Measurement and Analysis of Variability in CMOS

circuits

Liang Teck Pang

Electrical Engineering and Computer Sciences

University of California at Berkeley

Technical Report No. UCB/EECS-2008-108

http://www.eecs.berkeley.edu/Pubs/TechRpts/2008/EECS-2008-108.html

August 29, 2008

 Copyright 2008, by the author(s).
All rights reserved.

Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission.

Acknowledgement

I wish to acknowledge the contributions of the students, faculty and

sponsors of the Berkeley Wireless Research Center, the National Science
Foundation Infrastructure Grant
No. 0403427, wafer fabrication donation of STMicroelectronics, and the
support of the Center for Circuit & System Solutions (C2S2) Focus Center,
one of five research centers funded under the Focus Center Research
Program, a Semiconductor Research Corporation program.
 Measurement and Analysis of Variability in CMOS circuits

by

Liang Teck Pang

Diplôme D’Ingénieur (Ecole Centrale de Paris, France) 1997

Master of Philosophy (University of Cambridge, United Kingdom) 1997

A dissertation submitted in partial satisfaction of the

requirements for the degree of
Doctor of Philosophy

in

Electrical Engineering and Computer Sciences

in the

GRADUATE DIVISION
of the
UNIVERSITY OF CALIFORNIA, BERKELEY

Committee in charge:
Professor Borivoje Nikolić, Chair
Professor Andrew Neureuther
Professor Robert Leachman

Fall 2008
 The dissertation of Liang Teck Pang is approved:

Chair Date

Date

Date

University of California, Berkeley

Fall 2008
Measurement and Analysis of Variability in CMOS circuits

Copyright 2008

by

Liang Teck Pang

 1

Abstract

Measurement and Analysis of Variability in CMOS circuits

by

Liang Teck Pang

Doctor of Philosophy in Electrical Engineering and Computer Sciences

University of California, Berkeley

Professor Borivoje Nikolić, Chair

The scaling of CMOS technology into the deep sub-micron regime has resulted

in increased impact of process variability on circuits, to the point where it is considered a

major bottleneck to further scaling. In order to continue scaling, there is a need to reduce

margins in the design by classifying process variations as systematic or random. In this

work, a methodology to characterize variability through measurement and analysis has been

developed. Systematic and random, die-to-die (D2D) and within-die (WID) components of

variability are quantified and corresponding sources of variability are identified.

This methodology was developed for an early 90nm CMOS process and further

refined for an early 45nm CMOS process. Test-chips have been designed to study the

effects of layout, and characterize variability of delay and leakage current using an array of

test-structures. Delay is obtained through the measurement of ring oscillator frequencies,

and transistor leakage current is measured by an on-chip analog-to-digital converter (ADC).

 2

In 90nm, it has been found that transistor performance depends strongly on polysil-

icon (poly-Si) gate density and that spatial correlation depends on gate orientation and the

direction of gate spacing. WID variation is small with three standard deviations over mean

(3σ/μ) ≈ 3.5%, whereas D2D and systematic layout-induced variations are significant, with

3σ/μ D2D variation of ≈ 15% and a maximum layout-induced frequency shift of 10%.

In 45nm, a process which features immersion lithography, strained-Si and more

restrictive design rules for gate spacing, it has been found that systematic layout-induced

variability has decreased. However, new sources of variability due to the dependence of

stress on layout were found. WID has increased to 3σ/μ ≈ 6.6% and can be attributed to

a smaller transistor area whereas D2D variation has remained at 3σ/μ ≈ 15%.

This methodology is effective in characterizing variability. It improves the accuracy

of statistical models and allows process corners to be set up for WID or D2D variations. In

addition, sources of systematic variations are identified and the impact of layout design rules

are measured. As scaling continues and variation increases, characterization of variability

will become an integral part of the IC design process.

Professor Borivoje Nikolić

Dissertation Committee Chair
 i

To my wife and our parents

 ii

Contents

List of Figures v

List of Tables ix

1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Research Goal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Related Work on Process Characterization . . . . . . . . . . . . . . . . . . . 5
1.3.1 Metrology Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.2 Measurement of transistor performance . . . . . . . . . . . . . . . . 6
1.3.3 Other circuit techniques for process characterization . . . . . . . . . 8
1.4 Dissertation Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.5 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2 Characterization of Variations 11
2.1 Characteristics of Variations . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Impact of Variability on Circuits . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 Sources of Process Variations . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3.1 Interface roughness . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.2 Line-edge roughness . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.3 Random Dopant Fluctuation . . . . . . . . . . . . . . . . . . . . . . 17
2.3.4 Subthreshold Voltage Variation . . . . . . . . . . . . . . . . . . . . . 17
2.3.5 Strain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

3 Process-induced Systematic Variations 21

3.1 Variations in the Exposure System . . . . . . . . . . . . . . . . . . . . . . . 22
3.2 Variations in Patterning and Other Manufacturing Steps . . . . . . . . . . . 26
3.3 Layout Dependent Variations . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
 iii

4 Measuring CMOS Performance Variability 30

4.1 Leakage Current . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
4.1.1 Leakage Transistor Array . . . . . . . . . . . . . . . . . . . . . . . . 31
4.1.2 Leakage Current Measurement . . . . . . . . . . . . . . . . . . . . . 32
4.2 Frequency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.2.1 Ring Oscillator Array . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.2.2 Ring Oscillators with Different Number of Stages . . . . . . . . . . . 50
4.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

5 Measuring SRAM Performance Variability 53

5.1 SRAM Static Noise Margins Measurement . . . . . . . . . . . . . . . . . . . 54
5.2 SRAM Dynamic Write Performance Measurement . . . . . . . . . . . . . . 56
5.3 SRAM Dynamic Read Performance Measurement . . . . . . . . . . . . . . . 57
5.4 SRAM Access Times Measurement . . . . . . . . . . . . . . . . . . . . . . . 61
5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

6 Measurement Results and Analysis 63

6.1 Layout Structures and Test-Chips Overview . . . . . . . . . . . . . . . . . 64
6.2 Effects of Layout on Frequency and Leakage . . . . . . . . . . . . . . . . . . 75
6.2.1 Proximity of Poly-Si Gates . . . . . . . . . . . . . . . . . . . . . . . 76
6.2.2 Gate Rotation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
6.2.3 Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
6.2.4 Metal Coverage Over Gate . . . . . . . . . . . . . . . . . . . . . . . 83
6.2.5 Fixed Poly-Si Gate Pitch . . . . . . . . . . . . . . . . . . . . . . . . 86
6.2.6 Shallow Trench Isolation (STI) . . . . . . . . . . . . . . . . . . . . . 86
6.2.7 Source/Drain Diffusion area . . . . . . . . . . . . . . . . . . . . . . . 87
6.3 Characteristics of D2D and WID Variations . . . . . . . . . . . . . . . . . . 89
6.3.1 Statistics of D2D and WID Variations . . . . . . . . . . . . . . . . . 89
6.3.2 WID Spatial Correlation of ROs . . . . . . . . . . . . . . . . . . . . 91
6.3.3 Spatial Correlation of Gates in a RO . . . . . . . . . . . . . . . . . . 95
6.4 Inferring Process Parameters (90nm) . . . . . . . . . . . . . . . . . . . . . . 96
6.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

7 Conclusion 103
7.1 Key Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.2 Methods to Mitigate Variability . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

Bibliography 108

A Mathematics 117
A.1 Integrator equation derivations. . . . . . . . . . . . . . . . . . . . . . . . . . 117
A.2 Least squares estimate, error estimate. . . . . . . . . . . . . . . . . . . . . . 119
 iv

B Layouts 121
B.1 Arrangements of the Layouts in a Tile. . . . . . . . . . . . . . . . . . . . . . 121
B.2 Library cells in 90nm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
B.2.1 90nm Library Setup. . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
B.2.2 90nm Library. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
B.3 Library cells in 45nm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
B.3.1 45nm Library Setup. . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
B.3.2 45nm Library. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

C Test Setup 126

C.1 GPIB-ENET/100 Guide . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
C.2 DG2020A Pattern Generator . . . . . . . . . . . . . . . . . . . . . . . . . . 130
C.2.1 Configuring the DG2020A . . . . . . . . . . . . . . . . . . . . . . . . 130
C.2.2 Programming the DG2020A . . . . . . . . . . . . . . . . . . . . . . . 131
C.3 Measurement Setup and Procedures . . . . . . . . . . . . . . . . . . . . . . 133
 v

List of Figures

1.1 Evolution of the wavelength of light used in photolithography over the years.
Presently, in the 90nm and 45nm nodes, 193nm wavelength is used. Rounding
of printed features is shown on the right. [1] . . . . . . . . . . . . . . . . . . 3
1.2 a) Present design methodology. b) Categorizing variations. . . . . . . . . . 4

2.1 Correlation of path delays . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.2 Correlation of gate delays in a critical path . . . . . . . . . . . . . . . . . . 14

3.1 Step and scan photolithography using a slit of light. . . . . . . . . . . . . . 22

3.2 Formation of poly-Si gates. . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.3 Isolated and dense lines. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.4 Lens imperfections. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.5 Effect of flare. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

4.1 Standby leakage current measurement. . . . . . . . . . . . . . . . . . . . . . 32

4.2 45nm NMOS leakage current measurement. Inb refers to drain to bulk junc-
tion leakage current and Ing refers to drain to gate leakage current. Taking
the difference between Is1 and Is0 gives Is , the subthreshold leakage current of
the selected NMOS transistor. Both NMOS and PMOS leakage currents are
measured in the 45nm test-chip and only NMOS leakage current is measured
in the 90nm test-chip. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.3 Single slope ADC for current measurement (90nm test-chip). . . . . . . . . 34
4.4 Timing diagrams of the integrator control signals and the output signals. . 35
4.5 Current measurement procedure. . . . . . . . . . . . . . . . . . . . . . . . . 35
4.6 Schematics of the folded cascode amplifier for measuring NMOS leakage cur-
rent. Gain is optimized for input bias of VDD=1V. Transistor widths are in-
dicated in μm next to the individual transistors. All gate lengths are 0.35μm
and In = 90μA. All transistors whose bulk connections are not indicated
have it connected to VDDA(PMOS) or GNDA(NMOS). . . . . . . . . . . . 41
 vi

4.7 Schematics of the folded cascode amplifier for measuring PMOS leakage cur-
rent. Gain is optimized for input bias of GND=0V. Transistor widths are in-
dicated in μm next to the individual transistors. All gate lengths are 0.35μm
and Ip = 120μA. All transistors whose bulk connections are not indicated
have it connected to VDDA(PMOS) or GNDA(NMOS). . . . . . . . . . . . 41
4.8 BODE plots of the amplifier used for NMOS ILEAK measurement in the
90nm test-chip. Top plot shows a DC gain of ≈ 60dB and bottom plot shows
a phase margin of ≈ 80◦ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.9 Schematics of the comparator for generating the start and stop signals when
the output of the integrator crosses the threshold voltages. . . . . . . . . . . 43
4.10 Simulation results for the current measurement circuit. Plots of leakage cur-
rent measurement error vs leakage current. a) shows post-layout simulation
results and b) shows pre-layout simulation results. . . . . . . . . . . . . . . 44
4.11 Repeated measurement of ILEAK to demonstrate accuracy in the 90nm pro-
cess. Plots 160 measurements of NMOS ILEAK of the same 3 layouts in the
same tile taken over 27 hours. A σ/μ of ≤ 1.3% is measured. . . . . . . . . 44
4.12 BODE plots of the amplifier for NMOS ILEAK measurement in the 45nm
test-chip. Top plot shows a DC gain of ≈ 60dB and bottom plot shows a
phase margin of ≈ 75◦ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.13 BODE plots of the amplifier for PMOS ILEAK measurement in the 45nm
test-chip. Top plot shows a DC gain of ≈ 60dB and bottom plot shows a
phase margin of ≈ 73◦ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.14 PMOS leakage current measurement for the 45nm test-chip. . . . . . . . . . 46
4.15 Repeated measurement of ILEAK to demonstrate accuracy in the 45nm pro-
cess. a) and b) plot 200 measurements of PMOS and NMOS ILEAK of the
same layout in the same tile. σ/μ of 1.4% and 11.5% are measured for PMOS
and NMOS ILEAK respectively. . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.16 RO with n stages. Address bits will enable the RO by applying VDD to the
enable signal. The output of the RO (out) goes to the input of the multiplexer. 48
4.17 RO frequency measurement. . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.18 Repeated measurement of RO frequency to demonstrate accuracy in the
90nm process. Plots 236 measurements of RO frequency of 3 layouts in
the same tile taken over 18 hours. A σ/μ of 0.15% is measured. . . . . . . . 49
4.19 Repeated measurement of RO frequency to demonstrate accuracy in the
45nm process. Plots 1000 measurements of RO frequency of the same layouts
in the same tile. A σ/μ of ≤ 0.02% is measured. . . . . . . . . . . . . . . . 50
4.20 RO with programmable number of stages using NAND gates . . . . . . . . 51

5.1 A SRAM column showing the bitline switches. . . . . . . . . . . . . . . . . 55

5.2 Bitline IV plot showing write margin indicators . . . . . . . . . . . . . . . . 55
5.3 Bitline IV plot showing read margin indicators . . . . . . . . . . . . . . . . 56
5.4 RO to measure write access time and the differential buffer used in the RO. 58
5.5 Pulse RO for characterizing cell read timing. . . . . . . . . . . . . . . . . . 60
5.6 Current starved buffers for widening the pulse width and creating delay for
the regenerated pulse. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
 vii

5.7 SRAM access time measurement . . . . . . . . . . . . . . . . . . . . . . . . 61

6.1 Die photo of the 90nm test-chip . . . . . . . . . . . . . . . . . . . . . . . . . 64

6.2 12 layout structures in 90nm . . . . . . . . . . . . . . . . . . . . . . . . . . 65
6.3 Model of spatial correlation coefficient. Plot of correlation coefficient ρ vs
spacing i. k is a model parameter. . . . . . . . . . . . . . . . . . . . . . . . 66
6.4 Arrangement of ROs of different lengths. . . . . . . . . . . . . . . . . . . . . 67
6.5 Layout of a bigtile and the 10 × 4 array of bigtile in the 90nm test-chip. . . 68
6.6 17 layout configurations in 45nm . . . . . . . . . . . . . . . . . . . . . . . . 72
6.7 Illustrates (a)horizontally vs (b)vertically placed inverters in a RO. . . . . . 73
6.8 Die photo of the 45nm test-chip . . . . . . . . . . . . . . . . . . . . . . . . . 73
6.9 Frequency distribution for single-finger configurations in the 90nm test-chip.
Vertical lines correspond to typical and fast corner simulation results. Bar
plots correspond to the WID distribution of the fastest and slowest chip. . . 77
6.10 Log(ILEAK ) distribution for single-finger configuration in the 90nm test-chip. 78
6.11 RO frequency distribution for stacked gates plotted next to its corresponding
layout in the 90nm test-chip. Vertical lines represent simulation corners from
extracted layout. Distribution plots of a) RO frequency normalized to a
constant, b) RO frequency normalized to its corresponding TT simulation
corner. In b), less dense gates show a 5% increase in frequency compared to
dense gates. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6.12 Effect of nearest neighbor poly-Si pitch on RO frequency and transistor leak-
age currents in the 45nm test-chip. Plots of RO frequency distribution
normalized to the SS corners, NMOS log(ILEAK ) distribution and PMOS
log(ILEAK ) distribution for 22 dies. . . . . . . . . . . . . . . . . . . . . . . 80
6.13 Bottom left plot shows the frequency distribution for 90◦ rotated ROs with
horizontal gates in the 90nm test-chip. Non-rotated ROs in the top left plot
show more variation. 99% confidence interval of σ/μ and σ 2 shown on the
right verify that the top plot has more variations. . . . . . . . . . . . . . . . 81
6.14 ILEAK distributions for non-rotated and rotated ROs in the 90nm test-chip. 82
6.15 Effect of placement of gates horizontally and vertically in the 45nm test-chip. 83
6.16 Effect of symmetrical layouts which are mirror images of each other in the
45nm test-chip. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.17 Effect of M1 coverage over gates in the 90nm test-chips. . . . . . . . . . . . 85
6.18 Effect of M1 coverage over gates in the 45nm test-chips. . . . . . . . . . . . 85
6.19 Effect of fixed poly-Si gate pitch in the 45nm test-chips. . . . . . . . . . . . 86
6.20 Effect of STI on RO frequency and transistor leakage currents in the 45nm
test-chips. NSTI uses gate isolation [7] instead of STI. . . . . . . . . . . . . 87
6.21 Effect of a longer S/D diffusion on RO frequency and transistor leakage cur-
rents in the 45nm test-chips. . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.22 Wafer maps of mean RO frequency and mean Log(ILEAK) in the 90nm
process. X marks a defective chip. Location of dies on only half the wafer
are known. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.23 Comparison of RO frequency D2D variations for 2 wafers and the overall
D2D variations in the 45nm process. . . . . . . . . . . . . . . . . . . . . . . 90
 viii

6.24 45nm WID statistics of RO frequency for 22 dies. The frequency is normal-
ized to the SS corner frequency. . . . . . . . . . . . . . . . . . . . . . . . . . 91
6.25 Spatial correlation coefficient (ρ) vs a) horizontal column spacing and b)
vertical row spacing, for vertical and horizontal gates in the 90nm process.
Dotted lines represent 99% confidence bounds for ρ. . . . . . . . . . . . . . 92
6.26 Surface plot of normalized data of 36 chips in the 90nm process. RO of layout
with vertically oriented gates show a 1% frequency shift between rows 4 and 5. 93
6.27 Plot of 45nm RO frequency normalized to zero mean and unit variance show-
ing mask characteristics. Small systematic variation of ≈ 0.2 × σ. . . . . . . 94
6.28 Plots of correlation coefficient (ρ) vs spacing between inverters in a RO ob-
tained from fitting measurement data to a exponential model. . . . . . . . . 95
6.29 90nm scatter plot of Log(ILEAK ) vs frequency showing a) mean Log(ILEAK )
and mean RO frequency for 36 dies, b) all Log(ILEAK ) and RO frequency
for 3 layouts on 1 die. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.30 Illustrates the estimation of the nominal value of L (Lm ) in the case where
only one parameter (L) is inferred. I0 and F0 are the measurement data.
Lm is used to obtain the gradients and to find Im and Fm which are used to
normalize the data in the least squares estimation. . . . . . . . . . . . . . . 98
6.31 De-embedded distribution of L. . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.32 Top plots: surface plots of within-die RO frequency (F ) and Log(ILEAK )
(I) of a typical die; Bottom plot: corresponding surface plots of estimated
normalized L. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

A.1 Small signal model of the integrator. . . . . . . . . . . . . . . . . . . . . . . 118

B.1 Arrangement of the 12 layouts in a tile for the 90nm test-chip. . . . . . . . 122
B.2 Arrangement of the 17 layouts in a tile for the 45nm test-chip. . . . . . . . 123

C.1 DG2020A data generator from Tektronix. . . . . . . . . . . . . . . . . . . . 127

C.2 Keithley 2400 source meter. . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
C.3 Infiniium 54855A DSO 20GSPS oscilloscope from Agilent. . . . . . . . . . . 128
C.4 TLA5202 logic analyzer from Tektronix. . . . . . . . . . . . . . . . . . . . . 129
C.5 GPIB-ENET/100 box from National Instruments. . . . . . . . . . . . . . . 129
C.6 Test setup for collecting data. . . . . . . . . . . . . . . . . . . . . . . . . . . 133
C.7 PCB board for measuring the 90nm test-chips. . . . . . . . . . . . . . . . . 134
C.8 PCB board for measuring the 45nm test-chips. . . . . . . . . . . . . . . . . 135
C.9 Test setup for collecting leakage measurement data for the 45nm test-chips. 136
 ix

List of Tables

2.1 Random and systematic sources of process variations . . . . . . . . . . . . . 20

6.1 Summary of the 45nm process. [2, 3, 4] . . . . . . . . . . . . . . . . . . . . 69

6.2 Characteristics of the layout configurations. . . . . . . . . . . . . . . . . . . 71
6.3 Dimensions of the transistors in the ILEAK arrays. . . . . . . . . . . . . . . 74
6.4 Dimensions of the transistors in the inverters of the ROs. . . . . . . . . . . 74
6.5 Spatial correlation coefficient of neighboring ROs for different gate orienta-
tion and spacing direction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.6 Measurement conditions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.7 Sensitivities used in the 90nm process parameters extraction for one chip and
one layout structure. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.8 Comparison of 45nm and 90nm technology results. . . . . . . . . . . . . . . 102
 x

Acknowledgments

I would like to acknowledge my friends and colleagues who have helped me through-

out the last six years of my life in U.C. Berkeley.

First of all, I would like to acknowledge my dear Singaporean friends. Starting

from the beginning, Engling and Ailee, whose hospitality I abused when I stayed with them

for over a month during my first semester in Berkeley. I’m glad they are now happily

installed in Silicon Valley with two happy kids. Yuenhui, Siewleng, Shaokun and Xiaofang,

for their company and help in getting me through the prelims and quals, and Voon Bin,

Ann, Bingru, Kah Cheong, Shawn, Mien Ling, Yingshi, Chuohao and Wanling for their

patience in tolerating my less than exciting company.

My colleagues in the DCDG group are the most impressive bunch of people I

have ever worked with. Both intelligent and fun loving, these people make graduate school

interesting and less stressful. I will definitely miss our lunch outings and the long lunch time

discussions that provide much entertainment as well as information. Engling, Radu, Bill,

Josh and Sean are part of the previous generation and have all left school. Farhana, Zhengya,

Renaldi, Zheng Guo, Seng-Oon, Ji-Hoon, Dusan, Kenny, Lauren, Jason and Vinayak are

now the main actors in the DCDG group. All of them have created and continue to create

a lively and fun environment in the group and made life tolerable, even during tapeouts.

Throughout the years, I have come to know the staff at BWRC and realized what great

people they are. Gary, Tom, Brian, Kevin, Brad, Brenda and Susan are all experts in their

own areas and work hard to keep BWRC a center of excellence for research and an ideal

oasis for graduate students to work and hang out.

 xi

Finally, the friends I made at IBM Austin Research Labs during my three summer

internships played a very important part in my Ph.D endeavour. They were instrumental

both in developing my research as well as helping me find a research position in IBM.

J.B. Kuang is the nicest man I have ever met. His intelligence, wealth of experience and

knowledge and his willingness to go out of his way to help made my stay in Austin very

pleasant and accounts for my repeated internship there. ARL staff Fadi, Tuyet, Jay and

others, other interns I met, Jerry, John, Vicky and Praveen made Austin, Texas feel fun

and enjoyable. In addition, Kevin Nowka is unanimously voted as the best manager by the

people in ARL and I agree totally.

I am sure I forgot a few friends or colleagues in this acknowledgment. Do send

me an email to remind me so that I can save some time looking for your name. I recognize

faces better than remember names and rest assured I have you in my head.

Thank you all for being part of this.

 1

Chapter 1

Introduction

Scaling of CMOS technology into the sub-100nm range has introduced new design

challenges in the form of increased variability. Studies have predicted the end of CMOS

scaling if this trend continues and if the present design methodology remains unchanged.

