by grind-
ing and polishing and any thermal deposition and anneal process-related
thermal stresses. The typical alignment overlay performance of the bond-
ing as determined by infra-red through-wafer microscopy measurements
remains unchanged after the bonding anneal and is below 2 μm of the target
throughout the bonded wafer chips, in many cases even below 1 μm. This
is important as it enables the design of relatively small landing pads for the
interwafer TSV connections, therefore keeping the real-estate requirements
for these structures at a minimum, which is beneficial in terms of pitch and
interconnect density [7] (Figure 6.4).
The specific purpose of this work is to showcase the suitability of oxide
bonding as a flexible approach to build prototype-stacked devices and to
evaluate key device and interconnect performance elements. Figure 6.5 fea-
tures a schematic representation of the summarized wafer-level 3D inte-
gration process implemented, showing the use of a silicon carrier wafer to
stack a thinned device wafer on a bottom full-thickness device wafer [4]. As
a first step, a blank silicon wafer, intended to be used as handler for a device
wafer, was bonded to a first device wafer by using the low-temperature
oxide-bonding method described earlier. This enabled the subsequent thin-
ning of the device wafer from its backside to a thickness of 10–13  μm by

2.00

0.10

(a) (b)

FIGURE 6.4
(a) Scanning acoustic microscopy image showing the absence of voids in the bonding between
two wafers with an edge area of the bonded wafer system magnified to show virtually no
defectivity. (b) Wafer-bonding overlay map based on measurements from cross-in-box struc-
tures from 13 chips showing bonding overlay performance <2 μm. The insert is an image of
the overlay structures used.
3D Integration Stacking Technologies 133

Carrier

Bulk Si Bulk Si Bulk Si

(a) (b) (c)

Carrier

Bulk Si Bulk Si

(f ) (e) (d)

FIGURE 6.5
Schematic representation of wafer-level 3D integration process, showing the use of a silicon
carrier wafer to stack a thinned device wafer on a bottom full-thickness device wafer. (a) FEOL,
(b) BEOL, (c) carrier attach, (d) thinning, (e) bond, carrier removal, and (f) TSV and BEOL.

using a serial combination of mechanical grinding and polishing, reactive

ion etching (RIE), and wet chemical thinning. All these processes were exe-
cuted by using conventional silicon fabrication equipment and processes.
Immediately after that, another oxide layer was prepared on the thinned
backside of the device wafer and this wafer stack was bonded to another
full-thickness device wafer, with the blanket handler wafer now at the top of
the stack, therefore making sure that the thinned device wafer was bonded
face up versus the full-thickness device wafer, with both device wafers in
the same orientation, that is, face up. In the next step, the handler wafer was
removed from the stack by a similar combination of mechanical grinding
and polishing, RIE, and wet chemical thinning as the one described earlier.
Once this stacking and handler-removal processes were completed, cop-
per TSV interconnects of a diameter of 5 μm were formed through standard
lithographic definition, and a deep RIE etching process was executed from
the front side of the thinned device wafer, which was designed to specifi-
cally address the ILD stacks, bond interface, and thinned silicon strata. After
a post-RIE wet clean, the TSVs were lined with a conformal oxide insulator
liner using a subatmospheric chemical vapor deposition (SACVD) process.
Then another RIE step was utilized to ensure that the TSV insulator at the
bottom of the TSVs would be removed to access the metal-landing pad under-
neath. The wafers then underwent the TSV metallization process, which
includes metal liner and Cu seed deposition using physical vapor deposition
(PVD) processes and a specially designed bottom–up copper-plating process
134 3D Integration in VLSI Circuits

TSVs

45 nm CMOS
stratum-1
showing
BEOL wire 45 nm CMOS
levels stratum-2
showing
BEOL wire
levels

FIGURE 6.6
SEM cross section of donor and acceptor device wafers after wafer-level bonding and removal
of handle wafer and subsequent TSV formation.

to fill the TSVs without voids. As a result, void-free TSVs were formed
successfully with a 5  μm CD at a 13  μm pitch, as shown in Figure 6.6 [4],
where a scanning electron microscopy (SEM) cross section of the joined and
TSV-interconnected donor and acceptor device wafers is shown. This was
followed by the fabrication of three additional copper-wiring levels on the
top thinned device wafer. It is essential for high-performance, low-voltage
applications to use low-resistivity Cu TSVs to limit the IR voltage drop.
Throughout the whole process flow that enabled the wafer-level integra-
tion, all processing and tools utilized were compatible with high-volume
manufacturing.
We should also note that we have demonstrated that this face-to-back
wafer-level bonding integration process can be repeated to create intercon-
nected wafer multistacks by adding more thinned wafers in the same fash-
ion as the second wafer. With respect to the extendibility of this technology
to include smaller diameter TSVs and thinner Si strata, we have successfully
demonstrated four-wafer strata structures that feature 1  micron diameter
interwafer TSVs chain structures at a Si strata thickness of ~6  μm, where
each wafer also features intrawafer TSVs of 0.25 μm diameter, to connect the
front and back side of the same wafer [8]. As can be seen in this prototype
structure, SEM cross-section image shown in Figure 6.7a, a handle silicon
wafer, four strata of wafers with both interwafer and intrawafer TSVs, and
associated TSV chain structures were demonstrated. For each stratum, the
intrawafer TSVs that eventually connect its front to its back are fabricated
when the wafer is at full thickness from the front side. These intrawafer
TSVs are of 0.25  μm in diameter. Once these are connected to the respec-
tive thin wire levels on the front side of the wafer, the bonding layer is pre-
pared and then bonding ensues to the handler wafer (for the first stratum)
3D Integration Stacking Technologies 135

Strata 4

Strata 3 A2
E2
B2
Strata 2
F1

Strata 1

Handle wafer
(a) (b)

FIGURE 6.7
(a) SEM cross section of donor and acceptor device wafers after wafer-level bonding and
removal of handle wafer and subsequent TSV formation. (b) SEM cross sections showing the
formed TSV chain structures from wafer to wafer for evaluation for TSV performance.

or on stratum n−1 for the n stratum. In each case the newly bonded wafer is
thinned to <7 μm to reveal the thin intrawafer TSVs from its back side. For
n >  1 strata, interwafer TSVs of 1  μm in diameter are fabricated from the
backside, using a process that etches the remaining silicon of the stratum
and its complex front-end-of-line (FEOL)/BEOL dielectric stack on the front
side, the bonding layer, and the ILD stack on the back side of the n−1 stra-
tum. Subsequently, backside wiring is defined to complete test chains that
are formed for the intra- and interstrata connections, respectively, as shown
in the SEM cross-section image in Figure 6.7b. The low resistivity of Cu
achieved through the special TSV-filling processes, especially the intrawafer
TSVs, is extremely powerful and essential for the performance requirements
with regard to the key potential application of multistacked DRAM memory.
This TSV design system can allow for an unprecedented interconnection
density, with very tight allowable pitches. This opens a very promising inte-
gration route toward high-volume manufacturing low-power, high-density
3D DRAM memory stacks.

6.5 Key Electrical Metrics Performance Results for

Embedded Dynamic Random Access Memory Stacking
The wafer-stacking processes involve thermal processes up to 350°C; due to the
differential thermal expansion mismatch present among metal/dielectric/
silicon layers, this results in significant stress on the devices and structures,
a serious concern for structural and parametric integrity. Furthermore, mod-
ern CMOS technologies involve strain-engineered elements that are sensitive
136 3D Integration in VLSI Circuits

(a) (b) (c)

FIGURE 6.8
To accurately capture the effect of 3D stacking and TSV processing (a) TSV chains, (b) TSV
banks, and (c) FETs in proximity to TSV were used.

to stress. So any significant perturbation resulting from 3D integration needs

to be characterized thoroughly via structures that are specifically designed to
capture the effect of these perturbations. Specifically, we utilize TSV chain
structures that interconnect different strata (Figure 6.8a), TSV bank struc-
tures that join TSVs in parallel (Figure 6.8b), and structures that place field
effect transistors (FETs) in proximity to TSV (Figure 6.8c). These structures,
while necessary for capturing the impact of the bonding and TSV process-
ing, are not sufficient to capture the full performance characteristics of the
3D stack that requires functional testing of memories at speed.
For this work, we focused on the electrical characterization of the two-
wafer stack prototype featuring high-performance POWER7™ cache cores
that were built based on 45  nm SOI technology with EDRAM [9]. As the
resistance of one TSV is quite small, ranging in the milli-ohms, to accurately
characterize the TSV resistance, chain structures comprising of multiple
links, with each link containing two TSVs and interconnect wiring, were
used. The median of the resistance for a specific chain structure was plot-
ted against the number of links in the TSV chains as shown in Figure 6.9a.
The linear behavior exhibited in the plot indicates that there is a consistent
ohmic behavior, which is expected, with an average resistance of 120  mΩ
per link. This value incorporates both the TSV and the local wire circuit
resistance contributions. This resistance value is quite compatible with

12.0
Measured resistance (Ω)

10.0

8.0
0.4
Probability

6.0
0.3
4.0 0.2
2.0 0.1
0.0
0 20 40 60 80 100 0 5 10 15 20 25 30 35 40
(a) Number of links in the chain (b) Resistance (Ω)

FIGURE 6.9
(a) Plot of measured median chain resistance versus number of links in TSV chain. The typi-
cal resistance per link was 120 mΩ, including the link wiring resistance. (b) Frequency plot of
number of dies measured versus measured resistance for a TSV chain consisting of 82 links.
3D Integration Stacking Technologies 137

high-performance applications that typically require current density values

in the order of 1 Amp/mm2 or even higher [25,26]. The wiring impact from
the necessary keep-out zones was estimated to be less than 1%. The distribu-
tion of resistance from a chain comprising of 82 links from several dies is
shown in Figure 6.9b. The observed distribution is characterized by a stan-
dard deviation of roughly 10  mΩ per link, indicative of a well-controlled
TSV-formation process.
In a 3D stack the TSV is utilized for power and signal transmission. The
ability of the TSV to carry high-frequency signals is characterized by its
impedance that is mainly capacitive. TSV banks comprising of different num-
bers of TSV’s, joined in parallel, were constructed and their leakage and the
capacitance characteristics were measured. Figure 6.10 shows the cumulative
distribution of TSV banks with different numbers of TSV links. The capaci-
tance scales with the number of links with an individual link capacitance
measured is in the range of ~80  fF, a value which is significantly less than
the case of bump-bonding technology [26,27]. The RC characteristics of these
TSVs are consistent with the typical 100–200  μm length wiring load needs,
which means that 3D macro-to-macro signaling can be achieved easily in this
RC performance space and without having to resort to any buffering solu-
tions. The leakage-to-substrate of the various TSV arrays was also measured
and plotted for different numbers of TSVs in the respective capacitance banks
as shown in Figure 6.11a and was also found to scale linearly with num-
ber. The observed leakage is very close to the lower limit of the tester capa-
bilities, which is a very good sign regarding overall leakage performance.

100 No. of TSV links

40
80
Cumulative distribution (%)

75 120
160

50

25

0
2.5E−12 5E−11 7.5E−12 1E−11 1.25E−9 1.5E−11
Capacitance (F)

FIGURE 6.10
Plot of measured TSV bank capacitance versus number of TSVs in the bank. The capacitance
measured per TSV was extracted as 80 fF per link.
138 3D Integration in VLSI Circuits

100
Cumulative distribution (%)

No. of TSV links

40
80
75 120

Probability
160 0.10

50 0.05

25

0
−1

−1

−1

−1

−1

−1

−1
e

1e

3e

5e

7e

9e
0

−4

−1
(a) Leakage current (b) Leakage current (A)

FIGURE 6.11
(a) Plot of measured leakage current versus number of TSVs in bank. The worst leakage per TSV
was well below the pA level, suitable for most demanding CMOS applications. (b) Frequency
plot of the distribution of the leakage current from a bank with 120 TSV.

The calculated current leakage per TSV is 1.18 pA at a 2V load, which means

that for a 400 mm2 chip design, which may feature more than 50 thousand
TSVs, the total TSV leakage current would be approximately 59 nA. The dis-
tribution of leakage current measured, captured in Figure 6.11b, confirms
consistent values in a very low range and practically negligible when it
comes to most high-performance applications.
One key metric to verify that the devices are not perturbed by wafer
bonding and the TSV integration process is to evaluate Idlin/Ioff char-
acteristics for the FET devices in close proximity to TSV. In this respect,
Ioff (Figure 6.12a) and Idlin (Figure 6.12b) were measured respectively for
devices in proximity to TSV and devices far from TSV (control devices).
The data were repeated before and after bonding. As it is apparent in
Figure 6.12a and b, the FET devices Idlin/Ioff show no significant changes
post wafer stacking and TSV formation and metallization versus the mea-
surements on the single 2D wafers [22]. The TSV-induced Ion/Ioff variation

−2.60e−4
Device
Device
−2.50e−4 with
with
Device Control TSV
3.00e−7 Device Control TSV
with −2.40e−4
with
Control TSV Control
TSV −2.30e−4
2.00e−7
Idlin (A)
Ioff (A)

−2.20e−4
1.00e−7 −2.10e−4

−2.00e−4
0.00e+0
−1.90e−4

−1.00e−7 −1.80e−4
Prebond Postbond Prebond Postbond
(a) (b)

FIGURE 6.12
(a) Plot of Ioff for FET devices, control and devices with TSV in proximity. The data before
bonding and after bonding showed no effect from either introduction of TSV or bonding.
(b) Same as (a) but Idlin, a sensitive metric of TSV, process effects is shown.
3D Integration Stacking Technologies 139

1.54
Passives

1.4

1.25

Clock frequency (GHz)

1.15
eDRAM SRAM
1.05

0.975
Reliability 0.9

0.85

HSS PTN 0.8

0.8 0.85 0.9 0.95 1 1.05 1.1 1.15 1.2 1.25 1.3
(a) (b) Vdd (V)

FIGURE 6.13
(a) The chip layout of a one strata. This stratum is joined to another identical stratum via 3D
stacking (b), which shows the lowest functional frequency versus Vdd shmoo for eDRAM read
and write operations demonstrating the functionality of a stacked eDRAM.

for NMOS transistors was not measured for the purpose of this work but it
is expected to be smaller than P-type metal–oxide–semiconductor (PMOS)
devices as demonstrated previously [27,28].
Through this wafer-level integration process, a high-performance 45 nm
SOI stacked chip (layout is shown in Figure 6.13a) with stacked EDRAM cache
prototype was built with more than 11,000 integrated TSVs. The chips in the
two different strata are designed to emulate a stacked processor and cache
chip assembly, as described in a previous work [14]. A built-in-self-test (BIST)
engine was designed for the purpose of testing the strata-to-strata commu-
nication performance and to test the memory functionality of these struc-
tures for each stratum. The BIST engine successfully accessed the EDRAM
on both strata of this prototype. The resulting shmoo plots, Figure 6.13b, of
clock frequency versus supply voltage demonstrate 16 Mb EDRAM function-
ality, which is fixable, and points to possible strata-to-strata communication
frequencies of up to 2.1 GHz at 1.3 V. The memory patterns for these tests
were written in this stacked EDRAM prototype using four different configu-
rations, as described in the following, and are also shown in Figure 6.12 [22]:

1. Single 2D-thick wafer mode where the memory on the thick wafer
was activated.
2. Bonded 2D-thin wafer mode where the memory on thinned wafer
S1 was activated and the test patterns were loaded using the TSVs.
3. 3D mode where the BIST engine on the thick wafer controls the
memory on both wafer strata.
4. 3D mode where the BIST on the thin wafer controls the memory on
both wafer strata.
140 3D Integration in VLSI Circuits

By evaluating modes 3 and 4, the ability to write and read data from alter-
nating strata memory in a single cycle was demonstrated, which confirms
the quality of the clocks, power, control, and data signals across the chip
boundary. This is necessary to be able to transfer the data at speed for the
entire memory assembly without any errors. Furthermore, it is obvious
from the shmoo plots in Figure 6.13 that the failure mechanism is shared by
all  the modes evaluated, all plots are practically identical, which indicates
that the EDRAM performance was not affected by the wafer-level bonding
and integration of the TSVs to achieve interconnection. The pattern shmoo
that was used was the march 9 pattern, which by design forces cycle-to-cycle
simultaneous switching patterns across the strata boundary for this evalu-
ation. In addition, an equivalent column march pattern was also evaluated,
with the results being very similar as well. The maximum allowed fre-
quency at wafer test was limited by the voltage drop, which is inherent in
the cantilever-probing method used, compared to socket-based module test.
The retention signature obtained for the EDRAM indicates retention times
achieved that were over 200 μs.

6.6 Summary
Two different technologies of 3D technology for chip stacking were demon-
strated, both featuring very promising results. The first is a die-level stacking
integration scheme featuring 22  nm CMOS technology with ULK BEOL,
which enabled die assembly on a laminate with seven layers of build-up
circuitry on each core side. These featured BEOL-integrated TSVs, micro-
bumps, BEOL wiring, and C4s used for the bonding and interconnection
were low-volume SnAg solder with Cu pillars for the top die to the bottom
die and C4 for the bottom die to the laminate. Interconnect characterization
showed high yield and high integrity of microbumps, TSVs, and C4 joints.
Very good thermal performance was exhibited, with the test structures pass-
ing TC-K, TC-G, THB, and HTS, thermal stress test comfortably, which makes
this a very good candidate for die-level stacking and packaging.
As the main focus of this work and aiming at high-volume manufactur-
ing of high-performance memory, a wafer-level integration technology was
developed through oxide bonding stacking of high-performance POWER7™
45 nm SOI technology cache cores with EDRAM. The measured TSV capaci-
tance and resistance are compatible with high-bandwidth chip–chip com-
munication and the TSV leakage performance of 1 pA/TSV is very good. The
FET Ion/Ioff shows no significant change post stacking and TSV processing,
and functionality of 3D-stacked EDRAM cache cores has been confirmed by
successful writing of memory patterns at up to 2.1 GHz at 1.3 V. This technol-
ogy is highly compatible with existing high-performance logic technology
3D Integration Stacking Technologies 141

and thus it is a very promising candidate for high-volume production of

3D-scaled logic and memory applications, especially since the technology
elements for (a) further aggressive scaling of TSVs have been demonstrated
and (b) n > 2 wafer stacking, up to four-wafer stacking have been demon-
strated, promising higher performance and adequate extendibility.

Acknowledgments
The authors would like to extend their gratitude to all current and former
colleagues at IBM who were involved in the work of 3DI technology devel-
opment. Special thanks for their work and support to Jonathan Faltermeier,
Wei Lin, Troy Graves-Abe, John Golz, Pooja Batra, Douglas LaTulipe, Alex
Hubbard, Richard Johnson, Allan Upham, Toshiaki Kirihata, Jeffrey Zitz,
Eric Perfecto, William Guthrie, Marcus Interrante, Richard Langlois, Koushik
Ramachandran, Matthew Angyal, Vamsi Paruchuri, Thomas Gow, Mukesh
Khare, Daniel Berger, John Knickerbocker, Subramanian Iyer, and T.C. Chen.

References
1. R.H. Dennard, F.H. Gaensslen, V.L. Rideout, E. Bassous, and A.R. LeBlanc,
Design of ion-implanted MOSFET’s with very small physical dimensions, IEEE
Journal of Solid-State Circuits, 9: 256–268, 1974.
2. S.S. Iyer, T. Kirihata, M.R. Wordeman, J. Barth, R.H. Hannon, and R. Malik,
Process-design considerations for three dimensional memory integration.
In  Proceedings of the Symposium on VLSI Technology, Honolulu, HI, June 16–18,
2009, pp. 60–63.
3. W. Arden, M. Brillouët, P. Cogez, M. Graef, B. Huizing, and R. Mahnkopf, More
than Moore white paper. In International Roadmap Committee for the International
Technology Roadmap for Semiconductors, 2010. www.itrs2.net.
4. M. Koyanagi, H. Kurino, K.W. Lee, K. Sakuma, N. Miyakawa, and H. Itani,
Future system-on-silicon LSI chips, IEEE Micro, 18 (4): 17–22, 1998.
5. K. Sakuma, P.S. Andry, C.K. Tsang et  al., 3D chip-stacking technology with
through-silicon vias and low volume lead-free interconnections, IBM Journal of
Research & Development, 52 (6): 611–622, 2008.
6. K. Sakuma, P.S. Andry, C.K. Tsang et al., Characterization of stacked die using
die-to-wafer integration for high yield and throughput. In Proceedings of the
IEEE Electronic Components and Technology Conference (ECTC), Lake Buena Vista,
FL, 2008, pp. 18–23.
7. S. Skordas, D.C.L. Tulipe, K. Winstel et al. Wafer-scale oxide fusion bonding
and wafer thinning development for 3D systems integration. In Proceedings of
the 3rd IEEE International Workshop on Low Temperature Bonding for 3D integration
(LTB-3D), Tokyo, Japan, May 22–23, 2012, pp. 203–208.
142 3D Integration in VLSI Circuits

