UCLA

UCLA Electronic Theses and Dissertations

Title
High Frequency Multiphase Clock Generation Using Multipath Oscillators and Applications

Permalink
https://escholarship.org/uc/item/75g8j8jt

Author
Abou-El-Sonoun, Amr Amin

Publication Date
2012

Peer reviewed|Thesis/dissertation

eScholarship.org Powered by the California Digital Library

University of California
 UNIVERSITY OF CALIFORNIA

Los Angeles

High Frequency Multiphase Clock Generation Using

Multipath Oscillators and Applications

A dissertation submitted in partial satisfaction

of the requirements for the degree

Doctor of Philosophy in Electrical Engineering

By

Amr Amin Abou-El-Sonoun

2012
 © Copyright by

Amr Amin Abou-El-Sonoun

2012
 ABSTRACT OF THE DISSERTATION

High Frequency Multiphase Clock Generation Using

Multipath Oscillators and Applications

by

Amr Amin Abou-El-Sonoun

Doctor of Philosophy in Electrical Engineering

University of California, Los Angeles, 2012

Professor Chih-Kong Ken Yang, Chair

Oscillators and frequency dividers are core building blocks in communications systems and

processors used to provide proper synchronization for the flow of information. Different

variations of the conventional ring oscillator that involve coupling of different oscillator-stages

or different oscillators have been introduced to provide multiple-phases without penalizing the

oscillation frequency. These variations, known as multipath ring oscillators, enable system

designers to relax the performance-power trade-off through parallelism. These oscillator

ii
structures, however, introduce additional degrees of freedom and expand the design space

considerably which makes the process of designing them optimally a very difficult task.

This dissertation introduces an accurate analytical model and comprehensive analysis for

multipath ring oscillators and frequency dividers. The results of the analysis are incorporated into

an optimization algorithm that allows a designer to arrive at the desired optimal design at a very

short time. The analysis explains the factors that affect the different performance metrics

including the number of phases, oscillation frequency, phase noise, and oscillation-mode

stability.

As an example application, a 48 Gb/s serializing transmitter is designed in 65nm CMOS

technology using superharmonic injection-locked multipath ring oscillators to generate

multiphase sampling clock signals for the various stages of the serializer. The ability to generate

multiple clock phases at relatively high frequencies and low power cost allows significant power

and area savings in the overall Transmitter.

iii
The dissertation of Amr Amin Abou-El-Sonoun is approved.

Milos Ercegovac

Dejan Markovic

Sudhakar Pamarti

Chih-Kong Ken Yang, Committee Chair

University of California, Los Angeles

2012

iv
To my parents

v
 Table of Contents

ABSTRACT OF THE DISSERTATION ............................................................................... ii 

Table of Contents .................................................................................................................... vi 

Acknowledgments ................................................................................................................... ix 

VITA ......................................................................................................................................... xi 

CHAPTER 1 Introduction ................................................................................................... 1 

1.1. Examples of Multiphase Clocking Applications .......................................................... 2 

1.2. Motivation .................................................................................................................... 3 

1.3. Thesis Organization ...................................................................................................... 7 

CHAPTER 2 Background .................................................................................................... 9 

2.1. Oscillators Basics ......................................................................................................... 9 

2.2. Multipath Ring Oscillators ......................................................................................... 12 

2.3. Multiphase Frequency Dividers.................................................................................. 16 

2.4. Serial Link Transmitters ............................................................................................. 19 

CHAPTER 3 Multipath Ring Oscillators ......................................................................... 25 

3.1. Oscillator Modeling .................................................................................................... 25 

3.1.1.  General Form of the Multipath Ring Oscillator .................................................. 25 

3.1.2.  Linear Delay Stage Model................................................................................... 27 

3.1.3.  Nonlinear Delay Stage Model ............................................................................. 29 

3.2. Oscillation Frequency and Amplitude ........................................................................ 29 

vi
 3.2.1.  Oscillation Modes ............................................................................................... 29 

3.2.2.  Oscillation Amplitude and Nonlinear Time Constant ......................................... 31 

3.2.3.  Oscillation Frequency ......................................................................................... 37 

3.2.4.  Maximum Oscillation Frequency ........................................................................ 41 

3.2.5.  Mode Stability ..................................................................................................... 47 