This thesis describes a technique to characterize and quantify systematic, random, die-to-

die (D2D) and within-die (WID) transistor variability in order to manage it both from the

manufacturing side and the design side. It demonstrates the viability of this methodology

through the design, measurement and analysis of two test-chips in early 90nm and 45nm

processes [5, 6, 7].

1.1 Motivation

CMOS device scaling has increased the impact of process variability to the point

where it is now regarded as a major roadblock to further scaling [8]. The control of process

fluctuations has not kept pace with rapidly shrinking device dimensions [9]. Furthermore,
 2

the drive to improve performance has enticed device and circuit designers to operate at

conditions which are more sensitive to variability.

Transistor oxide thickness (Tox ), threshold voltage (Vth ) and gate length (L) vari-

ations have become more significant as the manufacturing process enters sub-100nm tech-

nology nodes. Currently, the technology is enhancing the capabilities of photolithography

in volume production with sophisticated resolution enhancement techniques (RET) and in-

creasing constraints in the form of restrictive layout design rules to continue printing ever

smaller sub-wavelength features. The evolution of the wavelength of light used in pho-

tolithography over the years and the difficulty in printing sub-wavelength features is shown

in Figure 1.1. Optical proximity corrections (OPC) and phase shift masks are being used

in the sub-100nm technology nodes and immersion lithography will enter production at the

45nm node. Double patterning will soon be necessary at 32nm node and beyond. Even with

shorter gate-lengths, the difficulty in scaling oxide thickness and supply voltage results in

transistor performance that does not scale sufficiently, leading to the use of strained silicon

and high-K dielectrics in order to continue the exponential growth in transistor perfor-

mance. These techniques have, however, contributed to an already complex manufacturing

process and compounded the sources of process variability.

Presently, the foundry measures a number of test-structures on the wafers and fit

the I-V data into a model such as the BSIM SPICE model. Variability is captured in the

statistics of the model parameters. This information is then used to generate the process

corners whereby certain parameters are varied by a number of standard deviations from

their nominal values. Monte Carlo simulations can also be performed using this statistical
 3

450

400 Min Feature Size

365nm
350 Wavelength O
350nm
300

nm 250 O 248nm
R k1 250nm
200 NA 180nm
193nm
R Rayleigh's constant
150 130nm
O wavelength of light 90nm
100
NA numerical aperture  0.85 65nm
50
k1 0.3 - 0.4 for high end process
0
1990 1995 2000 2005

Figure 1.1: Evolution of the wavelength of light used in photolithography over the years.
Presently, in the 90nm and 45nm nodes, 193nm wavelength is used. Rounding of printed
features is shown on the right. [1]

information. In a typical VLSI design process, satisfying design corners is often necessary

and sufficient to validate a design (Figure 1.2a). This approach regards all variations as D2D,

with all devices on a chip having correlated process parameters. Deep sub-micron scaling

has compounded the impact of variability and increased the amount of design margin to

cope with worst-case scenarios. Since the tolerances on the process parameters do not track

the scaling of their nominal values, designers are experiencing a proliferation of various

design corners. Characterizing variability in a more detailed way (Figure 1.2b) would allow

designers to reduce systematic variations and use the right amount of margins to obtain an

optimal design that maximizes performance, power and yield.

1.2 Research Goal

The goal of this research is to help circuit designers reduce the impact of variations

on circuit performance. This can be achieved in the following ways:

 4

Foundry Circuit Designer Corners Chip

corners simulation
33ı Total
Total Within Die (WID)
Worst
WorstScenario
Scenario
Variations
Variations Correlated
Correlated

a) Present design methodology

Variations
Variations

Die
DietotoDie
Die(D2D)
(D2D) WID
WID Spatial
SpatialCorrelation
Correlation

b) Circuit designers can manage variations with this information

Figure 1.2: a) Present design methodology. b) Categorizing variations.

• Reduce the total amount of variations.

This could be done by removing as much systematic variations as possible, selecting

the right design parameters that minimize variations, and using circuit techniques to

reduce variations. This research focuses on the impact of systematic layout-dependent

variations and the design of circuits to measure variations.

• Design for the right amount of variations.

This involves quantifying D2D, WID, systematic, random variations and quantifying

the spatial correlation characteristics of variations and applying the correct amount

of variations and correlation to the process parameters used in simulations.

• Exploiting the characteristics of spatial correlation.

By using the fact that randomness can be reduced by averaging, and that correlated

paths will be more likely to meet timing/power constraints, the placement of gates

and paths can be maneuvered to minimize the impact of variations.

 5

In order to do the above, a methodology has been developed whereby test chips

to characterize the manufacturing process have been designed and fabricated. These chips

contain arrays of closely spaced test-structures such as ring-oscillators, transistors in the off

state and SRAM arrays. Individually addressable array of test-structures and circuit tech-

niques to provide high resolution measurements were developed for characterizing transistor

variations whereas non-invasive, high resolution techniques were developed for measuring

variability in the individual SRAM cells of memory arrays. Measurement results provide

quantitative information on layout dependent systematic effects, D2D variations, WID vari-

ations and spatial correlation. Finally, analysis of these results will lead to a set of guidelines

for circuit designers to reduce the impact of variations.

1.3 Related Work on Process Characterization

1.3.1 Metrology Techniques

Electrical metrology techniques are used to collect large amounts of data on ma-

terial properties and process information. These techniques measure current, voltage or

charge and extract from them the physical dimensions of the shapes or the properties of

the material. The quality of lines produced through photolithography has been evaluated

in this way. ELectrical line-width metrology (ELM) measures the resistance of polysilicon

(poly-Si) lines and uses that as a measure of line widths [10, 11, 12, 13]. Orshansky et al

[12] used this method to investigate the effects of proximity and rotation on the printed

gatelength for different configurations of poly-Si lines distributed over the reticle field and

found layout-dependent behavior and systematic variation within the reticle field. Resis-
 6

tance measurements can also be used to find the placement of the feature, metal step cover-

age over topography, wire edge-taper width and layer to layer alignment [14]. Capacitance

measurement can be used to find wire to wire spacing and dielectric thickness.

Non-electrical metrology techniques include optical interferometry, profilometry,

atomic force microscopy (AFM), and scanning electron microscopy (SEM) metrology. Op-

tical interferometry is often used to measure film thickness, and profilometry is used to

measure the height profile of a surface. High resolution SEM and AFM images can also be

used to measure feature sizes such as line-widths. The advantage is that there is no need

for special pads or circuitry for measurement, but the disadvantages are that throughput is

low and the process is electrically destructive to the device.

These measurements are accurate but may not correlate strongly with the param-

eter that determines the electrical characteristics of the circuit. For example, measurement

of resistance of poly-Si lines reflect the resolution and image quality of the regularly spaced

poly-Si lines printed with lithography but does not directly measure the effective gatelength

that affects transistor performance.

1.3.2 Measurement of transistor performance

Increased variability has prompted the development of more accurate process char-

acterization techniques. In order to obtain accurate statistical models, a large number of

transistors, in a wide variety of layout configurations with different combinations of tran-

sistor dimensions have to be measured. Several work have been published which measure

a large array of transistor accurately in a short time. Prior to this work, Boning and his

students [15, 16, 17] measured the frequency of ring oscillators (RO) specially laid out to
 7

investigate the effects of layout on gates and interconnects. The RO array is made of small

tiles that can be placed together to form a large array. The RO frequency is routed through

multiplexers to one output where it is divided down and measured off-chip. Scan chains

provide address bits to enable a single RO in the array and generate control signals for the

multiplexers. In this way, the measurement process is simple, fast and accurate.

Other work involves inferring process parameters through measurement of transis-

tor performance characteristics. Ouyang et al [18] measured transistor saturation current

(IDSAT ) for an addressable array of transistors and inferred the transistor gate length L.

Ohkawa et al [19] used a device matrix array which contains an array of transistors, capaci-

tors, resistors and ring oscillators to evaluate WID variation in device parameters. Agarwal

et al [20, 21] measured gate leakage and Vth in an addressable array of devices in IBM’s 65nm

SOI technology. Levacq et al [22] constructed a long array to measure the spatial frequency

of intra-die variations using current measurements. A wide range of spatial frequency can

be measured using the long array with closely spaced devices.

The traditional test structures for characterizing MOSFETs consist of measuring

I-V data for different gate dimensions. Hence the most accurate way to statistically charac-

terize the MOSFET is to collect I-V data for a large array of devices. This method requires

huge amount of measurement and achieving efficiency is the main challenge. Wang et al

[23] has designed a circuit that allows for rapid measurement of I-V data for an array of

densely placed devices. This design consists of a dual-slope ADC and uses current mirrors

to keep the integrated current at the same range. It achieves very good accuracy and a

short measurement time.

 8

Using the test structure in [20], Zhao et al [24] demonstrated that only 3 points

in the I-V curve are needed to extract the process parameters. Wang et al [25] attempts to

bridge the gap between statistical metrology and circuit design by modeling random process

variability using current distribution measured at saturation and linear region.

1.3.3 Other circuit techniques for process characterization

Other interesting circuit techniques to characterize single inverter delays and SRAM

performance have been published. To measure inverter delays, programmable ring-oscillators

which allow for the selection of two ROs with overlapping inverter stages have been designed.

By taking the difference between the periods of oscillation of the two ROs, the delay of the

non-overlapped inverters can be extracted [26, 27]. Kim et al [28] uses a fixed RO to sample

another RO that has been subjected to negative bias temperature instability (NBTI). By

measuring the resulting beat frequency, the shift in frequency due to NBTI can be measured

very accurately.

For SRAM measurement, besides measuring static currents from the bit lines as

described in [29], work on connecting the SRAM columns together to form a RO whose

frequency gives a measure of the performance of the SRAM array has also been shown in

[30].

1.4 Dissertation Organization

Chapter 2 describes the characteristics of variability and how variations impact

circuit design. It explains how spatial correlation affects performance and gives an overview
 9

of the sources of process variations. Chapter 3 describes the sources of variations in the

manufacturing process. It starts with an overview of the lithography process and its re-

lated sources of variations and goes on to cover variations from other manufacturing steps.

Chapter 4 describes the circuits developed for measuring variability. These circuits were

implemented in a 90nm and a 45nm test-chip that are described in chapter 6. Chapter 5

describes circuits for measuring SRAM performance. These circuits have not been imple-

mented in test-chips and are part of the subject of future work. Chapter 6 presents the

measurement and analysis results. In the recent 45nm process, restricted design rules and

more complex OPC have been introduced in order to mitigate systematic poly-Si density

and other layout-induced effects. The effectiveness of these measures will be evaluated.

Finally, Chapter 7 concludes with a a set of guidelines for mitigating variability and a brief

summary of our research accomplishments.

1.5 Acknowledgements

I wish to acknowledge the contributions of the students, faculty and sponsors of the

Berkeley Wireless Research Center, the National Science Foundation Infrastructure Grant

No. 0403427, wafer fabrication donation of STMicroelectronics, and the support of the

Center for Circuit & System Solutions (C2S2) Focus Center, one of five research centers

funded under the Focus Center Research Program, a Semiconductor Research Corporation

program. I would also like to thank Ernesto Perea and the engineers at STMicroelectronics,

Crolles for their support, Profs Andrew Neureuther, Tsu-Jae King-Liu and Costas Spanos

and their students for helpful discussions, and Christopher Siow and Emmanuel Adeagbo
 10

for their help with the measurements. Profs Jan Rabaey and Robert Leachman for agreeing

to be in my qualifying examination and dissertation committees. Most importantly, I would

like to thank my advisor Prof Borivoje Nikolić for his guidance and support.
 11

Chapter 2

Characterization of Variations

Variability has to be measured and characterized in order to reduce its magnitude

and impact on circuit performance. Categorizing variability into D2D and WID will allow

for circuit designers to design for WID variations and use other techniques to reduce D2D

variations. Categorizing variability into systematic and random will allow us to identify

significant sources of variation. Systematic variation can then be reduced through tuning of

the process, introduction of new process techniques or new design rules. Understanding the

magnitude of the different sources of variations will allow process engineers to focus their

resources on targeting the more important ones. Systematic layout-induced variations can

also be reduced with the use of restricted design rules.

This chapter describes the characteristics of variability and gives an overview of the

sources of transistor process parameter variations. Section 2.1 describes the characteristics

of variability and how it is classified. Section 2.2 shows how spatial correlation affects the

yield of a chip, and section 2.3 gives an overview of the sources of random and systematic
 12

process parameter variability. Finally section 2.4 gives a summary of the characteristics of

variations.

2.1 Characteristics of Variations

Variations in CMOS performance can be dynamic or static [31]. Dynamic vari-

ations such as electromigration, hot-electron effect, negative bias temperature instability

(NBTI), self heating/temperature, SOI history effect, voltage drops across wire resistance,

capacitive coupling or residual source/drain channel charge are time varying and change

with the operations. Static variations are non time-dependent and are due mainly to the

manufacturing process. Other causes of static variations such as voltage drop due to static

current across a resistance can be accounted for in the design. In this work, the character-

istics of static process variations will be studied.

Process variations can be systematic or random, and are generally characterized

as within-die (WID), die-to-die (D2D) and wafer-to-wafer (W2W) [32]. WID and D2D

classifications reflect some of the spatial characteristics of the variations. Those which vary

rapidly over distances smaller than the dimension of a die result in WID variations whereas

variations that change gradually over the wafer will cause D2D variations. D2D variations

can be due to intra-field variations or field to field variations over the same wafer and also

include other inter-die variations such as W2W variations. W2W variations reflect both the

spatial as well as temporal characteristics of the process and cause different wafers to have

different properties. In a typical design methodology, designs are made to satisfy the worst

case corners which consist of the total WID and D2D variations.
 13

Systematic variations are deterministic shifts in process parameters, whereas ran-

dom variations change the performance of any individual instance in the design in an un-

predictable manner. In practice, although many of the systematic variations have a deter-

ministic source, they are not known at the design time, or are too complex to model and

are thus treated as random. The resulting random variation component will have a varying

degree of spatial correlation.

2.2 Impact of Variability on Circuits

Variability affects the yield of the chips. Yield is defined as the probability that

a chip will meet the constraints of timing and power. A circuit with more design margin

will have a higher yield. The challenge is in finding the smallest margin necessary to give

the required yield so that performance is not overly constraint. In order to model the

statistics of the circuit performance accurately both the amount of parameter variation and

the spatial correlation of the these parameters between different gates have to be known.

Characterizing the amount of variation involves making measurements of many devices

and obtaining the standard deviation. This has traditionally been done in order to obtain

corner information. However, spatial correlation has not been captured and incorporated

into existing design flows.

By characterizing spatial correlation, excessive margins can be avoided and the

effects of variations can be mitigated. Figure 2.1 illustrates how the correlation of path

delays affects the yield. In a circuit with P parallel critical paths, if all the path delays

are correlated, the probability of all the paths satisfying a timing constraint is equal to
 14

Yield = Prob(Max delay of P paths < clock period) Mean delay increases as P
increases for uncorrelated paths
Correlated paths give highest yield

P=1
P=2

P uncorrelated paths

Normalized PDF
P=10000
a1 b1 c1

aP bP cP
0 0.8 0.9 1 1.1 1.2
Normalized Critical Path Delay
Max Delay of P paths

Figure 2.1: Correlation of path delays

Variation remains constant

with correlated gates,
Yield = Prob(Sum of N gate delays < clock period)
Uncorrelated gates give highest yield through 20%
increased averaging
̒/mean of total delay 15% Variation is reduced with
N stages
non-correlated gates,
10%

d1 d2 dN
5% Monte carlo
1/¥N
0%
0 2 4 6 8 10
Number of stages (N)

Figure 2.2: Correlation of gate delays in a critical path

the probability of one path meeting the constraint. On the other hand, if the paths are

uncorrelated, the yield will be the product of the probabilities of the individual paths

meeting the timing constraint, making it much lower and dependent on the number of

paths. Figure 2.2 illustrates how the correlation of gate delays in a single path affects its

yield. If the gate delays in a critical path are all uncorrelated (ρ = 0), the total path delay

will have less variation due to averaging. Random variations can also be suppressed using

longer logic paths and larger gate area transistors [33].

 15

2.3 Sources of Process Variations

The primary sources of variability are the transistors themselves, interconnects,

and the operating environment (supply and temperature) [31]. Many sources of systematic

variability can be attributed to the different steps of the manufacturing process. The pho-

tolithography and etching processes contribute significantly to variations in nominal lengths

and widths due to the complexity required to fabricate lines that are much narrower than

the wavelength of light used to print them [1]. Variation in film thicknesses (e.g., oxide

thickness, gate stacks, wire and dielectric layer height) is due to the deposition and growth

process, as well as the chemical-mechanical planarization (CMP) step. Additional electrical

properties of CMOS devices are affected by variations in the dosage of implants, as well

as the temperature of annealing steps. In recent technologies, overlap error, mask error,

shift in wafer scan speed, rapid thermal anneal and the dependence of stress on layout have

become notable sources of systematic variations. Systematic variations will be discussed in

detail in Chapter 3.

The most important reason for increased random device parameter fluctuations is

that CMOS technology is scaling into atomic-scale dimensions [34]. The size of an atom

is in the order of ≈ 1 Å or 0.1nm. In this regime, small variation of discrete numbers

is significant when the total number is small. Smooth, continuous and distinct interfaces

become granular and pebbled with atoms. Quantum mechanical properties of these discrete

entities come into play and the classical models become increasingly inaccurate.

Random device parameter fluctuations stem mainly from Si/SiO2 and poly-Si/SiO2

interface roughness, line-edge roughness (LER) and doping fluctuations. Other effects in-
 16

clude random strain distribution and non uniform temperature during annealing.

2.3.1 Interface roughness

The dielectric that separates the transistor gate from the substrate strongly affects

the electrical properties of the device. The thickness of this layer (Tox ) and its dielectric

constant are used to model the device and a thinner layer with higher dielectric constant

gives better transistor performance. SiO2 has been the gate dielectric of choice for CMOS

technology due to its excellent interface properties and manufacturability. Scaling of Tox

has reached its limit of ≈ 2nm or 10 atomic layers of SiO2 [35]. This limit is dictated by the

exponential increase in gate leakage current due to quantum mechanical tunneling across

the thin layer of dielectric. Tox variation is around 1-2 atomic layers causing significant

gate leakage variation. The impact of Tox variation has been analyzed using 3D numeri-

cal simulation and incorporating quantum mechanical effects [36]. Results show that the

impact of Tox variation on Vth is small and strongly dependent on the correlation length.

A pessimistic estimation of its effects for longer correlation length generates a variation of

σVth ≈ 35mV.

2.3.2 Line-edge roughness

Line-edge roughness is due to statistical variation in the incident photon count

during exposure, contrast of the aerial image, absorption rate, chemical reactivity and

molecular composition of the resist [37, 31]. Atomic-scale behavior of the manufacturing

process creates missing chunks of atoms from the surface of the gate along the width giving

rise to the edge roughness. The impact of LER is stronger than that of interface roughness.
 17

Subthreshold leakage current (ILEAK ) is strongly affected due to its exponential dependence

on gate length [38]. LER becomes significant below the 50nm node and will severely affect

performance at the 32nm node if gate-length control does not improve [39].

2.3.3 Random Dopant Fluctuation

Dopant atoms are injected into the channel region through ion implantation. This

is followed by an annealing step to activate the dopant atoms and repair the damaged

surface. The volume of channel region and the total number of dopant atoms are both

so small that the number and position of dopant atoms in the channel of the transistor is

effectively random. The impact of random dopant fluctuation (RDF) is exhibited through

a large Vth variation. [40, 41, 42, 43, 44, 45, 46, 47]. It accounts for most of the variations

observed in SRAM arrays and in analog circuits where systematic variation is small and

random uncorrelated variation can cause mismatch that results in reduced noise margins.

Studies have shown that retrograde body doping has the least amount of Vth fluctuation

while uniformly doped case is worst.

2.3.4 Subthreshold Voltage Variation

The amount of Vth variation is reflected in σVth which varies with √1 [48] where
W ·L

W is the width of the gate. σVth due to RDF and thin film interface roughness also varies

with √1 [42, 36]. Other effects such as non-uniform temperature distribution during
W ·L

annealing and random distribution of biaxial strain also cause Vth variation. These are

described in Sections 2.3.5 and 3.2.

 18

2.3.5 Strain

The impact of strain on the electrical resistance of silicon was discovered soon after

Si was widely accepted as the material of choice for solid state devices [49]. Strain alters

the band structure of Si, causing changes to properties such as bandgap, effective mass,

mobility (μ), diffusivity of dopants, and oxidation rates [50, 51, 52].

The impact of strain on transistors depends on the crystal orientation (<100>or

<110>direction), the type of strain (compressive or tensile), the direction of strain (one

direction [53] (uniaxial) along the channel or orthogonal to the channel or in both directions

(biaxial)), and the type of device (NMOS or PMOS). In the <110>channel direction tensile

strain increases NMOS mobility by lowering the effective mass and the scattering rate of

electrons. Compressive strain increases PMOS mobility by lowering the effective mass of

holes [54, 53, 55]. In the <100>direction, PMOS mobility is insensitive to strain [56].

Threshold voltage (Vth ) is affected by biaxial strain and less so by uniaxial strain [57].

Experimental studies have shown improvements in drive current of up to 60% due to strain

[58, 59, 60, 61, 62].

Starting from the 90nm node, devices are intentionally strained to improve perfor-

mance [63, 64, 65]. Shallow trench isolation(STI) using SiO2 creates compressive strain on

the substrate that varies with the distance from the edge of the STI/diffusion interface to

the channel region. This effect is systematic and has been modeled in recent BSIM models.

Other standard process-induced strain [50] include the use of silicide layers, and the contact

etch stop layer (CESL). Silicide layers deposited onto the active and poly-Si region of the

transistor introduce a compressive strain on the Si channel that is sandwiched between the
 19

silicides due to differences in thermal expansion coefficient. Nitride films are deposited on

Si to act as the CESL. These layers can also be used to generate large stresses which can be

tensile or compressive. The amount of stress increases with the thickness of the nitride layer

that can fill between the gates and the amount of contact of the CESL with the source/drain

region. The latter depends on poly-Si pitches and spacer sizes. As these dimensions get

smaller with scaling, the effectiveness of this method is reduced. Additional process steps

have also been introduced to generate strain. A stress memorization technique consists of

depositing a stressed film onto the device, annealing it at high temperature and stripping

the stressed film. Re-crystallization of the poly-Si at high temperature under stress leads

to the generation of residual stress [66]. Finally, strain is generated when SiGe is grown

on Si because both materials have slightly different lattice constants. These epitaxial films

generate global strain (when used as the substrate for the channel region), or local strain

(when used only in the source/drain diffusion region of the transistor). Similarly to STI,

the amount of strain depends on the length of source/drain diffusion.

The manufacturing process subjects the Si wafers to various temperatures and

chemical, mechanical interactions. Different materials are grown or deposited on the wafer

to create devices and interconnects. Strain is created on the substrate due to differences

in thermal expansion coefficients or lattice constants of the different materials. This causes

the electrical properties of devices to vary randomly [67]. Simulations show that the use of

strain to improve performance also exacerbates the impacts of random process variability

[68, 69].
 20

Parameter Random Systematic

Gate Length L Line edge roughness (LER) Lithography and etching:
proximity effects, orientation.
[10]

[38]
Gate Oxide Thickness Si/SiO2 & SiO2 /Poly-Si Non uniformity in the process
Tox interface roughness [36] of oxide growth.