8. W. Lin, J. Faltermeier, S. Skordas et  al., Prototype of multi-stacked memory

wafers using low-temperature oxide bonding and ultra-fine dimension cop-
per through-silicon via interconnects. Proceedings of SOI-3D-Subthreshold
Microelectronics Technology Unified Conference (S3S), San Francisco, CA, 2014,
pp. 152–154.
9. J. Barth, W.R. Reohr, P. Parries et al., A 500 MHz random cycle, 1.5 ns latency,
SOI embedded DRAM macro featuring a three-transistor micro sense ampli-
fier, IEEE Journal of Solid-State Circuits, 43: 86–95, 2008.
10. K. Sakuma, S. Skordas, J. Zitz et  al., Bonding technologies for chip level and
wafer level 3D integration. In Proceedings of the IEEE Electronic Components and
Technology Conference (ECTC), Lake Buena Vista, FL, 2014, pp. 647–654.
11. K. Sakuma, K. Tunga, B. Webb, K. Ramachandran, M. Interrante, H. Liu,
M.  Angyal, D. Berger, J. Knickerbocker, and S. Iyer, An enhanced thermo-
compression bonding process to address warpage in 3D integration of large
die on organic substrates. In Proceedings of the IEEE Electronic Components and
Technology Conference (ECTC), San Diego, CA, 2015, pp. 318–324.
12. K. Sakuma, K. Sueoka, Y. Orii et  al., IMC bonding for 3D interconnection.
In  Proceedings of the IEEE Electronic Components and Technology Conference
(ECTC), Las Vegas, NV, 2010, pp. 864–871.
13. K. DeHaven and J. Dietz, Controlled collapse chip connection (C4)-an enabling
technology. In Proceedings of the IEEE 44th Electronic Components and Technology
Conference (ECTC), 1994, pp. 1–6.
14. P.A. Totta, S. Khadpe, N.G. Koopman, and M.J. Sheaffer, Chip-to-package inter-
connections. In Microelectronics Packaging Handbook, Part II, 2nd ed., Tummala,
R.R., Rymaszewski, E.J., and Klopfenstein, A.G. (Eds.). Springer: New York,
1997, pp. 129–283.
15. H. Oppermann and M. Hutter, Au/Sn solder. In Handbook of Wafer Bonding,
Ramm, P., Lu, J.J.-Q., and Taklo, M.M.V. (Eds.). Wiley-VCH Verlag GmbH & Co.:
Weinheim, Germany, 2012, p. 119.
16. K. Sakuma, S. Kohara, K. Sueoka, Y. Orii, M. Kawakami, K. Asai, Y. Hirayama,
and J.U. Knickerbocker, Development of vacuum underfil technology for 3-D
chip stack, Journal of Micromechanics and Microengineering, 21: 035024, 2011.
17. L. Bernstein, Semiconductor joining by solid-liquid-interdiffusion (SLID) pro-
cess, Journal of the Electrochemical Society, 113: 1282–1288, 1966.
18. N. Hoivik and K. Aasmundtveit, Wafer level solid-liquid interdiffusion
bonding. In Handbook of Wafer Bonding. Wiley-VCH Verlag & Co.: Weinheim,
Germany, 2012, p. 181.
19. K.N. Chen and C.S. Tan, Thermocompression Cu-Cu bonding of blanket and
patterned wafers. In Handbook of Wafer Bonding. Wiley-VCH Verlag & Co.:
Weinheim, Germany, 2012, p. 161.
20. L. Di Cioccio, Cu/SiO2 hybrid bodning. In Handbook of Wafer Bonding. Wiley-
VCH Verlag & Co.: Weinheim, Germany, 2012, p. 237.
21. M. Nimura, J. Mizuno, K. Sakuma, and S. Shoji, Solder/adhesive bonding using
simple planarization technique for 3D integration. In Proceeding of the IEEE
Electronic Components and Technology Conference (ECTC), Lake Buena Vista, FL,
2011, pp. 1147–1152.
22. P. Batra, D. LaTulipe, S. Skordas et al., Three-dimensional wafer stacking using
Cu TSV integrated with 45 nm high performance SOI-CMOS embedded DRAM
technology, S3S, 2013.
3D Integration Stacking Technologies 143

23. W. Lin, L. Shi, Y. Yao, A. Madan, T. Pinto, N. Zavolas, R. Murphy, S, Skordas, and
S.S. Iyer, Low-temperature oxide wafer bonding for 3-D integration: Chemistry
of bulk oxide matters, IEEE Transactions on Semiconductor Manufacturing, 27 (3):
426–430, 2014.
24. W.P. Maszara, G. Goetz, A. Caviglia, and J.B. McKitterick, Bonding of silicon
wafers for silicon-on-insulator, Journal of Applied Physics, 64 (10): 4943, 1988.
25. M.G. Farooq, T.L. Graves-Abe, W.F. Landers et al., 3D copper TSV integration,
testing and reliability. In Proceedings of the IEEE International Electron Devices
Meeting (IEDM), Washington, DC, December 5–7, 2011.
26. J. Golz, J. Safran, B. He et  al., 3D stackable 32  nm High-K/Metal Gate SOI
embedded DRAM prototype. In Proceedings of the Symposium on VLSI Circuits,
Honolulu, HI, June 15–17, 2011, pp. 228–229.
27. A. Mercha, G. Van der Plas, V. Moroz et  al., Comprehensive analysis of
the  impact  of single and arrays of through silicon vias induced stress on
high-K/metal gate CMOS performance. In Proceedings of the IEEE International
Electron Devices Meeting (IEDM), San Francisco, CA, December 6–8, 2010,
pp. 2.2.1–2.2.4.
28. L. Yu, W.-Y. Chang, K. Zuo, J. Wang, D. Yu, and D. Boning, Methodology for
analysis of TSV stress induced transistor variation and circuit performance.
In Proceedings of the 13th International Symposium on Quality Electronic Design
(ISQED), Santa Clara, CA, March 19–21, 2012, pp. 216–222.
7
Toward Three-Dimensional High Density

S. Cheramy, A. Jouve, C. Fenouillet-Beranger, P. Vivet, and L. Di Cioccio

CONTENTS
7.1 Introduction ................................................................................................ 146
7.2 Cu/SiO2 Hybrid Bonding.......................................................................... 147
7.2.1 Cu/SiO2 Hybrid Bonding Principle ............................................. 147
7.2.2 Technical Challenges Linked to Hybrid Bonding .................... 148
7.2.2.1 Surface Preparation......................................................... 148
7.2.2.2 Bonding Interface Characterization ............................. 149
7.2.2.3 Bonding Energy Evolution with Temperature............ 152
7.2.2.4 Copper Modification with Temperature ...................... 152
7.2.2.5 Wafer Alignment Consideration ................................... 154
7.2.2.6 Investigation on Copper Diffusion ............................... 154
7.2.2.7 Pad Dimension Reduction ............................................. 155
7.2.3 Electrical Performances Evaluation of Hybrid Bonding .......... 155
7.2.3.1 Electrical Structures Presentation and
Performances ................................................................ 155
7.2.3.2 Environmental Reliability Study .................................. 157
7.2.3.3 Electromigration.............................................................. 158
7.2.4 Hybrid Bonding Maturity Increase: Moving toward
Production....................................................................................... 159
7.2.5 Specificity of the Die-to-Wafer Process Variation ..................... 160
7.2.6 Die-to-Wafer Process Throughput Increase with
Self-Assembly ................................................................................. 161
7.3 3D Sequential: 3D Very-Large-Scale-Integration CoolCube™ ............ 162
7.3.1 3D Sequential: Principle ................................................................ 163
7.3.2 3D Sequential: State of the Art ..................................................... 164
7.3.3 3D Sequential: Integration Process Flow and
Low-Temperature Top FETs.......................................................... 165
7.3.4 3D Sequential: Intermediate Back-End-of-Line ......................... 167
7.3.5 3D Sequential: 300 mm Electrical Demonstration .................... 169
7.4 3D Technologies Comparative Thermal Analysis................................. 172
7.4.1 3D Technology Parameters for Thermal Comparison.............. 172

145
146 3D Integration in VLSI Circuits

7.4.2
Comparative Study Thermal Results .......................................... 175
7.4.2.1 For Hot-Spot Dissipation Scenarios .............................. 176
7.4.2.2 For Uniformly Distributed Power Dissipation
Scenarios........................................................................... 177
7.4.3 Thermal Comparison Conclusion ............................................... 177
7.5 General Conclusion ................................................................................... 179
References............................................................................................................. 179

7.1 Introduction
3D (three-dimensional technology) high-density integrations have gained an
increasing interest in the objective of maintaining high performances and/or
low-power consumption as the Moore’s law was slowing down.
If high-end advanced packaging solutions are developed, mainly for
heterogeneous packaging, those alternatives are not suitable to fulfill the
requirements for power-efficient applications, such as CIS (CMOS [comple-
mentary metal–oxide–semiconductor] image sensor), high-performance
computing (HPC), gaming, and data center.
As an example, back-side illuminated (BSI) imagers players have released
since 2012 many 3D prototypes and some are already in production, mainly
driven by mobile phone applications. 3D technology used for 3D BSI is
based on direct hybrid bonding, a pitch below 10  μm is reachable, pitch
that is not achievable with conventional advanced packaging techniques.
The idea is to dedicate top and bottom wafer to image sensor and logic
function, respectively, rather than integrating both functions on the same
floor plan.
We can predict that this specific application will pave the way for other
products; the 3D approach that will be used will depend on the granularity
scale of the product partitioning, as shown in Figure 7.1.
Below 10 μm pitch of chip-to-chip interconnection, two complementary 3D
solutions are taking over

1. A back-end-of-line (BEOL)-type technology, based on Cu/SiO2 direct

hybrid bonding in the range of a few micrometers of pitch.
2. A front-end-of-line (FEOL)-type technology, named CoolCube™ at
CEA-Leti, in the range of few tens of nanometers of pitch.

The basis of these principles, process integration, challenges, and perspec-

tives are described in this chapter. Besides, 3D parallel and 3D sequential
both are coming with new and common challenges, among which thermal
dissipation. Last paragraph of this chapter dedicated to 3D high-density inte-
gration will focus on thermal simulation.
Toward Three-Dimensional High Density 147

Wafer stacking
3D Parallel Alignment 50−100 nm

1-Entire core
Granularity scale

Die 1

2-Logic bloc Die 2

3D sequential

3-Logic gates

4-Transistors
CoolCubeTM
Alignment <3 nm

FIGURE 7.1
3D integration pitch roadmap, depending on design granularity.

7.2 Cu/SiO2 Hybrid Bonding

This paragraph will detail processes and challenges of the BEOL-type high-
density technology, based on the Cu/SiO2 direct bonding.

7.2.1 Cu/SiO2 Hybrid Bonding Principle

The Cu/SiO2 hybrid bonding process is one variety of direct bonding tech-
nique, which refers to a process by which two mirror-polished wafers are put
into contact and held together at room temperature by adhesive forces, with-
out any additional materials [1]. For hybrid bonding, the interfacial materials
are heterogeneous to create the electrical contact right after assembly and
successive annealing.
This technology has become more and more attractive for microelectronic
applications for its advantages, such as throughput and possibility of fine
pitch of interconnection, compared to conventional flip-chip technologies. It
can fit with 3D integration using wafer-to-wafer (WtW) bonding and die-to-
wafer (DtW) bonding, which makes this process suitable for a large variety
of applications.
Several metallic materials have been evaluated for hybrid bonding such as
Au, W [2,3], but among the evaluated materials, Cu is the best candidate as it
is already widely used in semiconductor device industry.
Direct bonding of patterned oxide/copper surfaces at room temperature
and ambient air was developed by CEA-Leti since several years ago. The
following paragraphs describe the characteristics of this integration illus-
trated thanks to most relevant results published in this field.
148 3D Integration in VLSI Circuits

7.2.2 Technical Challenges Linked to Hybrid Bonding

This paragraph focuses on the main challenges to overcome to obtain a
perfect direct bonding.

7.2.2.1 Surface Preparation

Wafer surface preparation before bonding is the key step for a successful
direct bonding, whatever this bonding is patterned or not and whatever
the 3D integration scheme (WtW or DtW). Therefore, to obtain defect-free
hybrid bonding, the best choice is to use a standard BEOL process including
chemical–mechanical polishing (CMP) step and then target a surface as flat
as possible. Main challenges are elimination of typical defects such as dish-
ing or recess of the copper pads on patterned surfaces, or array erosion as
illustrated on Figure 7.2.
At CEA-Leti, an optimized CMP-based process was implemented on pat-
terned surfaces to reach a high-surface planarization level [4]. The CMP opti-
mization (pad/slurry matching) aims to minimize copper dishing and oxide
erosion with respect to the layout. Number of CMP steps can be tuned to
improve Cu/SiO2 selectivity and reduce copper dishing: A three-step CMP
process demonstrated excellent WtW-bonding performances with a CMOS
imager application layout [5] but complementary work has shown that add-
ing a step of copper protrusion could even improve topology management to
enlarge process window and relax design rules for bonding levels [4].
Silicon oxide and copper roughness have been found to be lower than 5Å
(Figure 7.3), which is in specification for excellent direct bonding [4–6].
After surface preparation, topology of pads with dimensions ranging from
3 to 6 μm depending on layout [7,8] has been characterized by atomic force
microscopy (AFM) and shows a recess value that does not depend on pad
dimension or shape (square or octagon).
It has been published that the CMP process could be completed by a
plasma-prebonding treatment (N2, Ar, H2…) to increase adhesion energy

Dielectric
erosion
Dishing Edge over
Oxide erosion
SiO2 Cu TiN
loss Recess

(a) (b) (c)

FIGURE 7.2
Principle of the damascene CMP. (a) Wafer post plating, (b) wafer post-CMP idealistic case, and
(c) wafer post-CMP realistic case.
Toward Three-Dimensional High Density 149

500.0 nm Cu roughness ≈ 5Å
0.0 nm
30.0 nm 500
400
300
0.0 nm
200
8 100

300 400 500 nm

6 100 200

2
SiO2
SiO2 Copper

6 8 μm
2 4

FIGURE 7.3
Post-CMP 3D AFM images showing flat oxide/copper interface and below 5Å roughness
copper surface.

if required [9,10]. However at CEA-Leti, this treatment was not found to be

mandatory and bonding was achieved right after CMP process.

7.2.2.2 Bonding Interface Characterization

After CMP process, both substrates are bonded together at room tempera-
ture, without applied pressure. It is then highly important to check the bond-
ing interface quality to ensure following stages feasibility (such as backside
steps for CMOS image sensors). Bonding quality is assessed through the fol-
lowing parameters:

• Interfacial voids: No interfacial voids should be detected.

• Bonding strength: Adhesion energy should be sufficient to support
the steps following processing.
• Top and bottom electrical pads alignment: Mandatory to respect electri-
cal specifications.

Interfacial voids in bonding interface can lead to electrical connection issues

and then yield loss, as well as wafer breakage. Interfacial voids can be
checked thanks to infra-red light or acoustic microscopy. This second char-
acterization technology is widely used at CEA-Leti and in production as it
enables a fast-complete scan of the wafer. The resolution of the void detection
depends on the frequency of the scanning head, and with a 110 MHz trans-
ducer, the effective lateral resolution is 7 μm and defects as small as 5 nm
can be detected. With this characterization mode, the variation in bonding
quality is revealed by color contrast.
150 3D Integration in VLSI Circuits

Nonbonded zones named voids may be observed at the interface:

• Voids due to particles: They are due to insufficient cleaning before

bonding; they are in the range of few millimeters to several centi-
meters depending on particle size. For example, it has been demon-
strated that a 1 μm particle can lead to a 5 mm width bonding void
for a 525 μm thick substrate [11]. In the Figure 7.4, two acoustic scans
coming from 300 mm bonding, presenting copper pads in the range
of 3–6 μm, are shown. The bonding on the left presents some bond-
ing defects, whereas bonding on right is perfect. The particle defect
is the bright circle on the wafer.
• Voids due to copper dishing: We mentioned in our previous paragraph
that non optimized CMP process may generate copper and/or oxide
recess in the range of few nanometers (Figure 7.2). Yet this recess
could lead to remaining voids at bonding interface and poor electri-
cal connection as shown by b-type defect on Figure 7.4.

However, it has been demonstrated that bonding process could be tolerant

to low topologies remaining after CMP process. This tolerance is obtained,
thanks to copper expansion with temperature that enables the copper-to-
copper connection (Figure 7.5).
Thus, the annealing step that occurs after bonding to strength the SiO2/
SiO2 bonding is also used to compensate copper recess. Evolution of bonded
structures with temperature can also be observed, thanks to scan acoustic
microscopy (SAM). 200 mm-patterned wafers have been processed by pre-
viously described CMP leading to copper dishing and were subsequently
bonded on blanket oxide wafers. These pairs were then successively annealed
at 200°C and 400°C for 2 hours.
After the wafer bonding and successive annealing, the bonding inter-
face quality is observed using a SAM. In Figure 7.6, the evolution with

(a)

(b)

(A) (B)

FIGURE 7.4
(A) Acoustic microscopic images of 300 mm Cu/SiO2 bonding with defects: (a) large circular
defect due to a particle and (b) smaller defects due to excessive copper recess. (B) 300 mm Cu/
SiO2 bonding without interfacial defects before annealing.
Toward Three-Dimensional High Density 151

0.04

0.035
Height increase at the center

0.03
of the pads (µm)

0.025 H_Cu = 0.05

H_Cu = 0.10
0.02 H_Cu = 0.20
H_Cu = 0.50
0.015 H_Cu = 1.00
H_Cu = 1.50
0.01
H_Cu = 2.00
0.005

0
0 10 20 30 40 50
Copper pads diameter (μm)

FIGURE 7.5
Ansys simulation of the vertical displacement of copper pads at 200°C, as a function of pad
geometry. H is the height of the copper line in micron.

Room
temperature

Anneal at
200°C 2 h

Anneal at
400°C 2 h

FIGURE 7.6
SAM images showing the evolution of half Kelvin cross patterns with temperature, the width
of the branch is respectively from top to bottom 25, 25, 20, and 15 μm.

temperature, between room temperature to 400°C, of a SAM signal on a set

of crosses of sizes ranging from 15 to 25 μm can be seen. At room tempera-
ture, the residual CMP dishing makes the copper areas unclosed: The whole
crosses appear bright white. When temperature is increasing, the copper
structures expand, filling the gap left by the CMP process. When the copper
surface comes into contact with the facing oxide surface, bonding occurs and
SAM contrast disappears (anneal at 400°C for 2 hours).
This experiment confirms the vertical copper displacement during post-
bond anneal process and assess SAM relevancy for interfacial bonding
qualification.
152 3D Integration in VLSI Circuits

7.2.2.3 Bonding Energy Evolution with Temperature

Bonding strength can be determined using the double cantilever beam
(DCB) technique [12]. This technique consists of measuring the debonding
length induced by the insertion of a blade at the bonding interface. Figure 7.7
presents SiO2/Cu wafers (copper density 20%) bonding toughness evolu-
tion depending on temperature in the case of 200 mm wafers. The bonding
energy ranges from 1 J/m2 at 200°C to 6.6 J/m2 at 400°C, which is sufficient to
successfully achieve the following stages of process.

7.2.2.4 Copper Modification with Temperature

Electrical performances will directly depend on the copper pads connec-
tion quality. Therefore, many studies have been carried out to investigate the
copper-to-copper direct bonding mechanism [13–15], from the bonding inter-
face at room temperature to its evolution with temperature. Therefore, copper-
to-copper connection was firstly analyzed after bonding and annealing by
focussed ion beam-scanning electron microscopy (FIB-SEM). An example of a
cross-section of a 3.6 μm copper pad after bonding and annealing is shown in
Figure 7.8.

7
SiO2 direct bonding
6
Bonding toughness (J/m2)

Cu/SiO2 pattern direct bonding

5
Cu direct bonding
4

0
10 100 200 300 400
Annealing temperature (°C)

FIGURE 7.7
Comparison of bonding toughness for various wafer configurations. (From Di Cioccio, L. et al.,
An overview of patterned metal/dielectric surface bonding: Mechanism, alignment and char-
acterization, JECS, 81–86, 2011.)

Bonding
interface

3.6 μm

FIGURE 7.8
FIB/SEM cross-section of bonded Cu pads.
Toward Three-Dimensional High Density 153

Neither large void nor copper recess can be detected.

For better understanding of copper-to-copper bonding mechanisms, com-
plementary characterizations such as X-ray diffraction (XRD), X-ray reflec-
tivity (XRR), and transmission electron microscopy (TEM) characterizations
have been realized on wafer fully deposited with copper prior to bonding as
described by P. Gueguen et al. Fully 200 mm copper-plated wafers have been
treated with the previously described CMP recipe to ensure a low roughness
and high hydrophilic behavior. TEM observations in Figure 7.9 have shown
a sharp interface at room temperature into which a 4  nm crystalline layer
could be seen at high magnification. Electron Energy-Loss Spectroscopy
(EELS) analyses have detected the presence of oxygen in this layer which is
thought to be the signature of the presence of a copper oxide at the interface.
At 200°C, the copper oxide becomes thermodynamically unstable. Its dif-
fusion along the bonding interface is observed and diffusion bonding can
happen.
Nanovoids partially filled with oxide are created in parallel to a copper-to-
copper grain boundary growth, which is even more significant after 400°C
post-bond anneal, as shown in Figure 7.10.

5 nm
5 nm

FIGURE 7.9
High-resolution TEM cross section of direct copper bonding without post-bonding anneal.
A 4 nm thick crystalline layer is present at the bonding interface.

Bonding interface Bonding interface Bonding interface

Cu Cu Cu

Cu Cu Cu (b)
After bonding After 200°C anneal After 400°C anneal
500 nm 500 nm 500 nm

(a)

FIGURE 7.10
(a) STEM cross sections of direct bonding with successive post-bond annealing. (b) TEM obser-
vation of a typical interfacial void.
154 3D Integration in VLSI Circuits

More recently, L. Di Cioccio et al. published a complete study to describe

into details the void formation mechanisms in the case of three different
metallic bonding: copper, tungsten, and titanium [15].

7.2.2.5 Wafer Alignment Consideration

As mentioned in introduction, Cu/SiO2 hybrid bonding is shown to address
interconnection pitch close to 1 μm; therefore, to recover vertical intercon-
nect in 3D technology it is mandatory to bond with an alignment accuracy
coherent with the desired interconnection dimension. Indeed, misalign-
ment between top and bottom copper pads would lead to contact resistance
increase.
To reach this goal, the WtW bonding equipment suppliers have been work-
ing for several years to develop manufacturing tools that match the intercon-
nection roadmap. It is now demonstrated that alignment with 200 nm (3σ)
over a full 300 mm wafer can be reached with a perfect control of translation,
rotation, and scaling as shown in Figure 7.11 [6,7].

7.2.2.6 Investigation on Copper Diffusion

It is important to notice that a pad misalignment leads to areas where copper
is directly in contact with oxide material without any diffusion barrier mate-
rial, therefore one could fear a copper diffusion in the dielectric after the
post-anneal bake that can go up to 400°C/2 hours. However, work achieved
by Y. Beilliard et al. [16] has confirmed, thanks to energy-dispersive X-ray
spectroscopy-analytical scanning transmission electron microscopy (EDX–
STEM), that copper diffusion remains under detection limit, even after a
thermal storage of the bonding at 300°C for 336 hours (Figure 7.12).
The Cu/SiO2 interface stability has also been observed by S. Lhosthis et al.
[8] after a 400°C 2 hours annealing.

FIGURE 7.11
Vector map of overlay data quantified from a bonded 300 mm wafer pair using the EVG®40 NT
measurement system. Max overlay here is 250 nm.
Toward Three-Dimensional High Density 155

Bonding interface
Thermal storage: 300°C during 336 h
SiO2 Cu
SiO2 Cu 100
Cu O Si
80

Proportion %
SiO2 y 60
Cu
40

x 20
SiO2 Cu No copper in SiO2 layer
0
x 90 y
Cu 500 nm Position (nm)

FIGURE 7.12
STEM cross-section image and EDX–STEM mapping of the misalignment area of hybrid bond-
ing interface. The graph presents the Cu, Si, and O concentration profiles across the bonding
interface.