3.3. Phase Noise Analysis.................................................................................................. 51 

3.4. Conclusion .................................................................................................................. 61 

CHAPTER 4 Superharmonic Injection-Locked Multipath Ring Oscillators ............... 62 

4.1. General Model ............................................................................................................ 62 

4.2. Locking Ranges .......................................................................................................... 69 

4.2.1.  Phase-Limited Locking Range ............................................................................ 71 

4.2.2.  Gain-Limited Locking Range.............................................................................. 72 

4.2.3.  Mode-Limited Locking Range ............................................................................ 74 

4.2.4.  Quadrature-Phase Divider Example .................................................................... 75 

4.3. Power Consumption and Effect of Loading ............................................................... 78 

4.4. Design Procedure ........................................................................................................ 81 

4.5. Octal-Phase Divider.................................................................................................... 84 

4.6. Hexadecimal-Phase Divider ....................................................................................... 87 

4.7. Conclusion .................................................................................................................. 88 

CHAPTER 5 A 32-48 Gb/s Serial Link Transmitter ....................................................... 90 

vii
 5.1. Transmitter System Design ........................................................................................ 90 

5.1.1.  Final Multiplexing Stage Order........................................................................... 91 

5.1.2.  Multiphase Sampling........................................................................................... 92 

5.1.3.  Transmitter Architecture ..................................................................................... 94 

5.1.4.  On-Chip Eye Monitor.......................................................................................... 95 

5.2. Circuit Details ............................................................................................................. 97 

5.2.1.  Quadrature LC-VCO ........................................................................................... 97 

5.2.2.  Frequency-Divider Chain .................................................................................... 99 

5.2.3.  Phase/Frequency Detector ................................................................................. 100 

5.2.4.  Charge-Pump ..................................................................................................... 102 

5.2.5.  Multiplexer Design ............................................................................................ 102 

5.2.6.  Comparator Design ........................................................................................... 106 

5.3. Conclusion ................................................................................................................ 108 

CHAPTER 6 Experimental Results ................................................................................ 110 

6.1. Testing Setup ............................................................................................................ 110 

6.2. Measurement Results................................................................................................ 111 

CHAPTER 7 Conclusion .................................................................................................. 117 

Appendix Oscillation Modes in Multipath LC Oscillators ............................................ 120 

References ............................................................................................................................. 123 

viii
 Acknowledgments

I would first like to thank my advisor, Professor Ken Yang, for the continuous help, support,

and guidance that he gave to me throughout this journey. I find it really hard to express my

gratitude for all the time he spent with me discussing my ideas, reviewing my papers, listening to

my design problems, and always being ready with advice and the right feedback.

I would also like to thank my dear friend and colleague, Ming-Shuan Chen, who without his

help and great effort, I would have never been able to get my chip out in time and get it back

working.

I spent very enjoyable three months in Maxim-IC in Hillsboro, OR working with a group of

very talented RFIC designers whom I learnt so much from. I would like to thank Scott Williams,

Joel Birkeland, Radu Fetche, Alexei Shatalov, Susan Tibbettes, and my dear friend Farhad

Farahbakhshian for this wonderful experience. I would also like to thank Prof. Robert Meyer of

UC Berkeley who I was lucky to work with and learn from during my internship at Maxim-IC.

I consider myself so lucky having got the chance to spend the Summer of 2010 working in

Broadcom with one of the best RFIC design teams in the world. It was a great experience

working with Hooman Darabi, Mohyee Mikhemar, Ahmad Mirzaei, and my friend, David

Murphy. Thank you all guys for this great experience.

During my stay at UCLA, I was blessed to be among a group of wonderful people who made

my stay in Los Angeles a wonderful memory and who were always there for me, helping me get

over every technical and non-technical obstacle I have met. Thanks go to my dear friends, Ramy,

ix
Sameh, Tamer, Ismail, Omar, Aboudina, Henry, Wonho, Preeti, Michael, Wessam, Karam, Said,

Wael, Mantawy, Sherif, Hussein, Yousr, Bassel, Hamza, and Yasser.