Channel Dopant Affects σVth [43] Non uniformity in the pro-

Concentration Nch cess of dopant implantation,
dosage, diffusion.

Threshold Voltage Vth Random anneal temperature Non-uniform annealing tem-

(non Nch related) and strain effects perature [70] (metal coverage
over gate). Biaxial strain.
Mobility μ Random strain distributions Systematic variation of strain
in the Si due to STI, S/D area,
contacts, gate density, etc.

Table 2.1: Random and systematic sources of process variations

2.4 Summary

The sources of transistor process parameter variations can be classified as random

or systematic, D2D or WID. This classification together with information on the spatial

correlation of these parameters can help designers manage variability in circuit performance.

Random variations are mostly due to physical limits whereas systematic variations are due

to the manufacturing process. A summary of the sources of random and systematic process

variations is shown in Table 2.1.

 21

Chapter 3

Process-induced Systematic

Variations

Systematic sources of variations have been dominant in sub-100nm technologies.

In many cases they can be reduced or accounted for in the design process. This chapter

describes important sources of variation in the manufacturing process. Section 3.1 describes

the sources of variations in a standard exposure system. This includes the mechanical

system, the source of light, the optics and the mask. Section 3.2 describes other process

steps involved in patterning of the transistor gate and formation of the gate channel. Section

3.3 gives a list of effects that are dependent on the layout of the circuits and finally, section

3.4 gives a summary of this chapter.

 22

light source
Slit of light

Mask moved to the right

mask

optics

Wafer moved to the left

wafer

Figure 3.1: Step and scan photolithography using a slit of light.

3.1 Variations in the Exposure System

Present lithography systems employ a step-and-scan method, where the stepping

is used to move the wafer between exposure fields. Within an exposure field, a narrow slit

of light illuminates the mask and projects the mask pattern optically onto a wafer [71]. The

mask and the wafer are moved simultaneously in opposite directions such that the slit of

light scans the entire mask and projects the image onto the wafer. This is illustrated in

Figure 3.1.

Modified illumination [72], refers to the use of partially coherent light sources

and off-axis illumination schemes such as dipole, quadrupole and annular illumination.

This technique improves resolution but also causes line-widths to vary according to their
 23

orientation [73] and exacerbates the effects of lens aberrations [74]. Non uniformity and

asymmetry of the light source can cause systematic variations in the reticle field. Movement

of the wafer stage is controlled by look-ahead sensors. Changes in speeds during the scan

may cause systematic variations in effective channel lengths over the reticle. Mask overlay

error due to misalignment of the different masks can cause shifts in the position of poly-Si

gate with respect to diffusion and source-drain contacts.

In sub-wavelength lithography, the effective line width depends on the surrounding

features [72, 10, 12]. The process step of fabricating polysilicon (poly-Si) gates is shown in

Figure 3.2. When exposed beyond a certain light intensity, positive resist will dissolve in

the developer fluid. The exposed poly-Si gate stack layer will then be etched away leaving

behind the transistor gates. Narrow poly-Si lines with varying pitch will have different

channel lengths when exposed with the 193nm wavelength light, as illustrated in Figure

3.3. Dense lines also have higher depth of focus, and are more immune to defocusing of the

optical system [75]. Optical proximity correction techniques in the mask processing add

sub-lithographic assist features to control the printed critical dimensions (CD). However,

their effect is limited due to the shallow depth of focus.

Lens imperfections are often characterized through aberrations, Figure 3.4. Aber-

rations create optical path differences for each pair of rays through the imaging system.

Effects such as astigmatism and spherical aberrations cause differences in exposed patterns

at the level of a reticle. Coma effect [74] is an aberration due to lens imperfection, which

causes a gate surrounded by non-symmetrical structure to print differently from its mirror

image [76].
 24

Light
Mask Developer fluid dissolves
the exposed resist
Gate length
Positive
resist
SiN etch etch
Poly_Si
SiO2
Si

Figure 3.2: Formation of poly-Si gates.

Isolated

Dense Masks

Resist
Ldense exposure
threshold
Liso

Figure 3.3: Isolated and dense lines.

Source

Condenser Reticle Aberration Projection Wafer

lens plate lens

Figure 3.4: Lens imperfections.

 25

Surface
scattering Intensity
CD Resist
exposure
threshold

With flare

Reflections No flare
Lens

Figure 3.5: Effect of flare.

Flare results from the scattering and reflection of light through the projection

system and causes variations in the effective CDs, Figure 3.5. In general, the amount of flare

is dependent on the local pattern density in the mask [74]. Proximity effects, aberrations,

and flare are usually not captured in the design process, and they induce layout-dependent

systematic variations in the design.

Another source of variability comes from e-beam mask stitching [77] discontinuity.

E-beam lithography is employed in writing the optical mask. In order to cover the large

reticle field of the mask with a small e-beam field, it is necessary to construct the mask

by drawing smaller e-beam fields and stitching the whole image together. Discontinuity at

the boundaries between the smaller e-beam fields will result in stitching errors in the mask.

Even though optics will reduce imperfections on the mask by a few times, aggressive scaling

has made stitching discontinuities more significant.

 26

3.2 Variations in Patterning and Other Manufacturing Steps

After the resist is spun onto the wafer and exposed, it undergoes post-exposure

bake (PEB). This step is essential to activate the photo-active compound and set the resist

exposure threshold [78]. If the temperature over the entire wafer is not even, this will result

in different resist exposure thresholds over the wafer and cause D2D gate-length variation.

This effect is mostly systematic, and wafers have exhibited a radial temperature profile

during baking.

Resist development can also be a source of systematic wafer scale variation. The

application of developer fluid, if not uniform over the wafer surface, can result in some parts

of the wafer being more developed than others. This will also cause systematic gate-length

variation.

Dry etching is employed in patterning of the poly-Si gates. This process suffers

from microscopic loading effects [71] in which densely spaced poly-Si gates experience a lower

etch rate than isolated poly-Si, resulting in dense poly-Si gates having longer gate-lengths

than isolated gates.

Rapid thermal annealing (RTA) is used in activating implanted dopants and an-

nealing defects without causing significant dopant diffusion [71]. Temperature distribution

over the wafer surface could be non-uniform resulting in wafer-level systematic variations.

For example, the edge of the wafer could experience more cooling and hence a lower tem-

perature, resulting in a radial temperature gradient across the wafer. This effect could also

cause wafer warping that results in variation in substrate stress. Finally, local variation in

reflectivity of the wafer surface could cause local non-uniform heating, resulting in variation
 27

in Vth and in external resistance (REXT )[79]. The later includes mainly source-drain and

contact resistances. It has also been shown in [70] that differences in anneal temperature

can cause shifts in Vth due to non-passivation of interface states.

Shadowing during pocket implantation can result in different properties in the

source and drain of transistors. These effects, like the coma effect, could cause symmetrical

structures to have non-symmetrical properties.

Finally, well proximity effects (WPE) in which transistor performance depends on

its distance from the edge of the well in which it sits, has been observed to cause significant

variations [80].

3.3 Layout Dependent Variations

Layout-dependent effects that will be studied are shown below:

• Proximity effects

Proximity effects refer to the dependence of L on transistor gate density or gate pitch.

In the previous sections, proximity effects due to optical and etch loading effects have

been described. These effects are also present in the mask fabrication since the same

type of process steps are employed.

• Direction

The step and scan lithography creates different properties in the two orthogonal direc-

tions of slit and scan. In the slit direction, variations are due to lens aberrations and

result in more correlated features. In the scan direction, variations are due to dosage
 28

and scan speed which are usually better controlled. Hence there is less systematic

variation and features are less correlated. However, sudden discontinuities in scan

speed can occur, resulting in the formation of lines along the slit direction, indicating

the locations where there is a sudden change in scan speed. Mask stitching errors also

contribute to systematic variations that are dependent on the x-y directions.

• Metal coverage over gate

RTA can cause local non-uniformity in annealing temperature, resulting in changes

in Vth with metal coverage over gate as described in section 3.2. Existing layout

extraction tools only capture the capacitance of metal layers over the transistor gate.

Shifts in Vth will be observed through measurement of the transistor ILEAK .

• Strain effects

In 45nm, the use of strained Si introduces mobility variations due to non-uniform

stress in the Si substrate. The stress induced can be layout dependent. Sources of

stress come from the shallow trench isolation (STI), a capping layer, silicide layers,

stress memorization techniques and SiGe epitaxial films and they are affected by the

layout such as gate pitch, contact separation, source/drain area, etc [81, 82, 83]. In the

case of SiGe films in the source/drain diffusion region, a shorter source/drain length

creates less stress. In the case of CESL nitride capping layer, the amount of strain

depends on the thickness of CESL layer and the area of source/drain region covered

with CESL which varies with the poly to poly pitch and the size of the gate spacers.

A smaller contact area between the CESL and the source/drain region results in less

strain on the transistors. The amount of stress starts to roll off quickly when the
 29

sidewalls of the CESL starts touching each other. This occurs for thicker CESL layer.

In the design of analog circuits whereby matching of transistors and passives are

very important, techniques for improving matching have been developed. These include the

use of fixed gate pitches, dummy gates, common centroid layout, uniform distribution of

metal over gate and larger design rule margins. Digital circuits have been slow to adopt

them as the problem of matching was small and the area penalty was important. However,

as variability begins to dominate the design, layouts are beginning to show more regularity.

Work on design for manufacturability (DFM) involves setting design rules to ensure that

variations in transistor process parameters are small. This has resulted in the implemen-

tation of fixed gate pitches, non-rotation of gates and regular layouts. In this work, the

impact of layout on transistor performance in the presence of design rules can be evaluated

by measuring an array of test circuits. These circuits, their implementation on test-chips

and the measurement results and analysis of those chips will be described in Chapters 4, 5

and 6.

3.4 Summary

In summary, many sources of systematic variations are derived from the different

steps in the manufacturing process as described in this chapter. Differences in transistor

layouts and its proximity also contribute to systematic variation in performance.

 30

Chapter 4

Measuring CMOS Performance

Variability

Circuits that measure CMOS performance rapidly and accurately were designed to

collect statistical data for a large array of devices. This chapter describes circuits that have

been developed to make this type of measurements. It focuses on investigating the properties

of transistors and the performance of CMOS logic circuits. Section 4.1 describes the circuits

used to measure the subthreshold leakage current (ILEAK ) of an array of transistors. Section

4.2 describes the circuits for measuring the frequencies of an array of ring-oscillators (ROs)

and for characterizing the spatial correlation of neighboring gate delays. Finally section 4.3

summarizes this chapter.

 31

4.1 Leakage Current

Since ILEAK varies exponentially with Vth , measuring ILEAK gives a good idea of

Vth variation. ILEAK measurement of an array of transistors has been implemented in two

test-chips in a 90nm and a 45nm technology.

4.1.1 Leakage Transistor Array

Figure 4.1 shows how the leakage current of a NMOS device in an array is mea-

sured. In the figure, the gates of the NMOS devices are connected to ground. The selection

of the device to be measured is done with row and column bits obtained from two shift

registers in order to reduce pad count. These address bits will supply either VDD or GND

to the source of the transistor using a large inverter. The selected NMOS will have GND

applied to its source and the other NMOS transistors will have VDD applied to their source.

This will enable a subthreshold current to flow from the drain terminal to the source of the

selected NMOS. Parasitic currents from the drain of the NMOS that are not selected will

be measured and subtracted from the final measurement. Similarly for PMOS leakage, the

address bits will apply VDD to the source of the selected PMOS transistor and GND to

the source of the other non-selected PMOS transistors. The current will then be measured

using an on-chip single-slope analog-to-digital converter (ADC) that will be described in

the next section.

Figure 4.2 shows how the parasitic drain-to-bulk junction leakage (Inb ) and drain-

to-gate leakage (Ing ) are removed from the measurements. Only NMOS leakage measure-

ment is shown in the figure. The gates of the NMOS transistors are tied to GND and there
 32

ILEAK
Single
Slope ADC
ILEAK

Gnd Vdd Vdd Gnd Gnd Vdd

row

column0 column1 column2

Figure 4.1: Standby leakage current measurement.

are parasitic Inb and Ing for all devices whether they are selected or not. Initially, none

of the NMOS devices are selected and the parasitic current is measured as Is0 . Then one

NMOS is selected and the current is measured as Is1 . Taking the difference between Is1

and Is0 gives the subthreshold leakage current of the selected NMOS, Is . The size of the

inverter is large enough so that even when subthreshold leakage is high, the voltage drop

across the large inverter when GND is applied to the source of the selected NMOS is small.

This is important as VGS for the selected NMOS is the voltage across the large inverter.

4.1.2 Leakage Current Measurement

In the test-chips, an on-chip single-slope analog-to-digital converter (ADC) is used

to measure transistor off-currents from 1nA to 1μA with 1nA resolution. The ADC (Figure

4.3) consists of a high-gain folded-cascode amplifier implemented with high voltage, thick

oxide devices, a large on-chip metal fringe capacitor (10pF), and comparators. The am-

plifier uses large devices in order to increase gain and reduce the offset voltage. A large

capacitance is necessary in order to remove the effects of parasitic capacitances and increase

the integration time so that it can be measured more precisely. In addition, a fringe ca-
 33

Is0
Is0

Inb1 Ing1 Inb2 Ing2

Is1 = Is + Is0
Is1

Inb1 Ing1 Inb2 Ing2

Is

Figure 4.2: 45nm NMOS leakage current measurement. Inb refers to drain to bulk junction
leakage current and Ing refers to drain to gate leakage current. Taking the difference between
Is1 and Is0 gives Is , the subthreshold leakage current of the selected NMOS transistor. Both
NMOS and PMOS leakage currents are measured in the 45nm test-chip and only NMOS
leakage current is measured in the 90nm test-chip.
 34

Vdd C V2=1.8V

Leaky NMOS +
transistor P1 P1 stop
P2 2.5V -

out
Iref Vdd
Is -
P1b start
Select +
0.7V
V1=0.8V

Figure 4.3: Single slope ADC for current measurement (90nm test-chip).

pacitor was used to reduce capacitance leakage current that is more significant with MOS

capacitors.

The switches, controlled by signals P1, P2, and P1b provide the correct bias to

the input and output of the amplifier and determine whether to measure the reference

current (Iref ) only or the sum of Iref and Is . The latter is the sum of all parasitic gate and

substrate leakage currents and the ILEAK of the selected device when a device is selected.

As mentioned in the previous section, by not selecting any of the devices, it is possible to

measure the parasitic leakage currents and subtract them from the measurement. Iref is

obtained from an external source through current mirrors that divide the current down by

a factor of ten.

During current integration, the output of the op-amp will ramp up, and as it

crosses the voltages V1 and V2, the start and stop signals will be generated. The timing

diagram for the signals in the integrator is shown in Figure 4.4. Measuring the time interval

between start and stop signals will give the current (Figure 4.5).

In the integrator the relationship between the integrator output and the integrated
 35

Figure 4.4: Timing diagrams of the integrator control signals and the output signals.

out

't ref  't s Is

~
't s I ref
't ref integration time when measuring I ref
't s integration time when measuring I ref  I s

Figure 4.5: Current measurement procedure.

 36

current is given in Equation 4.1.


1 1
V0 = ·I − · I · dt (4.1)
Gm0 C

where

V0 = integrator output voltage

Gm0 = transconductance gain of the amplifier for amplifier output voltage at V0

I = integrated current

C = integrating capacitance

t = time

Integrating from V1 to V2 for constant currents Iref and Iref + Is , the Equations

4.2 are obtained.

1 Iref
ΔV = Iref · Δ − · Δtref
Gm C
1 (Iref + Is )
ΔV = (Iref + Is ) · Δ − · Δts (4.2)
Gm C

where

ΔV = V1 - V2

Δ G1m = 1
Gm1 - 1
Gm2

Iref = integrator reference current

Is = current from the array

 37

Δtref = time to integrate Iref from V1 to V2

Δts = time to integrate Iref + Is from V1 to V2

When the gain of the amplifier is sufficiently large, these two equations will give

the linear relationship shown in Figure 4.5. This is shown in Equation 4.3. By measuring

the integration times Δtref and Δts , the measured current can be obtained in terms of Iref .

Is
Δtref − Δts Iref Is
= ≈ (4.3)
Δts Δ G1 Iref
1 − (Ib + Iref ) · ΔV
m

The accuracy of the current measurement can be improved by removing uncer-

tainty in the value of Iref . Iref is generated by dividing down an external current source

(Iext ) by 10 times using a current mirror. Mismatch in the current mirror and parasitic

currents in the diffusion of the transistors in the current mirror contribute to errors in the

current measurement. In a simple measurement procedure, a current of 1uA is supplied

to the current mirror and divided down to 100nA to act as Iref . Mismatch in the current

mirror and parasitic junction and gate currents in the current mirror will change this value

as shown in Equation 4.4.

Iref = ((Iext + Ipad ) × 0.1 × M + Imirr (4.4)

where

Ipad = parasitic currents from the I/O pad

Imirr = parasitic currents at the diffusion terminal of the mirror that connects to the input
 38

of the current integrator

M = mismatch of the current mirror

The parasitic currents are fixed and can vary from die to die. Mismatch is small

since the transistors used are long and wide (M ≈ 1). Since the external current source is

very accurate, a better measurement procedure to remove the effect of parasitic currents is to

make a measurement with Iext = 1μA followed by a second measurement with Iext = 1.1μA.

Knowing the difference in mirrored current is 10nA, an accurate value of Iref can be obtained

and used in the measurement. An improved current measurement circuit was presented by

Sato et al [84] which reduces the contribution of parasitic leakage current from non-selected

transistors in the array and virtually eliminates the constraint on the number of transistor

allowed.

The following steps describes the current measurement procedure. The first two

steps describe how Iref is determined accurately and the other steps describe how parasitic

leakage currents is removed from the measurement.

1. Unselect all devices and the array and make a measurement with Iext = 1μA, obtain

Δtref . Iref contains the mirrored current and all the parasitic currents.

2. Unselect all devices and the array and make a measurement with Iext = 1.1μA, obtain

Δts0 . The external current Iext is 100nA larger in this case. After going through the

current mirror this difference is divided down by 10 to 10nA. Since the current is now
 39

10nA larger than Iref , the corresponding measured current Is0 = 10nA.

Δtref − Δts0 Δts0

Is0 = · Iref ⇔ Iref = · Is0 (4.5)
Δts0 Δtref − Δts0

3. Unselect all devices but select the array and make a measurement with Iext = 1μA,

obtain Δtparasite
Δtref − Δtparasite
Iparasite = · Iref (4.6)
Δtparasite

substitute Iref from Equation 4.5 into Equation 4.6, we obtain

Δtref − Δtparasite Δts0

Iparasite = · · Is0 (4.7)
Δtparasite Δtref − Δts0

4. Select one device and make a measurement with Iext = 1μA, obtain Δts

Δtref − Δts Δts0

Is = · · Is0 (4.8)
Δts Δtref − Δts0

5. Subtract parasitic currents from the measurement to obtain ILEAK of the measured

device.

ILEAK = Is − Iparasite (4.9)

Two folded cascode amplifiers were designed, one for measuring NMOS leakage

(Figure 4.6) and the other for PMOS leakage (Figure 4.7). Each amplifier is optimized

for its operations at nominal conditions and has a DC gain of more than 60dB over the

voltage range of interest. This is more than sufficient to achieve the required measurement
 40

resolution. Figure 4.8 shows the simulated BODE plot for the amplifier used for NMOS

ILEAK measurement in the 90nm process.

The first amplifier for PMOS leakage measurement has an input bias voltage of

0V and the second amplifier for measuring NMOS leakage has an input bias voltage of

VDD. In both 90nm and 45nm technologies, nominal VDD is around 1V. The transistor

sizes in the amplifier remains the same and layout is also very similar in both technologies.

Differences are due to design rules for the dimensions of contacts and certain spacing rules

that differ between the design kits from the two technologies. A single stage amplifier

was chosen in order to avoid problems with loop stability, and a folded-cascode design was

chosen for its simplicity and its ease in providing the required gain. Long gate-lengths

were used to increase the gain, and large transistors were used to improve matching and

reduce offset voltages. Much effort was made in the layout of the transistors in order to

maximize matching. Common centroid layout and dummy poly-Si lines were used to ensure

close matching. In Monte Carlo simulations using the 45nm models, a 3 × σ input offset

voltage of 2.4mV was obtained. Since the offset voltage affects the voltage at the drain of

the leakage transistors, it does not affect ILEAK significantly.

The schematics of the comparator is shown in Figure 4.9. These comparators

are used to generate the start and stop signals when the output of the integrator crosses

the threshold voltages. The start and stop signals are then buffered and output to pads.

The time interval between the two signals are measured off-chip with a logic analyzer, the

TLA5202 from Tektronix, which is capable of 125ps timing resolution. The long integration

time of more than 50 μs easily permits a current resolution of less than 1nA. Thick oxide
 41

VDDA

M12 M16 M13 M3 M4

19.2
19.2 192 38.4 38.4

M17 M5 M6
192 38.4 38.4
M11 M15
In 4 19.2
M1 M2 Vout
Vin 1V
768 768
M7
M8
96

M100 M14 M9 M10

4.8 192
4.8

GNDA

Figure 4.6: Schematics of the folded cascode amplifier for measuring NMOS leakage cur-
rent. Gain is optimized for input bias of VDD=1V. Transistor widths are indicated in μm
next to the individual transistors. All gate lengths are 0.35μm and In = 90μA. All tran-
sistors whose bulk connections are not indicated have it connected to VDDA(PMOS) or
GNDA(NMOS).

VDDA

M100 M18 M14 M13

M3 M4
24 24 24 192 192 192

M17 M5 M6
192 192 192
M19
Ip 24 Vout
M1 M2 0V
Vin
768 768

M21 M11 M15 M7 M8

24 5 24 48 48

M20 M12 M16 M9 M10

24 24 24 144 144

GNDA

Figure 4.7: Schematics of the folded cascode amplifier for measuring PMOS leakage cur-
rent. Gain is optimized for input bias of GND=0V. Transistor widths are indicated in
μm next to the individual transistors. All gate lengths are 0.35μm and Ip = 120μA. All
transistors whose bulk connections are not indicated have it connected to VDDA(PMOS)
or GNDA(NMOS).
 42

70.0 V(out)

40.0

20.0
Magnitude (dB)

0.07828
0.0

-20.0

-40.0
-60.0

20.0 V(out)
-20.0

-60.0 -101.68051
Phase (degrees)

-100.0

-140.0

-180.0

-220.0
-260.0
C1: 9.76251e+7
1.0e+
0e+1 1.0e+2 1.0e+3 1.0e+4 1.0e+5 1.0e+6 1.0e+7 1.0e+8 1.0e+9 1.0e+
Frequency (Hz)

Figure 4.8: BODE plots of the amplifier used for NMOS ILEAK measurement in the 90nm
test-chip. Top plot shows a DC gain of ≈ 60dB and bottom plot shows a phase margin of
≈ 80◦ .
 43

VDDA VDD
24
24 24 24 24

in1 out
18 18 18 18
I

4 4 in2 4 4 4 4

GNDA GND

Figure 4.9: Schematics of the comparator for generating the start and stop signals when
the output of the integrator crosses the threshold voltages.

transistors were also used in the comparator and transistor matching was improved with

common centroid layout and dummy poly-Si lines.