However, it is important to note that specific pads can be designed to

reduce alignment criterion. For example, Wang et al. recently communi-
cated on a specific 25 μm bonding pad designed as a grid, which allows a
10 μm misalignment tolerance [9]. This is a very interesting strategy as long
as the device connection density can be relaxed to this innovative larger
pad size.

7.2.2.7 Pad Dimension Reduction

For industrial applications, density of interconnection will have to be contin-
uously increased. Some recent publications have investigated feasibility of
hybrid bonding at very fine pitch below 2 μm for simple integration (1 level
of copper per wafer). Two recent publications have confirmed that even if
the pad dimension is getting as large as conventional copper grain size (1.1
measured on 4.4 μm large pads) [8], hybrid bonding surface preparation pro-
cesses and tool performances enable ultrafine pitch hybrid bonding [9,17].

7.2.3 Electrical Performances Evaluation of Hybrid Bonding

7.2.3.1 Electrical Structures Presentation and Performances
Over the past years, several publications from CEA-Leti and ST Micro-
electronics assessed the electrical behavior of different test structures by
measuring their resistance under direct DC current [8,16,18–21]. Along with
time and process maturity growing, the complexity of the evaluated struc-
tures has increased to be better representative of future industrial products.
156 3D Integration in VLSI Circuits

The cross sections after CMP, bonding, and annealing of these structures
are presented in Figure 7.13. The most recently published ones (case 3) consist
of copper pads and via achieved thanks to dual damascene process.
Before electrical tests, all the bonding pairs have been characterized with
the previously described protocol: No void has been detected, as shown
on the perfectly bonded 300 mm wafer presented in Figure 7.13. This con-
firms the robustness of CMP-based surface preparation process.
In order to enable the electrical tests, a backside electrical contact has to
be created on top wafer. Therefore, the bonded pairs have all been thinned
successfully at various thicknesses: 50  μm with through-silicon via (TSV)
and under bump metallization (UBM) in case 1; 40 μm in case 2 with fur-
ther Ti/TiN/AlSi redistribution layer (RDL) and electrical pads, and more
recently down to only 3 μm on a 300 mm wafer with following aluminum
pads creation process (case 3).
Table 7.1 presents a comparison of the specific contact pad resistances mea-
sured on the three integration schemes presented in Figure 7.13.

400°C (C2)
BSI (C1) (C3)
BSI
Logic

Logic

(A) (B) (C)

FIGURE 7.13
Contact pad cross-sections of the three different hybrid bonding stacks evaluated by CEA-Leti
and STMicroelectronics. Case 1 and 2 are 200 mm wafers, Case 3 is 300 mm wafers. (A) Case 1:
1+1 copper level (From Taïbi, R. et al., ECTC proceedings, 219–225, 2010; Taïbi, R. et al., IEEE IEDM,
2011), (B) Case 2: 2+2 copper level (From Beilliard, Y. et al., IEEE 3DIC, 2014) and (C) Case 3: 3+3
copper level. (Lhostis, S. et al., IEEE 66th ECTC, Las Vegas, NV, 2016; Moreau, S. et al., IEEE 66th
ECTC, Las Vegas, NV, 2016; Jourdon, J. et al., IEEE IRPS, 2017.)

TABLE 7.1
Specific Contact Resistance Comparison of Electrical Structures with Different
Numbers of Copper Layers
Case 1 Case 2 Case 3

Number of copper levels (top + bottom) 1 + 1 2 + 2 3 + 3

Seed layer material TiN TiN TaN/Ta
Smallest pad size 3 × 3 μm 3 × 3 μm 3.6 × 3.6 μm
Wafer diameter 200 mm 200 mm 300 mm
Specific contact resistance @400°C 22.55 mΩ·µm2 154 mΩ·µm2 130 mΩ·µm2
Longest DC chain 29,422 30,160 30,000
Toward Three-Dimensional High Density 157

The specific contact resistance ρc is representative of the resistance induced

by the hybrid bonding interface and can be calculated with the following
formula:

ρc = R c × A c

where:
Rc is the contact resistance
Ac is the contact area

This resistance is extracted experimentally from the difference between

theoretical calculations and experimental measurements of the daisy chain
resistances.
It can be seen from Table 7.1 that the contact resistance induced by direct
copper-to-copper bonding is very low. For example, R. Taïbi et al. have dem-
onstrated that the contact resistance of 3 × 3 cm2 pad can be as low as 2.5 mΩ
[18], which confirms that a copper-to-copper connection obtained with hybrid
bonding is less resistive than more classical μbumps (copper + solder), which
presents a total resistance of more than 50 mΩ (without considering the con-
necting copper lines) [22].
Process maturity increase has recently confirmed that connection yield
could be excellent with hybrid bonding on 300  mm wafers [8]. For exam-
ple, the mean unitary resistance of a 30,000 connection daisy chain with
3.3 × 3.3 μm2 square pads is only 82 mΩ (including connection copper lines)
with a standard deviation of 6 mΩ over a set of more than 100 tested wafers.
Connection yield is close to 100%. This value is lower than the unitary resis-
tance of a daisy chain with 10 μm width μbumps connection, which has been
found to be over 100 mΩ [22].

7.2.3.2 Environmental Reliability Study

In order to investigate the environmental reliability of hybrid bonding inter-
face, all the bonded wafers have been exposed to a variety of reliability tests
including environmental reliability as unbiased Highly Accelerated Stress
Test (HAST) (uHAST, JESD22-A118A), Temperature Cycling (TC, JESD22-
A104D), Temperature Storage (TS, JESD22-A103C), and ElectroMigration
(EM) tests.
Unbiased HAST tests have been achieved on a 416-connection daisy chain
packaged and wire-bonded after dicing [16] or at wafer level on backside
aluminum pads [8]. In both cases, no significant variation of daisy-chain
resistance has been observed after unbiased HAST test (85°C/85% RH/168h),
TC (−50°C/+150°C), and thermal storage (1000 hours +1000 hours).
High-temperature thermal storage or thermal cycling resistance have been
evaluated at wafer level on 300 mm diameter wafers [8]. In this case the per-
formance of the 30,000-connection daisy chain with 3.6  μm pads (7.2  μm
pitch) has been followed. A very small decrease in the electrical resistance
158 3D Integration in VLSI Circuits

of the daisy chains is observed after High-Temperature Storage (HTS). The

maximum difference is −2% after 2000 hours storage, far from the −10% dif-
ference of the reliability specification. This resistance decrease is attributed
to a better copper rearrangement in the lines and via or at the hybrid bond-
ing interface with temperature. No significant evolution of the leakage cur-
rent and capacitance of the combs is observed after thermal storage.

7.2.3.3 Electromigration
EM tests were performed on three structure types to identify possible failure
related to hybrid bonding. This paragraph will focus on results with more
complex structure including copper via and pads (case 3 Figure 7.13). In this
case, EM has been studied using NIST (National Institute of Standards and
Technology)-like or 100-connections daisy chain (DC100).
For NIST-like test structure, failure analysis reveals voids that are always
in the single damascene line of the top/bottom wafer depending on the elec-
tron flow direction (up/downstream).
For DC100, no EM-induced void is found along the daisy chain. Voids are
only localized at the cathode side in the feed line that is the metal line in the
top die (Figure 7.14).
These results support the fact that an electron flow flowing perpendicu-
larly to the hybrid bonding interface is favorable compared to a parallel one
as, in this case, voids and extrusions can be observed in bonding interface
[20]. In addition, one must notice that intrinsic bonding voids originating
from interfacial copper oxide do not move with the electrical current show-
ing that the hybrid bonding process is mature.
Complementary work also investigated the impact of seed layer type on
EM, confirming the higher Cu/TaN/Ta adhesion energy and longer resis-
tance to EM [16].
In conclusion, the hybrid bonding module has no impact on the EM resis-
tance and presents excellent environmental reliability. The weakest spot is
always the BEOL level (top or bottom depending on the electron flow).

(a) (b)

FIGURE 7.14
Characterization of daisy-chain (100 connections) after electromigration tests. (a) Lock-in ther-
mography results (amplitude, 15×, 0–0.15 V, 10 Hz, 60 seconds. White arrow locates a possible
failure) and (b) FIB-SEM cross-section in the area indicated by the white arrow in (a).
Toward Three-Dimensional High Density 159

More recently we have shown that final passivation annealing conditions

could significantly influence the EM lifetime [21]. In the case of high-pressure
Deuterium (HPD2) annealing, EM lifetime is reduced and the degradation is
attributed to the presence of deuterium accumulated at the barriers and cap-
ping layers evidenced by physicochemical characterizations.

7.2.4 Hybrid Bonding Maturity Increase: Moving toward Production

Over the past years, the strong interest of industry for direct hybrid bonding
has enabled to push forward this process through even more complex but
realistic integrations with increasing number of metal levels. We have men-
tioned in the previous paragraph how this process can be sensitive to topol-
ogy variation and that uniformity after CMP is essential to obtain perfect
bonding pairs. Therefore, by increasing the number of underneath layers it
can potentially degrade the final interfacial flatness.
Nevertheless, since 2015, several demonstrations have proven that the
BEOL topology could be successfully managed to obtain perfect WtW bond-
ing despite more than six BEOL levels in each wafer.
First work published in 2015 by L. Benaïssa et al. [5] reported the 200 mm
wafer-level assembly of an advanced image sensor with control logic units
and memories. The stack includes all back-end levels of a 0.13  μm double
damascene. Thanks to a three-step CMP process applied at the bonding level
and a proper wafer conditioning (control of each buried metal-level topol-
ogy), the bonding was achieved at room temperature and under atmospheric
pressure. Despite the total of 12 metal levels (5 + 5 BEOL and 2 added hybrid
bonding levels), no defect was observed by SAM after final 400°C anneal
(Figure 7.15) and alignment accuracy was measured at less than 400  nm.
Excellent copper-to-copper connection has been observed and this final
structure, which was the first demonstration of its kind at that time, confirms
the potential of direct hybrid bonding.

Top wafer: BSI

Bottom wafer: Control logic

3/25/2015 HV mag tilt mode 4 μm 3/26/2015 HV CUIT mag 5 μm
11:55:47 AM 2.00 kV 15.000 x 38° BSE CEA Leti 7:17:19 PM 30.00 kV 24 pA 15.000 x CEA Leti

FIGURE 7.15
FIB SEM characterization of a 3D image sensor stack containing 12 metal levels.
160 3D Integration in VLSI Circuits

Since then, SONY announced in March 2016 the production of the IMX260
3D Imager achieved, thanks to WtW hybrid bonding [23]. It is a 12MP camera
constituted of a five-metal (Cu) CIS die and a seven-metal (6 Cu + 1 Al) image
signal processor (ISP) die. The copper-to-copper pads are 3  μm wide and
present a 14 μm pitch in the peripheral regions and a 6 μm pitch in the pixel
array. This imager is already used as a rear-facing camera in the Samsung S7
mobile and confirms that WtW is mature for mass production.
The two examples mentioned are imager applications in which electrical
signal exit is achieved thanks to wire bonding after top imager wafer thin-
ning. However, to reduce the module size and to fasten data exchange with
other chips, it has been recently shown that hybrid bonding can also be com-
bined with a TSV–last process type [24].

7.2.5 Specificity of the Die-to-Wafer Process Variation

If, as said, the WtW hybrid bonding is already in production for the imager
industry, it naturally raises some questions about yield.
Indeed, stacking two wafers together may erase all possible advantages of
the integration, especially when considering insufficient yield from at least
one of the two wafers stacked.
A DtW approach has then to be considered as an alternative. This approach
has the advantage to stack known good dies (KGD) on known good dies.
Consequently, a presumed 100% yield wafer may be rebuilt. Furthermore, the
DtW integration would enable a multidie stacking with heterogenous func-
tionality, which could widen the application field.
Nevertheless, additional difficulties of this integration lead, so far, to a
reduced level of maturity: Proper handling of the dies is necessary, with a
continuous compatibility with further copper–copper bonding: No degra-
dation nor contamination of the top die surface after dicing is mandatory.
However, in the work published by Y. Beillard et  al., it has been demon-
strated that these difficulties could be overcome [25].
In this published work, a complete comparison between WtW and DtW elec-
trical performances has been achieved on 200mm wafers using daisy chains
with connection number ranging from 4,872 to 29,422. The copper-to-copper
area is 3 × 3 μm2 and the pitch is ranging from 7 to 30 μm. In addition to prove
that bonding interface is morphologically similar to whatever die- or wafer-
level approach, the measured contact resistance was found to be identical to the
one obtained in the WtW approach as well as the theoretical calculated value.
Nevertheless, if it is demonstrated now that pick-and-place equipment
designed for hybrid bonding of DtW can have an alignment accuracy in the
range of +/–1 μm (SET French Equipment company), then moving toward
mass production is still challenging. The current hybrid bonding equipment
throughput is very low, therefore far from the current throughput reached
with standard mass reflow bonding using larger μbumps (>10,000 dies per
hour). Nevertheless, we can predict that equipment suppliers may increase
Toward Three-Dimensional High Density 161

this throughput, a figure of 500 or 1000 dies by hour may be sufficient to

make this integration economically viable, especially for very large dies.
To meet the DtW hybrid bonding throughput required by the industry,
CEA-Leti has additionally developed since several years a self-assembly-
based process, using capillary force between the top die and bottom wafer
surfaces for aligning.

7.2.6 Die-to-Wafer Process Throughput Increase with Self-Assembly

Self-assembly process is based on the use of small water drops to align and
bond dies to a wafer substrate. Chip alignment results from surface tension
minimization of the water drop. Hybridation is achieved, thanks to direct
bonding after the drop evaporation as presented in Figure 7.16.
A high-fluid containment is required to obtain a high accuracy in DtW
self-assembly process. The aim of this containment is to avoid fluid over-
flowing beyond the die surface, which would lead to an alignment loss. The
confinement area is obtained by physical and/or chemical contrast. The
physical contrast consists of a topology modification at die edges. In this
case, the water overflowing is controlled, thanks to the canthotaxis effect.
Canthotaxis condition is detailed in Figure 7.17. The chemical contrast is cre-
ated by the deposition of a hydrophobic material on die edges. It can be seen

(a) (b) (c)

(d) (e)
(A) (B)

FIGURE 7.16
(A): Self-assembly process description: (a) bottom structure before stacking, (b) water drop
deposition and top die prepositioning, (c) self-assembly thanks to capillary restoring forces,
(d) die alignment, residual interfacial water film, and (e) water evaporation and dies direct
bonding. (B): 200 mm electrical wafer stacked, thanks to self-assembly process with 20 top dies.

θ1+α

α θ1 θ2
Liquid
θ1 ≤ ϕ ≤ α + θ2 Solid

FIGURE 7.17
Principle and illustration of canthotaxis phenomena.
162 3D Integration in VLSI Circuits

TABLE 7.2
Comparison between the Contact Resistances Obtained with Three Different
Integration Processes
NIST Structure Contact Resistance
Hybridation Process DtW (Self-Assembly) DtW (Pick and Place) WtW

Contact area (μm²) 340 × 5 340 × 5 340 × 5

Copper resistivity (Ω·μm) 2.72 × 10−2 2.06 × 10−2 2.07 × 10−2
Experimental resistance (Ω) 5.1 3.9 3.89
Theoretical resistance (Ω) 5.09 3.74 3.77
Variation (%) +0.19 +3.08 +4.1

that the best theoretical confinement will be reached by creating both topol-
ogy (high α value) and hydrophobic treatments (high θ2 value).
If interesting demonstrations have already been made with μbumps
assembly [26], CEA-Leti has adapted the technology to the constraints of
hybrid bonding.
The process has been optimized, thanks to simple 1 × 1 cm nonpatterned
dies with both topology and hydrophobic contrast on their edges. These
structures enabled to demonstrate high alignment accuracy yield (>90%) and
high bonding quality (below 1 μm) [27,28].
Finally, S. Mermoz et al. [29] demonstrated that the presence of interfa-
cial water did not modify the electrical performances of the connected daisy
chains (as shown in Table 7.2) for a Kelvin NIST structure. Furthermore, this
work was continuously completed by theoretical simulation analysis using
a numerical software to explain how outer perturbations could impact the
capillary restoring force [30].
To conclude on this paragraph, hybrid bonding Cu/SiO2 has gained many
interest from industry players since few years, whereas first publications
were released in early 2000. In between, impressive understanding studies of
the bonding principle were carried out, many progresses about CMP and its
material were achieved. Finally, production in a WtW scheme is now a reality
for CIS and there is no doubt that other application fields will now continue
the trend. DtW is the natural (but challenging) continuation of the roadmap.

7.3 3D Sequential: 3D Very-Large-Scale-Integration CoolCube™

To increase the density of layers’ interconnects even more and to reach some
pitches in the range of few tens of nanometers, an alternative technology
to parallel stacking is needed. This paragraph will detail process and chal-
lenges of the FEOL-type high-density technology, the sequential integration
CoolCube™.
Toward Three-Dimensional High Density 163

7.3.1 3D Sequential: Principle

An alternative approach to conventional device scaling for future nodes is
the 3D sequential integration (also referred as 3DVLSI (3D Very-Large-Scale-
Integration) CoolCube™ or monolithic 3D integration) [31]. CoolCube™ pro-
cess flow where devices are built one above the other in a sequential manner
offers the possibility to stack devices with a lithographic alignment precision
(few nm), enabling via density >100 million/mm2 between transistors tiers
(for 14 nm design rules). 3DVLSI can be routed either at gate or at transistor
levels (Figure 7.18).
The partitioning at the gate level allows IC performance gain without
resorting to scaling, thanks to wire length reduction [32]. Partitioning at the
transistor level by stacking n-FET (field-effect transistor) over p-FET (or the
opposite) enables the independent optimization of both types of transistors
(customized implementation of performance boosters: channel material/
substrate orientation/raised source and drain, strain, and so on. [31,33]) with
reduced process complexity compared to a planar cointegration. This inte-
gration enables to merge several technologies together; for the bottom layer,
it can be any CMOS from bulk planar to (fin field-effect transistor) FinFET
or fully-depleted silicon-on-insulator (FDSOI). To benefit from the full 3D
opportunities and avoid global routing congestion [33], there is a need to
implement local routing of the bottom tier: Intertier metal layers need to be
incorporated in the technology (Figure 7.19).
As a consequence, intermediate back-end-of-lines (iBEOL) levels need to
support top FET thermal budgets. A reasonable maximum thermal budget for
top FET has been determined around 500°C [34]. Indeed, the silicide has been
identified as the main contributor for the bottom metal–oxide–semiconductor
FET (MOSFET) instability. The summary of thermal limitation is shown in
the graph given in Figure 7.20. The maximum top FET temperature is fixed
around 500°C for 2 hours, however this temperature can clearly be increased
when the anneal duration is reduced [35].

3DVLSI gate level 3DVLSI MOS level

LT MOSFET
CMOS III-V
LT MOSFET
nMOS

Ge
CMOS PMOS

(a) (b)

FIGURE 7.18
CoolCube™ integration: Partitioning schemes for 3DVLSI, namely (a) 3DVLSI at the gate level
and (b) 3DVLSI at the transistor level.
164 3D Integration in VLSI Circuits

Top tier

Bottom tier

FIGURE 7.19
3DVLSI structure with two levels of intertiers interconnections.

1000

Stable FET performance with this TB

800
Temperature (°C)

600

400

200
1n 1µ 1m 1 1k 1M
Anneal duration (s)

FIGURE 7.20
Summary of thermal budgets tested on FDSOI technology with SiGe channel for pFET
and SiGe:B/SiC:P RSD with preserved N &PFET ION-IOFF performance. (From Fenouillet-
Beranger, C. et al., New insights on bottom layer thermal stability and laser annealing prom-
ises for high performance 3D VLSI, IEEE IEDM, 2014.)

7.3.2 3D Sequential: State of the Art

Depending on the partitioning scheme and the technologies used for bottom
and top layers, CoolCube™ addresses challenges associated to CMOS scaling
and to More than Moore applications. One of the main difference in the 3D
sequential integration demonstrations is the top silicon layer formation method
Toward Three-Dimensional High Density 165

(Seed window [36], Poly-Si and laser-crystallized Epi-like Si [39] or oxide direct
bonding [32]). Direct bonding clearly stands out with respect to the other tech-
niques: Thanks to the high crystalline quality of top silicon layer, the devices’
performance outperforms the other ones. Regarding CMOS scaling and espe-
cially SRAM integration, there have been great demands for higher density in
all areas of SRAM applications such as network and cache standalone mem-
ory, and embedded memory of the logic devices. 3DVLSI integration is very
promising for this type of applications as evidenced by the number of pub-
lications in literature [3,33,36–39]. Here again, as evoked previously, the best
transistor performances are always achieved in case of top silicon crystalline
layer obtained by oxide wafer bonding [3,31,33]. The ultimate example of high-
performance CMOS at low process cost is the stacking of III–V nFETs above
SiGe pFETs [40,41]. These high-mobility transistors are well suited for 3DVLSI
because their process temperatures are intrinsically low. 3D sequential integra-
tion, with its high contact density, can also be seen as a powerful solution for
heterogeneous cointegrations requiring high 3D via densities such as compu-
tation immersed in memory [42], heterogeneous IoT (Internet Of Things) chip
[43,44], nano-electromechanical systems (NEMS) with CMOS for gas-sensing
applications [45,46] or highly miniaturized imagers [47].