I would finally like to express my greatest gratitude toward my family; my parents, my

brother, and my little sister for the unlimited love and support they gave to me throughout my

entire life without which I would have never been even close to where I am today.

x
 VITA

2004 B.Sc, Electrical Engineering, Ain Shams University, Egypt

Teaching Assistant, Electronics and Electrical Communications Department,

2004-2008
Ain Shams University, Egypt

2007-2008 Senior Member of Technical Staff, Silicon Vision, Cairo, Egypt

2008 M.Sc, Electrical Engineering, Ain Shams University, Egypt

Research and Teaching Assistant, Electrical Engineering Department,

2008-2012
University of California, Los Angeles, California, USA

2009 RFIC Design Intern, Maxim Integrated Products, Hillsboro, Oregon, USA

2010 RFIC Design Intern, Broadcom, Irvine, California, USA

PUBLICATIONS

Amr A. Hafez, Ming-Shuan Chen, and C.-K. Ken Yang, “A Multi-Phase Multi-frequency
Clock Generator Using Superharmonic Injection Locked Multipath Ring Oscillators as
Frequency Dividers,” Accepted in IEEE Asian Solid State Circuits Conference (ASSCC), Nov.
2012.

Amr A. Hafez, C.-K. Ken Yang, “Analysis and Design of Superharmonic Injection-Locked
Multipath Ring Oscillators,” Submitted to IEEE Trans. Circuits and Systems—I: Regular
Papers.

xi
 Amr A. Hafez, C.-K. Ken Yang, “Design and Optimization of Multipath Ring Oscillators,”
IEEE Trans. Circuits and Systems—I: Regular Papers, vol. 58, no. 10, Oct. 2011, pp. 2332-
2345.
David Murphy, Amr A. Hafez, Ahmad Mirzaei, Mohyee Mikhemar, Hooman Darabi, M.-C.
Frank Chang, Asad Abidi, “A Blocker-Tolerant Wideband Noise-Cancelling Receiver with a
2dB Noise Figure,” ISSCC, Feb. 2012, pp. 74-75.

Tamer A. Ali, Amr A. Hafez, Robert Drost, Ronald Ho, C.-K. Ken Yang, “A 4.6GHz MDLL
with -46dBc Reference Spur and Aperture Position Tuning,” ISSCC, Feb. 2011, pp. 466-467.

xii
 CHAPTER 1

Introduction

Oscillators and frequency dividers are core building blocks in communication systems and

processors to provide proper synchronization of the flow of information.

A multiphase clock generator is a core building block in many applications that allows a

system designer to relax the performance-power trade-off by exploiting parallelism [1].

Numerous examples of applications that use multiphase clock generators exist. Some are

discussed in Section 1.1. In such applications, the benefit of parallelism is largely based on the

assumption that multiphase generation can be achieved efficiently and with low overhead.

A common approach to generating multiphase clocks is to use either ring oscillators or

frequency divided outputs of an LC oscillator with high quality factor and low phase noise.

Complex structures known as multipath oscillators that involve coupling of multiple oscillators

and oscillator stages have been introduced in order to improve oscillator performance. These

structures, however, introduce additional degrees of freedom and expand the design space

considerably.

This dissertation introduces an analytical model and comprehensive analysis for multipath

ring oscillators and frequency dividers to aid designers in arriving at a near optimal design.

Section 1.2 of this chapter motivates this analysis by describing the design complexity of this

type of oscillators and dividers. The organization of the dissertation is then presented in Section

1.3.

1
 1.1. Examples of Multiphase Clocking Applications

Any application that involves the interleaving of multiple paths requires multiple clock

phases. Common examples include interleaved data-converters and sub-rate clock and data

recovery systems. Fig. 1 shows an example where a 4-way interleaved analog-to-digital

converter is used to achieve a sampling rate 4 times higher than that of a single unit converter.

Fig. 1.1. A four-way interleaved analog-to-digital converter.

Another important application where a multiphase clock is needed is software-defined radios.

These radios are expected to operate in different frequency-bands making it very expensive to

achieve band selection using off-chip filters. To prevent signals from out of the desired band

from interfering with the desired signal, a receiver has to use multiple signal paths driven by

different phases of the local oscillator to achieve harmonic rejection [2-3].