Simulation results that demonstrate the precision of the current measurement

circuit is shown in Figure 4.10. The measurement circuit as well as the leakage array

is modeled in SPICE and simulated. Is is the true subthreshold leakage current of the

transistor under test and Is cal is the leakage current that is calculated from the integration

times as described previously. A maximum error of ≈ 2% is obtained for currents greater

than 1nA. The accuracy of the current measurement is verified in the 90nm test-chip by

measuring the same transistors repeatedly. Figure 4.11 plots 160 leakage measurements

performed continuously over a 27 hour period for 3 layouts. A σ/μ of ≤ 1.3% is measured.

In the 90nm test-chip, 2.5V thick oxide transistors were used and VDDA = 2.5V.

However, in the 45nm test-chip, the process only had a 1.8V option and the analog circuits

were migrated to VDDA = 1.8V and verified to work according to specifications. The

BODE plots of the amplifiers for measuring NMOS leakage currents and PMOS leakage

currents in the 45nm test-chip are shown in Figures 4.12 and 4.13 respectively. A gain of
 44

0.5 0.5

Error (Is - Is_cal)/Is (%)

0 0

- 0.5 - 0.5

a) Post Layout b) Schematics

- 1.0 - 1.0

- 1.5 - 1.5

- 2.0 - 2.0
10-10 10-9 10-8 10-7 10-6 10-5 10-10 10-9 10-8 10-7 10-6 10-5
Is (NMOS) Is (NMOS)

Figure 4.10: Simulation results for the current measurement circuit. Plots of leakage current
measurement error vs leakage current. a) shows post-layout simulation results and b) shows
pre-layout simulation results.

8
7.5
Normalized ILEAK

7
6.5
6
5.5
5
4.5
0 20 40 60 80 100 120 140 160
90nm Measurements (160 total)

Figure 4.11: Repeated measurement of ILEAK to demonstrate accuracy in the 90nm process.
Plots 160 measurements of NMOS ILEAK of the same 3 layouts in the same tile taken over
27 hours. A σ/μ of ≤ 1.3% is measured.
 45

70.0 V(out)
50.0

30.0
0.08225
Magnitude (dB) 10.0

-10.0

-30.0

-50.0
-70.0

20.0 V(out)
-20.0

-60.0 -105.40920
Phase (degrees)

-100.0

-140.0

-180.0

-220.0
-260.0
C1: 8.37678e+7
1.0e+
0e+1 1.0e+2 1.0e+3 1.0e+4 1.0e+5 1.0e+6 1.0e+7 1.0e+8 1.0e+9 1.0e+
Frequency (Hz)

Figure 4.12: BODE plots of the amplifier for NMOS ILEAK measurement in the 45nm
test-chip. Top plot shows a DC gain of ≈ 60dB and bottom plot shows a phase margin of
≈ 75◦ .

≈ 60dB is achieved in both cases with ≥ 60◦ phase margin. Figure 4.14 shows the circuits

for measuring PMOS leakage current in the 45nm test-chip. Analog supplies, comparator

threshold voltages and amplifier output bias voltages are different from the 90nm test-chip.

For NMOS leakage measurement, the output of the integrator, Vout is biased to 0.4V at the

start of each measurement. Similarly to the 90nm test-chip, the accuracy of the current

measurement circuit is verified by measuring the same transistor repeatedly, and the results

are shown in Figure 4.15. This time, σ/μ of 1.4% and 11.5% are obtained for PMOS and

NMOS ILEAK respectively. The more significant noise in the NMOS ILEAK measurement

is due to a smaller mean NMOS ILEAK .

 46

70.0 V(out)
50.0
30.0
-0.03294
10.0
Magnitude (dB)

-10.0
-30.0
-50.0
-70.0
-90.0

20.0 V(out)
0.0
-20.0
-40.0
-60.0
Phase (degrees)

-80.0 -107.13613
-100.0
-120.0
-140.0
-160.0
-180.0
C1: 1.01410e+8
1.0e+
0e+1 1.0e+2 1.0e+3 1.0e+4 1.0e+5 1.0e+6 1.0e+7 1.0e+8 1.0e+9 1.0e+
Frequency (Hz)

Figure 4.13: BODE plots of the amplifier for PMOS ILEAK measurement in the 45nm test-
chip. Top plot shows a DC gain of ≈ 60dB and bottom plot shows a phase margin of ≈ 73◦
.

1.8V
ǻtref - ǻts Is
10W W ǻts ~ Iref
Select 10W W
Vdd
ǻtref = Integration time when measuring Iref
ǻts = Integration time when measuring Iref + Is
C
PMOS Iref
V1 = 1V
Vout

Is
row1 P2 1.8V
V2 = 0.5
Vout t
P1 P1
row2 ǻts
P1b
gnd
1.3V

Figure 4.14: PMOS leakage current measurement for the 45nm test-chip.
 47

b) Normalized NMOS ILEAK

a) Normalized PMOS ILEAK
7.5
3.9
7
3.8 6.5
6
3.7 5.5
5
3.6
4.5
3.5 4
0 50 100 150 200 0 50 100 150 200
45nm Measurements (200 total) 45nm Measurements (200 total)

Figure 4.15: Repeated measurement of ILEAK to demonstrate accuracy in the 45nm process.
a) and b) plot 200 measurements of PMOS and NMOS ILEAK of the same layout in the
same tile. σ/μ of 1.4% and 11.5% are measured for PMOS and NMOS ILEAK respectively.

4.2 Frequency

Ring-oscillators (RO) are commonly used to characterize transistor performance.

They are small, easy to implement and can be measured easily, accurately and rapidly.

Long ROs average random variations of the individual stages exposing systematic effects

whereas short ROs have less averaging and can be used to measure random variations.

Modified ROs can also be used to measure delay of individual gates with good accuracy

as mentioned in chapter 1 [26, 27]. In this section, circuits that use ROs to characterize

transistor performance variability and spatial correlation of gate delays will be described.

4.2.1 Ring Oscillator Array

In the test-chips, frequency measurements of a large array of ROs are made and

used to characterize performance variations and systematic effects. Two scan chains for

row and column addresses are used for addressing the array. Two address bits per row

and a column bit are used to generate two control bits for each RO. One bit will enable
 48

Figure 4.16: RO with n stages. Address bits will enable the RO by applying VDD to the
enable signal. The output of the RO (out) goes to the input of the multiplexer.

the selected RO through a NAND gate placed in the ring ((Figure 4.16)), and the other

will select the multiplexer to output its own RO frequency instead of the previous. The

multiplexer output goes through a series of multiplexers to a frequency divider that outputs

to a pad (Figure 4.17) [15, 16, 17, 85]. This design makes it very easy to place ROs together

to form an array. Frequency is measured off-chip with a 20GSPS oscilloscope and averaged

over ≈ 100 periods. Accuracy of the measurement is shown in Figure 4.18 and Figure 4.19.

The same RO is repeatedly measured and noise is presented as the σ/μ of the RO frequency.

A σ/μ of 0.15% and 0.02% was obtained from the 90nm and 45nm test-chips respectively.

In both 90nm and 45nm test-chips, 13-stage ROs are used for the single finger

transistor layouts. For the layouts with a stack of 3 gates in the 90nm test-chip (described

in chapter 5), 7-stage ROs are used. The number of stages is chosen such that the NAND

gate in the RO contributes less than 10% to the total RO delay. This ensures that the
 49

F Divide
F/N
by N

RO RO
sel sel

row0a
row0b

column0 column1

Figure 4.17: RO frequency measurement.

10.2
Normalized RO frequency

10
9.8
9.6
9.4
9.2
9
8.8
0 50 100 150 200 250
90nm Measurements (236 total)

Figure 4.18: Repeated measurement of RO frequency to demonstrate accuracy in the 90nm

process. Plots 236 measurements of RO frequency of 3 layouts in the same tile taken over
18 hours. A σ/μ of 0.15% is measured.
 50

Figure 4.19: Repeated measurement of RO frequency to demonstrate accuracy in the 45nm

process. Plots 1000 measurements of RO frequency of the same layouts in the same tile. A
σ/μ of ≤ 0.02% is measured.

RO frequency is more sensitive to the inverter delays than the NAND gate delay. It is

also important to keep the RO short so that fine spatial resolution can be achieved. If the

RO oscillates too fast to propagate through the multiplexers to the output pad, a local

divide-by-two circuit can be used to reduce the frequency. Wide transistors are used to

make variations less sensitive to fluctuations in transistor gate widths.

4.2.2 Ring Oscillators with Different Number of Stages

A set of test structures in the 90nm test-chip was designed to characterize cor-

relation of delays between inverters that are connected to each other in a RO. This test

structure consists of ROs with different number of stages (Figure 4.16). Details will be

given in Section 6.1.

The delay and spatial correlation of individual gates can also be obtained using
 51

1
SA SA SA SA SA sel1b SA SA
1 sel0b 0
Dummy
0 0 0 sel0 sel1 sel2 load

SA
SA
SA
SA
SA
SA
Vdd

out
A
SA
c B
Gnd

Figure 4.20: RO with programmable number of stages using NAND gates

a RO with programmable number of stages shown in Figure 4.20. This RO is made up of

NAND gates and by applying the correct control signals to the NAND gates, RO of different

lengths are generated. The gates of the longer RO will contain the gates of the shorter RO

and the difference in delay will be due to the extra gates that are added plus the mismatch

of one NAND gate. By taking the difference in delay, the delays of the extra gates can be

obtained. Since these are placed next to each other, high resolution spatial correlation of

gate delays can be measured.

4.3 Summary

In summary, circuits to measure CMOS performance accurately and rapidly have

been proposed. Measurement of leakage current and RO frequency of an array of devices

have been described in detail and implemented in a 90nm and a 45nm test-chip. RO circuits

with different number of stages to characterize spatial correlation of gate delays have been

presented and but have not been implemented. Future work consists of implementing these
 52

circuits.
 53

Chapter 5

Measuring SRAM Performance

Variability

Measurement of SRAM cell noise margins and dynamic performance in large arrays

with minimum perturbation of the layout is necessary in order to accurately characterize

SRAM variations. This chapter proposes several circuit ideas to measure both static noise

margins(Section 5.1) and the read and write access times of individual SRAM cells with

minimum or no perturbation of the array layout [86]. Section 5.2 describes a ring-oscillator

circuit whose frequency correlates with the write access time of the SRAM cell, and Section

5.3 describes a pulse ring-oscillator circuit whose frequency correlates with the read access

time of the SRAM cell. Section 5.4 describes how the access times for read and write can

be measured by adjusting the wordline pulse width and measuring it. Finally, Section 5.5

summarizes this chapter.

 54

5.1 SRAM Static Noise Margins Measurement

SRAM cell read and write margins are important performance parameters [87].

Traditional noise margin measurements are obtained from butterfly curves or N-curves.

These require measurement of the internal storage nodes of the SRAM cell which, in sim-

ulations, is easily done but in silicon, requires perturbing the layout. Instead, the IV

characteristics of the bitlines can be measured and used to deduce read and write margins

as well as read and write strength of the cell [86, 88]1 .

Figure 5.1 shows the schematic of a SRAM column and the switches connected to

the bitlines that will allow us to measure the bitline current for given voltages. A plot of

the bitline IV curve is given in figure 5.2 and figure 5.3 showing indicators for the write

margin and read margins respectively. The write margin IV plot is obtained by sweeping

VL from VDD to 0V and measuring the current IL with a ’1’ state stored in the left node

of the cell. IL drops initially and rises suddenly when the SRAM cell flips. The amount

of current required to flip the cell and the voltage at which this occurs can be used as the

write margin indicators. The read margin IV plot is obtained by sweeping VL from 0V to

VDD and measuring IL with a ’0’ state stored in the left node of the cell. The current when

VL = V DD and the slope at this point can be used as read margin indicators. This circuit

has been implemented in [88]. Using various ways of sweeping supply, wordline and bitline

voltages and measuring the bitline current, different noise margin parameters are obtained

and shown to correlate strongly with the tradition noise margins obtained from butterfly

curves.
1
This work was done in collaboration with Zheng Guo, U.C. Berkeley.
 55

Figure 5.1: A SRAM column showing the bitline switches.

IL BLL BLR

WL
VDD
PL PR

VL ‘1’ ‘0’ NAXR

NAXL
0 IL
Vdd NL NR
GND

Write margin
indicators IL VL IR VR = Vdd

Figure 5.2: Bitline IV plot showing write margin indicators

 56

BLL BLR
IL WL
VDD
PL PR

‘0’ ‘1’ NAXR

Read NAXL
IL
margin NL NR
VL GND
0 Vdd

IL VL IR VR = Vdd

Figure 5.3: Bitline IV plot showing read margin indicators

5.2 SRAM Dynamic Write Performance Measurement

Static noise margins of the SRAM cell are useful for verifying functionality. How-

ever, the more important performance metrics are the read and write access time as these

determine if the cell works during real-time operations. The read access time is the time

required to perform a read, and the write access time is that required to perform a write.

These dynamic parameters depend both on the SRAM cell strengths as well as the capac-

itances and leakage currents of the cell and the bitline which cannot be obtained through

static measurement alone. As mentioned in chapter 1, [30] has shown that by connecting

SRAM columns together to form a ring, the oscillation frequency thus obtained can be used

to evaluate the average dynamic performance of an SRAM array. Based on similar ideas,

two ring oscillators have been designed that will allow for evaluating the read and write

access times of SRAM cells in an array with minimum perturbation of the layout.

Figure 5.4 shows the RO for characterizing write access time. The differential

inverter forces the internal storage nodes of the SRAM cell that is being accessed to switch
 57

at each cycle. The SRAM cell acts as a load to the RO. This changes the RO frequency

which is divided down and measured. The pull-down strength of the differential inverter is

controlled by Vbias. It is adjusted such that pull-down is strong enough to flip the SRAM

cells across all corners but sufficiently weak such that the delay in flipping the SRAM cell is

a significant portion of the total delay through the RO. In that case the RO will be sensitive

to variations in the write access time of the cell. The switching threshold of the differential

inverter that is driven by the bitlines should be tuned to be low so that the bitline delay

reflects the writing of the SRAM cell. The number of stages should be small for increased

sensitivity to the bitline delay. The RO is connected to the column bitlines of the SRAM

array through addressable switches that allow to select any particular column. The write

strength obtained is an average of the write ”0” and write ”1”’ strength of the SRAM cell.

The RO frequency is first measured with all wordlines turned off to measure the intrinsic

frequency of the RO due to the bitline load only. The wordline of the SRAM cell under test

is then turned on and a slower oscillation frequency is obtained. The difference between the

two frequencies gives a measure of the average write access time of the SRAM cell.

5.3 SRAM Dynamic Read Performance Measurement

Figure 5.5 shows the RO for read access time measurement. A pulse is propagated

through the ring and regenerated at each cycle. The pulse width is reduced due to the

slow pull-down of the bitline (blc), enlarged by the pulse widening circuit and regenerated

by the pulse generator. The oscillation frequency will vary as a function of the pull-down

strength of the SRAM cell. The RO frequency is first measured with all wordlines turned
 58

Odd number of inverters col0 col1

SRAM Array
WL0
RO output

WLn

enable

c col0 col1

Vdd

out outb
in out
in inb
inb outb
Vbias

Gnd

Figure 5.4: RO to measure write access time and the differential buffer used in the RO.
 59

off. The bitline will be pulled down by the NMOS stack whose strength is controlled by

voltage Vctr. The wordline of the SRAM cell under test is then turned on and a faster

oscillation frequency is obtained as the NMOS pull-down stack is now aided by the SRAM

cell. The difference between the 2 frequencies gives a measure of the read access time of the

SRAM cell. Several knobs are necessary in this circuit to ensure the oscillations occur. In

the pulse widening circuits, Vc1 and Vc2 are used to control the amount of pulse widening.

The pulse has to be widened sufficiently so that the pulse generation circuits can regenerate

a pulse. Next, in the pulse generation circuits, the pulse width of the generated pulse can

be adjusted with Vc3 and Vc4. This pulse must be sufficiently large so that the bitline

has sufficient time to pull-down the voltage to 0V. These voltage controlled buffers can be

implemented as shown in Figure 5.6. If Vc2 is low to weaken the pull-down of the second

stage, the falling edge of the pulse will be slowed down resulting in a widening of the pulse.

Also, both Vc1 and Vc2 can be used to weaken both PMOS and NMOS of the second stage

in order to increase the delay in the pulse generation circuit. This will increase the width

of the regenerated pulse. Finally Vctr controls the pull-down strength of the NMOS stack.

This pull-down strength must be weak relative to the SRAM pull-down strength so that

the delay in pulling down the bitline is more sensitive to the SRAM cell than the NMOS

stack. Overall the tuning of the oscillator has to ensure that oscillations occur and that

the frequency of oscillations is sensitive to the pull-down strength of the SRAM cell. The

single ended RO is connected to the column bitlines of the SRAM array through addressable

switches.
 60

Pulse regeneration Vc3, Vc4

init odd
b
c

a
a
WL vctr b
c

f/N

WL
Pulse width is shortened Vc1,Vc2
even
blt blc

LE will hold this bitline

to vdd when reading
Increases the pulse
Skewed strong N inverter width to > original pulse

Figure 5.5: Pulse RO for characterizing cell read timing.

Figure 5.6: Current starved buffers for widening the pulse width and creating delay for the
regenerated pulse.
 61

blt blc

wl
lclk

Row select
row
address prch
dellclk

Precharge gblt
and latch the data data
Write select

write

write gblt
data

Figure 5.7: SRAM access time measurement

5.4 SRAM Access Times Measurement

Figure 5.7 illustrates how a SRAM in a 45nm SOI technology works [89]. The local

bitline (blc) will trigger the global bitline (gblt) and the data read is latched. lclk generates

the wordline pulse (wl) and local precharge (prch) signals whereas dellclk generates the

global bitline precharge and data-latching signals. In that way, lclk and dellclk determine

several important timing parameters of the SRAM. The minimum wordline pulse width

for correct operation is equivalent to the access times of the SRAM cell. This pulse width

is the same as the lclk pulse width. By generating a programmable lclk pulse width and

measuring the minimum width for correct operation a measure of the dynamic performance

of the SRAM cell can be obtained.

 62

5.5 Summary

In summary, circuits to measure SRAM cell variability accurately and rapidly with

minimum perturbation to the array layout have been proposed. However, they have either

not been implemented or not been measured yet. Future work consists of implementing

these circuits.
 63

Chapter 6

Measurement Results and Analysis

Two test-chips, the first in an early 90nm process and the second in an early 45nm

process, have been designed and measured. These chips contain the ring-oscillator and

leakage transistor arrays and the leakage current measurement circuit described in chapter

4. Measured data show several important trends which are analyzed in detail in this chapter.

The impact of layouts on transistor performance and the statistics of transistor D2D and

WID variations has been measured and characterized.

In this chapter, section 6.1 will describe the layout structures that were studied

and the circuits in the test-chips. Several properties of the 45nm process will also be

presented. Section 6.2 will quantify the impacts of the different layout configurations and

provide analysis of the causes of these systematic variations. Section 6.3 will describe the

measurement results for D2D and WID variations. Spatial correlation of RO delays in the

test-chips will be analyzed. Section 6.3.3 describes the measurement result of an experiment

in which an array of ring oscillators with different number of stages was measured in the
 64

1mm
ADC

1mm
Tile Array

Figure 6.1: Die photo of the 90nm test-chip

90nm test-chip in order to study the spatial correlation of closely spaced gates in the RO.

Section 6.4 describe the process of extracting process parameters from measurement in the

90nm process. Finally, section 6.5 summarizes the results of the measurement.

6.1 Layout Structures and Test-Chips Overview

A test-chip has been implemented in a general-purpose 90nm CMOS technology

to evaluate the effects of lithography-induced variations, and to measure WID and D2D

variations, as well as WID spatial correlation. This is done by measuring ring-oscillator

(RO) frequencies and transistor source-drain leakage currents (ILEAK ) of an array of test-

structures [5]. These circuits were described in Sections 4.1 and 4.2.

The chip contains 10 × 16 tiles, with a total array area of 1mm × 1mm (Figure 6.1).

The tiles are separated horizontally by 62.5um and vertically by 100um. Each tile has twelve

13-stage ROs and twelve transistors in the off-state, each with a different layout (Figure
 65

Figure 6.2: 12 layout structures in 90nm

6.2). Some transistors in both the ROs and those in the off-state consist of a single poly-

Si finger, while the rest are constructed with a stack of three fingers. The poly-Si pitch of

neighboring dummy poly-Si lines is varied in the test structures to capture proximity effects.

Poly-Si orientations, together with the properties of the two-dimensional tile array, are used

to characterize spatial correlation. Rotated poly-Si ROs are obtained by rotating standard

ROs by 90o . Non-symmetrical structures and their mirror image target measurement of

the coma effect. The first layer metal coverage over certain gates are also varied in the

layout to investigate the effects of anneal [70]. The chip also contains the leakage current

measurement circuit described in section 4.1.2. A 2.5V analog supply was used together

with thick oxide transistors to implement the analog components in this circuit.

The 90nm test-chip also contains RO of different number of stages for characteriz-

ing the correlation of delays between inverters that are connected to each other (mentioned

in Section 4.2.2). Using a simple model of spatial correlation coefficient (ρ) shown in Figure

6.3, the relationship between the variance of the ROs (s2D ) and the number of stages (n)
 66

Figure 6.3: Model of spatial correlation coefficient. Plot of correlation coefficient ρ vs

spacing i. k is a model parameter.

is obtained (Equation 6.1). The unknown parameters are the variance of the delay of a

single inverter (s2 ) and the exponential constant that characterizes the spatial correlation

coefficient (k). By measuring s2D for ROs of known n, the measurement data can be fitted

to the equation to obtain s2 and k.


n−1
s2D = n · s2 + 2 · s2 · (n − i) · e−k·i (6.1)
i=1

where

i = spacing, in this case the spacing between adjacent inverters ≈ 2.1μm

k = parameter that determines the characteristics of the correlation coefficient model

sD = standard deviation of the RO

s = standard deviation of the delay of a single inverter

n = number of stages of the RO

This test structure is implemented as an array of VRO tile, each of which contains 8

ROs with 13, 19, 25, 31, 37, 43, 49 and 55 number of stages arranged as shown in Figure 6.4.

The transistors have their gates oriented vertically and the gates are spaced horizontally.
 67

13 55

19 49

43

37

Figure 6.4: Arrangement of ROs of different lengths.

In the layout, 4 RO and off-transistor tiles combine with one VRO tile to form a bigtile

that is stacked to form the 10 × 4 bigtile array shown in Figure 6.5. A simple exponential

model of spatial correlation coefficient (ρ) shown in Figure 6.3 will be used to derive the

relationship between the variance of the ROs (s2D ) and the number of stages (n) (Equation

6.1). The unknown parameters s2 and k will be derived by fitting the measurement results

to the model equation.

The chip also contains frequency dividers and a few other control logic. The left

side of the array contains a vertical column of shift registers that supply the row address

bits and the bottom of the array contains a horizontal row of shift registers that supply the

column address bits. Both the input and output of each shift register chain are connected

to I/O pads in order to ensure that they work properly during debugging. The tiles are

identical and are abutted to each other in the layout to form the array.