7.3.3 3D Sequential: Integration Process Flow

and Low-Temperature Top FETs
The process flow principle is detailed in Figure 7.21 [48].
Bottom layer is firstly processed up to a few metal layers within a standard
CMOS flow. In case of thin-film integration, a blanket wafer is subsequently
transferred at low temperature on top of the bottom layer. It is realized through
direct bonding either by a silicon-on-insulator (SOI) wafer [49] or by using

(a) (b) (c) (d)

FIGURE 7.21
Process flow principle of monolithic 3D integration. By resorting to a unique alignment flow
throughout the whole process, layers are stacked on top of each other within a lithographic
alignment precision. (a) Bottom-layer processing with plugs down to the CMOS, (b) fabrication
of the inter level lines to ensure short distance connection with bottom layer, (c) high-quality
top film transfer by direct oxide–oxide bonding, and (d) top-layer fabrication and connection.
(From Vinet, M. et al., Monolithic 3D Integration: A powerful alternative to classical 2D scaling,
IEEE S3S, 2014. © 2014 IEEE. With permission.)
166 3D Integration in VLSI Circuits

hydrogen implantation and splitting [50]. Top-layer lithography relies on a

single-stream alignment flow, leading to the excellent alignment precision
between top and bottom transistors. As mentioned before the top devices’
process flow is limited to 500°C for a couple of hours to not degrade the bot-
tom one’s performance. The temperature of four key process modules needs
to be lowered to match the thermal budget specifications to ensure bottom
MOSFET and interconnects preservation. Analysis of the top MOSFET pro-
cess flow shows that the dopant activation is the most critical thermal budget
contributor. Indeed, in a standard high-temperature process Rapid Thermal
Annealing (RTA), activation anneal steps are performed above 1000°C with a
time scale around 1 second. The three other process modules to address are
the offset spacers (~630°C), the raised source drain epitaxy (SiGe 30% at 650°C
or Si at 750°C), and the gate-stack stabilization. Solid-Phase-Epitaxy-Regrowth
(SPER) at 600°C combined with extension first implantation for dopant activa-
tion yields FDSOI devices with similar ION/IOFF performances than devices
activated at standard temperatures [51]. Nanosecond laser anneal with some
nanoseconds (ns) pulse duration (with low in depth thermal diffusion) is effec-
tive to activate dopants in FDSOI devices with sheet resistance below the high-
temperature reference [52]. Finally, low-k spacers (for parasitic capacitance
reduction) with lower thermal budgets than for SiN atomic layer deposition
(ALD) liners could be used (SiCO) [53]. The epitaxy step has been success-
fully demonstrated at 500°C [54] by using cyclic deposition and etch process.
Regarding the gate stack, the reliability has been demonstrated at 525°C [55]
but further studies toward decreasing temperature while preserving reliability
are needed especially in case of P-type metal–oxide–semiconductor (PMOS)
devices. In order to alleviate this problem and then to relax the maximum
thermal budget allowed for top-level processing in 3DVLSI integration, bot-
tom transistors’ silicide stability must be increased. Thanks to process boosters
(introduction of silicon capping and pre-amorphizing implantation [PAI]), the
thermal stability of NiPt can be extended [56]). In parallel the introduction of a
novel silicide, Ni0.9Co0.1, seems very promising [57]. Thus, the main changes
with respect to standard fabrication flow are summarized in Figure 7.22,
highlighting that, thanks to the bottom silicide stability improvement, all dif-
ferent high-temperature modules have been optimized to fulfill the targeted
temperature criteria for 3D sequential integration [58].
Focusing on contamination issue, all the CoolCube™ process should be
realized within an ultraclean environment respecting stringent contamina-
tion constraints imposed by an industrial clean room. The different species
can be monitored along the process through different steps, especially on the
bevel edges, using Vapour Phase Decomposition (VPD), Droplet Collection
(DC), and analysis by Inductively Coupled Plasma Mass Spectrometry
(ICPMS). An example of results is summarized in Table 7.3 and confirms
the very low Ni concentration fully compatible with a front-end environ-
ment demonstrating the efficiency of the contaminants’ containment strat-
egy developed at CEA-Leti [3].
Toward Three-Dimensional High Density 167

4 modules for
which a
temperature
decrease is
required
Ni0.9Pt0.1
Low-temperature solid-phase epitaxy regrowth/laser
Dopant activation 500°C
1000°C
Desired 800°C
Gate oxide
∼525°C
stabilization
Si @ 750°C
Process window SiGe @ 650°C
Epitaxy ∼525°C

Spacer Low-k 630°C

450°C
deposition

Temperature of the thermal budget

450 500 550 600 650 ... 1000°C

FIGURE 7.22
Bottom-layer thermal stability has been increased, thanks to silicide process optimization,
(From Vinet, M. et al., IEEE ESSDERC, 2016.) and top-layer thermal budget has been decreased
by optimization of hot-temperature modules.

TABLE 7.3
VPD-DC-ICPMS Monitoring of Ni through the Process
on the Bevel. Low Limit of Detection = 6.5E7 at/cm²
Ni (at/cm 2)

Cleaning before bonding <9.4E8

Bonding annealing <9.4E8
High-k deposition 2.9E9
Gate stack etch 8.3E9
Epitaxy 6.4E+09
Dopant activation annealing 2.8E9

7.3.4 3D Sequential: Intermediate Back-End-of-Line

As for the bottom devices, iBEOL levels need to endure top FET thermal
budgets implying to find solutions for implementing back-end material
stable beyond 400°C. Currently the combination of copper with ultra
low-k (ULK) materials is widely used for standard BEOL (low resistivity
and capacitance, and thus speed improvement). However, the integration
of such materials in the iBEOL of a CoolCube™ integration faces a number
of challenges. Indeed, copper metallization can cause contamination issues
in the case of wafer break during the process of the top transistor where FEOL
contamination environment is required. An alternative to copper can be the
168 3D Integration in VLSI Circuits

use of tungsten (W) as it has already been integrated in the FEOL of several
products. CEA-Leti studies [59] highlighted that Cu and W interconnections
combined with ULK are stable up to 500°C for 2 hours and 550°C for 5 hours,
respectively, in case of line 1 integration with 28 nm design rules. However,
W resistance is still larger as compared to the copper one (by a factor 6).
Moreover, as presented in the literature, the ULK stability is still questioned
beyond 500°C [60]. Indeed, the modification of ULK structure and permittiv-
ity during a thermal anneal at temperature higher than 500°C may increase
the leakage and delays of the iBEOL and thus may degrade the circuit perfor-
mances. Thermal stability of dielectrics has been studied to select the most
appropriate ones. Depending on the thermal budget set by the top MOS layer,
various couple of materials are possible and summarized in Table 7.4 [61].
Regarding the barrier layer, SiCO seems to be the most promising material
due to its robust composition, low permittivity (4.5) (lower than the state-
of-the-art barrier layer SiCNH (5.6)), and its high thermal stability. However,
for a top thermal budget limited to 500°C the state of the art SiCNH is still
suitable. Regarding the oxide-based material, several of them seem suitable
depending on the top feasible thermal budgets. For a thermal budget limited
to 500°C, 2 hours, the state-of-the art ULK (2.5) material is still possible. On
the other hand, at 600°C, 2 hours, only the SiO2 is suitable. Finally, the permit-
tivity of these materials is crucial to avoid circuit performance degradations.
The reliability of ULK dielectrics after a relatively high thermal budget
has never been demonstrated up to now. Using the standard extrapolation
parameters, the extrapolated lifetime is extracted for W/ULK interconnec-
tions [55] (Table 7.5). Although time to failure (TTF) decreases with increasing
thermal budget, a lifetime (line to line) of 109 years is calculated for the high-
est thermal budget, larger as to Cu/ULK intermetal dielectric (IMD) lifetime
TABLE 7.4
Possible Dielectrics for the iBEOL as a Function of the Top FET Thermal Budget and
Associated Permittivity
State-of-the-Art 500°C, 2 hours 550°C, 2 hours 600°C, 2 hours

Barrier layer SiCNH (5.6) SiCO (4.5)/SiCN(5.6) SiCO (4.5) SiCO (4.5)
Oxide-based ULK (2.5) ULK (2.5) Dense LK (3) SiO2 (4.1)
Source: Xu, H. et al., Design, Automation & Test in Europe (DATE), 1–6, 2011.

TABLE 7.5
Extrapolated Lifetime (TTF) of Intermetal Diel ULK with Ti/TiN/W before and
after Anneals. Years Needed to Reach the Failure Rate (<1%) with an Operating
Voltage at 1.115 V
No Anneal 500°C 550°C

Ti/TiN/W (years) 4.10

15 (5 hours) 5.10
11 (5 hours)1009
Source: C-M. V. Lu et al., Key process steps for high performance and reliable 3D sequential
integration, IEEE VLSI 2017.
Toward Three-Dimensional High Density 169

without thermal budget. Therefore, W/ULK reliability is not an issue for the
3D sequential integration due to the high initial lifetime, as top transistor
thermal budget is limited to 500°C for a couple of hours.
In spite of contamination issue, it could be very interesting to integrate Cu/
ULK lines in the iBEOL of 3D sequential integration to use standard bottom tier
process, however the reliability versus thermal budget should be studied.
Integration of dense W lines introduces a new complexity degree not only
in terms of contamination but also in terms of CMP. Indeed, in spite of the
high metal lines density, the planarity of the structure should be preserved to
ensure a good bonding quality. A layer transfer by wafer bonding has been
realized above a 28 nm industrial metal 1 short loop with line densities up to
70% [3]. The schematic of the experiment is illustrated in Figure 7.23. W filling
is used instead of standard copper, coupled with a Ti/TiN barrier. A high-
quality layer transfer as observed with acoustic microscopy observation is
presented on Figure 7.24. Figure 7.24a shows a SiO2/SiO2 reference bonding
(without any defect that would appear as white dot), whereas Figure 7.24b
shows the bonding above W lines with very few bonding defects at the
wafer edges. These residual defects are explained by a nonfully optimized
W lines CMP process that will be easily adjusted on a real lot. The bonded
structure was then thinned down to the top buried oxide (BOX) with both
grinding and tetramethylammonium hydroxide (TMAH) etching. A SEM
cross-section of the structure after thinning with the Si layer highlighted in
red is presented in Figure 7.24c.

7.3.5 3D Sequential: 300 mm Electrical Demonstration

For the first time in 2016 [3], a full CMOS over CMOS 3D sequential integra-
tion has been demonstrated with a top level compatible with state of the art
high-performance FDSOI process requirements such as high-k/metal gate

Direct bonding
200 nm 90 nm

40 nm

W W Si 9 nm
Ti/TiN Ti/TiN BOX
Bulk Si
350 nm

SiO2
SiN
Bulk Si

FIGURE 7.23
Description of the Si layer transfer above W metal 1 level. Ti/TiN diffusion barrier is used.
170 3D Integration in VLSI Circuits

9 nm Si layer

W lines Bonding interface

500 nm

(a) (b) (c)

FIGURE 7.24
(a) Acoustic microscopy on a reference SiO2/SiO2 bonding. Perfect bonding is observed (no
defect) and (b) acoustic microscopy on the studied structure. High-quality bonding reached.
Some defects on the edge due to W over-polishing. (c) SEM cross-section of the bonded struc-
ture, with bottom W lines, after thinning. The 9 nm Si layer is highlighted in dark gray.

and raised source and drain. On the bottom level, standard high-temperature
NMOS is N-type metal–oxide–semiconductor (NMOS) and PMOS tran-
sistors with HfSiON/TiN gate stack are fabricated on 300  mm SOI wafers
(tBOX = 145 nm/tSi = 7 nm) with Si raised source and drain junctions activated
at 1050°C. For the top level, a HfO2/TiN gate stack was formed followed by
a spacer zero deposition at 630°C and a selective SiGe27% raised source drain
epitaxy at 650°C for both NMOS and PMOS transistors. The junctions were
activated by SPER technique during 1 minute at 600°C. Figure 7.25 shows the
TEM cross section of the 3D sequential contacted structure with a focus on
two stacked transistors. The Id-VG characteristics for both NMOS and PMOS

200 nm 50 nm

FIGURE 7.25
TEM cross-section of the 3D sequential structure.
Toward Three-Dimensional High Density 171

104 104
103 Top NMOS 103
Top PMOS
102 102
Drain current ID (µA/µm)

Drain current ID (µA/µm)

101 101 L = 60 nm
100 VD = 0.04 V 100 W = 10 μm
10−1 L = 60 nm 10−1 VD = 0.04 V
10−2 W = 10 μm VD = 1 V 10−2
10−3 10−3 VD = −1 V
10−4 10−4
10−5 10−5
10−6 ION = 419 μA/μm 10−6 ION = 228 μA/μm
10−7 IOFF = 8.5 × 10−12 A/μm 10−7 IOFF = 5.8 × 10−11 A/μm
10−8 DIBL = 31 mV/V 10−8 DIBL = 61 mV/V
10−9 10−9
−0.5 0.0 0.5 1.0 1.5 −1.5 −1.0 −0.5 0.0 0.5
Gate voltage VG (V) Gate voltage VG (V)

FIGURE 7.26
ID-VG characteristics of both top-level NMOS and PMOS transistors. W = 10 μm, Lg = 60 nm.
(From Brunet, L. et  al., First demonstration of CMOS over CMOS 3D VLSI CoolCubeTM
integration on 300 mm wafers, IEEE VLSI, 2016. © 2016 IEEE. With permission.)

transistors on the top level are presented in Figure 7.26. Finally, for the first time
the voltage transfer characteristics of 3D sequential inverters with either NMOS
or PMOS on the top level, on 300 mm wafer, are presented in Figure 7.27.
As a conclusion, 3DVLSI sequential 3D integration CoolCube™ success-
fully achieved in the past 10  years to prove the feasibility of the top layer
manufacturing and its robustness. This innovative integration, in a blaze of
many high-level worldwide publications, could be the answer for the end of
Moore’s law that some specialists predicted in a near future. In addition, in
the same manner as for advanced packaging or 3D high density, heteroge-
neous integration may be also a key advantage offered by CoolCube™, in a
range of pitch that is more aggressive than parallel stacking.

1.4 1.4
NMOSTOP VDD = 1.2 V VDD
1.2 VDD = −1.2 V 1.2
1V PMOSTOP
1.0 −1 V VIN 1.0 VOUT
VOUT 0.8 V VIN
VOUT (V)

VOUT (V)

0.8 −0.8 V 0.8

PMOSBOT V 0.6 V
0.6 −0.6 V DD 0.6 NMOSBOT
0.4 0.4 0.4 V
0.2 W = 1 μm 0.2 W = 1 μm
L = 100 nm L = 100 nm
0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 1.2 0.0 0.2 0.4 0.6 0.8 1.0 1.2
(a) VIN (V) (b) VIN (V)

FIGURE 7.27
(a) Voltage transfer characteristics of a 3D sequential inverter with NMOSFET on top and
PMOSFET on the bottom level. (b) Voltage transfer characteristics of a 3D sequential inverter
with PMOSFET on top and NMOSFET on the bottom level.
172 3D Integration in VLSI Circuits

7.4 3D Technologies Comparative Thermal Analysis

One the main challenge of 3D technologies, specifically high-density 3D inte-
gration, is power density and associated thermal dissipation. This is often
put ahead as a potential strong show stopper. Indeed, due to the increased
number of die layers, a too large power density may arise, that would imply
a large thermal flux, which might be difficult to dissipate using standard
packaging technologies. In an extreme manner, as a result, thermal runaway
would arise that would potentially destroy the 3D circuits.
The thermal dissipation of commonly used 3D-packaging technologies
using TSV has already been widely studied [61–64]. For instance, at CEA-Leti,
a systematic thermal exploration [64] has been carried out to investigate the
key differentiating parameters of TSV-based 3D ICs: power density, die thin-
ning, die-to-die thermal coupling, and the impact of TSVs. The study reveals
that nonthinned dies in a 3D stack may act as heat spreaders, whereas TSVs
may even provoke exacerbated hotpots due to the dielectric layer used for
TSV isolation. TSVs induce a decreased horizontal Heat Transfer Coefficient
(HTC) cause of the dielectric layer for a light increase in the vertical heat
path. TSVs lead to larger thermal hotspots in case of lateral dissipation block-
age, contrarily to the common believes [65]. Measurements on a dedicated
test-chip and systematic simulations have been carried out leading to the
same conclusions [64,66].
Nevertheless, little work has been carried out on thermal study of denser
3D technologies such as the ones described in the previous paragraphs
[67]. In this section a thermal comparative study is carried out between the
respective new advanced 3D technologies presented in the above-mentioned
sections: Cu/SiO2 hybrid bonding and CoolCube™ [68]. This comparative
thermal study between standard μbumps, Cu/SiO2 hybrid bonding, and
CoolCube™ is performed for multiple-die stacks and considering a full
range of technology parameters and different application scenarios.
Section 7.4.1 presents the respective technology parameters to setup a ther-
mal comparative study and the associated applicative scenarios, whereas the
following one presents the associated thermal results.

7.4.1 3D Technology Parameters for Thermal Comparison

CEA-Leti has monitored a comparative thermal study to compare three
different 3D technologies: two parallel 3D technologies (TSV based with
μbumps or with Cu/SiO2 bonding) and the sequential approach CoolCube™,
for a different number of parameters: number of active layers, die thick-
ness, interconnect density, and power scenarios. In order to allow fair com-
parison, as a hard constraint the total die thickness and package are kept
fixed. Figure 7.28 shows an example of stacking diagram used for thermal
simulations and indicates circuit and 3D technology parameters, details are
Toward Three-Dimensional High Density 173

25 W/K·m2 (low-power)
• Stack size: 2, 4 or 8 tiers (F2B)
2000 W/K·m2 (HPC)
• Top-die thickness: Variable to keep
total die stack height equal to 722 μm

• Die thickness: 80/50/25/10 μm (TSV-

TM
based) and 160 nm (CoolCube )

F • BEOL: 7+1-layer Cu/Al (TSV-based)

I
1.2 mm

and four-layer Tungsten (CoolCubeTM)

X
E
• Interdie layer: 25 μm (Cu-pillar), 1.7 μm
D TM
(Cu-Cu DB), 60 nm (CoolCube )

• Stack-to-package layer: Underfill with

C4-bumps (40 μm)

• Package: 12 mm × 12 mm × 1.2 mm
300 W/K·m2 with a four-layer substrate (288 μm)

FIGURE 7.28
Example of stacking diagram indicating the main circuit and 3D integration technology
parameters considered for thermal simulations.

TABLE 7.6
3D Technology Parameters: Material and Thickness
Sequential 3D
Parallel 3D Technologies Technology
µbumps Cu/SiO2 Direct Bonding CoolCube™
Layer Material Thickness Material Thickness Material Thickness

BEOL Cu/Al+SiO2 7+1 layers Cu/Al+SiO2 7+1 layers W+SiO2 4 layersa

7 μm 7 μm 1.72 μm
Interdie Cu/SnAg+ 25 μm Cu+SiO2 1.7 μm W+SiO2 60 nm
layer polymer
underfill
Dieb Si 80/50/25 μm Si 50/20/10 μm Mostly 160 nm
substrate SiO2
a Except for bottommost tier, which has a standard 7+1 Cu/Al BEOL configuration.
b Except the topmost tier, which is as thick as the die stack height limit permits.

given in Table 7.6. The 3D stack contains 1–8 tiers, which are assembled in
a face-to-back (F2B) manner [69]. In the case of TSV-based integration, the
thickness of the die substrate spans from 25 to 80 μm for assemblies using
μbumps and from 10 to 50 μm for Cu/SiO2 hybrid bonding. The layout of
each tier is based on a real scalable 65 nm 3D circuit where identical tiers
are stacked one on top of each other. An exception is made for CoolCube™,
which uses tungsten for the metallization layers (BEOL) of the intermediate
tiers due to thermal limitations and contamination risks during the fab-
rications process, as explained in the previous paragraph. The height and
diameter of the μbumps in the interdie layer are, respectively, 25 and 20 μm,
thus in the typical range for this technology. The thickness of the interdie
174 3D Integration in VLSI Circuits

layer is 1.7 μm for the Cu/SiO2 hybrid bonding and only 60 nm in the case of
CoolCube™. As already mentioned, the impact of the TSVs on the thermal
conductivity of the silicon substrate has been demonstrated to be very lim-
ited [65] and therefore will not be considered in this study.
The 3D die is stacked on a flip-chip organic package with the bottom tier
connected to a four-layer ball grid array (BGA) package (288  μm height)
through C4-bumps (40 μm height). To be compatible with typical low-power
applications, such as mobile, the total package height was fixed at 1.2 mm,
thus resulting in a max height of 722 μm for the 3D die stack. Provided this
constraint, the topmost tier is made as thick as possible as it plays as a heat
spreader for hot-spot mitigation. Specific HTCs were applied to the boundary
surfaces according to the considered power application scenario to emulate
the behavior of typical Printed Circuit Board (PCB) and package components
such as Thermal Interface Material (TIM) heat spreader and heat sink.
Figure 7.29 illustrates the power profiles used to emulate multiple appli-
cation scenarios and to provoke distinct thermal behaviors depending on
the application use case. Heat dissipation at hot spots is primarily diffused
through the silicon substrate and usually spreads in a semispherical direc-
tion, rapidly decreasing the heat density and lowering the peak temperature.
But for 3D ICs, the thinned silicon substrate reduces the lateral heat spread-
ing capability and the interdie layers (mainly polymer) act as vertical thermal
barriers, which results in exacerbated hot spots. When power is evenly dis-
tributed, the temperature gradient over the tier is much smaller and the heat
flows mostly perpendicular to the die substrates in the 3D stack.
Most of the heat generated in applications with intensive power dissipa-
tion, such as HPC, flows toward efficient heat sinks usually mounted on top
of the package. On the contrary, for applications with limited power dissipa-
tion, circuit is mainly cooled from the bottom and heat flows toward the PCB
through the package substrate.
This study also considers applications with similar power budget allo-
cated either in a single or in multiple tiers. Taking into account voltage and

(a) (b) (c) (d)

FIGURE 7.29
Examples of power application scenarios. (a) Strong hot spots in a single tier, (b) limited
uniform power in a single tier, (c) limited uniform power distributed between tiers, and
(d) intensive uniform power dissipation.
Toward Three-Dimensional High Density 175

TABLE 7.7
Power Dissipation Profiles, according to Application Scenarios Described in
Figure 7.29
Hot Spot Single-Tier Multitiers Multitiers HPC
Total Power Total Power Total Power Total Power
Quantity Power Densitya Power Densitya Power Densitya Power Densitya
of Tiers (W) (W/cm 2) (W) (W/cm 2) (W) (W/cm 2) (W) (W/cm 2)
2 2.64 365 2.657 3.65 2.657 1.825 5.3 3.65
4 2.64 365 2.657 3.65 2.657 0.913 10.6 3.65
8 2.64 365 2.657 3.65 2.657 0.456 21.2 3.65
a In active regions.

frequency scaling, to distribute tasks in multiple CPUs running at a lower

frequency may result in higher throughput for same power consumption
than running in a single CPU at high frequency [70]. Total power dissipation
and power density for the different applications scenarios in Figure 7.29 are
listed in Table 7.7.
Thermal modeling and simulation for this large set of technology
parameters and power applications have been carried out using a thermal
analysis prototype provided by Mentor Graphics Calibre® [71]. This pow-
erful thermal analysis solution, which relies on the FloTHERM simulation
engine, has an automated effective thermal property extraction (EFFP) fea-
ture that allows to easily account for complex fine-grain structures such as
the BEOL, interlayer vias, and μbumps arrays used for 3D connections.