In an application such as high-speed serial link transceivers, not only multiple phases but

multiple divided clock frequencies are usually needed to perform data serialization and

deserialization. These clock signals are typically generated using a chain of frequency-dividers
2
that is driven by the output of a high-frequency oscillator [5-9]. For serial links operating at tens

of Gb/s, these dividers consume a considerable amount of power to meet the target operation

frequencies over all process corners. In many cases, bandwidth extension techniques using

expensive passive inductors must also be used to meet the target speeds. This dissertation uses a

48 Gb/s serial link transmitter as a design example to illustrate the area and power benefits of

using injection-locked multipath oscillators to frequency divide and generate multiphase clocks.

1.2. Motivation

To generate a multiphase clock, a ring oscillator can be used in a phase-locked loop as shown

in Fig. 1.2(a). A conventional inverter-based ring oscillator, as shown in Fig. 1.3, consists of a

single loop of an odd number of inverters. While compact, easy to design, and tunable over a

wide frequency range, this oscillator suffers from several limitations. For example, it is not

possible to increase the number of phases while maintaining the same oscillation frequency since

the frequency is inversely proportional to the number of inverters in the loop. In other words, the

time resolution of the oscillator is limited to one inverter delay and cannot be improved below

this limit. In addition, the number of phases that can be obtained from this oscillator is limited to

odd values. Otherwise, if an even number of inverters is used, the circuit remains in a latched

state and does not oscillate.

To overcome the limitations of conventional ring oscillators, several variations have been

proposed. These variations all share the property that instead of having each phase in the

oscillator driven by a single inverter, or a single path, each phase can be driven by two or more

inverters, or multi-paths. Each of these paths is driven by a different phase in the oscillator [10-

3
19]. These variations are collectively known as multipath ring oscillators or MPROs as opposed

to the conventional single-path ring oscillators, SPROs.

(a)

(b)

Fig. 1.2. Multiphase generation: (a) Using a multiphase ring oscillator. (b) Using an LC oscillator and a
multiphase frequency-divider.

Fig. 1.3. A conventional N-phase inverter-based ring oscillator.

In MPROs, the selection of which phase to use to drive each path and the relative sizing of

these different paths add additional degrees of freedom to the design of the oscillator. These

degrees of freedom, if properly used, can allow a designer to overcome the limitations of

conventional SPROs. However, these degrees of freedom also expand the design space of the

oscillator considerably making its design a more difficult task.

Previous MPRO analyses have been either limited to a qualitative argument or otherwise

restricted to certain special cases. Since the number of possible coupling structures is very large
4
and it keeps increasing as the number of output phases increases, the existing analysis does not

allow a designer to explore the whole design space and find the optimum design. Accordingly, a

new analytical framework that explains the operation of these oscillators is needed if the full

potential of such structures is to be utilized.

One thing that makes the MPRO design problem even more complicated is its property of

having multiple possible oscillation modes. Without a clear understanding of what makes one of

these modes dominant, it is very likely that a designer might end-up having an oscillator that can

start-up each time in a different oscillation mode depending on the initial state of the oscillator.

This is of course unacceptable in any practical application. This problem, which is characteristic

of multipath oscillators, is one that has been almost totally overlooked in all previous

publications.

In this dissertation, we introduce an accurate analytical model and a comprehensive analysis

for multipath ring oscillators. We explain the effect of the oscillator coupling structure on the

different performance metrics including the number of phases, oscillation frequency, phase

noise, and oscillation-mode stability. We use the presented model to formulate an optimization

algorithm that allows a designer to find the optimum coupling structure of an MPRO, needed to

achieve certain design goals, in a very short time.

While ring oscillators are very compact and can be tuned over a wide frequency range, they

are typically suitable for applications that require relatively less stringent phase noise

specifications as compared to high-quality factor oscillators such as LC oscillators. For

applications like RF and high-speed serial-link transceivers, better phase-noise performance than

can be provided by ring oscillators is needed. Accordingly, LC oscillators are typically used. To

5
generate multiple phases, it is possible in principle to use a multiphase LC oscillator. However,

due to the high cost of high-Q on-chip passive inductors in terms of chip area, it is very rare to

generate more than 4 phases using a multiphase LC oscillator. Alternatively, a multiphase clock

with high spectral purity can be generated from an LC oscillator using frequency-dividers as

shown in Fig. 1.2(b). This method has been shown to have less jitter as compared to other

approaches such as delay-locked loops [4] for the same power budget.