The methodology used to characterize variations in the 90nm process was also

applied to a 45nm low power strained-Si process. A 45nm test-chip containing an array of

ROs and both NMOS and PMOS leakage transistors was used to characterize the impact

of layout on transistor performance and measure D2D and WID variability. In addition,
 68

LEAK LEAK LEAK LEAK

Figure 6.5: Layout of a bigtile and the 10 × 4 array of bigtile in the 90nm test-chip.
 69

PROCESS FEATURE 45NM PROCESS EFFECT

Si substrate <100>- oriented Higher PMOS mobility
channel
Shallow trench isolation Sub-atmospheric Reduce STI stress
(STI) deposited oxide
Contact etch stop layer Nitride layer create high Higher NMOS mobility
(CESL) tensile strain
Immersion NA >1 Improved resolution
Backend dielectric Low k 2.5 Low RC delay

Table 6.1: Summary of the 45nm process. [2, 3, 4]

this chip contains SRAM structures that allow for the measurement of IV characteristics of

bitlines in large SRAM arrays as described in Section 5.1.

Table 6.1 summarizes several techniques used in the low-power 45nm process [2, 3,

4]. Transistor channels are oriented in the <100>direction, which increases PMOS transistor

mobility and makes it insensitive to stress. The use of sub-atmospheric chemical vapor

deposition oxide (SACVD) for trench isolation further reduces stress effects. Instead of a

strong compressive strain, these trenches now exert a weak tensile strain on the transistors.

Strong uniaxial tensile strain is created by the nitride layer in order to increase NMOS

mobility. Resolution is enhanced with immersion lithography and low-k dielectric is used

for the copper interconnects. Strain induced by the shallow trench isolation (STI) and the

contact etch stop layer (CESL) nitride are layout dependent and are investigated in the test

chip.

In the 45nm process, many strict design rules have been introduced. Most notable

is the rule on minimum length poly-Si gate orientation and spacing. Orientation of gates

are now fixed and only a small continuous range followed by a discrete jump in poly-Si

spacing is allowed. Taking into account the new rules and the strained process, 17 layout
 70

styles were created to study the effects of layout. These are shown in Figure 6.6 and 6.7 and

described in Table 6.2. Layouts P1, P2, P3, P4 vary the spacing of the poly-Si nearest to

the transistor’s gate. Measurement results from the 90nm test-chip suggest that proximity

effects could dominate again. Having a range of poly-Si pitch would improve the analysis.

The significance of proximity effects in the 90nm test-chip also led to the structures SP1,

SP2, SP3 which vary the distance of the poly-Si that is the second nearest neighbor to

the gate. S1, S2 are layouts that are symmetrical to each other. Optical aberrations can

cause these structures to print differently, giving rise to different gate-lengths. D1 has a

longer source/drain (S/D) diffusion area than P3, which has been observed to cause different

strain in a transistor [90]. M1 has metal-2 coverage over its gate which has been shown to

cause different annealing temperatures [70]. Transistors with metal covered gate tend to

experience less annealing, resulting in increase Vth and reduced ILEAK . T1 has neighboring

poly-Si at the ends of its gate which could cause the gate poly-Si to print differently at the

edges. R1, R2 and R3 have regular poly-Si pitches that vary from minimum to maximum.

The benefits of regular poly-Si pitch has been observed in lithography but not measured

in circuits. These structures permit to measure the impact of regular pitches and pitch

spacing on variability. NSTI is the same as R1 except that there is no STI and isolation

is achieved by turning off the adjacent transistors [91]. STI causes strain in the substrate

which changes transistor mobility. It also creates unevenness on the surface of the substrate,

resulting in systematic changes in transistor properties. Finally, in layout V1 the inverters of

a RO are placed in the vertical direction instead of the horizontal direction as illustrated in

Figure 6.7. In the 90nm test-chip, variations were dependent on the horizontal and vertical
 71

Layouts Targeted Effect

P1,P2,P3,P4 Primary proximity
SP1,SP2,SP3 Secondary proximity
S1,S2 Symmetry
D1 Larger S/D area
M1 Metal coverage over gate
T1 Poly-Si st extremity of gate
R1,R2,R3 Regular pitch
NSTI No STI
V1 Vertically placed gates

Table 6.2: Characteristics of the layout configurations.

directions. Even though 90◦ gate rotation are not allowed by the design rules in this 45nm

process, certain properties of the two orthogonal directions can be investigated with these

structures.

The die photo of the 45nm test-chip is shown in Figure 6.8. The array contains 18

x 16 tiles, each tile contains 17 ROs and 17 NMOS and PMOS transistors with VGS = 0 in

each of the 17 layouts. In the RO array, a local divide-by-2 circuit within each RO allows for

the use of small number of stages by reducing the frequency of the signal that is multiplexed

out. The leakage NMOS and PMOS transistors have their gate connected to GND and VDD

respectively so that VGS=0 only when that particular transistor is selected. Details on the

circuit implementation of RO frequency and ILEAK measurement can be found in Section

4.1 and 4.2.1.

In the 90nm test-chip, only NMOS leakage current is measured. Each NMOS

transistor has a width of 1.92μm and is made up of 4 smaller transistors in parallel, each

small transistor has a width of 0.48μm. Transistor width is chosen to have sufficient ILEAK

such that measurement accuracy is not compromised. As the precision target is 1nA, gate

widths are chosen to give ILEAK of ≈ 10nA in the slow corner. Individual gate fingers are
 72

min max2 min max2

P1 S1
T1

Lx Lx
Lx
mid1
P2 min max2 min
SP1 S2

Lx
max1 mid2 Lx
P3 SP2 D1
max1

Lx
max2 Lx
max2 SP3 Longer diffusion
P4
min max2
M1

Lx
R1 min
Lx
M2 over gate

mid
R2
Poly-Si

max Contact
R3
Diffusion

min
NSTI

Figure 6.6: 17 layout configurations in 45nm

 73

Figure 6.7: Illustrates (a)horizontally vs (b)vertically placed inverters in a RO.

SRAM SRAM
SRAM
array array
padout

SRAM padout
SRAM padout

LEAK

SRAM SRAM
array array

LEAK

Figure 6.8: Die photo of the 45nm test-chip

 74

Test-chip MOS Total Width No. Fingers Finger Width

90nm NMOS 1.92μ(m) 4 0.48μ(m)
45nm NMOS 3.9μ(m) 13 0.3μ(m)
PMOS 5.46μ(m) 13 0.42μ(m)

Table 6.3: Dimensions of the transistors in the ILEAK arrays.

Test-chip Gate-stacks No. Stages PMOS NMOS

90nm single gate 13 1μ(m) 0.48μ(m)
3 gate stack 7 1μ(m) 0.48μ(m)
45nm single gate 13 0.42μ(m) 0.3μ(m)

Table 6.4: Dimensions of the transistors in the inverters of the ROs.

chosen to be much longer than the lengths in order to avoid variation in gate widths. In the

45nm test-chip, both NMOS and PMOS leakage currents are measured. The total widths

of each PMOS and NMOS transistor measured are 5.46μm and 3.9μm respectively. Each

transistor consists of 13 smaller transistors in parallel, each with a width of 0.42μm and

0.3μm for PMOS and NMOS respectively.

In both 90nm and 45nm test-chips, 13-stage ROs are used for the single finger

transistor layouts. For the layouts with a stack of 3 gates in the 90nm test-chip, 7-stage

ROs are used. Section 4.2.1 explains how the number of stages is chosen. A shorter RO

achieves good spatial resolution but longer RO is necessary for good sensitivity to the

inverter stages and for the signal to propagate through the multiplexers and dividers. The

widths of the 90nm transistors in the inverter are 1um and 0.48um for PMOS and NMOS

respectively. In the 45nm test-chip, they are 0.42um and 0.3um for PMOS and NMOS

respectively.

The dimensions of the transistors in the arrays are summarized in Table 6.3 and

Table 6.4.
 75

In the analysis of the 90nm measurement results, tiles at the perimeter of the

array are ignored to eliminate edge effects. The ROs with single isolated poly fingers often

exceed the fast corner and are excluded in the measurements. In the 45nm test-chip, no

edge effects were observed and hence all the tiles were used in the analysis. All the RO and

leakage transistors were analyzed.

To distinguish systematic from the random effects, the results for different layouts

are compared for each die from die to die. Measurements for identical structures within a

die are compared to obtain WID correlations. Averages for various dies are used to indicate

the D2D spread. Averages across several dies are used to analyze the mask writing effects.

6.2 Effects of Layout on Frequency and Leakage

In order to investigate the systematic layout effects, the RO frequency and ILEAK

distribution of different layout structures are compared. In the following analysis, the mean

value for each layout configuration for each die is compared to observe systematic effects.

RO frequencies are normalized to the nearest corner in order to remove the differences in

parasitics that are captured by the layout extraction when these differences are significant.

This is the case for all RO frequency measurements in 45nm and the stacked-gates RO

frequency measurement in 90nm. Leakage currents are not aligned with the corners as they

are independent of parasitics. The impact of layout on random variations is studied by

comparing the standard deviation over mean (σ/μ) of the measurement data.

Measurement results show systematic variations due to layout that are not cap-

tured by the layout extraction tool. The impact of layout on performance is more significant
 76

in the 90nm test-chip than in the 45nm test-chip. Proximity of poly-Si causes more than

10% shift in RO frequency (ΔF = 10%) in 90nm and around 2% shift in 45nm. The most

significant effects in 45nm were due to a larger S/D area (ΔF = 5%) and the removal of

STI (ΔF = 3%).

Possible causes of these systematic effects are proposed through analysis of the

measurement results. Accurate analysis in the 45nm results is difficult as the variations are

relatively small.

6.2.1 Proximity of Poly-Si Gates

Figures 6.9 and 6.10 plot the distribution of normalized RO frequency and normal-

ized leakage current in the log domain respectively for the 90nm test-chips. Each colored

histogram represents the distribution for a die and the overall distribution is plotted as a

continuous blue curve. 90nm measured results are obtained from 36 dies located on half a

wafer. Only the histograms of the fastest and slowest dies are shown. Vertical lines labeled

TT and FF represent simulation results from the extracted layout for TT and FF corners.

These plots show that proximity effects due to different poly-Si gate pitch cause a

shift in frequency of over 10%, and a 20x shift in ILEAK . More isolated gates have shorter

gate lengths and hence higher RO frequencies and higher leakage currents. This is much

larger than the 1.1% frequency shift predicted by SPICE simulations of the extracted lay-

out, which capture only changes in parasitic capacitances. D2D and WID leakage current

variations are reduced with increased poly-Si density. There is also a similar but smaller

effect of poly-Si density on stacked gates. Figure 6.11a plots the RO frequency distribu-

tion for stacked gates whereby the frequency is normalized to a fixed constant. Extracted
 77

36 chips, 4032 data

fastest
count simulation
40 slowest
TT FF
20

0
40

20

0
40

20

Figure 6.9: Frequency distribution for single-finger configurations in the 90nm test-chip.
Vertical lines correspond to typical and fast corner simulation results. Bar plots correspond
to the WID distribution of the fastest and slowest chip.

simulation results from the TT and FF corners show that RO frequencies for the different

gate densities differ due to parasitics that is captured by the layout extraction. In order

to show the proximity effects that are not captured by the layout extraction, the measured

frequencies are normalized to the TT corner and the distributions are plotted in Figure

6.11b. These plots show that the frequency of the densest stacked gate configuration is 5%

slower than the fastest layout.

Figure 6.12 shows the distributions for 4 layouts with different poly-Si gate pitches

from the 45nm test-chips. Maximum systematic shift in frequency is ≈ 2%. Leakage

currents also experience small systematic shifts. This effect could be due to a combination

of small gate-length(L) variations and layout dependent variation of the strain caused by

the CESL. PMOS transistors have a sharper VT roll-off and hence PMOS leakage currents
 78

count 36 chips
40 fastest
20 slowest

0
40
20
0
40
20
0
40
20
0
leak

Figure 6.10: Log(ILEAK ) distribution for single-finger configuration in the 90nm test-chip.

are more sensitive to L variation. An isolated gate will generally experience more strain

from the CESL than a poly-Si gate with dense pitch. However, only NMOS is affected by

stress in this process.

In the plot of normalized PMOS logILEAK , poly density induced L variation is

observed. The P2 layout on the second row likely has a shorter gate length than the

others resulting in increased PMOS leakage. The effect on NMOS leakage is weaker and is

completely compensated by the CESL stress. As the poly-Si pitch increases, more tensile

stress is applied, increasing the mobility of NMOS transistors and raising the amount of

NMOS leakage thereby offsetting the effect of increased gate length. PMOS leakage is not

affected since it is insensitive to stress in a <100>-oriented channel. At the same time, RO

frequency also increases and this offsets the effect of a longer L.

 79

40
20
0
40
20
0
40
20
0
40
20
0
3.5 4 4.5 5

40
20
0
40
20
0
40
20
0
40
20
0
1 1.05 1.1 1.15 1.2 1.25 1.3 1.35 1.4

Figure 6.11: RO frequency distribution for stacked gates plotted next to its corresponding
layout in the 90nm test-chip. Vertical lines represent simulation corners from extracted
layout. Distribution plots of a) RO frequency normalized to a constant, b) RO frequency
normalized to its corresponding TT simulation corner. In b), less dense gates show a 5%
increase in frequency compared to dense gates.
 80

min max2
SS
P1
50

01
P1 Lx 0
0.95 1 1.05 1.1 1.15 1.2
mid1
P2
50

02
P2 Lx 0
0.95 1 1.05 1.1 1.15 1.2
max1
P3
50

03
P3 Lx
0
0.95 1 1.05 1.1 1.15 1.2
max2
P4

04
50

P4 Lx 0
0.95 1 1.05 1.1 1.15 1.2
Normalized frequency
100 SS TT TT FF
P1
P1 100
01

50 50
0 0
2.5 3 3.5 4 4.5 5 5.5 4 5 6 7
100 P2
P2 100
02

50 50
0 0
2.5 3 3.5 4 4.5 5 5.5 4 5 6 7
100 100 P3
P3
03

50 50
0 0
2.5 3 3.5 4 4.5 5 5.5 4 5 6 7
100
P4 100 P4
04

50 50
0 0
2.5 3 3.5 4 4.5 5 5.5 4 5 6 7
normalized NMOS logILEAK normalized PMOS logILEAK
Slowest chip Fastest chip 22 chips from 2 wafers

Figure 6.12: Effect of nearest neighbor poly-Si pitch on RO frequency and transistor leakage
currents in the 45nm test-chip. Plots of RO frequency distribution normalized to the SS
corners, NMOS log(ILEAK ) distribution and PMOS log(ILEAK ) distribution for 22 dies.
 81

count 36 chips fastest slowest

40 99% Confidence Intervals
TT FF
Mean ˰/˩ ˰
0.9 1 1.1 0 2 4
20

0
TT FF
40

20

0
3.5 4 4.5 5
RO frequency

Figure 6.13: Bottom left plot shows the frequency distribution for 90◦ rotated ROs with
horizontal gates in the 90nm test-chip. Non-rotated ROs in the top left plot show more
variation. 99% confidence interval of σ/μ and σ 2 shown on the right verify that the top
plot has more variations.

6.2.2 Gate Rotation

Figure 6.13 shows the impact of rotating a RO by 90◦ in the 90nm test-chip for

stacked gates layout. All the layers below metal-2 were rotated in order to keep the parasitics

as similar as possible. Variation (3σ/μ) is smaller by ≈ 0.4% for rotated gates across all

chips. This is verified with a high level of confidence as shown in the confidence intervals

plots. The mean value of frequency stays around the same in the measurement. However,

in the simulation corners, the RO frequency is significantly slower for the rotated layout.

Leakage current of single gate layout for the non-rotated and the rotated cases is shown

in Figure 6.14 and verified to be similar, with the rotated layout having less leakage. This

shows that there is no significant difference in the properties of the transistors in the two

layouts. The difference in the simulation results is likely due to an overestimation of the

parasitics in the rotated layout due to the higher level metals by the layout extraction tools.

Figure 6.15 shows the impact of placing the gates horizontally vs vertically in the
 82

36 chips fastest slowest

30

20

10

30

20

10

0
-4 -3 -2 -1 0
LEAK

Figure 6.14: ILEAK distributions for non-rotated and rotated ROs in the 90nm test-chip.

45nm test-chip. ROs were not rotated in the 45nm process because rotation of minimum

length poly-Si gates is not allowed. The distribution of normalized RO frequency before

normalizing to the SS corner shows that the extracted frequency and measured data are

significantly different for the two layouts. After normalizing with the SS corner, a significant

shift is observed and the RO frequency of the vertically placed gates is significantly lower

than horizontally placed ones. However, the leakage currents remain almost the same. This

is likely due to the parasitics in the metal interconnects which have not been accurately

extracted. In this case, comparison of the mean values of frequency is not valid.

6.2.3 Symmetry

In the 90nm test-chips, the difference between structures that are mirror image of

each other is small and is likely due to the coma effect. Roughly 1-2% shift in the mean

RO frequency was observed with single gate stages as shown in the second and third plots
 83

P1 SS TT SS
80 80
40 40
0 0
V1
80 80 SS TT
40 40
0 0
0.9 1 1.1 1.2 0.75 0.8 0.85 0.9 0.95 1 1.05 1.1
Frequency normalized to SS corner Normalized frequency
P1
100 80
50 40
0 0
V1
100 80
50 40
0 0
4 5 6 7 8 2.5 3 3.5 4 4.5 5 5.5 6
Normalized PMOS logILEAK Normalized NMOS logILEAK
P1 min max2 V1

Slowest chip

Lx
Fastest chip

22 chips from 2 wafers

Figure 6.15: Effect of placement of gates horizontally and vertically in the 45nm test-chip.

of Figure 6.9. This was verified with a high level of confidence.

In the 45nm test-chips, this difference has also been observed as shown in Figure

6.16. However, although RO frequency increased by around 1%, both PMOS and NMOS

leakage currents decreased. This could be due to a combination of the coma effect and

assymmetry in the pocket doping process. However, since the difference is small, it is

difficult to infer the exact cause.

6.2.4 Metal Coverage Over Gate

In the 90nm test-chips, M1 coverage over gates has negligible effects on RO fre-

quency and causes a small reduction in leakage current, indicating a small increase in Vth

(Figure 6.17). It has been reported [70] that metal-covered gates suffer from larger number
 84

min max2 100 SS TT

S1
S1
50

Lx 0
100
max2 min
S2 S2
50

0
Lx 1 1.05 1.1 1.15 1.2 1.25
Frequency normalized to SS corner
100 SS TT SS TT
S1 100
S1
50 50
0 0
100
100
S2 S2
50 50
0 0
3 4 5 6 4 5 6 7 8
Normalized NMOS logILEAK Normalized PMOS logILEAK

Slowest chip Fastest chip 22 chips from 2 wafers

Figure 6.16: Effect of symmetrical layouts which are mirror images of each other in the
45nm test-chip.

of interface states because the metal coverage lowers the temperature on the gate during an-

nealing, which resulted in Vth increase of ≈ 10%. In our layout, metal-1 is in close proximity

to non metal-covered gate since the gate-length is small and the source and drain are fully

covered with metal-1 contacts. Hence, it is likely that the temperature difference during

annealing between metal-covered gates and non metal-covered gates is too small to create

a significant shift in Vth . It is also possible that the process has sufficiently high annealing

temperature to passivate the interface states of metal-covered gates resulting in a small Vth

shift due to different annealing temperature. The effect of Vth shift on RO frequency at

high voltage is less significant.

In the 45nm testchips, the gates were covered with M2 instead due to the space

limitation. In this case the impact of metal coverage is negligible for both RO frequency

and ILEAK measurements (Figure 6.18).

 85

40
30
30
20 20

10 10
0 0
40 30
30
20
20
10 10
0 0
1 1.1 1.2 1.3 -17 -16 -15 -14 -13
normalized frequency Log(ILEAK)
36 chips fastest slowest

Figure 6.17: Effect of M1 coverage over gates in the 90nm test-chips.

min max2
P1 100 SS TT
P1
50

Lx 0
min max2 100
M1
50 M1

0
Lx 1 1.1 1.2
M2 over gate Frequency normalized to SS corner
100 SS TT SS TT
P1 100
P1
01

50 50
0 0

100
100
M1 M1
09

50 50
0 0
3 4 5 6 4 5 6 7 8
Normalized NMOS logILEAK Normalized PMOS logILEAK

Slowest chip Fastest chip 22 chips from 2 wafers

Figure 6.18: Effect of M1 coverage over gates in the 45nm test-chips.

 86

100 SS TT
P1
50
min max2
P1

0
100
Lx
R1 min R1
50

0
0.95 1 1.05 1.1 1.15 1.2 1.25
Frequency normalized to SS corner
100 SS TT SS TT
100 P1
P1
50
50

0 0
100
100
R1 R1
50 50

0 0
3 4 5 6 4 5 6 7 8
Normalized NMOS logILEAK Normalized PMOS logILEAK

Slowest chip Fastest chip 22 chips from 2 wafers

Figure 6.19: Effect of fixed poly-Si gate pitch in the 45nm test-chips.

6.2.5 Fixed Poly-Si Gate Pitch

The use of fixed gate pitch has negligible impact on variability as shown in Figure

6.19. RO frequency of fixed gate pitch layout(R1) has a σ/μ of 2.2% which is the same as

for non-fixed gate pitch layout(P1). This shows that, as long as there is regularity, the use

of fixed poly-Si pitches in a grid-like layout does not reduce variability significantly.

6.2.6 Shallow Trench Isolation (STI)

Figure 6.20 compares the impact of layouts with and without STI in the 45nm

process. The layout without STI is slower by around 3% and has higher PMOS leakage

current likely due to the effect of L variation on PMOS leakage. As this process uses SACVD

trench oxide that generates a low tensile strain, STI stress increases the mobility of NMOS
 87

min 80 SS R1

40

R1 0
0.95 1 1.05 1.1 1.15 1.2
min 80
NSTI

40

NSTI 0
0.95 1 1.05 1.1 1.15 1.2
normalized frequency
TT
100 SS TT 100
R1 R1
50 50

0 0
2 2.5 3 3.5 4 4.5 5 4 4.5 5 5.5 6 6.5 7

100 100
NSTI NSTI
50 50

0 0
2 2.5 3 3.5 4 4.5 5 4 4.5 5 5.5 6 6.5 7
normalized NMOS logILEAK normalized PMOS logILEAK
Slowest chip Fastest chip 22 chips from 2 wafers

Figure 6.20: Effect of STI on RO frequency and transistor leakage currents in the 45nm
test-chips. NSTI uses gate isolation [7] instead of STI.

transistors and causes the layout with STI to be faster, thereby compensating the effect of

L. Variation in L could be due to the STI step that causes unevenness on the surface of the

chip.