7.4.2 Comparative Study Thermal Results

This section brings thermal simulation results for the Cu/SiO2 hybrid
bonding, CoolCube™, and μbumps technologies with regard to die thick-
ness, die-to-die connectivity density, and the number of stacked tiers. Due to
their very distinct thermal behaviors, hot spot and uniform power applica-
tion profiles are investigated in separate subsections. The thermal conduc-
tivity of the interdie layer is highly dependent on the number of die-to-die
3D connections, more precisely the ratio of metal and insulator within this
layer. This study considers two distinct corners for this metal density param-
eter either (1) very few die-to-die connections (ratio <1%) or (2) a high die-
to-die connectivity density (ratio of 16% for CoolCube™ and ratio of 25% for
μbumps and hybrid direct bonding). In most real circuits, the actual die-to-
die connectivity density will lie in between these two corners due to design
and layout constraints.
176 3D Integration in VLSI Circuits

7.4.2.1 For Hot-Spot Dissipation Scenarios

Figure 7.30 brings the peak temperature results for hot-spot dissipation sce-
narios (case [a] in Figure 7.29). The interdie interface layer and die thickness
play important roles as they directly affect the thermal coupling between
adjacent tiers and the lateral heat dissipation, respectively. CoolCube™ pres-
ents very tight thermal coupling and heat can easily reach the top tier to
benefit from its thickness to spread out. For μbump assembly technology,
the underfill (nonconductive polymer, which mechanically bonds the dies
together) layer acts as a strong thermal barrier between the hot spots and the
heat sinks and provoke higher peak temperatures. Cu/SiO2 hybrid bonding

Very few die-to-die connections

140 μbumps 8 dies

μbumps 4 dies
Peak temperature (°C)

130
μbumps 2 dies
Cu/SiO2 8 dies
120
Cu/SiO2 4 dies

110 Cu/SiO2 2 dies

CoolCubeTM 8 dies
Cu-Cu DB
100 Cu-pillar CoolCubeTM 4 dies
CoolCubeTM 2 dies
CoolCube
90
0 10 20 30 40 50 60 70 80
(a) Die thickness (μm)

Max few die-to-die connections

140 μbumps 8 dies

μbumps 4 dies
Peak temperature (°C)

130
μbumps 2 dies
Cu/SiO2 8 dies
120
Cu/SiO2 4 dies
Cu/SiO2 2 dies
110
CoolCubeTM 8 dies
Cu-pillar
100 Cu-Cu DB CoolCubeTM 4 dies
CoolCubeTM 2 dies
90 CoolCube
0 10 20 30 40 50 60 70 80
(b) Die thickness (μm)

FIGURE 7.30
(a) Peak temperature in the case of hot-spot dissipation considering very few die-to-die con-
nections (ratio <1%) or (b) max connection density between tiers (ratio of 25% for TSV-based
and 10% for CoolCube™). Hot spots are active only in midmost tier to account for similar heat
flow paths in both vertical directions.
Toward Three-Dimensional High Density 177

100°C

70°C
(a) (b) (c)

FIGURE 7.31
Thermal maps of the middle and topmost tiers in the case of hot-spot dissipation (Figure 7.29a)
in a eight-dies stack: (a) parallel stacking with μbumps, (b) parallel stacking with Cu/SiO2
direct bonding, and (c) sequential stacking CoolCube™.

benefits from both thermal coupling and lateral heat dissipation in the inter-
mediate tiers, therefore presenting the best thermal performance in most 3D
stacks with more tiers. This difference in the thermal profiles is depicted in
the temperature maps in Figure 7.31, where the hot spot outlines reveal the
lateral heat dissipation through the die substrates.

7.4.2.2 For Uniformly Distributed Power Dissipation Scenarios

Figure 7.32a and b show results for uniform power dissipation scenarios
where the different 3D technologies have similar peak temperatures. In this
case, most of the heat flows vertically and the 3D technology parameters
have reduced thermal impact as external components such as the package
structures dominate the vertical thermal resistance of the circuit. In case the
total power dissipation allocated in a single tier or distributed in multiple
tiers leads to about the same temperature results for most of the 3D con-
figurations. This brings a perspective of increasing the processing capability
of the circuit by allocating tasks into different tiers of the 3D stack while
keeping the same power consumption and preserving the thermal perfor-
mance. In the case where every additional stacked tier results in extra power
dissipation, and therefore the resulting increase in the peak temperature is
nearly linear with regard to the number of stacked tiers.

7.4.3 Thermal Comparison Conclusion

The study presented in this section reports a comparison of the thermal
performance of the CoolCube™ and Cu/SiO2 hybrid bonding technologies
versus widely used μbump technology. It highlights the contribution of indi-
vidual technology parameters to the overall thermal performance of the 3D
stacks considering various power application scenarios. For the sake of com-
parison, the respective technologies are compared for different scenarios
and for fixed packaging. As conclusion, the μbump technology shows poor
results for thermal hot spots, due to underfill layers that have poor thermal
178 3D Integration in VLSI Circuits

Very few die-to-die connections

140
μbumps 8 dies
120 8 dies μbumps 4 dies
Peak temperature (°C)

μbumps 2 dies
100
Cu/SiO2 8 dies
80
Cu/SiO2 4 dies
4 dies
60 Cu/SiO2 2 dies
CoolCubeTM 8 dies
40 2 dies
CoolCubeTM 4 dies
20
CoolCubeTM 2 dies
0
0 10 20 30 40 50 60 70 80
(a) Die thickness (μm)

Max few die-to-die connections

140
μbumps 8 dies
120 8 dies μbumps 4 dies
Peak temperature (°C)

100 μbumps 2 dies

Cu/SiO2 8 dies
80
Cu/SiO2 4 dies
4 dies
60 Cu/SiO2 2 dies
CoolCubeTM 8 dies
40 2 dies
CoolCubeTM 4 dies
20
CoolCubeTM 2 dies
0
0 10 20 30 40 50 60 70 80
(b) Die thickness (μm)

FIGURE 7.32
Peak temperature resulting from a power application scenario with intensive power dissipa-
tion uniformly distributed over the tiers (Figure 7.29d) and in the case of (a) very few connec-
tions between tiers and (b) max density connections between tiers.

conductivity. On the other hand, Cu/SiO2 hybrid bonding and CoolCube™

technologies present a very good thermal coupling between the different cir-
cuit layers, where the impact of interconnect density is rather limited. In the
case of uniform power dissipation, the thermal result is more or less linear
with power density, then with the number of active dies. It is known that 3D
integration offers the possibility of running circuits at different performance
modes [70], where some of them are not practical in 2D.
Toward Three-Dimensional High Density 179

7.5 General Conclusion

3D high-density integrations (typically below a 10 μm pitch of interconnects)
may require some alternative technologies to widely used μbumps. In this
section, CEA-Leti presents two alternative but complementary approaches.
The first one is similar to a BEOL approach, using Cu/SiO2 hybrid bonding
well adapted for a pitch of interconnect superior to 1 μm. The second one is
nearer to FEOL processes, named CoolCube™, well adapted for pitch below
1 μm. This chapter gives all details on technology and results, proving that
those alternatives offer high performances without the need of More than
Moore advanced nodes. Meantime, Cu/SiO2 hybrid bonding and CoolCube™
have the potential to solve the interconnection density limitation of the
TSV-based integration processes. Furthermore, despite active devices being
integrated in extremely thin layers, and contrarily to common believes, Cu/
SiO2 hybrid bonding and CoolCube™ technology process exhibit thermal
performance similar to or even better than that of the 3D integration using
μbumps and associated underfill considering the same power applications.
The very tight die-to-die thermal coupling allows easy heat dissipation of
internal hot spots without any noticeable temperature difference between
the stacked layers.

References
1. L. Di Cioccio et al., An overview of patterned metal/dielectric surface bonding:
Mechanism, alignment and characterization, JECS, 2011, pp. 81–86.
2. M. Goto et al., Three-dimensional integrated CMOS image sensors with pixel-
Parallel A/D converters fabricated by direct bonding of SOI layers, IEEE,
Electron Device Meeting IEDM, 2014.
3. L. Brunet et al., First demonstration of CMOS over CMOS 3D VLSI CoolCubeTM
integration on 300 mm wafers, IEEE VLSI, 2016.
4. V. Balan et al., CMP process optimization for bonding applications, ICPT 2012,
October 15–17, 2012, Grenoble, France, pp. 177–183.
5. L. Benaissa et  al., Next generation image sensor via direct hybrid bonding,
2015 IEEE 17th Electronics Packaging Technology Conference, December 2–4, 2015,
Singapore.
6. B. Rebhan et al., <200 nm wafer-to-wafer overlay accuracy in wafer level Cu/
SiO2 hybrid bonding for BSI CIS, Electronics Packaging Technology Conference,
December 2–4, 2015, Singapore.
7. M. Okada et  al., High-precision wafer-level Cu-Cu bonding for 3DICs, IEEE,
Electron Device Meeting IEDM, 2014.
180 3D Integration in VLSI Circuits

8. S. Lhostis et  al., Reliable 300  mm wafer level hybrid bonding for 3D stacked
CMOS image sensors, 2016 IEEE 66th Electronic Components and Technology
Conference, May 31–June 5, 2016, Las Vegas.
9. L. Wang et al., Direct bond interconnect (DBI) for fine-pitch bonding in 3D and
2.5D integrated circuits, Pan Pacific Microelectronics Symposium, 2017, IEEE.
10. Z.Y. Liu et al., Detection and formation mechanism of micro-defects in ultrafine
pitch Cu-Cu direct bonding, Chinese Physics B, 25(1), 018103, 2016.
11. Q.Y. Tong and U. Gösele, Semiconductor Wafer Bonding Science and Technology,
1999, Wiley, New York, 320 p.
12. W.P. Maszara et al., Bonding of silicon wafers for silicon–on-insulator, Journal of
Applied Physics, 64(10), 4943–4950, 1988.
13. P. Gueguen et al., Copper direct-bonding characterization and its interests for
3D integration, Journal of the Electrochemical Society, 156(10), H772–H776, 2009.
14. P. Gueguen et  al., Direct bonding: An innovative 3D interconnect, ECTC
Proceeding, 2010, pp. 878–883.
15. L. Di Cioccio, F. Baudin, P. Gergaud et al., Modelling and integration phenomena
of metal-metal direct bonding technology, ECS Transactions, 64(5), 339–355, 2014.
16. Y. Beilliard et al., Advances toward reliable high density Cu–Cu interconnects
by Cu-SiO2 direct hybrid bonding, IEEE International 3D System Integration
Conference (3DIC), 2014.
17. Imec and EVG demonstrate for the first time 1.8 μm pitch overlay accuracy for
wafer bonding. https://www.evgroup.com/en/about/news/2017_01_imec.
18. R. Taïbi et al., Full characterization of Cu/Cu direct bonding for 3D integration,
ECTC Proceedings, 2010, pp. 219–225.
19. R. Taïbi et  al., Investigation of stress induced voiding and electromigration
phenomena on direct copper bonding interconnects for 3D integration, IEEE
Electron Device Meeting IEDM, 2011.
20. S. Moreau et al., Mass transport-induced failure of hybrid bonding-based
integration for advanced image sensor applications, 2016 IEEE 66th Electronic
Components and Technology Conference, May 31–June 5, 2016, Las Vegas, NV.
21. J. Jourdon et al., Effect of pasivation annealing on the electromigration properties of
hybrid bonding stack, IEEE International Reliability Physics Symposium (IRPS), 2017.
22. A. Garnier et al., Electrical performance of high density 10 μm diameter 20 μm
pitch Cu-Pillar with chip to wafer assembly, IEEE 67th Electronic Components and
Technology Conference, May 30–June 2, 2017, Orlando.
23. Sony IMX260 in Samsung Galaxy S7: Stacked or not?. http://image-sensors-
world.blogspot.fr/2016/03/sony-imx260-in-samsung-galaxy-s7.html.
24. C. Cavaco et al., Copper oxide direct bonding of 200 mm CMOS wafers with
five metal levels and TSVs: Morphological and electrical characterization, JECS,
2016, pp. 43–46.
25. Y. Beillard et al., Chip to wafer copper direct bonding electrical characterization
and thermal cycling, IEEE International Conference on 3D System Integration (3D IC),
2013.
26. T. Fukushima et al., Transfer and non-transfer stacking technologies based
on chip-to wafer self-assembly for high throughput and high precision align-
ment and microbumps bonding, IEEE International 3D Systems Integration
Conference, 2015.
27. L. Sanchez et al., Chip to wafer direct bonding technologies for high density 3D
integration, IEEE Electronic Components and Technology Conference, 2012.
Toward Three-Dimensional High Density 181

28. S. Mermoz et al., Impact of containment and deposition method on sub-micron

chip-to-wafer self-assembly yield, IEEE International Conference on 3D System
Integration (3D IC), 2012.
29. S. Mermoz et al., High density chip-to-wafer integration using self-assembly on
the performances of directly interconnected structures made by direct copper/
oxide bonding, IEEE 15th Electronics Packaging Technology Conference, December
11–13, 2013.
30. J. Berthier et al., Self-alignment of silicon chips on wafer: The effect of spread-
ing and wetting, Sensor & Transducers, 2012.
31. P. Batude et al., Advances in 3D CMOS sequential integration, IEEE IEDM, 2009.
32. Y.-J. Lee, P. Morrow, and S. K. Lim, Ultra high density logic designs using
transistor-level monolithic 3D integration, International Conference on Computer-
Aided Design, 2012.
33. P. Batude et al., Demonstration of low temperature 3D sequential FDSOI inte-
gration down to 50 nm gate length, IEEE VLSI, 2011.
34. C. Fenouillet-Beranger et  al., New insights on bottom layer thermal stability
and laser annealing promises for high performance 3D VLSI, IEEE IEDM, 2014.
35. P. Batude et al., 3DVLSI with CoolCube process: An alternative path to scaling,
IEEE VLSI, 2015.
36. S.-M. Jung et al., High speed and highly cost effective 72M bit density S3 SRAM
technology with doubly stacked Si layers, peripheral only CoSix layers and
tungsten shunt W/L scheme for standalone and embedded memory, IEEE VLSI,
2007.
37. C.C. Yang et al., Enabling low power BEOL compatible monolithic 3D nanoelec-
tronics for IoTs using local and selective far infrared ray laser anneal technol-
ogy, IEEE IEDM, 2015.
38. T. Naito et al., Worlds first monolithic 3D FPGA with TFT SRAM over 90 nm
9 layer Cu CMOS, IEEE VLSI, 2010.
39. T.T. Wu et al., Sub 50nm monolithic 3D IC with low power CMOS inverter and
6T SRAM, IEEE VLSI-TSA, 2015.
40. T. Irisawa et  al., Demonstration of ultimate CMOS based on 3D stacked
InGaAs-OI/SGOI wire channel MOSFETs with independent back gate, IEEE
VLSI, 2014.
41. V. Deshpande et al., Advanced 3D monolithic hybrid CMOS with sub 50nm gate
inverters featuring replacement metal gate (RMG)–InGaAs nFETs on SiGe-OI
fin pFETs, IEEE IEDM, 2015.
42. Max M. Shulaker et al., Monolithic 3D integration of logic and memory: Carbon
nanotube FETs, resistive RAM, and silicon FETs, IEEE IEDM, 2014.
43. Tsung-Ta Wu et  al., Low cost and TSV free monolithic 3D-IC with heteroge-
neous integration of logic, memory and sensor analogy circuitry for Internet of
Things, IEEE IEDM, 2015.
44. Fu-Kuo Hsueh et al., First fully functionalized monolithic 3D+ IoT chip with
0.5V light-electricity power management, 6.8 GHz wireless-communication
VCO, and 4-layer vertical ReRAM, IEEE IEDM, 2016.
45. I. Ouerghi et al., Polysilicon nanowire NEMS fabricated at low temperature for
above IC NEMS mass sensing applications, IEEE IEDM, 2014.
46. P. Batude et al., 3-D Sequential Integration: A key enabling technology for het-
erogeneous co-integration of new function with CMOS, Journal on Emerging and
Selected Topics in Circuits and Systems, 2, 714–722, 2012.
182 3D Integration in VLSI Circuits

47. P. Coudrain, Setting up 3D sequential integration for back-illuminated CMOS

image sensors with highly miniaturized pixels with low temperature fully-
depleted SOI transistors, IEEE IEDM, 2008.
48. M. Vinet et al., Monolithic 3D Integration: A powerful alternative to classical 2D
scaling, IEEE S3S, 2014.
49. L. Brunet et al., Direct bonding: A key enabler to CoolCubeTM (sequential) inte-
gration, IEEE ECS, 2014.
50. R. Choi et al., Bonding based channel transfer and low temperature process for
monolithic 3D integration platform development, IEEE S3S, 2016.
51. L. Pasini et al., High performance low temperature activated devices and opti-
mization guidelines for 3D VLSI integration of FD, Trigate, FinFET on insulator,
IEEE VLSI, 2015.
52. C. Fenouillet-Beranger et  al., Recent advances in low temperature process in
view of 3D VLSI integration, IEEE S3S, 2016.
53. D. Benoit et al., Interest of SiCO low k = 4.5 spacer deposited at low temperature
(400°C) in the perspective of 3D VLSI integration, IEEE IEDM, 2015.
54. C.-M.V. Lu et al., Dense N over CMOS 6T SRAM cells using 3D sequential inte-
gration, IEEE VLSI-TSA, 2017.
55. C.-M.V. Lu et  al., Key process steps for high performance and reliable 3D
sequential integration, IEEE VLSI, 2017.
56. F. Deprat et al., Technological enhancers effect on NiCo silicide stability for 3D
sequential integration, EMRS Spring, 2016.
57. F. Deprat et al., NiCo 10%: A promising silicide alternative to NiPt 15% for ther-
mal stability improvement in 3DVLSI integration, AMC, 2015.
58. M. Vinet et al., Opportunities brought by sequential 3D CoolCubeTM integra-
tion, IEEE ESSDERC, 2016.
59. C. Fenouillet-Beranger et  al., W and Copper interconnection stability for 3D
VLSI CoolCube intergration, SSDM, 2015.
60. F. Deprat et al., Dielectrics stability for intermediate BEOL in 3D sequential,
MAM, 2016.
61. A. Jain, R.E. Jones, R. Chatterjee, and S. Pozder, Analytical and numerical
modeling of the thermal performance of three-dimensional integrated circuits,
IEEE Transactions on Components and Packaging Technologies, 33(1), 56–63, 2010.
62. H. Xu, V.F. Pavlidis, and G. De Micheli, Analytical heat transfer model for thermal
through-silicon vias, in Design, Automation & Test in Europe (DATE), 2011, pp. 1–6.
63. C. Santos, P. Vivet, J.P. Colonna, P. Coudrain, and R. Reis, Thermal performance
of 3D ICs: Analysis and alternatives, in IEEE International 3D Systems Integration
Conference (3DIC), 2014, pp. 1–7.
64. P. Coudrain, P. Momar Souare, J.P. Colonna, H. Ben-Jamaa, P. Vivet, R. Prieto,
V. Fiori et al., Experimental insights into thermal dissipation in TSV-based 3D
integrated circuits, IEEE Design & Test, 33, 21–36, 2015.
65. C. Santos, P.M. Souare, P. Coudrain, J.P. Colonna, F. de Crecy, P. Vivet, A. Borbely
et al., Using TSVs for thermal mitigation in 3D circuits: Wish and truth, 3DIC,
December 2014, Cork, Ireland.
66. P.M. Souare et al., A comprehensive platform for thermal studies in TSV-based
3D integrated, IEEE IEDM, 2014.
67. S.K. Samal et  al., Fast and accurate thermal modeling and optimization for
monolithic 3D ICs, in Design Automation Conference (DAC), 2014, pp. 1–6.
Toward Three-Dimensional High Density 183

68. C. Santos, P. Vivet, S. Thuries, O. Billoint, J.-P. Colonna, P. Coudrain, and

L. Wang, Thermal performance of CoolCube™ monolithic and TSV-based 3D
integration processes, 3DIC’2016 Conference, December 2016, San Francisco, CA.
69. P. Vivet, et al., A 4 × 4 × 2 homogeneous scalable 3D network-on-chip circuit
with 326 MFlit/s 0.66 pJ/bit robust and fault tolerant asynchronous 3D links,
IEEE Journal of Solid-State Circuits, 2016, 1–17.
70. P. Emma et  al., 3D stacking of high-performance processors, in International
Symposium on High Performance Computer Architecture (HPCA), 2014, pp. 500–511.
71. Physical Verification with CalibreTM. https://www.mentor.com/products/
ic_nanometer_design/verification-signoff/physical-verification/.
8
Novel Platforms and Applications
Using Three-Dimensional and
Heterogeneous Integration Technologies

Kuan-Neng Chen, Ting-Yang Yu, Yu-Chen Hu, and Cheng-Hsien Lu

CONTENTS
8.1 Introduction ................................................................................................ 186
8.2 Stacked Terahertz Optical Component ................................................... 186
8.2.1 THz Wave Applications ................................................................ 186
8.2.2 Issues of Common THz Polarizers .............................................. 187
8.2.3 Fundamentals and Fabrication Methods.................................... 188
8.2.3.1 Structure Design of THz Polarizer ............................... 188
8.2.3.2 Fabrication Methods and Low-Temperature
Eutectic Liquid Bonding ................................................ 189
8.2.4 Performance of Staked THz Polarizer ........................................ 190
8.2.4.1 Bonding and Etching Qualities..................................... 190
8.2.4.2 High Transmittance THz Polarizer .............................. 191
8.2.4.3 Broadband THz Polarizer .............................................. 192
8.2.5 Comparison between Common and Stacked THz
Polarizers .............................................................................. 193
8.3 Pressure-Sensing System .......................................................................... 194
8.3.1 Pressure-Sensing Platform ........................................................... 195
8.3.2 Pressure Sensor .............................................................................. 196
8.3.3 Micropin-Fin Heat Sink Interposer ............................................. 196
8.3.4 Integration of Micropin-Fin Heat Sink Interposer and Chips
with Double-Self-Assembly Approach ...........................................198
8.3.5 Achievements and Outlook .......................................................... 199
8.4 Neurosensing Systems .............................................................................. 200
8.4.1 Fabrication, Scheme, and Reliability of Neural Sensing
Biosensor ......................................................................................... 201
8.4.1.1 Three-Dimensional System-in-Packaging Neural
Sensing Biosensor ........................................................... 201
8.4.1.2 Chip-Level Heterogeneous Integration Scheme......... 201
8.4.1.3 Electroplating Solution Improvement .......................... 201
8.4.1.4 Electrical Reliability Tests .............................................. 203

185
186 3D Integration in VLSI Circuits

8.4.2
2.5D-Silicon Interposer Neural Sensing Biosensor ................... 203
8.4.2.1 Fabrication of Silicon Interposer and
Through-Silicon Via-Embedded μ-Probe .................... 204
8.4.2.2 Electrical Reliability Test ................................................ 204
8.4.3 2.5D-Flexible Interposer Neural Sensing Biosensor ................. 205
8.4.3.1 Fabrication of Flexible Interposer ................................. 205
8.4.3.2 Novel Flexible Bonding Approach ............................... 206
8.4.3.3 Electrical Reliability Test of Novel Thin
Film-Bonding Approaches............................................. 206
8.4.4 Demonstration ................................................................................ 207
References............................................................................................................. 208

8.1 Introduction
With the growing demands of high computing and Internet of Things (IoT)
applications, high I/O counts with fast signal transmission speed and vari-
ety integration ability become equally significant in current semiconductor
development. Platforms based on 3D-advanced packaging and integration can
provide such solutions to high-speed, low-power, small form factors, and het-
erogeneous integration requirements. Advanced applications that have been
successfully manufactured such as graphics processor units (GPU) and mobile
processers adopt key technologies of 3D integration and advanced packaging
such as fan-out scheme, whereas advanced complementary metal–oxide–
semiconductor (CMOS) image sensors use through-silicon via (TSV) and fine-
pitch Cu direct bonding. Other than these applications, 3D integration and
heterogeneous technologies create many opportunities and a whole new world
for advanced systems, which were difficult to be fulfilled in the past. This
chapter describes examples of novel platform and application demonstration,
including optical components, pressure sensors, and bioneural applications.