Frequency division at very high frequencies can be achieved efficiently at using injection-

locking [20-29]. While injection-locked LC oscillators can operate at higher frequencies than

ring oscillators, they suffer a narrow locking-range due to their high quality factor. This makes

them sensitive to process, supply-voltage, and temperature (PVT) variations. In addition,

injection-locked LC oscillators are not suitable for multiphase generation due to the need for

multiple inductors which would cost a lot of area.

Injection-locked ring oscillators, on the other, hand have wider locking-ranges but they

operate at lower frequencies. Similar to a ring oscillator, a conventional injection-locked ring

oscillator has a maximum locking frequency that decreases with the number of phases at its

output. This means that same trade-off between phase resolution, operation frequency, and

power consumption still exists in frequency dividers realized as injection-locked ring oscillators.

In this work, we show that conventional latch-based dividers can be analyzed as

superharmonic injection locked ring oscillators. Accordingly, the performance of these dividers

can be extended in the same way the performance of a ring oscillator can. We propose, for the

first time, the use of superharmonic injection-locked multipath ring oscillators (SHIL-MPROs) to

achieve frequency division at higher frequencies for the same number of output phases. The

6
theory we propose explains the different factors that limit the locking range of a SHIL-MPRO

and how it can be maximized for a certain number of output phases by choosing the optimum

coupling structure of the core MPRO and the input clocking configuration.

1.3. Thesis Organization

The dissertation consists of seven chapters. Chapter 2 gives the necessary background on

oscillators, MPROs, frequency dividers, and serializing transmitters. It begins with a brief

discussion for the important performance metrics of an oscillator followed with a review for the

different forms of MPROs that have been proposed in the literature before. The chapter then

reviews the different categories of multiphase frequency dividers. Finally, it gives a quick

overview of high-speed serial link transmitters and the factors that limit their operation range.

In Chapter 3, we present our analysis for MPROs. We derive accurate expressions for the

oscillation frequency and phase noise. We show that for the same power consumption, if the

oscillator is designed correctly, the number of phases can be increased without decreasing the

oscillation frequency or degrading the phase noise. We also explain how the oscillator should be

designed to oscillate at the desired oscillation mode independent on the initial conditions.

The analysis for MPROs is extended to dividers in Chapter 4. Theoretically, the maximum

division frequency of a frequency divider can be decoupled from the number of phases at its

output when multipath coupling is used. We derive accurate expressions for the locking range

and use them to find the optimum structure to maximize the division frequency. The results of

the model are incorporated into a design procedure that can rapidly explore various coupling and

sizing of the multipath structure.

7
 We explain the details of the serializer design in Chapter 5 and discuss the measurement

results in Chapter 6. Finally, conclusions and future work are discussed in Chapter 7.

8
 CHAPTER 2

Background

This chapter begins with a quick review of some basics of oscillators followed by the

evolution of multipath ring oscillators. We then give a brief summary of the different methods to

design a multiphase frequency divider. The definitions in this chapter are used in the next two

chapters where the analysis is explained in detail. Since a high-speed serial-link transmitter is

used an application to demonstrate the design of SHIL-MPROs, the last section in this chapter

introduces the components and design challenges of serial link transmitters. We explain the

timing constraints that limit its operating range and show how they can be relaxed using high-

speed multiphase frequency dividers. We show that if these dividers can be designed at a low

power cost, significant savings in power and complexity for the overall transmitter can be

achieved.

2.1. Oscillators Basics

An oscillatory system can be defined as an autonomous dynamical system that possesses an

asymptotically stable limit cycle in its state space [30]. A more circuit-oriented definition is a

circuit that generates one or more periodic waveforms at its output. These waveforms can be

characterized by a certain frequency and amplitude and each of them has a certain phase.

If the harmonic content of the oscillator output can be neglected compared to its fundamental

frequency component, the oscillator output can ideally be expressed as vout(t)=Aocos(ωot+Φo)

where Ao is the oscillation amplitude, ωo is the angular frequency, and Φo is a fixed phase angle.