6.2.7 Source/Drain Diffusion area

Figure 6.21 studies the impact of a longer S/D diffusion in the 45nm process. Dif-

ferences in S/D capacitance are captured by the layout extraction. After normalization, the

layout with larger S/D area is 5% faster, while the leakage currents remain approximately
 88

SS TT
max1 80
P3
40

P3 Lx
0
0.95 1 1.05 1.1 1.15 1.2 1.25

max1 80 D1

40
D1 Lx
Longer diffusion 0
0.95 1 1.05 1.1 1.15 1.2 1.25
normalized frequency
SS TT TT FF
80 P3 100 P3

40 50

0 0
2.5 3 3.5 4 4.5 5 5.5 6 3.5 4 4.5 5 5.5 6 6.5 7 7.5

80 D1 100 D1

40 50

0 0
2.5 3 3.5 4 4.5 5 5.5 6 3.5 4 4.5 5 5.5 6 6.5 7 7.5
normalized NMOS logILEAK normalized PMOS logILEAK
Slowest chip Fastest chip 22 chips from 2 wafers

Figure 6.21: Effect of a longer S/D diffusion on RO frequency and transistor leakage currents
in the 45nm test-chips.

unchanged. This can be explained by the fact that a larger S/D area will allow the CESL

to exert more tensile strain on the transistors, thereby increasing the mobility of the NMOS

transistors and resulting in faster RO. Since mobility varies linearly with leakage current,

its effect on leakage current in the log scale is small.

 89

6.3 Characteristics of D2D and WID Variations

In the 90nm process 36 chips from half a wafer were measured and in the 45nm

process, 22 chips from 2 wafers were measured. Less systematic variations were observed in

the 45nm test-chip for both D2D and WID variations. However, random WID variations

increased in 45nm due to reduction in channel area. Spatial correlation in 90nm was signif-

icant due to systematic WID variations whereas no significant spatial correlation was found

in 45nm.

6.3.1 Statistics of D2D and WID Variations

In the 90nm process, the wafer map of mean frequency and leakage of each die for

the densest single gate layout configuration is shown in Figure 6.22. It shows a typical radial

pattern that can be attributed to non-uniform resist development [11, 13] or systematic

temperature gradient during RTA [71]. Faster and leakier chips are located at the center

of the wafer. D2D variation is significant resulting in a 3σ/μ of 15% in RO frequency over

half a wafer for the densest single gate structure. For the other single gate structures, it

increases slightly to around 17%.

In the 45nm process, measured results of 22 dies from two wafers are studied. For

each layout, there is a shift of ≈ 2.5% in frequency between the mean of the 2 wafers. D2D

3σ/μ is shown in Figure 6.23. It varies from 6.5% to 7.5% for the 2 wafers and overall D2D

variation was 8.5% for all the layouts except for the vertically placed layout V1. Extensive

measurements on more chips even distributed over the wafer gives a D2D variation of ≈

15%1 .
1
Measurements by Kun Qian, U.C. Berkeley.
 90

ILEAK
.94
-4
.90
-4.4
.86

.82 -4.8

.78 -5.2

Figure 6.22: Wafer maps of mean RO frequency and mean Log(ILEAK) in the 90nm process.
X marks a defective chip. Location of dies on only half the wafer are known.

9
8.5
D2D 3* / (%)

8
7.5
7
6.5
6 wafer 1, 9 dies
wafer 2, 13 dies all 22 dies
5.5
P1 P3 SP1 SP3 S2 D1 R1 R3 V1
P2 P4 SP2 S1 M1 T1 R2 STI
Layout Configurations

Figure 6.23: Comparison of RO frequency D2D variations for 2 wafers and the overall D2D
variations in the 45nm process.
 91

Average value Single chip

1.2

Mean (
1.1
1

2.5

P1 P3 SP1 SP3 S2 D1 R1 R3 V1
P2 P4 SP2 S1 M1 T1 R2 STI
Layout Configurations

Figure 6.24: 45nm WID statistics of RO frequency for 22 dies. The frequency is normalized
to the SS corner frequency.

Figure 6.24 plots the WID mean and σ/μ of the RO frequency for each of the 22

dies in the 45nm process as a function of the layout configuration. The mean frequency

has been normalized to the SS corner in order to compare differences in layouts that have

not been captured by the layout extraction. WID 3σ/μ of RO frequency is approximately

6.6% which is more than twice that in 90nm (3σ/μ ≈ 3.5%). This is consistent with the

transistor channel area reduction by a factor of 4 between the two processes. No significant

dependence of WID variation on layout was observed.

6.3.2 WID Spatial Correlation of ROs

In the 90nm process, WID variation of identical layout structures is small (3σ/μ ≈

3.5%) and weakly dependent on the layout. For each layout structure, the data is normalized

to zero mean and unit variance for each chip before being used to compute its spatial correla-

tion. Confidence intervals for the correlation are computed using Fisher’s z-transformation

to convert Pearson’s correlation to a normally distributed random variable. In the fre-

 92

0.4 0.4

0.3 0.3

0.2 0.2

0.1 0.1

0 0

- 0.1 -0.1
2 4 6 8 10 1 2 3 4 5

RO
RO RO
RO

Figure 6.25: Spatial correlation coefficient (ρ) vs a) horizontal column spacing and b)
vertical row spacing, for vertical and horizontal gates in the 90nm process. Dotted lines
represent 99% confidence bounds for ρ.

quency measurements, the spatial correlation depends only on the direction of spacing and

orientation of the gates.

Spatial correlation is higher in the horizontal direction of the chip than in the

vertical. In particular, for horizontally spaced ROs (Figure 6.25a), correlation is higher

for vertically oriented gates, whereas for vertically spaced ROs (Figure 6.25b), correlation

is higher for horizontally oriented gates. This data is summarized in Table 6.5, showing

that horizontally spaced ROs with vertically oriented gates have a spatial correlation of 0.4

between immediate neighbors whereas horizontally spaced ROs with horizontally oriented

gates have a spatial correlation of 0.25.

This dependence on direction could be explained by the step-and-scan photolithog-

raphy. The horizontal direction in this chip is along the slit of light in the stepper and is
 93

Slit direction = horizontal Horizontal Vertical

spacing spacing
Scan direction = vertical

Vertical Gates 0.4 0.1

RO Freq.
Horizontal Gates 0.25 0.2

ILEAK 0 0

Table 6.5: Spatial correlation coefficient of neighboring ROs for different gate orientation
and spacing direction.

RO Frequency RO Frequency ILEAK

2 2 2 2
4 4 4 4
row

6 6 6 6
8 8 8 8
10 10 10 10
5 10 15 5 10 15 5 10 15 5 10 15
column

Figure 6.26: Surface plot of normalized data of 36 chips in the 90nm process. RO of layout
with vertically oriented gates show a 1% frequency shift between rows 4 and 5.

subject to lens aberrations and curvature, resulting in more correlated features. The verti-

cal direction is along the scan direction, which is subject to variation in scan speeds, stage

vibration that affects focus, and light dosage, resulting in less correlated features.

Other manufacturing effects such as e-beam mask stitching [77] discontinuity or

systematic shift in wafer stage scan speed could also account for the observed dependence

on direction. In order to investigate the systematic reticle field effects, normalized data from

36 chips is averaged and Figure 6.26 shows the resulting surface plot of RO frequency and

ILEAK . There is a sharp systematic shift in frequency between the 4th and 5th rows for RO
 94

17 Layouts Linear fit y = 0.0017*x - 0.04

normalized RO freq
0.4 Mean of 17 layouts 0.2
0.3 Linear fit y = - 0.011*x + 0.07 0.1
0.2
0.1 0
0 -0.1
-0.1 -0.2
-0.2
-0.3
-0.3 (a) (b)
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16 18
column row

Figure 6.27: Plot of 45nm RO frequency normalized to zero mean and unit variance showing
mask characteristics. Small systematic variation of ≈ 0.2 × σ.

structures with vertically oriented gates and a small shift in ILEAK . This could be attributed

to a discontinuity at the level of the mask reticle which suggests e-beam mask stitching

defects. Another reason could be a shift in the scan speed of the wafer/mask stage during

scanning of the reticle image onto the wafer. Since the scan speed is adjusted simultaneously

during scan based on look-ahead sensors, it is possible that there is discontinuity in the

sensor readings causing the scan speed to shift from the 4th to 5th row. Leakage current,

which is less sensitive to gate-length (L) but more sensitive to threshold voltage than RO

frequency, has much smaller spatial correlation.

In the 45nm process, systematic variations in the mask are also investigated by

normalizing the data of each chip to zero mean and unity standard deviation and aver-

aging the normalized data from 22 chips to remove random variations. Figures 6.27a and

6.27b plot the mean frequency for each column and each row respectively. No significant

systematic variation was found, partially due to the fact that random WID variation has

increased. Finally, there is no significant spatial correlation in the WID measurements.

 95

1 1 1

: chip 1 : chip 2 : chip 3

0 0 0
50 100 50 100 50 100

1 1 1

: chip 4 : chip 5 : chip 6

0 0 0
50 100 50 100 50 100

Figure 6.28: Plots of correlation coefficient (ρ) vs spacing between inverters in a RO obtained
from fitting measurement data to a exponential model.

6.3.3 Spatial Correlation of Gates in a RO

The 90nm test-chip contains an array of VRO tiles as described in Section 6.1.

Each tile contains 8 ROs with 13, 19, 25, 31, 37, 43, 49 and 55 number of stages for

characterizing the correlation of delays between inverters that are connected to each other

(Section 6.1).

In the analysis, systematic variations in the vertical direction are significant and

are removed by shifting each measurement by the mean of the ROs of the same row. In this

way the variances observed are due to random variations instead of systematic RO to RO

variations that will render inverters within a RO strongly correlated. The measured data is

fitted to the model and the extracted correlation plots for 6 chips are shown in Figure 6.28.

Except for chip 6, correlation goes to almost zero after ≈ 40μm.

 96

Scatter Plot for Mean of 36 dies Scatter Plot for 1 die

-17

Log(ILEAK)
Log(ILEAK)
-16
-17.5

-17
-18

8 8.5 9 9 9.5 10
frequency frequency

Figure 6.29: 90nm scatter plot of Log(ILEAK ) vs frequency showing a) mean Log(ILEAK )
and mean RO frequency for 36 dies, b) all Log(ILEAK ) and RO frequency for 3 layouts on
1 die.

6.4 Inferring Process Parameters (90nm)

In the 90nm process, the scatter plot of Log(ILEAK ) vs frequency shows a strong

positive correlation between ILEAK and frequency for D2D and layout-to-layout (L2L) vari-

ations (Figure 6.29). This means that these variations are dominated by process parameters

which cause both Log(ILEAK ) and frequency to vary in the same manner. WID scatter plot

shows no significant correlation, suggesting that it is caused by a combination of process

parameters.

In order to relate these variations to process parameters, the least squares method

and the Berkeley Short-channel IGFET Model (BSIM3) model are used to infer the variation

in gate length (L), gate oxide thickness (T ox) and channel dopant concentration (N ch)

from the frequency (F ) and leakage current (I = Log(ILEAK )) measurements. Since the

sensitivities of the measured data to process parameters vary with different condition of
 97

No. Vdd (V) Vbs (V) T (o C)

1 1 0 25
2 1.4 0 25
3 1 -0.2 25
4 1 0 60

Table 6.6: Measurement conditions.

supply voltage (Vdd), substrate voltage (Vbs) and temperature (T), the accuracy of the

inference has been improved by taking the measurements under the different conditions

listed in Table 6.6.

If index i represents the four different measurement conditions, then:

• Fm , Im : Estimated mean F and I

• Lm , T oxm , N chm : Estimated mean process parameters

• F0 , I0 : Mean measurement data for 1 die and 1 layout

• Ltyp , T oxtyp , N chtyp : Typical values of L, T ox and N ch

⎡ ⎤
⎡ ⎤
⎡ ⎤ L−Lm
∂ FF ∂ FF F
⎢ ⎥ ∂ F0
F −Fm ⎢ Ltyp
⎥ ⎢ 0 0
⎥
⎢ F0 ⎥ ⎢ ⎥ ⎢ ∂ Ltyp
L T ox
∂ T ox ∂ Nch ⎥
Yi = ⎢ ⎥ , X = ⎢ T ox−T oxm ⎥ , Mi = ⎢ ⎥
typ Nchtyp
⎣ ⎦ ⎢ T oxtyp ⎥ ⎢ ∂ I ∂ II ∂ II ⎥
I−Im ⎢ ⎥ ⎣ I ⎦
I0 ⎣ ⎦ 0
L
0
T ox
0
Nch
i N ch−N chm ∂ Ltyp
∂ T oxtyp
∂ Nchtyp
N chtyp

⇒ Yi = Mi · X

⎡ ⎤ ⎡ ⎤
⎢ Y1 ⎥ ⎢ M1 ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ Y ⎥ ⎢ M ⎥
⎢ 2 ⎥ ⎢ 2 ⎥
⎢ ⎥=⎢ ⎥ · X ⇒ X̂ = M + · Y
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ 3 ⎥ ⎢ 3 ⎥
Y M
⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦
Y4 M4
 98

Figure 6.30: Illustrates the estimation of the nominal value of L (Lm ) in the case where
only one parameter (L) is inferred. I0 and F0 are the measurement data. Lm is used to
obtain the gradients and to find Im and Fm which are used to normalize the data in the
least squares estimation.

where

M + = pseudo-inverse of M

X̂ = Least squares estimate of X

By using SPICE (Simulation Program with Integrated Circuit Emphasis) simula-

tions, F and I are plotted vs. L, and the overlap region of L which gives a ±15% change

in the mean measured data F0 and I0 is found. The mean of this overlap region gives

Lm which will be used to obtain the gradients and the normalization constants Im and Fm .

The parameters are then normalized and process parameters are inferred using least squares

estimation. Figure 6.30 illustrates this procedure.

This method is used for three process parameters, L, T ox and N ch. By using
 99

Default Vdd = 1.4V Vbs = -0.2V T = 60o C

∂ FF
0
Nch
-0.052 -0.053 -0.039 -0.049
∂ Nch
typ
∂ FF
0
T ox
-0.35 -0.35 -0.14 -0.39
∂ T ox
typ
∂ FF
0
-1.79 -1.83 -1.65 -1.86
∂ LL
typ
∂ II
0
Nch
0.061 0.066 0.090 0.050
∂ Nch
typ
∂ II
0
T ox
-0.045 -0.056 -0.098 -0.028
∂ T ox
typ
∂ II
0
1.01 1.07 1.23 0.92
∂ LL
typ

Table 6.7: Sensitivities used in the 90nm process parameters extraction for one chip and
one layout structure.

SPICE simulations to obtain plots of F and I vs L, T ox and N ch over a small range

around their typical values for the different conditions listed in Table 6.6. Using these

plots, the measured F and I data are mapped to an overlapping range of values of the

process parameters. The mean values of this range (Lm , T oxm , N chm ) are used as the

nominal value of X to obtain the gradients in M and the normalization constants (Fm ,

Im ). For a typical chip with the densest single gate layout configuration, the gradients (M )

are given in Table 6.7 for the four different conditions. The components in X and Y are

then normalized. Finally, least squares estimation is applied to estimate X (X̂). This is

repeated for each die and each layout configuration since the small amount of variation

assures a good linear approximation.

Figure 6.31 shows the distribution of normalized L, obtained from the inference

process. The sensitivities of F and I to L are larger than that of other process parameters as

shown in Table 6.7. This results in significantly more accuracy in the inferred L. The plots
 100

Figure 6.31: De-embedded distribution of L.

of normalized L show a strong correlation to the frequency and leakage plots in Figure 6.9

and 6.10. Faster and leakier ROs have shorter L. The mean values of the L distributions

depend significantly on the layout, showing that systematic layout-to-layout variation is

attributed primarily to L shifts. D2D spreads of L are significant as the bar-plots of the

fastest and slowest chip are spaced far apart.

Figure 6.32 shows a typical surface plot of a die showing the WID variation of

measured F and I. and the corresponding inferred normalized L. Systematic WID variation

observed in F is strongly reflected in L and varies by ≈ 1%.

 101

Figure 6.32: Top plots: surface plots of within-die RO frequency (F ) and Log(ILEAK ) (I)
of a typical die; Bottom plot: corresponding surface plots of estimated normalized L.

6.5 Summary

In 90nm, the largest impact of layout on performance comes from gate poly-Si den-

sity, which causes a systematic shift in frequency of up to 10%. D2D variation is significant

resulting in a 3× standard-deviation/median (3σ/μ) of 15% over half a wafer. Finally, WID

variation for identical structures is relatively small (3σ/μ ≈ 3.5%). WID spatial correlation

of RO frequency is significant and shows a dependency on the direction of spacing and the

orientation of the gates. Each of these observations can be attributed to a particular step

in the manufacturing process.

In 45nm, systematic layout-induced variations, in particular those related to poly-

Si density, are significantly reduced compared to the early 90nm process. Variations due to

strain and other factors are now dominant. Finally, random WID variation has increased
 102

90nm 45nm
Proximity Effects 10% 2%
D2D (3σ/μ) 15% 15%
WID (3σ/μ) 3.5% 6.6%
WID Spatial Correlation Significant Insignificant

Table 6.8: Comparison of 45nm and 90nm technology results.

proportionally to transistor area reduction, while systematic D2D variation has remained

the same. Table 6.8 compares the results of the two technologies.
 103

Chapter 7

Conclusion

Two test-chips have been designed as part of a methodology to measure and an-

alyze variability in CMOS processes. These chips contain circuits to measure an array of

RO and leakage transistors. The array is designed to investigate the impacts of layout on

transistor performance and allows for the measurement of systematic, random, D2D and

WID variations. This methodology was successfully applied to a 90nm and a 45nm pro-

cess. Several circuits for measuring SRAM performance variability have also been proposed.

These circuits permit the measurement of static and dynamic performance metrics of the

SRAM cell in an array with minimum perturbation to the array layout.

7.1 Key Contributions

The contributions of this work is two-fold. Firstly, we have developed a method-

ology to measure and analyze systematic layout dependent effects and both systematic

and random WID and D2D variations in the early stages of a CMOS process. This was
 104

achieved with a custom-designed test-chip containing an addressable array of devices under

test (DUT) and accurate on-chip measurement circuits. The DUT consists of ring oscillators

(RO) and transistors in the off-state with different layout configurations that investigate

layout effects. The array is made up of small scalable modules that easily abut to form the

big array. Scan chains are used to address the array, resulting in a small pad count. RO

frequency is measured off-chip after dividing the frequency down with on-chip dividers. An

on-chip current measurement circuit in the form of a single slope ADC was used to measure

leakage currents of the off-transistors with good accuracy. Finally, a complex setup using

a custom-designed PCB, easily available instruments, and GPIB connection was used to

automate the data collection.

Secondly, analysis of the measurement data shows that systematic layout depen-

dent effects are significant in 90nm. In 45nm, layout effects are weaker and dominated by

strain. Other less known systematic effects were also revealed. In 90nm, rotated structures

have smaller variation, and symmetrical structures that are mirror image of each other ex-

perience a small but systematic shift in frequency. Metal coverage over gate causes Vth to

increase resulting in lower ILEAK . In 45nm, strain effects due to an increase in source/drain

area dominate the systematic effects but overall effects are small. Effects due to STI strain

and poly-Si pitch are within 2%.

Characteristics of D2D and WID variations were also measured and analyzed. In

90nm, D2D variation was significant and random WID variation was small, whereas in

45nm, D2D variation has remained approximately the same and random WID variation has

increased, likely due to the reduction in channel area. The systematic component of D2D
 105

variation is important and exhibits a radial pattern on the wafer.

7.2 Methods to Mitigate Variability

Layout has a significant impact on lithography-induced variability in 90nm tech-

nology. Analysis of the inferred process parameters indicates that poly-Si gate pitch has

the strongest impact on the effective transistor gate length, resulting in up to 10% variation

in RO frequency for inverters laid out with single poly-Si lines. This systematic effect can

be compensated by using layout extraction tools that account for proximity effects [76] or

better OPC [92]. A simpler method would be to use regular layouts [93] or more restrictive

layout rules that only allow a few possible gate pitches, together with an extractor that

maps each gate pitch to its respective gate length.

The use of step-and-scan lithography induces stronger correlation between gates

placed parallel to the direction of slit of light than those parallel to the direction of scan.

This effect can be exploited in the layout of regular datapaths and memory [8, 94]. By

placing gates of the same path in the low correlation direction and by placing parallel paths

in the high correlation direction, a tighter performance spread can be obtained.

Analysis of the spatial correlation of ROs and inverters can be used to model the

spatial correlation coefficient of circuit paths and interconnected gates, allowing for more

accurate statistical timing analysis.

For the same layout, systematic D2D variations dominate the total variations for

small chips. Knowing the contribution of WID and D2D variation to the total variation will

allow circuit designers to design for WID variations and use die-level compensation schemes
 106

[32] to reduce D2D variations. Process engineers will also be able to focus on reducing the

dominant sources of variations.

More restricted design rules and likely better OPC in 45nm process reduce the

layout induced performance variations by limiting variations in poly-Si density. Only 2%

frequency shifts due to proximity effects were measured in the 45nm test-chips. However,

other layout-induced variations due to the CESL and STI stress have become more signifi-

cant and are added on to the total variation.

Measurements show a trend toward less systematic but more random variations.

From 90nm to 45nm, random WID variation has more than doubled whereas systematic

D2D variation has remained unchanged. One way to reduce random variations is to increase

channel area. As this increases area and power, only critical gates should be wider. Also,

averaging more gates in a critical path will reduce the uncertainty over the path delay.

7.3 Future Work

The existing methodology can be refined to measure a larger range of current.

Instead of measuring leakage current only, the array can be modified to make I-V measure-

ments which can be used to obtain more accurate parameter extraction.

Investigating other layout effects such as well-proximity effects, gate width vari-

ation and alignment errors can yield interesting results. TCAD tools that simulate the

gate formation process can be used to match measurement results to predicted behavior.

Layouts that are sensitive to certain aspects of lithography such as focus or aberrations can
 107

be used to monitor these effects1 .

This work focus on the variability of the transistor. However, interconnects vari-

ability can also have a significant impact on performance. Variability of the backend process

can be measured and analyzed using the RO array. New methods can also be developed to

measure the capacitance and resistance of wires and vias.

A methodology for SRAM measurement can be developed based on the circuits

presented in Chapter 5. These circuits measure the static and dynamic noise margins of the

SRAM cell in the array with minimum perturbation of the array layout. Measurements from

the SRAM write RO and the SRAM read RO correlate with access times measurements

but are simpler, more accurate, and faster to measure2 .

7.4 Conclusion

In conclusion, a methodology has been successfully applied, to characterize vari-

ability in a 90nm and a 45nm CMOS process. Sources of systematic variations have been

identified through analysis of the measurement results. Systematic and random components

of D2D and WID variations have been measured. Several new layout-dependent effects and

novel circuits for characterizing SRAM performance variability have been proposed for fu-

ture investigation. This will setup a framework for characterizing variability and help in

the mitigation of its effects.