8.2 Stacked Terahertz Optical Component

8.2.1 THz Wave Applications
The THz gap, which ranges between ~0.1 and ~10 THz, is a gap in the electro-
magnetic spectrum for generating and detecting the radiation that does not live
because of rare light source. Luckily, the bottleneck of producing THz has been
gotten through free-electron laser, nonlinear effect of electro-optic (EO) crystal,
quantum cascade, photoconductive switch, and so forth lately [1,2]. THz radia-
tions are nondestructive and intrinsically safe compared to X-ray because of low
Novel Platform and Applications Using 3D Integration Technologies 187

photon energy. Furthermore, it is highly sensitive with metals, some chemical

bond and function groups, and therefore THz wave is broadly utilized for mate-
rial characterizations, security, and bio-images. Apart from these, THz astro-
nomical observations are also a vital application for water and dust detection
from space. Also, THz is the carrier frequency of next generation for commu-
nications. Due to THz source breakthrough, the quantities of THz applications
have embraced a boom. However, researchers have not systematically explored
the relative optical components, such as polarizer and filter. In this endeavor,
the prior art of THz polarizer fabrication by 3D-IC technologies would be dis-
cussed and compared by other methods.

8.2.2 Issues of Common THz Polarizers

Structures of optical components are usually with scale in wavelength or
subwavelength. Therefore, THz wire-grid polarizers are usually with pitch
of microns. The commercial wire-grid polarizer stands in air or on the thin
film, and also known as free-standing or thin-film wire-grid polarizer. The
transverse magnetic (TM) mode of THz wave would infiltrate the polarizer
without power loss because of free-standing structure. However, the free-
standing wire-grid polarizers are very fragile without substrate supporting.
The mechanical strength of polarizer would be improved by fabricating on
substrate but enhances reflective loss. To align, the polarizer on substrate at
Brewster’s angle would increase the transmission power to 100%. However,
specific angle alignment is difficult to modulate in systems. Antireflection
(AR) coating is another option to reduce reflection loss. Single-layer AR coat-
ing with thickness of quarter wavelength and suitable refractive index would
enhance the transmittance to near 100%. Figure 8.1 demonstrates the reason-
ing stream and composed structure of new sort polarizer. The AR coating

New design of THz polarizer

Silicon wafers as
Seal wire-grid by
Mass production, substrate: Hide wire-grid into
wafer bonding and
large area, and robust (1) suitable material, wafers and fabricate
fabricate antireflection
(2) compatible for antireflection layers
structure by using layers on
semiconductor on both sides of the
thick substrate both sides of the
manufacture, and structure
structure
(3) high reflectance

Advantages:
• Low cost
Si Si Si • Robust structure
Wire-grid polarizer Hide wire-grid Antireflection • High transmittance
with Si substrate layers • No precision optical
alignment required

FIGURE 8.1
The designed diagram and advantages of new type THz polarizer.
188 3D Integration in VLSI Circuits

could be fabricated on both substrate surface, if the wire-grid was sealed in

to the substrate. The power transmittance would be theoretically 100%. This
section would demonstrate the seal of wire-grid polarizer by wafer bonding
and AR-coating process.

8.2.3 Fundamentals and Fabrication Methods

8.2.3.1 Structure Design of THz Polarizer
Silicon is a most suitable substrate for wire-grid polarizer, which has advan-
tages of robustness, low cost, low dispersion, and low absorption loss in
THz [3]. However, in THz region, silicon substrate usually accompanies high
reflectance because of the high refractive index. In order to avoid high reflec-
tance due to Si substrate, AR coating must be designed to have precise thick-
ness (tAR) of a quarter of wavelength and refractive index (nAR) of square root
of Si refractive index for a single frequency. The purpose of AR coating is
to make the first reflection beam from surface and second reflection beam
from interface between AR coating and Si destructive interference, as shown
in Figure 8.2. The relations between tAR and nAR are shown in the following
equations:

λ0
t AR = (8.1)
4n AR

=
n AR =
n si 3.4 = 1.84 (8.2)

where λ0 is the central wavelength in free space. According to Equations 8.1

and 8.2, the AR-coating material must be in the range of microns thickness
and refractive index of 1.84. But in THz region, it is really tough to find such
a material with exact refractive index and low absorption loss. Moreover,
microns thickness coating layer would induce large stress between AR

Long wavelength limited

zero-order effective medium approach
nair nAR nSi
AR film

Si

λ0 AR layer
LAR

nAR

lay
Si Si
de LAR
se
ha
π p1
T

15 μm
R

FIGURE 8.2
AR coating with quarter wavelength thickness leads to destructively interfere, and the index
could be controlled by whole array pitch under zero-order effective medium approach.
Novel Platform and Applications Using 3D Integration Technologies 189

coating and substrate, and also get peeled. Therefore, Si substrate is etched
by deep reactive-ion etching (DRIE) with cylinder holes and the index of
AR layer is tuned to specific index and also specific thickness according to
the effective medium theory. The effective index of two mixing materials is
equal to Equation 8.3:

n AR ≅ f × 3.4 + (1 − f ) × 1 = 3.4 (8.3)

where f is the filling factor of Si and air and is 0.35 to achieve the effective
index. As the pitch of etching holes is less than λ0/10, the AR layer seems like
a homogeneous layer with effective index, as shown in Figure 8.2.

8.2.3.2 Fabrication Methods and Low-Temperature Eutectic Liquid Bonding

The process flow of THz wire-grid polarizer fabrication is shown in
Figure 8.3a. Double-side polished wafers are ready for etching and deposi-
tion of AR layer soon after standard Radio Corporation of America (RCA)
clean. After that, AR layer is patterned through lithography and etched via
DRIE with the optimized parameters. The diameter and space between hole
array are etched to 12.6 and 2.4 μm with hexagonal arrangement according to

Si Si
PR patterning and AR
etching
Si
PR removing Si
Si Si
PR patterning on
Ti 200Å
PR patterning on top
wafer bottom wafer
Cu 3000Å
Thin buffer layer 100Å
Si Si
Thin buffer layer/Cu/Ti In/Sn/Ni/Ti sealing
grating deposition deposition

In 3000Å
Si Si Sn 3000Å
Thin buffer layer/Cu/Ti In/Sn/Ni/Ti sealing lift-off
grating lift-off
Ni 200Å
Ti 200Å
Si
Si
Si
(a) Wafer bonding (b)

FIGURE 8.3
(a) The process flow of stacked THz polarizer fabrication (From Chi, N.-C. et al., IEEE ECTC,
1793–1798, 2017.) and (b) the bonding structure for Cu wire-grid sealing. (From Chi, N.-C. et al.,
SPIE Optics + Optoelectronics, 102420Z–102420Z-6, 2017.)
190 3D Integration in VLSI Circuits

the designed filling factor and AR thickness. The Cu wire-grid gratings and
low-temperature In/Sn solder-sealed rings are deposited with relative thick-
ness, which are patterned on the other side of the respective Si wafers after
fabricating the AR layer, as shown in Figure  8.3b. Two wafers are bonded
face to face at 150°C for 30 minutes. The bonding mechanism uses low eutec-
tic point (118°C) of In/Sn alloy. Therefore the polarizer could be bonded at
low temperature [4]. Using ultrathin buffer layer could tentatively avoid
Cu-In/Sn interdiffusion and ensure that the residue of In/Sn would melt
even with submicron thickness [5]. The fabrication method of robust THz
polarizer with AR layers using 3D-IC technologies is demonstrated. The Cu
wire-grid is sealed into bonded wafers to prevent corrosion.

8.2.4 Performance of Staked THz Polarizer

8.2.4.1 Bonding and Etching Qualities
When the temperature rises to the eutectic point of In/Sn binary alloy, which
is 118°C, In interdiffuses with Sn at the interface and then melts at the inter-
face. The liquid layer of In/Sn improves the bonding yield because of large
tolerance of surface roughness. However, In/Sn solder also interdiffuses
with Cu to form intermetallic compound (IMC) immediately. The residue
of In/Sn solder is not enough to form the liquid layer if the solder layer is
too thin, which could cause bonding failure. A Ni buffer layer is capped on
Cu layer as a barrier to delay the interdiffusion of the solder and Cu. To test
the bonding yield of the bonding structure in this research, blanket wafers
with the same metal bonding layers as mentioned earlier are fabricated as
shown in Figure 8.3b and bonded with the same condition of the polarizers.
Figure  8.4a shows good bonding yields of Cu-In/Sn bonding structure by
scanning acoustic tomography (SAT) images of the blanket sample.

Size: 2 cm × 2 cm
12 polarizers for a bonded wafer
(a) (b)

FIGURE 8.4
(a) SAT image showing the excellent bonding quality and (b) the finished component of THz
polarizers. (From Yu, T.-Y. et al., SPIE Optical Engineering + Applications, 95850L–95850L-7, 2015.)
Novel Platform and Applications Using 3D Integration Technologies 191

AR1 20 um AR2 20 um AR3 20 um

FIGURE 8.5
SEM images and parameters set up of AR layers. (From Yu, T.-Y. et al., SPIE Optical Engineering +
Applications, 95850L–95850L-7, 2015.)

As discussed earlier, the depth and filling factor of the etching holes
control the central frequency and the refractive index of the AR layer.
Figure 8.5 demonstrates the designed parameters and SEM measurement
results of the three different AR layers. Comparing the etching results
with the designed parameters, all samples are close to the original design
except the hole diameter of AR3, which is a little larger. The large holes
make  the n AR smaller than the designed value because of smaller filling
factor according to Equation 8.3. Modified etching process would improve
the etching profile control.

8.2.4.2 High Transmittance THz Polarizer

THz time-domain spectroscopy (TDS) is used to measure the transmittance
of stacked THz polarizers. The samples are placed behind a commercial
free standing wire-grid polarizer that fixes polarization direction during
the measurement. Figure 8.6 shows the measured method of the THz-TDS

AR layers
Commercial wire
grid polarizer TE
E

90°

TM

k
THz TDS system under low-humidity
environment

FIGURE 8.6
Stacked THz polarizer measurement by THz-TDs through commercial polarizer. (From Yu,
T.-Y. et al., SPIE Optical Engineering + Applications, 95850L–95850L-7, 2015.)
192 3D Integration in VLSI Circuits

1.0 1

0.8
0.1
Transmittance

Transmittance
0.6
0.01
0.4
TM with AR1 TE with AR1
0.2 TM with AR2 TE with AR2
1E−3
TM with AR1 TE with AR1
TM with AR3 TE with AR3
TM with AR2 TE with AR2
0.0 1E−4 TM with AR3 TE with AR3

0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4
(a) Frequency (THz) (b) Frequency (THz)

FIGURE 8.7
(a and b) The transmittance spectrum of TE and TM mode of stacked THz polarizers in linear
and log scale. (From Yu, T.-Y. et al., SPIE Optical Engineering + Applications, 95850L–95850L-7, 2015.)

measurement. THz-TDS results demonstrate electric field variation in time

domain. Fast Fourier transform (FFT) transfers TDS signal to the frequency
domain with complex amplitude and phase information and then transfers
to transmittance spectrum after a simple calculation. Figure 8.7a shows the
transmittance spectra of transverse electric (TE) and TM mode of stacked
polarizers. Over 90% of TM-mode transmittance and minimum transmit-
tance about 0.01% of TE mode are demonstrated with each polarizer sample.
Especially, the AR2 sample reaches almost 100% TM transmittance at the
designed central frequency. However, the enlarged holes of AR3 sample
shift the central frequency to higher frequency as a result of smaller effective
index. Nevertheless, all three samples show high extinction ratio between 20
and 40 dB from 0.2 to 2.2 THz as shown in Figure 8.7b.

8.2.4.3 Broadband THz Polarizer

For the broadband THz polarizer, two wafers with different AR layers are
bonded and the Cu wire-grid is sealed by bonded wafers. The transmittance
spectrum is broadened due to combination of two different AR layers with
different central frequencies. Figure 8.8 shows the transmittance spectrum
of different stacked AR layers, and the transmittance is basically the product
of transmittance of each AR layer. Therefore, the bandwidth of transmittance
spectrum is extended more than 1  THz. Furthermore, the peak values of
AR1×AR2 and AR1×AR3 polarizers have uniform transmittance about 85%
and 70% with width more than 250 GHz and 1 THz, as shown in Figure 8.8b
and d. This means when passing through the broadband polarizer the
Novel Platform and Applications Using 3D Integration Technologies 193

1.0 1.0
0.9
0.8 AR1×AR2
0.9
0.7
Transmittance

Transmittance
0.6 Peak width: 1.1 THz 0.8
0.5
0.4 0.7 >250 GHz
0.3 AR1×AR1
0.2 AR2×AR2
0.6
AR1×AR2
0.1
0.0 0.5
0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
(a) Frequency (THz) (b) Frequency (THz)

1.0 1.0
0.9 0.9
0.8 0.8
0.7 0.7
Transmittance

Transmittance

0.6 0.6
0.5 0.5
0.4 0.4
0.3 AR3×AR3 0.3 AR3×AR1
0.2 AR1×AR1 0.2
AR3×AR1
0.1 0.1
0.0 0.0
0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
(c) Frequency (THz) (d) Frequency (THz)

FIGURE 8.8
Transmittance spectra of (a) TE and (b) TM mode of stacked THz with AR1×AR2 layers, and
(c) TE and (d) TM mode of THz with the AR1×AR3 layers. (From Chi, N.-C. et al., SPIE Optics +
Optoelectronics, 102420Z–102420Z-6, 2017; Chi, N.-C. et al., IEEE ECTC, 1793–1798, 2017.)

broadband signal is distortion less and still has high-power transmission.

Hence,  the THz polarizer fabricated by bonding two different AR layers
could have a wider usage rather than limited in central frequency owing to
its narrow bandwidth.

8.2.5 Comparison between Common and Stacked THz Polarizers

Table 8.1 demonstrates the previous works of THz polarizers and characteris-
tic of stacked THz polarizer [8]. Stacked polarizer has advantages with large
area, low cost, high performance, and convenient for using. Most materials of
THz polarizers listed in the table are not so easy to fabricate or easy to use.
However, stacked polarizer is almost fabricated on Si and really potential for
manufacturing because of mature semiconductor foundries. In conclusion,
stacked THz polarizer provides a better option for THz optical system, which
comparably requires robust structure but provides efficient performance.
194 3D Integration in VLSI Circuits

TABLE 8.1
Overview of Some Designed and Commercial THz Polarizer
Measured
Extinction Bandwidth
Group Transmittance Ratio (dB) (THz) Structure
Hsieh et al. 0.60 ~ 0.85 20 ~ 40 0.2 ~ 1.0 Feussner polarizer based on
a liquid crystal
Ren et al. 0.75 ~ 0.95 18 ~ 50 0.2 ~ 2.2 Carbon nanotube (CNT)
layers
Kyoung et al. 0.45 ~ 0.60 27 ~ 38 0.1 ~ 2.0 Reel-wound CNT
multilayers
Wojdyla and 0.90 ~ 0.99 37.8 (max) N/A Brewster’s polarizer using a
Gallot stack of silicon wafers
Yamada et al. 0.95 35 ~ 50 0.5 ~ 3.0 Al on thick Si substrate
angled at Brewster angle
MacPherson et al. 0.91 ~ 0.98 20 ~ 37 0.2 ~ 2.0 Thin-film aluminum on SiO2
Microtech 0.9 ~ 1.0 20 ~ 40 0 ~ 3 Free standing wire grid
Microtech 0.6 ~ 1.0 20 ~ 40 0 ~ 3 Wire-grid pattern on thin
substrate
Tydex 0.9 ~ 0.95 ~20 0 ~ 3 Wire-grid pattern on thin
substrate
Stacked THz 0.5 ~ 1.0 20 ~ 40 0.2 ~ 2 Wire-grid sealed into Si
polarizers wafer and AR processing
on surface
Source: Yan, F. et al., J. Infrared Millim. Terahertz Waves, 34, 489–499, 2013.

8.3 Pressure-Sensing System

Heterogeneous integration is an important development direction for IoT cloud
and smart life [9]. Furthermore, complementary metal–oxide–semiconductor–
microelectromechanical systems (CMOS–MEMS) integration is critical in the
future to sense the external environment. Advanced packaging technologies
are developed to integrate different functions of the chips including sensor,
logic, and memory. Among the packaging approaches, 3D/2.5D integration
technologies have attracted attention because they are highly integrated, low
weight, with high performance, and small form factor [10–11]. However, seri-
ous issues still exist in conventional sensor chip of 3D/2.5D integration. Due
to the functionality of sensor chip, the surface morphology is quite different
from circuit chip. The surface conditions of sensor chip usually have complex
topography and brittle surface, even transparent material. Hence, the han-
dling of sensor chip suffers throughput and alignment accuracy and becomes
a bottleneck during assembly. Therefore, an alternative approach is needed to
overcome the limitations.
Novel Platform and Applications Using 3D Integration Technologies 195

Pressure Readout
sensor circuit

Wireless

FIGURE 8.9
Concept of pressure-sensing system for monitoring external pressure variation. (From Hu,
Y.-C. et al., IEEE IEDM, 9.2.1–9.2.4, 2016.)

Self-assembly technology has been proposed for multichip integration to

improve the throughput [12–13]. This approach is also suitable for the hard-
to-handle sensor chip. Furthermore, we develop a double-self-assembly tech-
nology to implement chip-level stacking with more complicated topography.
A pressure sensing system is adopted to verify this technique as shown in
Figure 8.9. A heat dissipation micropin-fin interposer with uneven backside
profile is formed to carry a readout circuit chip and a complex topogra-
phy sensor chip. Cu/In metal bonding is used for low resistance and high
I/O density in this platform. The excellent reliability electrical tests prove
that the innovative double-self-assembly approach can efficiently integrate
CMOS–MEMS chips efficiently.

8.3.1 Pressure-Sensing Platform

The block diagram of the integrated pressure-sensing platform is mainly
composed of four parts: pressure sensor, readout circuit, wireless module,
and a receiver, which is shown in Figure 8.10. For small form factor and
lower power consumption, the pressure sensor and readout circuit chip

Sensor app demo platform

Integrated with double-

self-assembly approach

Gain Phone
Sensor Wireless
stage app

μ-pin fin interposer

FIGURE 8.10
Pressure-sensing system consists of sensor chip, readout circuit chip, wireless module, and
receiver. (From Hu, Y.-C. et al., IEEE IEDM, 9.2.1–9.2.4, 2016.)
196 3D Integration in VLSI Circuits

VCC Ref
−
G ADC

Sensor TSV Gain stage Wireless Phone app.

FIGURE 8.11
Block diagram of the proposed pressure-sensing system. The signal begins from the pressure
sensor to the receiver. (From Hu, Y.-C. et al., IEEE IEDM, 9.2.1–9.2.4, 2016.)

are integrated on the heat dissipation micropin-fin interposer. Double-self-

assembly approach is adopted to assemble the stacked chips. Figure 8.11
presents the simplified architecture and equivalent circuit of the proposed
pressure sensing system. The integrated platform transmits voltage signal
from the pressure sensor to readout circuit through TSV and redistribution
layer (RDL) on the interposer under pressure variation. The readout circuit
is mainly configured a gain stage, which amplifies weak signal to an analog-
to-digital converter (ADC) circuit. The wireless chip receives and transmits
information to a mobile device receiver. Users can be remotely informed the
environmental pressure variation on the screen.

8.3.2 Pressure Sensor

A piezoresistive pressure sensor is adopted in the pressure-sensing system
for sensing environment pressure difference. When the piezoresistive mate-
rial is deformed due to a change in stress, the resistance value also changes
with the characteristic of piezoresistive material. Therefore, we can measure
the pressure through the circuit connection and a designed Wheatstone
bridge. Figure 8.12 shows SEM image of the pressure sensor and a schematic
of the integrated pressure system. A membrane on the pressure sensor is fab-
ricated by silicon DRIE and ion implantation for sensing the pressure varia-
tion. However, the top side of the pressure sensor is fragile due to slender
membrane structure, and the backside of the pressure is hard to handle as
the surface is uneven with cavities.

8.3.3 Micropin-Fin Heat Sink Interposer

Interposer is adopted for carrying CMOS readout circuit chip and pressure
sensor chip. The particular design, micropin-fin, is developed for heat dissi-
pation due to heat accumulation, which is a critical issue in 3D IC integration.
Figure 8.13 shows the schematic and SEM image of the micropin-fin heat
sink interposer. A 200  mm silicon wafer is processed to fabricate Cu TSV
with the formation of 30 μm diameter and 250 μm pitch. On the front side of
Novel Platform and Applications Using 3D Integration Technologies 197

Membrance

Readout TSV
circuit Cavity
Bonded joint
RDL
TSV

μ-pin fin

ENIG

FIGURE 8.12
Schematic of the integrated readout circuit chip, pressure sensor chip, and micropin-fin inter-
poser. Cross-sectional and top view of the pressure sensor. (From Hu, Y.-C. et al., IEEE IEDM,
9.2.1–9.2.4, 2016.)