9
 In practice, due to the various sources of noise and interference in a real circuit, both the

amplitude and the phase of the oscillator output suffer from small fluctuations around their

nominal average values. Accordingly, the oscillator output can be more practically expressed as

vout  t    Ao  a  t   cos ot   o  n  t   (2.1)

where a(t) and φn(t) represent the amplitude and phase fluctuations, respectively.

While small amplitude fluctuations can be tolerated in most practical applications, phase

fluctuations, commonly known as phase noise, are usually problematic. Fig. 2.1 shows the effect

of phase noise on the output of the oscillator in both the frequency and time domains. In the

frequency domain, random phase noise causes the output spectrum to broaden in shape compared

to the ideal impulse-shaped spectrum of a noiseless oscillator. In applications like RF

transceivers, this broadening is highly undesirable since it causes signal interference. In the time

domain, phase noise causes the zero-crossing instants of the oscillator output to deviate from

their ideal positions at integer multiples of the nominal oscillation-period. These deviations,

known as timing jitter, are highly undesirable in applications like high-speed digital circuits and

serial link transceivers since they reduce the timing margin where the system operates correctly.

Fig. 2.1. Spectral broadening and timing jitter due to phase noise.

10
 Phase noise of an oscillator is measured in terms of the single-sideband phase noise spectrum

L(Δω). This is defined as the signal power in a 1-Hz bandwidth at a certain frequency offset, Δω,

from the nominal carrier frequency, ωo. The different mechanisms by which device noise is

converted into phase noise have been well studied by different researchers before [31-37]. In

general, for hundreds of Hz offset frequencies and above, it is fairly accurate to assume that

thermal noise causes phase noise L(Δω) that is proportional to (1/Δω)2 while flicker noise

produces phase noise skirts proportional to (1/Δω)3. Fig. 2.2 shows a typical single-sideband

phase noise spectrum for an electrical oscillator built in CMOS technology.

Fig. 2.2. Typical single-sideband phase noise spectrum.

Phase noise caused by random device noise can be reduced in a certain oscillator by scaling

the whole design so that it burns more power. Accordingly, if two different oscillator topologies

are to be compared in terms of their phase noise performance, the two oscillators must be

compared when scaled to burn the same power. In addition, oscillators running at higher

frequency are more sensitive to device noise due to their shorter oscillation periods. Accordingly,

to compare two different topologies, the oscillation frequency must also be the same for the two

oscillators.

11
 To help compare the performance of different kinds of oscillators, a figure of merit (FOM) is

typically used that factors out the effect of power consumption, oscillation frequency, and offset

frequency where the phase noise is measured. This FOM , given by (2.2), will be used in the next

chapter to compare the performance of different ring oscillators in terms of phase noise.

 o  2 
FOM  10  log   (2.2)
 L     P mW  

2.2. Multipath Ring Oscillators

Fig. 2.3 shows the schematic of a typical 5-stage inverter-based ring oscillator. For this

oscillator to operate correctly, the number of inverters in the loop and accordingly the number of

phases at the output has to be an odd number. The oscillation frequency of this oscillator is

inversely proportional to the number of stages. Accordingly, the time resolution of the oscillator

is limited to one inverter-delay.

Fig. 2.3. Typical 5-stage inverter-based ring oscillator.

To overcome the limitations of conventional single-path ring oscillators, several forms of

multipath ring oscillators have been introduced [10-19]. The simplest of these forms is the dual-

path ring oscillator from which three different forms have been proposed.

The first of these forms is the coupled oscillator proposed by Maneatis in 1993 [10]. In this

oscillator, the number of phases can be increased, according to Maneatis, without changing the

oscillation frequency. This is achieved by coupling a number of rings using a second path as

12
shown in Fig. 2.4 which causes the phases of the different rings to shift with respect to each

other, effectively increasing the phase resolution of the overall oscillator.

Fig. 2.4. A 3×5 coupled-oscillator.

The second of form of dual-path ring oscillators is the negative-delay-skewed oscillator

proposed by Lee in 1997 [11]. In this oscillator, the time resolution is improved by reducing the

delay per stage below 1 inverter-delay. This is achieved, as shown in Fig. 2.5, by driving each

output with an early signal, or a signal with a negative-delay, to accelerate the transition at the

output. The optimum skew has been derived analytically in [12] using a linear model and

verified by simulation.