1
Work in progress by Lynn Wang, U.C. Berkeley.
2
Work in progress by Jason Tsai, U.C. Berkeley.
 108

Bibliography

[1] A. B. Kahng and Y. C. Pati, “Subwavelength lithography and its potential impact
on design and eda,” in Proc. Design Automation Conference, New Orleans, LA, USA,
June 1999, pp. 799–804.
[2] E. Josse, S. Parihar, O. Callen, P. Ferreira, C. Monget, A. Farcy, M. Zaleski, D. Vil-
lanueva, R. Ranica, M. Bidaud, D. Barge, C. Laviron, N. Auriac, C. L. Cam, S. Har-
rison, S. Warrick, F. Leverd, P. Gouraud, S. Zoll, F. Guyader, E. Perrin, E. Bay-
lac, J. Belledent, B. Icard, B. Minghetti, S. Manakli, L. Pain, V. Huard, G. Ribes,
K. Rochereau, S. Bordez, C. Blanc, A. Margain, D. Delille, R. Pantel, K. Barla,
N. Cave, and M. Haond, “A cost-effective low power platform for the 45-nm tech-
nology node,” in 2006 IEEE Int. Electron Devices Meeting Tech. Dig., San Francisco,
CA, USA, Dec 2006, pp. 1–4.
[3] C. Cam, F.Guyader, C. Buttet, P.Guyader, G.Ribes, M.Sardo, S.Vanbergue, F.Boeuf,
F.Arnaud, E.Josse, and M.Haond, “A low cost drive current enhancement technique
using shallow trench isolation induced stress for 45-nm node,” in 2006 Symp. VLSI
Technology Dig. of Tech. Papers, 2006, pp. 82–83.
[4] B. L. Gratiet, P. Gouraud, E. Aparicio, L. Babaud, K. Dabertrand, M. Touchet, S. Kre-
mer, C. Chaton, F. Foussadier, F. Sundermann, J. Massin, J.-D. Chapon, M. Gatefait,
B. Minghetti, J. de Caunes, and D. Boutin, “Process control for 45nm cmos logic gate
patterning,” in Metrology, Inspection, and Process Control for Microlithography XXII,
ser. Proc. SPIE, J. A. Allgair and C. J. Raymond, Eds., vol. 6922, March 2008.
[5] L.-T. Pang and B. Nikolić, “Impact of layout on 90nm process parameter fluctuations,”
in 2006 Symp. VLSI Circuits Dig. of Tech. Papers, June 2006, pp. 69–70.
[6] B. Nikolić and L.-T. Pang, “Measurements and analysis of process variability in 90nm
cmos,” in 8th Int. Conf. Solid-State and Integrated Circuit Technology, Shanghai,
China, 2006, pp. 505–508.
[7] L.-T. Pang and B. Nikolić, “Measurement and analysis of variability in 45nm strained-
Si CMOS technology,” in Custom Integrated Circuits Conference Dig. of Tech. Papers,
September 2008, accepted to CICC 2008.
[8] K. Bowman, S. Duvall, and J. Meindl, “Impact of die-to-die and within-die parameter
fluctuations on the maximum clock frequency distribution for gigascale integration,”
IEEE J. of Solid-State Circuits, vol. 37, no. 2, pp. 183–190, February 2002.
 109

[9] S. Nassif, “Delay variability: sources, impacts and trends,” in ISSCC 2000, Dig. of
Tech. Papers, February 2000, pp. 368 – 369.

[10] M. Orshansky, L. Milor, P. Chen, K. Keutzer, and C. Hu, “Impact of spatial intrachip
gate length variability on the performance of high-speed digital circuits,” IEEE Trans.
on Computer-Aided Design of Circuits and Systems, vol. 21, no. 5, pp. 544–553, May
2002.

[11] J. Cain and C. Spanos, “Electrical linewidth metrology for systematic cd variation
characterization and causal analysis,” in Metrology, Inspection, and Process Control
for Microlithography XVII, ser. Proc. SPIE, D. J. Herr, Ed., vol. 5038, June 2003, pp.
350–361.

[12] M. Orshansky, L. Milor, and C. Hu, “Characterization of spatial intrafield gate CD

variability, its impact on circuit performance, and spatial mask-level correction,” IEEE
Trans. on Semiconductor Manufacturing, vol. 17, no. 1, pp. 2–11, February 2004.

[13] P. Friedberg, Y. Cao, J. Cain, R. Wang, J. Rabaey, and C. Spanos, “Modeling within-
field gate length spatial variation for process-design co-optimization,” in Design and
Process Integration for Microelectronic Manufacturing III, ser. Proc. SPIE, L. W. Lieb-
mann, Ed., vol. 5756, May 2005, pp. 178–188.

[14] E. G. Colgan, R. J. Polastre, M. Takeichi, and R. L. Wisnieff, “Thin-film-transistor

process-characterization test structures,” IBM J. of Research and Development, vol. 42,
no. 3/4, May/July 1998.

[15] D. Boning, J. Panganiban, K. Gonzalez-Valentin, C. MdCowell, A. Gattiker, and

F. Liu, “Test structures for delay variability,” in 2002 ACM/IEEE Tau Workshop,
Monterrey, CA, USA, 2002.

[16] K. M. G.-V. Gettings, “Extraction of variation sources due to layout practices,” Mas-
ter’s thesis, Massachusetts Institute of Technology, Cambridge, Mass., 2002.

[17] J. S. Panganiban, “A ring oscillator based variation test chip,” Master’s thesis, Mas-
sachusetts Institute of Technology, Cambridge, Mass., 2002.

[18] X. Ouyang, C. N. B. Milor, and R. F. W. Pease, “High-throughput mapping of short-

range spatial variations using active electrical metrology,” IEEE Trans. on Semicon-
ductor Manufacturing, vol. 15, no. 1, pp. 108–117, February 2002.

[19] S. Ohkawa, M. Aoki, and H. Masuda, “Analysis and characterization of device varia-
tions in an lsi chip using an integrated device matrix array,” IEEE Trans. on Semicon-
ductor Manufacturing, vol. 17, no. 2, pp. 155–165, May 2004.

[20] K. Agarwal, F. Liu, C. McDowell, S. Nassif, K. Nowka, M. Palmer, D. Acharyya, and

J. Plusquellic, “A test structure for characterizing local device mismatches,” in 2006
Symp. VLSI Circuits Dig. of Tech. Papers, 2006, pp. 67 – 68.
 110

[21] K. Agarwal and S. Nassif, “Characterizing process variation in nanometer CMOS,” in

Proc. Design Automation Conference, San Diego, CA, USA, June 2007, pp. 396–399.

[22] D. Levacq, T. Minakawa, M. Takamiya, and T. Sakurai, “A wide range spatial fre-
quency analysis of intra-die variations with 4-mm 4000 1 transistor arrays in 90nm
CMOS,” in Custom Integrated Circuits Conference Dig. of Tech. Papers, September
2007, pp. 257 – 260.

[23] V. Wang and K. Shepard, “On-chip transistor characterisation arrays for variability
analysis,” Electronics Letters, vol. 43, pp. 806 – 807, July 2007.

[24] W. Zhao, Y. Cao, F. Liu, K. Agarwal, D. Acharyya, S. Nassif, and K. Nowka, “Rigorous
extraction of process variations for 65nm CMOS design,” in 37th European Solid State
Device Research Conference, 2007, September 2007, pp. 89 – 92.

[25] V. Wang, K. Agarwal, S. Nassif, K. Nowka, and D. Markovic, “A design model for
random process variability,” in 9th Int. Symp. on Quality Electronic Design, 2008,
March 2008, pp. 734 – 737.

[26] A. Bassi, A. Veggetti, L. Croce, and A. Bogliolo, “Measuring the effects of process
variations on circuit performance by means of digitally-controllable ring oscillators,”
in 2003 Int. Conf. on Microelectronic Test Structures, March 2003, pp. 214–217.

[27] B. Zhou and A. Khouas, “Measurement of delay mismatch due to process variations by
means of modified ring oscillators,” in 2005 IEEE Int. Symp. on Circuits and Systems,
vol. 5, May 2005, pp. 5246–5249.

[28] K. Tae-Hyoung, R. Persaud, and C. Kim, “Silicon odometer: An on-chip reliability

monitor for measuring frequency degradation of digital circuits,” IEEE J. of Solid-
State Circuits, vol. 43, pp. 874 – 880, April 2008.

[29] X. Ouyang, T. Deeter, C. N. Berglund, R. F. W. Pease, J. Lee, , and M. A. Mc-

Cord, “High-throughput high-density mapping and spectrum analysis of transistor gate
length variations in sram circuits,” IEEE Trans. on Semiconductor Manufacturing,
vol. 14, no. 4, pp. 318–329, November 2001.

[30] J. B. Kuang, J. C. Kao, H. C. Ngo, and K. J. Nowka, “Cascaded test circuit with inter-
bitline drive devices for evaluating memory cell performance,” U.S. Patent 7,349,271,
March, 2008.

[31] K. Bernstein, D. J. Frank, A. E. Gattiker, W. Haensch, B. L. Ji, S. R. Nassif, E. J.

Nowak, D. J. Pearson, and N. J. Rohrer, “High-performance CMOS variability in the
65-nm regime and beyond,” IBM J. of Research and Development, vol. 50, no. 4/5,
July/September 2006.

[32] J. Tschanz, J. Kao, S. Narendra, R. Nair, D. Antoniadis, A. Chandrakasan, and V. De,

“Adaptive body bias for reducing impacts of die-to-die and within-die parameter vari-
ations on microprocessor frequency and leakage,” IEEE J. of Solid-State Circuits,
vol. 37, no. 11, pp. 1396 – 1402, November 2002.
 111

[33] H. Masuda, S. Ohkawa, A. Kurokawa, and M. Aoki, “Challenge: variability charac-

terization and modeling for 65- to 90-nm processes,” in Custom Integrated Circuits
Conference Dig. of Tech. Papers, 2005, pp. 593–599.

[34] M. Orshansky, S. Nassif, and D. Boning, Design for Manufacturability and Statistical
Design: A Constructive Approach, ser. Integrated Circuits and Systems, A. Chan-
drakasan, Ed. Springer, 2007.

[35] C. Liu, F. Baumann, A. Ghetti, H. Vuong, C. Chang, K. Cheung, J. Colonell, W. Lai,

E. Lloyd, J. Miner, C. Pai, H. Vaidya, R. Liu, and J. Clemens, “Severe thickness
variation of sub-3nm gate oxide due to Si surface faceting, Poly-Si intrusion, and corner
stress,” in 1999 Symp. VLSI Technology Dig. of Tech. Papers, 1999, pp. 75–76.

[36] A. Asenov, S. Kaya, and J. H. Davies, “Intrinsic threshold voltage fluctuations in

decanano MOSFETs due to local oxide thickness variations,” IEEE Trans. Electron
Devices, vol. 49, no. 1, pp. 112 – 119, January 2002.

[37] M. Ercken, L. Leunissen, I. Pollentier, G. P. Patsis, V. Constantoudis, and E. Gogolides,

“Effects of different processing conditions on line edge roughness for 193nm and 157nm
resists,” in Metrology, Inspection, and Process Control for Microlithography XVIII, ser.
Proc. SPIE, R. M. Silver, Ed., vol. 5375, February 2004, pp. 266–275.

[38] P. Oldiges, Q. Lin, K. Petrillo, M. Sanchez, and M. H. M. Ieong, “Modeling line edge
roughness effects in sub 100 nanometer gate length devices,” in Proc. SISPAD, Seatle,
WA, USA, 2000, pp. 131–134.

[39] J. Croon, G. Storms, S. Winkelmeier, I. Pollentier, M. Ercken, S. Decoutere, W. Sansen,

and H. Maes, “Line edge roughness: characterization, modeling and impact on device
behavior,” in 2002 IEEE Int. Electron Devices Meeting Tech. Dig., San Francisco, CA,
USA, 2002, pp. 307–310.

[40] R. W. Keyes, “Scaling, small numbers and randomness in semiconductors,” IEEE

Circuits and Devices Magazine, vol. 10, pp. 28 – 31, May 1994.

[41] M. Eisele, J. Berthold, D. Schmitt-Landsiedel, and R. Mahnkopf, “The impact of intra-

die device parameter variations on path delays and on the design for yield of low voltage
digital circuits,” IEEE Trans. on Very Large Scale Integration (VLSI) Systems, vol. 5,
pp. 360 – 368, December 1997.

[42] P. Stolk, F. Widdershoven, and D. Klaassen, “Modeling statistical dopant fluctuations

in mos transistors,” IEEE Trans. on Electron Devices, vol. 45, no. 9, pp. 1960–1971,
September 1998.

[43] D. J. Frank, Y. Taur, M. Ieong, and H.-S. P. Wong, “Monte Carlo modeling of threshold
variation due to dopant fluctuations,” in 1999 Symp. VLSI Circuits Dig. of Tech.
Papers, 1999, pp. 171–172.
 112

[44] A. Asenov, G. Slavcheva, A. R. Brown, J. H. Davies, and S. Saini, “Quantum mechan-

ical enhancement of the random dopant induced threshold voltage fluctuations and
lowering in sub 0.1 micron MOSFETs,” in 1999 IEEE Int. Electron Devices Meeting
Tech. Dig., San Francisco, CA, USA, December 1999, pp. 535–538.

[45] A. Asenov, G. Slavcheva, A. R. Brown, J. H. Davies, , and S. Saini, “Increase in the ran-
dom dopant induced threshold fluctuations and lowering in sub-100 nm MOSFETs due
to quantum effects: A 3-D density-gradient simulation study,” IEEE Trans. Electron
Devices, vol. 48, no. 4, pp. 722 – 729, April 2001.

[46] V.-Y. A. Thean, M. Sadd, and B. J. White, “Effects of dopant granularity on superhalo-
channel mosfets according to two- and three-dimensional computer simulations,” IEEE
Trans. on Nanotechnology, vol. 2, pp. 97–101, June 2003.

[47] T. Mizuno, J. Okumtura, and A. Toriumi, “Experimental study of threshold voltage

fluctuation due to statistical variation of channel dopant number in MOSFET’s,” IEEE
Trans. Electron Devices, vol. 41, pp. 2216 – 2221, November 1994.

[48] M. Pelgrom, A. Duinmaijer, and A. Welbers, “Matching properties of MOS transis-

tors,” IEEE Journal of Solid-State Circuits, vol. 24, pp. 1433 – 1439, 1989.

[49] C. S. Smith, “Piezoresistance effect in germanium and silicon,” Phys. Rev., vol. 94,
no. 1, pp. 42 – 49, April 1954.

[50] P. Chidambaram, C. Bowen, S. Chakravarthi, and R. W. C. Machala, “Fundamentals of

silicon material properties for successful exploitation of strain engineering in modern
cmos manufacturing,” IEEE Trans. on Electron Devices, vol. 53, pp. 994–964, May
2006.

[51] E. Ungersboeck, S. Dhar, G. Karlowatz, V. Sverdlov, H. Kosina, and S. Selberherr,

“The effect of general strain on the band structure and electron mobility of silicon,”
IEEE Trans. on Electron Devices, vol. 54, pp. 2183 – 2190, September 2007.

[52] H. Park, K. Jones, J. Slinkman, and M. Law, “The effects of strain on dopant diffusion
in silicon,” in 1993 IEEE Int. Electron Devices Meeting Tech. Dig., San Francisco, CA,
USA, December 1993, pp. 303 – 306.

[53] S. Thompson, G. Sun, K. Wu, J. Lim, and T. Nishida, “Key differences for process-
induced uniaxial vs. substrate-induced biaxial stressed Si and Ge channel MOSFETs,”
in 2004 IEEE Int. Electron Devices Meeting Tech. Dig., San Francisco, CA, USA,
December 2004, pp. 221 – 224.

[54] S. Mohta, N.; Thompson, “Mobility enhancement,” IEEE Circuits and Devices Maga-
zine, vol. 21, pp. 18 – 23, Sept.-Oct. 2005.

[55] Z.-Y. Cheng, M. Currie, C. Leitz, G. Taraschi, E. Fitzgerald, J. Hoyt, and D. An-
toniadas, “Electron mobility enhancement in strained-Si n-MOSFETs fabricated on
SiGe-on-insulator (SGOI) substrates,” IEEE Electron Device Letters, vol. 22, pp. 321
– 323, July 2001.
 113

[56] T. Komoda, A. Oishi, T. Sanuki, K. Kasai, H. Yoshimura, K. Ohno, A. Iwai, M. Saito,

F. Matsuoka, N. Nagashima, and T. Noguchi, “Mobility improvement for 45nm node
by combination of optimized stress and channel orientation design,” in 2004 IEEE Int.
Electron Devices Meeting Tech. Dig., San Francisco, CA, USA, December 2004, pp.
217 – 220.

[57] J.-S. Lim, S. Thompson, and J. Fossum, “Comparison of threshold-voltage shifts

for uniaxial and biaxial tensile-stressed n-MOSFETs,” IEEE Electron Device Letters,
vol. 25, pp. 731 – 733, November 2004.

[58] K. Mistry, M. Armstrong, C. Auth, S. Cea, T. Coan, T. Ghani, T. Hoffmann,

A. Murthy, J. Sandford, R. Shaheed, K. Zawadzki, K. Zhang, S. Thompson, and
M. Bohr, “Delaying forever: Uniaxial strained silicon transistors in a 90nm cmos tech-
nology,” in 2004 Symp. VLSI Technology Dig. of Tech. Papers, June 2004, pp. 50 –
51.

[59] C.-H. Ge, C.-C. Lin, C.-H. Ko, C.-C. Huang, Y.-C. Huang, B.-W. Chan, B.-C. Perng,
C.-C. Sheu, P.-Y. Tsai, L.-G. Yao, C.-L. Wu, T.-L. Lee, C.-J. Chen, C.-T. Wang, S.-
C. Lin, Y.-C. Yeo, and C. Hu, “Process-strained si (pss) cmos technology featuring
3d strain engineering,” in 2003 IEEE Int. Electron Devices Meeting Tech. Dig., San
Francisco, CA, USA, December 2003, pp. 3.7.1 – 3.7.4.

[60] S. Pidin, T. Mori, R. Nakamura, T. Saiki, R. Tanabe, S. Satoh, M. Kase, K. Hashimoto,

and T. Sugii, “MOSFET current drive optimization using silicon nitride capping layer
for 65-nm technology node,” in 2004 Symp. VLSI Technology Dig. of Tech. Papers,
June 2004, pp. 54 – 55.

[61] V. Chan, R. Rengarajan, N. Rovedo, W. Jin, T. Hook, P. Nguyen, J. Chen, E. Nowak,

X.-D. Chen, D. Lea, A. Chakravarti, V. Ku, S. Yang, A. Steegen, C. Baiocco, P. Shafer,
H. Ng, S.-F. Huang, and C. Wann, “High speed 45nm gate length cmosfets integrated
into a 90nm bulk technology incorporating strain engineering,” in 2003 IEEE Int.
Electron Devices Meeting Tech. Dig., San Francisco, CA, USA, December 2003, pp.
3.8.1 – 3.8.4.

[62] S. Pidin, T. Mori, K. Inoue, S. Fukuta, N. Itoh, E. Mutoh, K. Ohkoshi, R. Naka-

mura, K. Kobayashi, K. Kawamura, T. Saiki, S. Fukuyama, S. Satoh, M. Kase, and
K. Hashimoto, “A novel strain enhanced CMOS architecture using selectively deposited
high tensile and high compressive silicon nitride films,” in 2004 IEEE Int. Electron De-
vices Meeting Tech. Dig., San Francisco, CA, USA, December 2004, pp. 213 – 216.

[63] S. Thompson, M. Armstrong, C. Auth, S. Cea, R. Chau, G. Glass, T. Hoffman,

J. Klaus, M. Zhiyong, B. Mcintyre, A. Murthy, B. Obradovic, L. Shifren, S. Sivakumar,
S. Tyagi, T. Ghani, K. Mistry, M. Bohr, and Y. El-Mansy, “A logic nanotechnology
featuring strained-silicon,” IEEE Electron Device Letters, vol. 25, pp. 191 – 193, April
2004.
 114

[64] C. Mazure and I. Cayrefourcq, “Status of device mobility enhancement through

strained silicon engineering,” in Proc. 2005 IEEE Int. SOI Conference, October 2005,
pp. 1 – 6.

[65] G. Giusi, F. Crupi, E. Simoen, G. Eneman, and M. Jurczak, “Performance and reliabil-
ity of strained-silicon nmosfets with sin cap layer,” IEEE Trans. on Electron Devices,
vol. 54, pp. 78 – 82, Jan 2007.

[66] K. Ota, K. Sugihara, H. Sayama, T. Uchida, H. Oda, T. Eimori, H. Morimoto,

and Y. Inoue, “Novel locally strained channel technique for high performance 55nm
CMOS,” in 2002 IEEE Int. Electron Devices Meeting Tech. Dig., San Francisco, CA,
USA, December 2002, pp. 27 – 30.

[67] R. W. Keyes, “Explaining strain - the positive and negative effects of elastic strain in
n-silicon,” IEEE Circuits and Devices Magazine, vol. 18, pp. 36 – 39, September 2002.

[68] X. Wang, B. Cheng, S. Roy, and A. Asenov, “Simulation of strain enhanced variability
in nMOSFETs,” in 9th Int. Conference on Ultimate Integration of Silicon, 2008, March
2008, pp. 89 – 92.

[69] J.-S. Goo, Q. Xiang, Y. Takamura, F. Arasnia, E. Paton, P. Besser, J. Pan, and
L. Ming-Ren, “Band offset induced threshold variation in strained-Si nMOSFETs,”
IEEE Electron Device Letters, vol. 24, pp. 568 – 570, September 2003.

[70] H. Tuinhout, M. Pelgrom, R. P. de Vries, and M. Vertregt, “Effects of metal coverage

on mosfet matching,” in 1996 IEEE Int. Electron Devices Meeting Tech. Dig., San
Francisco, CA, USA, 1996, pp. 735–738.

[71] J. D. Plummer, M. D. Deal, and P. B. Griffin, Silicon VLSI Technology. Upper Saddle
River, NJ, USA: Prentice-Hall, 2000.

[72] A. K.-K. Wong, Resolution Enhancement Techniques in Optical Lithography.

Bellingham, WA, USA: SPIE Press, 2001.

[73] I. Matthew, C. E. Tabery, T. Lukanc, M. Plat, M. Takahashi, and A. Wilkison, “De-

sign restrictions for patterning with off-axis illumination,” in Optical Microlithography
XVIII, ser. Proc. SPIE, B. W. Smith, Ed., vol. 5754, May 2004, pp. 1574–1585.

[74] T. A. Brunner, “Impact of lens aberrations on optical lithography,” IBM J. of Research

and Development, vol. 41, pp. 57–67, Jan/Mar 1997.

[75] T. Brunner, D. Corliss, S. Butt, T. Wiltshire, C. P. Ausschnitt, and M. Smith, “Laser

bandwidth and other sources of focus blur in lithography,” in Optical Microlithography
XIX, ser. Proc. SPIE, D. G. Flagello, Ed., vol. 6154, March 2006, pp. 1574–1585.

[76] M. Choi and L. Milor, “Impact on circuit performance of deterministic within-die vari-
ation in nanoscale semiconductor manufacturing,” IEEE Trans. on Computer-Aided
Design of Circuits and Systems, vol. 25, no. 7, pp. 1350–1367, July 2006.
 115

[77] T. Kjellberg and R. Schatz, “The effect of stitching errors on the spectral characteristics
of dfb lasers fabricated using electron beam lithography,” J. Lightwave Technology,
vol. 10, no. 9, pp. 1256–1266, September 1992.

[78] D. Steele, A. Coniglio, C. Tang, and B. Singh, “Characterizing post exposure bake
processing for transient and steady state conditions, in the context of critical dimension
control,” in Metrology, Inspection and Process Control for Microlithography XVI, ser.
Proc. SPIE, D. Herr, Ed., vol. 4689, July 2002, pp. 517–530.