Bonded joint
RDL
TSV

Micropin-fin
ENIG

TSV-embedded micropin-fin

FIGURE 8.13
Schematic of the micropin-fin heat sink interposer and SEM image of the micropin-fin. (From
Hu, Y.-C. et al., IEEE IEDM, 9.2.1–9.2.4, 2016.)

interposer, two RDLs are fabricated for connecting sensor chip and circuit
chip. Cu/In-bonded pads are formed on the RDL. On the backside of inter-
poser, electroless nickel immersion gold (ENIG) is chosen as bonded pads.
DRIE process is adopted to etch deeply micropin-fins shape. The relation-
ship between micropin depth and temperature reduction is simulated. TSV
daisy chain electrical characteristics, thermal cycling test (TCT) with −55°C
to 125°C for 750 loops, and highly accelerated stress test (HAST) at 130°C for
96 hours have been investigated [14].
198 3D Integration in VLSI Circuits

8.3.4 Integration of Micropin-Fin Heat Sink Interposer and Chips

with Double-Self-Assembly Approach
Double-self-assembly approach is developed and adopted to integrate
CMOS readout circuit chip, pressure sensor chip, and micropin-fin heat
sink interposer. Figure 8.14 illustrates the 3D/2.5D assembly process flow
between the chips and interposer. On the carrier wafer, a 500 nm hydrophilic
oxide is deposited, followed by patterned perfluorodecyltrichlorosilane
(FDTS) hydrophobic film to define the self-assembly area of the interposer.
Afterwards, precisely controlled small volume of liquid is dropped on the
hydrophobic area and interposer is placed roughly on the liquid. The micropin-
fin side of interposer is also covered with oxide layer for self-assembly expect
ENIG-bonded pad. As the interposer is placed on the liquid, the liquid sur-
face tension pulls and drags the interposer to the defined hydrophilic area
immediately. Figure 8.14a demonstrates this procedure of the interposer first
self-assembly temporary bonding on the carrier wafer. The interposer front
side is also patterned FDTS hydrophobic film for the readout circuit and sen-
sor chip locations with bonded pads. After the chips are placed on top of the
liquid, the two chips are also self-aligned to the defined area with the same
method and mechanism. Next, the continued thermo-compression bonding
is adopted to accomplish the second self-assembly permanent bonding of
the readout circuit chip and sensor chip on the micropin-fin interposer, as
shown in Figure 8.14b. Finally, the 3D/2.5D pressure-sensing assembled chip
is detached from the carrier wafer once the liquid gets evaporated.
In order to measure the alignment accuracy, a designed misalignment
measurement mark with 0.5 μm accuracy is implemented to study the rela-
tionship with optimized liquid volume. The results of optimized liquid
volume of four different heights of micropins suggest that filling up each
10 μm depth of micropin gap must require 0.25 μL of liquid. The micropin-
fin interposer size depends on the liquid volume during first self-assembly

Readout
MEMS
circuit
Interposer Readout
circuit MEMS
TSV
Hydrophobic SiO2
Liquid
Carrier wafer Carrier wafer Carrier wafer

(a) (b) (c)

FIGURE 8.14
Process flow of double-self-assembly approach: (a) Interposer is temporarily attached on the
carrier wafer for the first self-assembly, (b) readout circuit and sensor chips are permanent
attached on the interposer for the second self-assembly, and (c) the integrated assembly chips
are detached from the carrier wafer. (From Hu, Y.-C. et al., IEEE IEDM, 9.2.1–9.2.4, 2016.)
Novel Platform and Applications Using 3D Integration Technologies 199

Cu/In IMC

Chip
μ-pin fin interposer 500 nm

FIGURE 8.15
Cross-sectional SEM image of Cu/In-bonded joint with no void or seam. (From Hu, Y.-C. et al.,
IEEE IEDM, 9.2.1–9.2.4, 2016.)

temporary bonding on the carrier wafer. The second self-assembly experi-

ments are performed with five different chip sizes to assess the correspond-
ing optimized liquid volume [14].
Figure 8.15 illustrates the cross-sectional SEM image of Cu/In between
upper die and lower interposer after double-self-assembly. From the image,
there are no crack and void at the bonding interface. In addition, SEM,
SAT, and EDX are used to investigate the bonding quality. Based on EDX
results, Cu-In is completely transferred into IMC. Shear test is adopted for
mechanical strength analysis. Cu-In bonding strength is over 20  MPa [14].
The material analysis and mechanical test prove the robust structure with
the double-self-assembly approach.
The electrical characteristics and reliability tests are evaluated using a
Kelvin structure consisting of Cu/In thin-film metal pad-bonded joint to
prove the feasibility of double-self-assembly method. This design is used to
measure the specific contact resistance of bonded joint. The obtained resis-
tance value from Kelvin structure is around 10−9 Ω·cm2. HAST and TCT are
carried out for 96 hours with 85% RH and 750 loops under −55°C to 125°C,
respectively. Results of stable specific contact resistance before and after
HAST imply good moisture resistance capability with double-self-assembly
approach. TCT also indicates good electrical property under significant
changes in temperature. The stable reliability results of HAST and TCT of
the integration between chips and interposer demonstrate the stability and
feasibility of the double-self-assembly approach. Figure 8.16 shows the pro-
posed 2.5D/3D integrated pressure-sensing system, which contains CMOS
readout circuit chip, pressure sensor chip, and micropin-fin heat sink inter-
poser with double-self-assembly approach.

8.3.5 Achievements and Outlook

A novel 3D integration technology based on liquid-surface tension mech-
anism to solve difficult-to-handling chip issue of double-self-assembly
approach is introduced. Using this advanced technique, we demonstrate
a pressure-sensing system with readout circuit chip, complex topography
200 3D Integration in VLSI Circuits

MEMS chip
Interposer
Circuit chip

Carrier wafer

FDTS film

FIGURE 8.16
Photograph of the proposed integrated pressure-sensing system with double-self-assembly
approach. (From Hu, Y.-C. et al., IEEE IEDM, 9.2.1–9.2.4, 2016.)

sensor chip, and micropin-fin interposer. Chips with Cu-In bonded pads are
successfully self-assembled on the interposer. Moreover, interposer is also
temporary attached on the carrier wafer and well debonded with first self-
assembly approach. Various sizes of interposer and chip are investigated
with the relationship of optimized liquid volume. The resulting electrical
characteristic shows high-accuracy alignment and stable reliability, which
implies the feasibility of double-self-assembly. Heterogeneous integration
can be further extended to various types by which difficult-handling chip
can be processed based on this novel technology.

8.4 Neurosensing Systems

In 1942, people first recorded the feeble brain signal from the scalp. Since
then, people realized the importance of relationship between electrical sig-
nal and the diseases using electrocorticography (ECoG) or electroencephalog-
raphy (EEG) to record the brain signal. The former president of the United
States of America, Barack Obama, has proposed the national research strategy
“Brain Research through Advancing Innovative Neuro technologies” in 2013.
Thus, the national research institutes throughout the world put lots of effort
to break through the clinical trial therapy. In order to complete understand-
ing the relationship between the brain signal and diseases, high spatial brain
signal extraction accuracy, low-signal transmitting loss, and high-resolution
neural sensing biosensor are very important. In recent years, different types
of neural sensing biosensors have been proposed, including electrical signal
analysis circuit chips, cables or wire bonding for signal transmission, and pas-
sive detection devices. However, there are some issues in the traditional neural
sensing biosensor [15]. First, wire-bonding connection and cables may cause
weak brain signal degradation and noise [16]. Second, spatial resolution limits
by the electrode channel numbers and its process [17–18]. Third, the yield is
low because of complicated process for electrode and circuit chip in one wafer.
Novel Platform and Applications Using 3D Integration Technologies 201

Three-dimensional integration technology is a potential packaging scheme

to overcome the above-mentioned issues to realize the novel neural sens-
ing biosensors. TSV and metal thin-film bonding techniques can substitute
the long cables and wire bonding to reduce signal degradation. Utilizing
standard semiconductor fabrication for silicon interposers and electrode
probes can achieve higher channel numbers and lower cost. Moreover, three-
dimensional heterogeneous integrations with bonding process can combine
optimized functional circuit chips, electrodes, and substrates, which can
increase yield, and the process is much more flexible.
In this section, three types of neural sensing biosensors, 3D-SiP neural
sensing biosensor, 2.5D-silicon interposer neural sensing biosensor, and
2.5D-flexible interposer neural sensing biosensor, are proposed for different
functional aspects.

8.4.1 Fabrication, Scheme, and Reliability of Neural Sensing Biosensor

8.4.1.1 Three-Dimensional System-in-Packaging Neural Sensing Biosensor
The scheme of 3D system-in-packaging (SiP) neural sensing biosensor con-
tains TSV-embedded electrode array, silicon interposer, and two stacked
circuit chips. The chip-level integration scheme is capable of integrating dif-
ferent substrates, chip sizes, and technology nodes by using suitable bonding
materials. Furthermore, it can be applied to wafer-level integration scheme.

8.4.1.2 Chip-Level Heterogeneous Integration Scheme

In this scheme, two stacked circuit chips are fabricated by Taiwan semicon-
ductor manufacturing company (TSMC) 40 nm and 0.18 μm CMOS node to
reveal the feasibility of chip-level Cu/Sn joint to ENIG bonding. As utilized
bump height of the joint must be less than 10 μm, the thicknesses of Cu and
Sn are 4 and 5  μm, respectively. The thickness of ENIG is 3.25  μm, which
contains 3 μm Ni and 0.25 μm Au.
There are 2035 μ-joints for the following face-to-face bonding process. The
two chips are bonded together at 270°C for 10  seconds. Figure 8.17 shows
the cross-sectional SEM image of chip level heterogeneous bonding without
having any voids and defects.

8.4.1.3 Electroplating Solution Improvement

Along with the fine pitch bump, the reduction of bump height becomes more
essential. Figure 8.18 illustrates the material phase transformation during the
bonding process. The concave–convex phenomenon caused by the Cu/Sn
bump formation and the passivation layer has been noticed in previous stud-
ies. However, the nonuniform surface causes degradation of the bonding qual-
ity in sub 10-μm-bumps. This issue can be improved by chemical–mechanical
202 3D Integration in VLSI Circuits

40 nm technology node chip

ENIG film

11.5 μm height
Pillar bump

0.18 μm technology node chip

20 μm

FIGURE 8.17
SEM cross-section image of chip-level heterogeneous integration scheme. (From Hu, Y.-C.
et al., IEEE Trans. Electron Devices, 62, 4148–4153, 2015.)

Upper chip Upper chip

ILD Top metal ILD ILD Top metal ILD
PD ENIG PD PD PD
ENIG
Sn squeeze Sn squeeze
Sn
Sn PD Cu PD
PD Cu PD
Top metal ILD ILD Top metal ILD
ILD
(a) Lower chip (b) Lower chip

Upper chip
ILD Top metal ILD
PD ENIG PD

IMC Sn

PD Cu PD
ILD Top metal ILD
(c) Lower chip

FIGURE 8.18
The schematic of material-phase transformation during bonding process. (a) ENIG film and
concave-shaped bump formation before bonding, (b) During bonding, Sn squeezes outside
and inside the concave-shaped bump, (c) Pure Sn is surrounded by the middle of IMC. (From
Hu, Y.-C. et al. IEEE Trans. Electron Devices, 62, 4148–4153, 2015.)

planarization (CMP) process, which is very expensive and dirty. In this study,
by adjusting the plating current and changing plating solution, the bonding
quality can be improved efficiently. The adopted methodology is compara-
bly lesser in cost, highly efficient, and also compatible to semiconductor pro-
cesses. The plating results under different plating current and solution shows
Novel Platform and Applications Using 3D Integration Technologies 203

that the height difference between the bump center and edge is less than 1 μm
[19]. During the plating process, the chloride in the improved solution links
with the Cu+ ion to form the suppression layer on the cathode. Thus, the Cu+2
on the surface of the cathode will not be inhibited.

8.4.1.4 Electrical Reliability Tests

The electrical performance and reliability of the bonded structure are
tested by Kelvin structure and daisy chain. One bonded joint resistance is
~0.22Ω by Kelvin structure measurement [20]. A 30-daisy-chain in series of
ENIG to Cu/Sn-bonded joints has a resistance of 24.6Ω [20]. The reliability
tests used here are TCT and unbiased HAST, which test the thermal reliabil-
ity of the bonded joint and the moisture resistance. Results of a 1000-loops
TCT test, which is under −55°C to 125°C with 15°C/min ramping rate show
that the resistance of the bonded joints gets better because of the thermal
treatment during the thermal cycling that enhances the diffusion to elimi-
nate the bonding interface and may also form low-resistance IMC [20]. The
96-hours unbiased HAST under 85% RH and 130°C results illustrate good
humidity resistance. Thus, the heterogeneous-bonded structure shows good
bonding quality and reliability [20].

8.4.2 2.5D-Silicon Interposer Neural Sensing Biosensor

In order to realize the high spatial and resolution neural sensing biosensor,
the channel numbers should be increased and cost must be scaled down.
The 2.5D-silicon interposer neural sensing biosensor can achieve all the
demands, as shown in Figure 8.19. This scheme illustrates the capability of
heterogeneous integration, different functional chips, high channel transi-
tion, and semiconductor fabrication compatibility.

Parylene-C covered
Connector
90 nm tech node neural chip 90 nm tech node neural chip

TSV
Interposer

TSV
μ-Probe

FIGURE 8.19
Schematic of 2.5D-silicon interposer neural sensing biosensor. (From Hu, Y.-C. et  al., IEEE
Trans. Electron Devices, 64, 1666–1673, 2017.)
204 3D Integration in VLSI Circuits

8.4.2.1 Fabrication of Silicon Interposer and Through-Silicon

Via-Embedded μ-Probe
Figure 8.20 shows the process flow of the TSV of silicon interposer. The
diameter and depth of TSV are 30 and 130 μm, respectively. First, utilizing
Bosch process to etch the silicon via and deposit 1 μm thickness oxide liner
to prevent the diffusion of atom and leakage current. Second, depositing TiN
adhesive layer and Cu seed layer for the following Cu electroplating process.
After the electroplating process, CMP process will be conducted to remove
the superfluous Cu film. The wafer is grinded to 130 μm, and then the RDL
and ENIG bonding pad are fabricated. Finally, two-layer RDL and 50-μm
μ-bump on another side for the functional circuit chips bonding joint are
accomplished.
TSV-embedded electrode μ-probe fabrication process flow is shown in
Figure 8.21. After TSV fabrication, photoresist defines the pitch and the size
of probe. The isotropic etching uses SF6 to shape the probe opening. Then
the nonisotropic inductively coupled plasma (ICP) etching is employed to
define the pitch depth. Pitch, diameter, and depth of the μ-probe are 140, 70,
and 80 μm, respectively. Finally, Ti and Pt thin film layers are deposited on
the probe to carry out the biocompatible electrode.

8.4.2.2 Electrical Reliability Test

The Cu TSVs and the bonded joints are verified under TCTs and unbias
HAST to check the electrical reliabilities, including the 400 Cu TSVs in series

(a) (b) (c) (d)

(e)
(f ) (g)

(h) (i) (j)

FIGURE 8.20
Process flow of silicon interposer. (a) TSV etching, (b) oxide linear and seedlayer deposit,
(c) TSV plating, (d) Cu CMP, (e) RDL and ENIG pad fabrication, (f) temporary bonding, (g) wafer
thinning, (h) 2 RDLs fabrication, (i) microbump plating, and (j) handle wafer debond. (From
Hu, Y.-C. et al., IEEE Trans. Electron Devices, 64, 1666–1673, 2017.)
Novel Platform and Applications Using 3D Integration Technologies 205

PR PR PR

(a)
SF6 PR SF6 PR SF6 PR SF6

(b)
PR PR PR

(c)

FIGURE 8.21
Process flow of TSV-embedded μ-probe. (a) PR is patterned as followed etching hard, (b) iso-
tropic etching for concave shape around probe opening, and (c) Bosch process is adopted for
probe formation. (From Hu, Y.-C. et al., IEEE Trans. Electron Devices, 64, 1666–1673, 2017.)

resistance and bonded joints resistance before and after the 1000 thermal
cycling loops, and the Cu TSVs in series and bonded joints resistance before
and after the unbias HAST under 130°C, 85% RH for 96 hours. All the reli-
ability test results show good reliability of the 2.5D neural sensing biosensor
scheme [18].

8.4.3 2.5D-Flexible Interposer Neural Sensing Biosensor

In this section, the polyimide, with biocompatible and bendable properties,
better for animal experiments, is used for interposer to replace the silicon
interposer. The polyimide is flexible along the surface of the cortex or other
tissues. Furthermore, the polyimide substrate is used widely in semiconduc-
tor fabrication. Figure 8.22 shows the integration schematic of 2.5D-flexible
interposer neural sensing biosensor, which contains μ-probe, PI interposer,
and circuit chips.

8.4.3.1 Fabrication of Flexible Interposer

The polyimide flexible interposer substrate, with a thickness of 75  μm, is
bendable and has high chemical resistance and high mechanical strength.
30-μm-diameter TSVs are Cu electrochemical deposited using a bottom-up
method. Then 50-μm-diameter and 1-μm-height In μ-bump arrays are fabri-
cated to carry the circuit chips.
206 3D Integration in VLSI Circuits

ENIG pad to Sn ball bonding ENIG pad to μ-bump bonding

Connector Parylene C
Die Die RDL
TSV (Au)
Flexible
interposer
Silicon
μ-probe
TSV (Cu)
Nitride TSV-embedded dissolvable μ-needle array Pt or Au

FIGURE 8.22
Schematic of 2.5D-flexible interposer neural sensing biosensor. (From Huang, Y.-C. et  al.,
Symposium on VLSI Technology, pp. 218–219, 2016.)

8.4.3.2 Novel Flexible Bonding Approach

In order to realize the flexible neural sensing biosensor, the key part is silicon
substrate to flexible substrate-bonding approach, which affects the electrical
properties and the sensor reliabilities. Traditional bonding approach uses
anisotropic conductive film (ACF) or nonconductive paste (NCP) in flexible
substrate bonding. However, the number of the conductive particles must
be well controlled for the good electrical connection without any leakage.
Furthermore, the fine pitch is also limited by the size of conductive particles.
Therefore, thin-film metal bonding approach is proposed to improve the
bonding quality and to achieve ultrafine pitch.
The 300 nm Cu thin film and 10 nm Ti thin film are deposited on 75 μm
polyimide flexible substrate, where Ti layer is the adhesive layer between the
Cu film and polyimide substrate. Next, 20 nm Ni film as the buffer layer is
deposited to prevent Cu oxidation before the bonding process. In order to
achieve low-temperature bonding, 300 nm Sn film is used for the bonding
material. On the other PI substrate, 300  nm In film is deposited. The two
PI substrates are bonded together under a force of 100  N and at 170°C for
30 minutes. The cross-sectional SEM image of In/Sn-Cu-bonded joint reveals
a good bonding quality [21].

8.4.3.3 Electrical Reliability Test of Novel Thin Film-Bonding Approaches

In this part, the electrical properties and reliabilities of two novel thin
film metal-bonding approach are tested. The specific contact resistances of
In/Sn-Cu thin film-bonding approach are measured under flat and bending
radius of 30 mm by Kelvin structure, respectively. The specific contact resis-
tance is about 2.5 × 10−7  Ω·cm2 in both cases. The reliability performance of
In/Sn-Cu thin film-bonding approaches under 500 thermal cycling loop of
Novel Platform and Applications Using 3D Integration Technologies 207

TCT tests and the un-bias HAST under 130°C, 85% RH for 96 hours, respec-
tively, illustrate the good bonded results, and the specific contact resistance
is better than the conventional bonding approaches [20]. Thus, this bonding
approach reveals the potential to improve the electrical properties of future
flexible packaging.

8.4.4 Demonstration
The overall schemes, fabrication procedures, electrical, and reliability
measurements of three types of the neural sensing biosensor schemes are
discussed. Demonstrations of a 2.5D-silicon interposer neural sensing bio-
sensor and a 2.5D-flexible interposer neural sensing biosensor are shown
in Figures 8.23 and 8.24, respectively. Key technologies of 3D IC are used in
these biosensors, such as heterogeneous integration, through silicon/flexible
via, wafer level thinning, and chip-level bonding. Circuit chips, interposer,
and TSV-embedded μ-probe are integrated within a small form factor. The
reliabilities of the bonded joints and TSVs are also investigated. These results
prove the potential of 3D IC technologies used in the biosensor scheme and
other products in the near future.

(b) Circuit chip

(a) Connectors Bonded joint

RDL
Interposer 40 μm
Chips
(c) Interposer
TSV
μ-probe Sn-solder
Interposer μ-probe
20

Polymer based μ-needle

Load (N)

10

0
μ-probe embedded 0.0 0.5 1.0 1.5 2.0
Displacement (mm)
(d) (e) (f )

FIGURE 8.23
Photos of 2.5D-silicon interposer neural sensing biosensor. (a) Side view, (b) Cross-sectional
view of circuit chip-interposer bonded joint, (c) Cross-sectional view of shunt-connected
interposer-probe bonded joint, (d) μ-needle SEM image, (e) Photo of neural sensing microsys-
tem, (f) Shear test diagram of the 2.5-D heterogeneous integrated neural sensing microsystem.
(From Hu, Y.-C. et al., IEEE Trans. Electron Devices, 64, 1666–1673, 2017.)
208 3D Integration in VLSI Circuits

Interposer
5 mm
die0
15 mm die1

15 mm
3 mm

5 mm
3 mm
die2
die3
TSV-embedded
Connector
μ-needles array

15 mm 15 mm
(a) (b)

(c) (d)

FIGURE 8.24
Photos of 2.5D-flexible interposer neural sensing biosensor. (a) Topview, (b) bottomview,
(c)  256-ch microsystem, and (d) bending. (From Huang, Y.-C. et  al., Symposium on VLSI
Technology, pp. 218–219, 2016.)