[79] I. Ahsan, O. Glushchenkov, R. Logan, E. J. Nowak, H. Kimura, G. Berg, J. Herman,

E. Maciejewski, A. Chan, A. Azuma, S. Deshpande, B. Dirahoui, G. Freeman, A. Ga-
bor, M. Gribelyuk, S. Huang, M. Kumar, K. Miyamoto, D. Mocuta, A. Mahorowala,
E. Leobandung, and H. Utomo, “Rta-driven intra-die variations in stage delay, and
parametric sensitivities for 65nm technology,” in 2006 Symp. VLSI Technology Dig. of
Tech. Papers, 2006, pp. 170–171.

[80] Y.-M. Sheu, K.-W. Su, S. Tian, S.-J. Yang, C.-C. Wang, M.-J. Chen, and S. Liu,
“Modeling the well-edge proximity effect in highly scaled MOSFETs,” IEEE Trans.
Electron Devices, vol. 53, no. 11, pp. 2792–279, November 2006.

[81] G. Eneman, M. Jurczak, P. Verheyen, T. Hoffmann, A. D. Keersgieter, and K. D.

Meyer, “Scalability of strained nitride capping layers for future cmos generations,” in
35th European Solid State Device Research Conference, 2005, September 2005, pp. 449
– 452.

[82] S. Eneman, P. Verheyen, R. Rooyackers, F. Nouri, L. Washington, R. Degraeve,

B. Kaczer, V. Moroz, A. D. Keersgieter, R. Schreutelkamp, M. Kawaguchi, Y. Kim,
A. Samoilov, L. Smith, P. Absil, K. D. Meyer, M. Jurczak, and S. Biesemans, “Layout
impact on the performance of a locally strained PMOSFET,” in 2005 Symp. VLSI
Technology Dig. of Tech. Papers, June 2005, pp. 22 – 23.

[83] G. Eneman, P. Verheyen, R. Rooyackers, F. Nouri, L. Washington, R. Schreutelkamp,

V. Moroz, L. Smith, A. D. Keersgieter, M. Jurczak, and K. D. Meyer, “Scalability of
the si/sub 1-x/ge/sub x/ source/drain technology for the 45-nm technology node and
beyond,” IEEE Trans. on Electron Devices, vol. 53, pp. 1647 – 1656, July 2006.

[84] T. Sato, T. Uezono, S. Hagiwara, K. Okada, S. Amakawa, N. Nakayama, and K. Masu,

“A MOS transistor-array for accurate measurement of subthreshold leakage variation,”
in 8th Int. Symp. on Quality Electronic Design, 2007, March 2007, pp. 21 – 26.

[85] K. M. G.-V. Gettings, “Study of cmos process variation by multiplexing analog char-
acteristics,” Ph.D. dissertation, Massachusetts Institute of Technology, Cambridge,
Mass., 2007.

[86] L.-T. Pang, J. B. Kuang, J. Sivagnaname, T. Nguyen, F. Gebara, J. Kao, G. Car-

penter, , and K. Nowka, “45nm low voltage SRAM characterization,” in 2007 Austin
Conference on Energy Efficient Design, 2007.
 116

[87] A. Bhavnagarwala, S. Kosonocky, C. Radens, K. Stawiasz, R. Mann, Q. Ye, and

K. Chin, “Fluctuation limits & scaling opportunities for CMOS SRAM cells,” in 2005
IEEE Int. Electron Devices Meeting Tech. Dig., San Francisco, CA, USA, Dec 2005,
pp. 659–662.

[88] Zhengguo, A. Carlson, L.-T. Pang, K. Duong, T.-J. K. Liu, and B. Nikolić, “Large-
scale read/write margin measurement in 45nm CMOS SRAM arrays,” in 2008 Symp.
VLSI Circuits Dig. of Tech. Papers, June 2008.

[89] J. Pille, C. Adams, T. Christensen, S. Cottier, S. Ehrenreich, T. Kono, D. Nelson,

O. Takahashi, S. Tokito, O. Torreiter, O. Wagner, and D. Wendel, “Implementation of
the CELL Broadband Engine in a 65nm SOI technology featuring dual-supply SRAM
arrays supporting 6GHz at 1.3V,” in ISSCC 2007, Dig. of Tech. Papers, February 2007,
pp. 322–606.

[90] R. A. Bianchi, G. Bouche, and O. R. dit Buisson, “Accurate modeling of trench isolation
induced mechanical stress effects on MOSFET electrical performance,” in 2002 IEEE
Int. Electron Devices Meeting Tech. Dig., San Francisco, CA, USA, Dec 2002, pp.
117–120.

[91] I. Ohkura, T. Noguchi, K. Sakashita, H. Ishida, T. Ichiyama, and T. Enomoto, “Gate

isolation - a novel basic cell configuration for CMOS gate arrays,” in Custom Integrated
Circuits Conference Dig. of Tech. Papers, May 1982, pp. 307–310.

[92] D. Beale, J. Shiely, L. M. III, and M. Rieger, “Advanced model formulations for optical
and process proximity correction,” in Optical Microlithography XVII, ser. Proc. SPIE,
B. W. Smith, Ed., vol. 5377, no. 12, May 2004, pp. 721–729.

[93] V. Kheterpal, V. Rovner, T. G. Hersan, D. Motiani, Y. Takegawa, A. J. Strojwas,

and L. Pileggi, “Design methodology for ic manufacturability based on regular logic-
bricks,” in Proc. Design Automation Conference, Anaheim, CA, USA, June 2005, pp.
353–358.

[94] H. Onodera, “Variability: Modeling and its impact on design,” IEICE Trans. Electron.,
vol. E89-C, no. 3, pp. 342 – 348, March 2006.
 117

Appendix A

Mathematics

A.1 Integrator equation derivations.

From Figure A.1, the following equations are obtained:

Vi Vi − Vo
+ = I (A.1)
ZCp1 ZC
Vo Vo Vo − Vi
Gm · Vi + + + = 0 (A.2)
Ro ZCp2 ZC

From A.2
1 1 1 1
Vi · (Gm − ) = −Vo ( + + ) (A.3)
ZC Ro ZCp2 ZC
From A.1
1 1 Vo
I = Vi ( + )− (A.4)
ZCp1 ZC ZC
Substitute A.3 into A.4
1 1 1
I Ro + ZCp2 + ZC 1 1 1
= 1 ( + )− (A.5)
Vo Gm − ZC
ZCp1 ZC ZC

After simplification, we get

1 1
Vo Gm− jwC
=  
I C+Cp1 C ·C +C(C +C )
1+ Gm ·Ro ·C + jw · p1 p2 C·Gmp1 p2
1 1
Vo Gm − jwC
⇔ = (A.6)
I 1 + A + jw · B
(A.7)

where
 118

C
Vi

I Cp1

Vo

GmVi Ro Cp2

Figure A.1: Small signal model of the integrator.

C+Cp1
A= Gm ·Ro ·C

Cp1 ·Cp2 +C(Cp1 +Cp2 )

B= C·Gm

Gm ≈ 10mS

Ro ≈ 1e5

C = 10pF ≥ Cp1 , Cp2

Thus,
A, B ≤ 1e − 3

and, 
1 1
Vo = ·I − · I · dt (A.8)
Gm C
where

Vo = integrator output voltage

Gm = transconductance gain of the amplifier for amplifier output voltage at Vo

I = integrated current

C = integrating capacitance
 119

t = time

A.2 Least squares estimate, error estimate.

In this section, error analysis for the least squares estimate(LSM) method to infer
the process parameters from measurement data is described.

In the LSM method,

Y =M ·X +E

where
⎡ ∂F1 ∂F1 ∂F1 ⎤
∂L ∂T ox ∂N ch
⎡ ⎤ ⎢ ⎥
F1 ⎢ ⎥
⎡ ⎤ ⎢ ∂I1 ∂I1 ∂I1 ⎥
⎢ I1 ⎥ ⎢ ∂L ∂T ox ∂N ch ⎥
⎢ ⎥ L ⎢ ⎥
⎢ F2 ⎥ ⎣ ⎦ ⎢ ⎥
Y =⎢ ⎥ , X = T ox , M = ⎢ ∂F2 ∂F2 ∂F2 ⎥,
⎢ I2 ⎥ ⎢ ∂L ∂T ox ∂N ch ⎥
⎣ ⎦ N ch ⎢ ⎥
.. ⎢ ⎥
. ⎢ ∂I2 ∂I2 ∂I2 ⎥
i ⎣ ∂L ∂T ox ∂N ch ⎦
.. .. ..
. . .
⎡ ⎤
eF1
⎢ eI1 ⎥
⎢ ⎥
⎢ eF2 ⎥
E=⎢ ⎥ ∼ N (0, Σ)
⎢ eI2 ⎥
⎣ ⎦
..
.

Y = Frequency and leakage current measurement data

X = Process parameters

M = Sensitivities

E = Errors associated with measurement. Assumed to be normally distributed with zero

mean and covariance matrix Σ

Given the measurement data Y and simulation data M , X is estimated with min-
imum error. The following steps describes the estimation process and how the estimation
errors can be obtained.

1. Whiten the noise

 120

In the general case, the noise may not be white which means that Σ may not be the
identity matrix(ID ). In order to minimize the error, the noise function E has to be
converted to white Gaussian noise. This is easily done using well-established methods
such as Graham-Schmitt decomposition.

If Σ = Σ1/2 · Σ1/2 , Y = Σ−1/2 · Y , M

 = Σ−1/2 · M , W
 = Σ−1/2 · E,

We now have Y = M
 · X + W where W ∼ N (0, ID )

2. Least means squares estimate

Using minimum least squares, an estimate of X is obtained.
⎡ ⎤

L −1
 =⎢
X ⎣ T
⎥  · M
ox ⎦ = M   · Y
·M

N ch


3. Covariance matrix of X
The error for each of the estimated process parameter is reflected in the covariance
 This is obtained as shown below.
matrix of X.
⎡ ⎤
V ar( 
L) Cov(  T
L, ox) Cov(  N
L,  ch)
−1 ⎢ ⎥
 = M
Cov(X)  · M
 = ⎣ Cov(T 
ox, L) V ar(Tox) Cov(T 
ox, N ch) ⎦ ,
Cov(N  Cov(N
ch, L) ch, Tox) V ar(Nch)

In the case of the 90nm measurements, we have

Y =M ·X +E

where E ∼ N(0, σ 2 ) and σ 2 is the variance of the white measurement noise.

We get
⎡ ⎤
−1
2.64 −7.44 −54.3
cov(X)  · M
= M  = ⎣ −7.44 23.5 149 ⎦ ,
σ2
−54.3 149 1155
We see that the error is smallest for L, ≈ 10 × bigger for T ox and ≈ 50 × bigger for N ch.
 121

Appendix B

Layouts

This appendix aims to provide information on the 90nm and 45nm test-chip designs
that would enable a circuit designer to examine the cadence libraries and understand the
structure of the designs.

B.1 Arrangements of the Layouts in a Tile.

Please refer to the cells testchip/bigtile for 90nm layouts and 45nmRO leak ST/bigtile2
for 45nm layouts. For testchip/bigtile (90nm), the floor-plan is given in Figure B.1. For
45nmRO leak ST/bigtile2 (45nm), the floor-plan is given in Figure B.2.

B.2 Library cells in 90nm.

This section describes the setup for the 90nm design kit and the location of the
90nm test-chip library. Several cells in the higher levels of hierarchy are described.

B.2.1 90nm Library Setup.

Sourced setup file: /vol/hitz/home/ltpang/cshrc.st90nm
Library paths are located in: /tools/designs/Dcdg/users/ltpang/project/st90nm/cds.lib

B.2.2 90nm Library.

Library Name: testchip

fullchip: Full chip minus the dummies.

padring2: Cell fullchip minus the sealring. This cell contains the I/O cells and their routing.

topcore: Contains the core of the chip. Main cells are tile array and topintbig2. Also con-
tains dummies.
 122

1a 2a 3a 4a 5a 6a

6b
1b 2b 3b 4b 5b

Figure B.1: Arrangement of the 12 layouts in a tile for the 90nm test-chip.
 123

A M2 over gate
min max2
1 9
17
mid1 max1

2 10

max1

3 11

max2 min
4 12

min max2 mid2

5 13

max2 min max2

6 14

min mid
7 15

max min
8 16

RO 1 Leakage 1 RO 9 Leakage 9

RO 2 Leakage 2 RO 10 Leakage 10
Bigtile2 (45nm)

RO 3 Leakage 3 RO 11 Leakage 11

Leakage 17
RO 4 Leakage 4 RO 12 Leakage 12
RO 17

RO 5 Leakage 5 RO 13 Leakage 13

RO 6 Leakage 6 RO 14 Leakage 14

RO 7 Leakage 7 RO 15 Leakage 15

RO 8 Leakage 8 RO 16 Leakage 16

Figure B.2: Arrangement of the 17 layouts in a tile for the 45nm test-chip.
 124

tile array: Contains array row.

array row: Contains a row of 4bigtiles and row dividers.

4bigtiles: Contains 4 tiles of RO and ILEAK transistors and a VRO tile.

bigtile: Contains tile1, tile2, tile3, tile4, tile5, tile8. tile1-5 correspond to layouts 1a/b to
5a/b and tile8 corresponds to layout 6a/b.

vro all: VRO tile.

topintbig2: Contains the integrators and logic for current measurement.

integrp: Integrator for PMOS current measurement.

integrn: Integrator for NMOS current measurement.

Library Name: finalchip

DUM fullchip: Full chip with dummy fills. This is the one that is taped out.

fullchip DUMMIES: Dummy fills for the whole chip.

fullchip: The full 90nm test-chip without dummy fills

B.3 Library cells in 45nm.

This section describes the setup for the 45nm design kit and the location of the
45nm test-chip library for the RO and leakage transistor array and the leakage measurement
circuits. Several cells in the higher levels of hierarchy are described.

B.3.1 45nm Library Setup.

sourced setup file: /tools/designs/Dcdg/users/ltpang/st45nm2/st45nm2/.cshrc cmos045
Library paths are located in: /tools/designs/Dcdg/users/ltpang/st45nm2/st45nm2/cds.lib

B.3.2 45nm Library.

Library Name: 45nmRO leak ST

top array leakmeas: Contains the RO and leakage transistor array with the leakage current
measurement circuits.
 125

top array leakmeas lvs: Contains the leakage measurement circuit and routing.

Ileakmeas: Contains the 2 integrators, multiplexers and routing.

integrp final: Integrator for PMOS current measurement.

integrn final: Integrator for NMOS current measurement.

row18x16 20: Contains the RO and leakage array made up of 18 rows of row1x16 20 and
the row multiplexer.

row1x16 20: Contains the a row of tiles and scan chain flip-flops at one end of the row and
multiplexers at the other end.

tile1x16 20: Contains 16 bigtile2 in a row.

bigtile2: contains tile1-9, tile12-14, tile20-23. tile1-6 correspond to layout1-6 in Figure ??,
tile20-21 to layout7-8, tile7-9 to layout9-11, tile12-14 to layout12-14, tile22-23 to layout15-
16, and tile vert to layout17.

Library Name: st45 padring

padring4: The 45nm padring.

 126

Appendix C

Test Setup

The chips are packaged in ceramic PGA packages and measured through a socket
mounted on a custom PCB. The PCB contains components that generate regulated supply
voltages and analog bias currents from external 5V and 10V supplies. Control signals are
sourced from a pattern generator(Tektronix DG2020A, Figure C.1) and accurate reference
currents are generated with a source-meter (Keithley 2400, Figure C.2). A 20 GS/s oscillo-
scope (Agilent Infiniium, Figure C.3) is used to make frequency measurements and a 150ps
resolution logic analyzer(Tektronix TLA5202, Figure C.4) is used to measure the delay be-
tween the start and stop signals of the ADC. In order to improve accuracy and reduce the
impact of noise, several hundred measurements are averaged for each data point.
Perl scripts and small programs in C are written to automate the process of data
collection. These scripts send commands to the instruments using a LAN/GPIB interface
(National Instruments GPIB-ENET/100, Figure C.5) and write measurement data onto a
file.
This appendix will briefly describe a few instruments used to carry out the mea-
surement procedure. Section C.1 gives a few pointers to how the GPIB-ENET/100 box is
set up. Section C.2 describes how the pattern generator is programmed to automate data
collection and section C.3 describes the procedures and set-up for data collection.

C.1 GPIB-ENET/100 Guide

Several instruments connect to each other via GPIB interface to automate the
measurement process. A computer sends out commands to the GPIB interface via the LAN
network. This is done via a GPIB-ENET/100 box from National Instruments shown in
Figure C.5. This section describes how the GPIB-ENET/100 box is set up.

1. Register the GPIB-ENET/100 box with Brad or Kevin. Obtain the following infor-
mation from them.

hostname: nigpib-3.EECS.Berkeley.EDU
IP address: 128.32.62.232
netmask: 255.255.255.0
 127

Figure C.1: DG2020A data generator from Tektronix.

 128

Figure C.2: Keithley 2400 source meter.

Figure C.3: Infiniium 54855A DSO 20GSPS oscilloscope from Agilent.

 129

Figure C.4: TLA5202 logic analyzer from Tektronix.

Figure C.5: GPIB-ENET/100 box from National Instruments.

 130

gateway: 128.32.62.2
DNS: 128.32.62.21

2. See chapter 3 of National Instruments GPIB Manual ’Getting Started with Your
GPIB-ENET/100 and NI-488.2 for Windows 2000/NT’.
Select ’Start>>Programs>>NI-488.2M for Windows NT>>
GPIB-ENET100 Utilities>>Device Configuration’

If the device is registered, you’ll see it (verify Ethernet address). It should have an
IP address/hostname. Currently the devices that have been registered are

128.32.62.232 00:80:2f:10:0b:5f 010663FE 1st setup

128.32.62.245 00:80:2f:10:0a:75 01065A05 Louis using
128.32.62.250 00:80:2f:10:0a:70 0106599A 2nd setup

If device is not configured, click properties and fill in the information according to
what is provided in 1.

3. See chapter 4. Select ’Start >>Settings>>Control Panel’ and double click the GPIB
icon. The GPIB configuration dialog box appears. Select GPIB0, click Board Type
and select GPIB-ENET/100, click ’ok’, Click ’Configure’ and fill in the IP address of
the device you want to control (eg. 128.32.62.250). Click ’ok’.

4. See chapter 5. Select ’Diagnostic’ item under ’Start>>Programs>>NI-488.2M’ for

Windows NT. Click ’Test All’. If all is good, it will indicate ’passed’ under ’Status’.

C.2 DG2020A Pattern Generator

The DG2020A data generator from Tektronix (Figure C.1) is used to provide the
digital input signals for the inputs and clocks of the scan chains, the control signals for the
current integrator and a few other control signals. This is described in Section C.3. This
section will briefly describe how the DG2020A is set up and programmed to automate the
collection of data. More information can be found in the DG2020A manual.

C.2.1 Configuring the DG2020A

1. Set to enhanced mode

2. Set voltage levels to the chip’s I/O pads

3. Set pod assignments

4. Set oscillation frequency

 131

C.2.2 Programming the DG2020A

1. putgrp GRPDEF.DAT
Defines the groups of signals and assigns them to the output pins. GRPDEF.DAT is a
text file that contains the assignments. An example of the contents of GRPDEF.DAT
is given below. p1a is assigned to pins 0 - 0, sclk to 7-7, etc.
sclk,7,7
nsel,6,6
incol,5,5
p2,4,4
inrow,3,3
p1b,2,2
p1,1,1
p1a,0,0

2. putword shiftblk.dat
Defines the bit pattern. An example of the contents of shiftblk.dat is given below. The
first line starts with a # symbol and indicate the number of lines in the bit pattern.
The second line also starts with a # symbol and indicate the starting address as
0. The next 64 lines show how the bit pattern changes over 64 cycles. There are
40 channels by default in the DG2020A and they are represented by 10 hexadecimal
numbers. Each hexadecimal number represents 4 bits of digital output. The example
below shows the last 8 outputs changing which is sufficient for the 45nm test-chip.
# size 64
# start 0
00,00,00,00,45
00,00,00,00,C5
00,00,00,00,45
00,00,00,00,45
|
|
00,00,00,00,45

3. putblk BLKDEF.DAT
Names the portion of waveform (block). In eg below, 0-63 time steps form a block
called SHIFT, 64-127 form INCOL1 and so on.
0,SHIFT
64,INCOL1
128,INROW2
192,COL2ROW
256,COL1ROW
320,REST
 132

384,NONE
448,MNMOS
512,MBNMOS
576,INROW2CK
640,INROW1
704,COL1ROW1

4. putseq file.dat
Defines the pattern that is to be used. In eg below, start by sending the block SHIFT
400 times, followed by INCOL1 1 time, and so on until MNMOS 1 time and that is
repeated until the DG2020A is stopped.
SHIFT,400,0,0,0,0
INCOL1,1,0,0,0,0
INROW1,1,0,0,0,0
MNMOS,1,0,0,0,1

The contents of the input file must express the following in ASCII with one step per
line.

• block name
• repeat count (1 to 65536)
• line number of the event jump destination
• trigger wait on/off setting(ON=1, OFF=0)
• event jump on/off setting (ON=1, OFF=0)
• infinite loop on/off setting (ON=1, OFF=0)

This format corresponds to the parameter block supplied to the DATA:SEQUENCE:DEFINE

command with the delimiter codes replaced by the newline code used in normal text
files. The first line in this file corresponds to sequence line number 0, and each fol-
lowing line to the sequence line number incremented by 1. See the seqdef.dat sample
data file below.
(seqdef.dat)
LEAD,1,5,OFF,0,0
LOOP1,256,5,OFF,ON,0
LOOP2,64,5,0,ON,0
LOOP1,256,5,ON,1,0
SUB1,1,5,ON,1,0
SUB2,1,5,ON,1,0
TRAIL,1,5,ON,0,0
TRAIL,1,5,0,0,0
 133

Pattern Generator
Results
LAN/GPIB
Logic Analyzer,
150ps resolution

LAN/GPIB

Test- chip: PGA package

mounted on a socket 20GSPS oscilloscope

Figure C.6: Test setup for collecting data.

C.3 Measurement Setup and Procedures

Figure C.6 shows the setup for data collection. The pattern generator generates
the address bits and the control signals while the logic analyzer measures the leakage current
integration time and the oscilloscope measures the frequency of the ROs. Commands are
sent to the instruments via the GPIB and LAN interfaces.
The PCB boards for measuring the 90nm and the 45nm test-chips are shown in
figure C.7 and figure C.8 respectively. The packaged chip is mounted on a socket which
is soldered onto the PCB. The socket has a lever that locks the packaged chip in place or
releases it. This allows for easy changing of chips and facilitates the measurement of many
chips. Components on the board include voltage regulators, current sources, capacitances,
resistances, inductances, connectors, pins, switches, etc.
In the BWRC laboratory, two work spaces are used to shorten measurement time
for the 45nm test-chips. One is used for leakage current measurement and the other for RO
frequency measurement. Figure C.9 shows the test-bench for leakage current measurement.
The TLA5202 is also a PC with its own hard disk and LAN connections. The Perl script is
executed from it and the data and setup files are stored in its disk space. In the setup for
RO frequency measurement, a PC is used instead and only the oscilloscope, the DG2020A
and the supplies are required.
 134

Figure C.7: PCB board for measuring the 90nm test-chips.

 135

Figure C.8: PCB board for measuring the 45nm test-chips.

 136

Figure C.9: Test setup for collecting leakage measurement data for the 45nm test-chips.