References
1. R. A. Lewis, A review of terahertz sources, Journal of Physics D: Applied Physics,
47, 374001, 2014.
2. X. Yin, B. W.-H. Ng, and D. Abbott, Terahertz sources and detectors, in
Terahertz Imaging for Biomedical Applications: Pattern Recognition and Tomographic
Reconstruction, New York: Springer, 2012, pp. 9–26.
3. T.-Y. Yu, H.-C. Tsai, S.-Y. Wang, C.-W. Luo, and K.-N. Chen, High transmittance
silicon terahertz polarizer using wafer bonding technology, in SPIE Optical
Engineering + Applications, 2015, pp. 95850L–95850L-7.
4. H.-W. Liang, T.-Y. Yu, Y.-J. Chang, and K.-N. Chen, Asymmetric low temper-
ature bonding structure using ultra-thin buffer layer technique for 3D inte-
gration, in 2016 IEEE 23rd International Symposium on the Physical and Failure
Analysis of Integrated Circuits (IPFA), 2016, pp. 312–315.
5. Y.-J. Chang, Y.-S. Hsieh, and K.-N. Chen, Submicron Cu/Sn bonding technol-
ogy with transient Ni diffusion buffer layer for 3DIC application, IEEE Electron
Device Letters, 35, 1118–1120, 2014.
6. N.-C. Chi, T.-Y. Yu, H.-C. Tsai, S.-Y. Wang, C.-W. Luo, and K.-N. Chen, High trans-
mittance and broaden bandwidth through the morphology of anti-relfective
layers on THz polarizer with Si substrate, in SPIE Optics + Optoelectronics, 2017,
pp. 102420Z–102420Z-6.
7. N.-C. Chi, T.-Y. Yu, H.-C. Tsai, S.-Y. Wang, C.-W. Luo, Y.-T. Yang, K.-N. Chen,
High transmittance broadband THz polarizer using 3D-IC technologies, in
IEEE Electronic Components and Technology Conference (ECTC), 2017, 1793–1798.
Novel Platform and Applications Using 3D Integration Technologies 209

8. F. Yan, C. Yu, H. Park, E. P. J. Parrott, and E. Pickwell-MacPherson, Advances

in polarizer technology for terahertz frequency applications, Journal of Infrared,
Millimeter, and Terahertz Waves, 34, 489–499, 2013.
9. W. Dehaene and A. S. Verhulst, New devices for internet of things: A circuit
level perspective, in 2015  IEEE International Electron Devices Meeting (IEDM),
2015, pp. 25.5.1–25.5.4.
10. Y. Takemoto, K. Kobayashi, M. Tsukimura, N. Takazawa, H. Kato, S. Suzuki
et  al., Multi-storied photodiode CMOS image sensor for multiband imaging
with 3D technology, in 2015 IEEE International Electron Devices Meeting (IEDM),
2015, pp. 30.1.1–30.1.4.
11. M.-H. Li, C.-Y. Chen, and S.-S. Li, A reliable CMOS-MEMS platform for titanium
nitride composite (TiN-C) resonant transducers with enhanced electrostatic
transduction and frequency stability, in 2015 IEEE International Electron Devices
Meeting (IEDM), 2015, pp. 18.4.1–18.4.4.
12. T. Fukushima, T. Konno, K. Kiyoyama, M. Murugesan, K. Sato, W.-C. Jeong
et al., New heterogeneous multi-chip module integration technology using self-
assembly method, in 2008 IEEE International Electron Devices Meeting (IEDM),
2008, pp. 1–4.
13. T. Fukushima, E. Iwata, Y. Ohara, A. Noriki, K. Inamura, K.-W. Lee et al., Three-
dimensional integration technology based on reconfigured wafer-to-wafer
and multichip-to-wafer stacking using self-assembly method, in 2009  IEEE
International Electron Devices Meeting (IEDM), 2009, pp. 1–4.
14. Y.-C. Hu, C.-P. Lin, H.-C. Chang, Y.-T. Yang, C.-S. Chen, and K.-N. Chen, An
advanced 3D/2.5D integration packaging approach using double-self-assembly
method with complex topography, and micropin-fin heat sink interposer for
pressure sensing system, in 2016  IEEE International Electron Devices Meeting
(IEDM), 2016, pp. 9.2.1–9.2.4.
15. Y.-C. Hu, Y.-C. Huang, P.-T. Huang, S.-L. Wu, H.-C. Chang, Y.-T. Yang et al., An
advanced 2.5-D heterogeneous integration packaging for high-density neu-
ral sensing microsystem, IEEE Transactions on Electron Devices, 64, 1666–1673,
2017.
16. N. K. Logothetis, J. Pauls, M. Augath, T. Trinath, and A. Oeltermann,
Neurophysiological investigation of the basis of the fMRI signal, Nature, 412,
150–157, 2001.
17. L.-C. Chou, S.-W. Lee, P.-T. Huang, C.-W. Chang, C.-H. Chiang, S.-L. Wu et al.,
A TSV-based bio-signal package with μ-probe array, IEEE Electron Device Letters,
35, 256–258, 2014.
18. C.-W. Chang, P.-T. Huang, L.-C. Chou, S.-L. Wu, S.-W. Lee, C.-T. Chuang et al.,
Through-silicon-via-based double-side integrated microsystem for neural
sensing applications, in IEEE International Solid-State Circuits Conference Digest
of Technical Papers (ISSCC), 2013, pp. 102–103.
19. Y.-C. Hu, C.-P. Lin, Y.-S. Hsieh, N.-S. Chang, A. J. Gallegos, T. Souza et al., 3D
heterogeneous integration structure based on 40  nm- and 0.18  μm- technol-
ogy nodes, in 56th Electronic Components and Technology Conference (ECTC), 2015,
pp. 1646–1651.
20. Y.-C. Hu and K.-N. Chen, A novel bonding approach and its electrical per-
formance for flexible substrate integration, IEEE Journal of the Electron Devices
Society, 4, 185–188, 2016.
210 3D Integration in VLSI Circuits

21. Y.-C. Hu, C.-P. Lin, Y.-J. Chang, N.-S. Chang, M.-H. Sheu, C.-S. Chen, and
K.-N. Chen, A novel flexible 3-D heterogeneous integration scheme using elec-
troless plating on chips with advanced technology node, IEEE Transactions on
Electron Devices, 62, 4148–4153, 2015.
22. Y.-C. Huang, Y.-C. Hu, P.-T. Huang, S.-L. Wu, Y.-H. You, J.-M. Chen et  al.,
Integration of neural sensing microsystem with TSV-embedded dissolv-
able μ-needles array, biocompatible flexible interposer, and neural recording
circuits, in Symposium on VLSI Technology, 2016, pp. 218–219.
Index

Note: Page numbers followed by f and t refer to figures and tables respectively.

2.5D-flexible interposer neural sensing 3D technologies, 1

biosensor, 205, 208f 2D using logic cell partitioning, 9, 9t
flexible bonding approach, 206 BSI imagers, 146
flexible interposer fabrication, 205 co-engineering, 87–90, 88f
integration schematic, 206f comparative thermal analysis, 172–178
thin film-bonding approaches, die, 174
206–207 drivers, 5
2.5D-silicon interposer neural sensing heterogeneous processor die, 11, 12f
biosensor, 203, 207f for hot-spot dissipation scenarios,
electrical reliability test, 204–205 176–177, 176f, 177f
TSVs of, 204 module, 121, 122f
2D vs. 3D memories, 9, 9t monolithic, 4–5, 165f
3D Chip Stack with Through-Silicon Vias NAND flash, 4
(TSVs), 34 parameters, 172–175, 173t, 175t
3D die-level integration technology, performance enhancement, 8–10
120–121 power efficiency, 10–11, 11f
3DIC. See Three-dimensional integrated profilometry imaging, 130f
circuit (3DIC) stacking technologies, 2–4, 3f
3D-IC. See Three-dimensional for uniformly distributed power
integration chip (3D-IC) dissipation scenarios, 177, 178f
3DI technology. See Three-dimensional 3D VLSI (3D Very-Large-Scale-
integration (3DI) technology Integration), 163, 164f, 165
3D large-scale integration (LSI) test
chip, 72, 72f
A
heterogeneous, 76–79, 76f, 77f, 78f
3D-packaging technologies, 172 Advanced driver assistance systems
3D sequential, 162 (ADAS), 73, 74f
300 mm electrical demonstration, AFM (atomic force microscopy),
169–171, 171f 130, 130f
iBEOL, 167–169, 168t, 169f, 170f Analog-to-digital converters (ADC),
integration process flow and 8, 74
low-temperature top FETs, Antireflection (AR) coating, 187
165–166, 167f, 167t layer, 189, 192
principle, 163 purpose, 188
state of the art, 164–165 with quarter wavelength thickness,
TEM cross-section, 170, 170f 188, 188f
3D-stacked image sensor, 72–73, 73f SEM images and parameters set up,
image signal processing, 74f 191, 191f
system module, 73–74, 74f Application-specific integrated circuits
three-dimensional display, 74, 75f (ASICs), 15–16

211
212 Index

Application specific modular block Comparative thermal analysis, 3D

(ASMBL™) architecture, 46, 46f technologies, 172, 175
AR (Antireflection) coating, 187 for hot-spot dissipation scenarios,
ASICs (application-specific integrated 176–177, 176f, 177f
circuits), 15–16 for uniformly distributed power
ASMBL™ (application specific modular dissipation scenarios, 177, 178f
block) architecture, 46, 46f Complementary metal–oxide–
Atomic force microscopy (AFM), semiconductor (CMOS), 6–7,
130, 130f 135–136
3D module with IBM’s 22 nm, 122f
3D-stacked structure, 71
B
chip-level 3DI with 22 nm, 121–125,
Back-end-of-line (BEOL) process, 2, 146 124f, 125t
Back-side illuminated (BSI) imagers, 146 CoolCube™, 166, 172–173
BIST (built-in-self-test) engine, 139 3DVLSI, 162–171
Board-level reliability (BLR), 62 integration, 163, 163f
20 nm 3D-IC, 64f Copper (Cu) pillar technology, 25
TVs, 63–65 Copper-to-copper bonding
Bosch process, 3, 103 mechanisms, 152–153
Broadband THz polarizer, 192–193, 193f Cu-diffusion phenomenon, 101–104,
BSI (back-side illuminated) imagers, 146 102f, 104f
Built-in-self-test (BIST) engine, 139 Cu/SiO2 hybrid bonding, 147, 150f
Bumpless interconnects, 90–91, 90f, 107 bonding energy evolution with
thermal resistance of, 106–109, temperature, 152, 152f
107f, 108f bonding interface characterization,
using TSVs, 89, 89f 149–151
and wafer-level 3DI, 90–92 copper diffusion, 154–155, 155f
copper modification with
temperature, 152–154, 153f
C
pad dimension reduction, 155
Canthotaxis phenomena, 161, 161f surface preparation, 148–149
CEA-Leti, 148–149, 161–162 wafer alignment, 154, 154f
Chemical–mechanical polishing/
planarization (CMP) process,
D
20, 121, 201–202
ECP-Cu planarization, 95, 95f DCB (double cantilever beam)
optimization, 148 technique, 152
principle of damascene, 148f Deep reactive-ion etching (DRIE)
Chip-level heterogeneous integration process, 3, 20, 197
scheme, 201, 202f Defect-monitoring vehicle (DMV), 61–62
CMOS. See Complementary metal– Die-to-wafer (DtW) process, 160–161
oxide–semiconductor (CMOS) Directed self-assembly (DSA), 79, 80f
CMP process. See Chemical–mechanical DMV (defect-monitoring vehicle), 61–62
polishing/planarization (CMP) Double cantilever beam (DCB)
process technique, 152
Coarse-pitch silicon interposer, 20–21 Double-self-assembly approach,
fabrication steps, 22f 198–199, 198f
vs. fine-pitch silicon interposers, DRAM. See Dynamic random-access
20, 20t memory (DRAM)
Index 213

DRIE (deep reactive-ion etching) with ASMBL™ architecture, 46, 46f

process, 3, 20, 197 excessive latency, challenges, 44
DSA (directed self-assembly), 79, 80f limited connectivity and bandwidth,
DtW (die-to-wafer) process, 160–161 challenges, 44
Dynamic random-access memory power penalty, challenges, 44
(DRAM), 8–10, 17–18 product, 54–55
cross-sectional TEM image, 101f Si-IP with TSVs, 47–48, 48f
density, 110f stacked silicon integration,
retention time, 101f 46, 46f
retention time change, 98–99, 98f Xilinx, 45–46
stack structure, 112f Finite element method (FEM),
92, 92f, 104
Flip Chip and Memory Package
E
(FCAMP), 16
ECP-Cu (electrochemical-plated Cu), Focussed ion beam-scanning electron
95, 95f microscopy (FIB-SEM),
EDRAM. See Embedded dynamic 152, 152f
random access memory FPGA. See Field-programmable gate
(EDRAM) arrays (FPGA)
Effective thermal property extraction Free-standing/thin-film wire-grid
(EFFP), 175 polarizer, 187
Electrical reliability tests Front-end-of-line (FEOL)-type
2.5D-silicon interposer neural technology, 146
sensing biosensor, 204–205 Fully-depleted silicon-on-insulator
neural sensing biosensor, 203 (FDSOI), 164, 164f, 166
novel thin film-bonding approaches,
206–207
Electrochemical-plated Cu (ECP-Cu), G
95, 95f
Gold/tin (Au/Sn) soldering
Electromigration (EM), 158–159, 158f
technology, 126
Embedded dynamic random access
memory (EDRAM), 121
electrical metrics performance,
H
135–140, 136f, 139f
Even-mode coupling, 53, 54f Heterogeneous integration, 7–8, 194
High bandwidth memory (HBM),
9, 17–18
F
insertion loss of, 27f
Failure analysis (FA) of 3D-IC, 60, 61f return loss of, 28f
FCAMP (Flip Chip and Memory warpages measured, 30, 30f
Package), 16 High-density interposer
FDSOI (fully-depleted silicon-on- organic interposer, 21–23, 24f
insulator), 164, 164f, 166 Si-IP, 19–21, 19f, 21f
FEM (finite element method), 92, 92f, 104 High transmittance THz polarizer,
FIB-SEM (focussed ion beam-scanning 191–192
electron microscopy), 152, 152f HMC (Hybrid Memory Cube),
Field-programmable gate arrays 9, 18
(FPGA), 42 Hot-spot dissipation scenarios, 176–177,
3D-IC, 43f 176f, 177f
214 Index

Hybrid bonding, 3, 6–7, 159–160 Microbumps, 123

bonding energy evolution with interconnect, 23, 25
temperature, 152, 152f stacked silicon integration, 46–47
bonding interface characterization, warpage behavior, 62f
149–151 Micropillar interconnection, 25
copper diffusion, 154–155, 155f Micropin-fin heat sink interposer,
copper modification with 196–197, 197f
temperature, 152–154, 153f and chips with double-self-assembly
Cu/SiO2, 147–162 approach, 198–199, 198f
electrical structures presentation Middle-end-of-line (MEOL), 55
and performances, 155–157, process flow, 19–20
156f, 156t silicon-footing improvement,
EM, 158–159, 158f 58–59
environmental reliability study, silicon-grinding quality
157–158 optimization, 57, 57f
pad dimension reduction, 155 soft reveal process flow, 56f
surface preparation, 148–149 wafer edge and bevel-cleaning
technical challenges, 148–155 optimization, 57–58, 58f
wafer alignment, 154, 154f Miniaturization, 5–6, 6f
wafer-bonding processes, 127–128 Monolithic 3D integration, 165, 165f
Hybrid Memory Cube (HMC), 9, 18 Monolithic 3D technology, 4–5
MOSFETs (metal–oxide–semiconductor
field-effect transistors), 72, 166
I Multichip module (MCM) technology,
i-ACF film, 80, 81f 2, 44
iBEOL (intermediate back-end-of-lines),
167–169, 168t, 169f, 170f N
Integration process flow and
low-temperature top FETs, NAND flash technology, 4
165–166, 167f, 167t Neural sensing biosensor
Interdie routing, 53–54 2.5D-flexible interposer, 205–207, 208f
Intermediate back-end-of-lines (iBEOL), 2.5D-silicon interposer, 203–205, 207f
167–169, 168t, 169f, 170f 3D SiP, 201
Interposer (2.5D) technology, 2, 2f chip-level heterogeneous integration
scheme, 201, 202f
demonstration, 207
L electrical reliability tests, 203
electroplating solution improvement,
Lithography, economic and technical 201–203, 202f
issues, 87–88 fabrication, scheme, and reliability,
201–203
Neurosensing systems, 200–208
M
MCM (multichip module) technology,
O
2, 44
MEOL. See Middle-end-of-line (MEOL) Organic interposer, 21–23, 24f
Metal–metal bonding, 126–127 design, 26–27
Metal–oxide–semiconductor field-effect vs. silicon interposer, 22, 23t
transistors (MOSFETs), 72, 166 warpages measured, 30, 30f
Index 215

Outsourced assembly and test (OSAT), chip-level heterogeneous integration

19, 42 scheme, 201, 202f
Oxide-bonding technology Cu/In-bonded joint, 199f
layer preparation, 129–131 Cu TSVs, 95, 95f
wafer-level 3DI with, 131–135, 132f, i-ACF film, 80, 81f
133f micropin-fin, 197f
Oxide wafer-to-wafer bonding, 128–129 seven-wafer stack, 97, 97f
Self-assembly process of DtW,
161–162, 161f
P
Self-assembly technology, 195
Packaging module, wafer-level three- SEM. See Scanning electron microscopy
dimensional process, 97 (SEM)
Piezoresistive pressure sensor, 196 Semi-additive process (SAP), 20–21, 37
Plasma-enhanced chemical vapor Si-IP. See Silicon interposer (Si-IP)
deposition (PECVD), 94, 103, 129 Silicide, 163
Polyimide, 205 Silicon, 188
Poly-Si film, 72 Silicon-footing improvement, 58–59, 59f
Post-CMP 3D AFM images, 148, 149f Silicon-grinding quality optimization,
Power application scenarios, 174, 57, 57f
174f, 178f Silicon interposer (Si-IP), 19–21, 19f, 21f, 45
Pressure-sensing system, 194–195 with die-to-die interconnect
block diagram of, 196f wiring, 47f
concept of, 195f FPGA, 48f
with double-self-assembly approach, organic interposer vs., 23t
199, 200f process flow of, 204, 204f
micropin-fin heat sink interposer, TSV modeling/optimization, Type I
196–197, 197f signal, 50–53, 51f, 52f, 53f
piezoresistive pressure sensor, 196 with TSVs, 47–48
platform, 195–196 Type II signal integrity, 50–51, 51f,
readout circuit, 196 53–54, 54f
Pressure sensor, 196, 197f Simulation program with integrated
circuit emphasis (SPICE), 48
SiP (system-in-package), 16, 86
R
SoC (System-on-chip) designs, 44
Radio-frequency identification (RFID), 6 Solid–liquid interdiffusion (SLID), 126
Redistribution layer (RDL) process, 20, 91 SPC (statistical process control), 56
Stacked silicon interconnect technology
(SSIT), 42, 45, 54–55
S
Stacked terahertz optical component,
SAP (semi-additive process), 20–21, 37 186–194, 191f, 192f
Scalloping, 96f, 97 Stacking module, wafer-level three-
Scan acoustic microscopy (SAM), dimensional process, 93–94, 93f
150–151, 151f Statistical process control (SPC), 56
Scanning acoustic tomography (SAT), Stress-Test-Driven Qualification of
97, 98f, 190, 190f Integrated Circuits, 34
Scanning electron microscopy (SEM), Substrate process flow, 19–20
134, 135f Superchip, 76
AR layers, 191, 191f System-in-package (SiP), 16, 86
bump and bumpless TSVs, 105f System-on-chip (SoC) designs, 44
216 Index

T rationale, 118–120
system challenges, 73–75, 76f
Terahertz (THz) thermal simulations, 173f
gap, 186 Three-dimensional SiP, 15–17, 33f
polarizers issues, 187–188, 187f 3D-stackable memory, 17–19
wave applications, 186–187 ASICs and HBM, 25–29
Test vehicles (TVs), 59 assembly, 29–31
20 nm 3D-IC, 61–63, 63f cross-sectional view of, 25, 26f, 32f
28 nm 3D-IC, 60–61, 60f high-density interposer, 19–23
BLR, 63–65 microbump interconnect, 23, 25
Tezzaron DiRAM4, 10 module, 33f
Thermal comparison, 177–178 neural sensing biosensor, 201
3D technology parameters, 172–175 organic interposer, 21–23, 24f
Thin film-bonding approaches, reliability challenge, 33–37
electrical reliability test, Si-IP, 19–21, 19f, 21f
206–207 test and characterization, 32–33
Thinning module, wafer-level three- top view of, 25, 26f, 31f
dimensional process, 92–93 Through-silicon vias (TSVs), 2, 20,
Three-dimensional integrated circuit 72, 172
(3DIC) 3D LSI, 72f
cost reduction, 6–7 bank capacitance, 137, 137f, 138f
elements by 2.5D and, 33, 34t Bosch and direct etching methods,
heterogeneous integration, 7–8 96, 96f
reliability evaluation, 35, 35f bump vs. bumpless interconnects
reliability test board assembly, 35f using, 89, 89f
SOI wafers, 4, 5f coarse-pitch, 48
stacking technologies, 2 cross-sectional TEM image, 103, 103f
technology set, 2–5 DSA, 79, 80f
Three-dimensional integration chip electrical characteristics, 106
(3D-IC), 42 embedded μ-probe, 204, 205f
challenges, 67–68 fabrication steps, 3, 4f
FPGA, 43f formation and damascene Cu plug
module-based netlist, 49f processes, 94f
physical FA working flow, 60, 61f interposer wafer, 57, 59f
reliability assessment, 65–66, 65f low-aspect ratio, characteristics,
with SPICE, 48–50, 50f 103–106
TVs, 59–65, 59f module, 94–97
Xilinx, 55 module, wafer-level three-
Three-dimensional integration (3DI) dimensional process, 94–97
technology, 71–72, 86, 201 silicon etching time vs. aspect
22 nm CMOS technology, 121–125, ratio, 95f
124f, 125t stresses in Cu, 104, 105f, 106f
bumpless interconnecting and thermal resistance, 109f
wafer-level, 90–92 THz. See Terahertz (THz)
challenges, 72–81 THz polarizers, 187–188, 194t
delay of, 87 broadband, 192–193, 193f
design granularity, 147f common vs. stacked, 193
Index 217

fabrication methods and hybrid-bonding, 127–128

low-temperature eutectic metal–metal bonding, 126–127
liquid bonding, 189–190, 189f oxide bonding, 128–129
high transmittance, 191–192 rationale, 120
structure design of, 188–189 Wafer edge and bevel-cleaning
Time-domain analysis, 29, 29f optimization, 57–58, 58f
Total thickness variation (TTV), Wafer-level three-dimensional process
99–100, 99f packaging module, 97
TSVs. See Through-silicon vias (TSVs) stacking module, 93–94, 93f
thinning module, 92–93
TSVs module, 94–97
U
Wafer-on-wafer (WOW) process
Ultrathinning co-engineering, benefits,
characteristics, 98–102 112–113, 113f
and estimation of critical thickness, development, 91
99–100, 100f memory capacity enhancement, 110
thickness of wafer, 91–92 yield in wafer stacking, 111–112, 111f
Uniformly distributed power Wide I/O, 10
dissipation scenarios, 177, 178f WOW process. See Wafer-on-wafer
(WOW) process
W
X
Wafer-bonding technology
3DI stacking, 125–129 Xilinx, 42, 45, 55, 62