Fig. 2.5. A 5-stage ring-oscillator using negative delay-skewed inverters.

Fig. 2.6 shows the third form of dual-path ring oscillators which uses cross-coupling to

achieve differential operation [13]. While the three proposed forms qualitatively seem different,

the three of them can be analyzed using one unified theory as will be shown in the next chapter.

13
 Fig. 2.6. An 8-phase ring oscillator using cross-coupling to achieve differential operation.

The triple-path ring oscillator can be considered the second step in the evolution of multipath

ring oscillators. Fig. 2.7 shows an example where both delay-skew and cross-coupling are used

to reduce the delay per stage and achieve differential operation, simultaneously [14-16].

Fig. 2.7. An 8-phase triple-path ring oscillator using cross-coupling to achieve differential operation and
negative delay skew to reduce the delay per stage.

In general, we can add more and more paths per stage. As the total number of phases

increases, the degrees of freedom increase very rapidly so that it becomes very hard for a

designer to find the optimum configuration that meets his design targets just by qualitative

analysis. And moreover, it becomes very difficult to verify this configuration just by simulation.

As an extreme case, the oscillator used by Straayer in 2008 to provide a time base for a time-

to-digital converter is a 47-stage, penta-path ring oscillator [17]. Of course, it can be imagined

how hard it is to come up with the best coupling configuration for such an oscillator using only

qualitative analysis and simulation.

14
 In all the above examples, the oscillation frequency of the MPRO does not just depend on the

number of stages as in SPROs but rather on the coupling structure of the oscillator as well. In

fact, if a designer tries to explore the design space by simulation, he will see that for some

coupling structures, as the sizing ratios are varied, the oscillation frequency can experience

discrete jumps accompanied by a change in the phase arrangement of the oscillator outputs. In

some cases, the oscillator may just stop oscillating at all.

The reason for this phenomenon is that multipath oscillators have different possible modes of

oscillation. Each of these modes is characterized by a different frequency, phase shifts, and phase

noise versus power tradeoff. Without clear understanding of what determines the dominant

mode, it is very possible that a designer will end up having an oscillator where the oscillation

mode is only marginally stable; meaning that it depends on the oscillator initial conditions [17].

The problem becomes even more severe as the number of phases in the oscillator increases.

Factors that determine the dominant oscillation mode have been studied in [18][19]. However,

the analysis in [18] is constrained to dual-path oscillators while that in [19] is constrained to even

number of phases with cross coupling inverters.

In the next chapter, we introduce a general model for multipath ring oscillators with arbitrary

coupling structures. We present a general analysis that explains the operation of a generic

multipath ring oscillator. We show that all the previously presented designs are special cases of

the general case we study. We explain how to design the oscillator for a certain number of phases

and maximize its oscillation frequency and phase noise performance while guaranteeing the

stability of the desired oscillation mode.

15
 2.3. Multiphase Frequency Dividers

Frequency dividers are important building blocks in clock generators performing frequency

multiplication and in low-jitter multiphase generators. The most widely used type of frequency

dividers is the latch-based divider. These dividers are composed of latches or flip-flops as unit

building blocks. The latches can be of any logic family like CMOS, true single-phase clock

(TSPC), or current-mode logic (CML). To optimize the performance of these dividers, the only

degree of freedom is optimizing the latch design. In [38] for example, all possible

implementations of TSPC latches and flip-flops are explored using an exhaustive search to find

the optimum flip-flop implementation. In [39], a modified CML latch circuit is proposed that can

operate at speeds up to tens of GHz. This latch is used in [40] to design a 40 GHz divide-by-2

circuit. Inductive peaking can be viewed as a way of optimizing the latch design by partially

tuning out its load capacitance using inductors [41].

Regenerative frequency dividers (RFDs), also known as Miller-type dividers [20], and

injection-locked frequency dividers (ILFDs) have been used for frequency division at very high

frequencies. Although both terms are sometimes used interchangeably [21], the most common

convention is that regenerative dividers do not have a self-oscillation frequency in the absence of

the input signal. In other words, the self-oscillation frequency of regenerative dividers is equal to

zero. On the other hand, injection-locked dividers oscillate at a certain self-oscillation frequency

if the input signal is not applied. In [22], the authors show that both types of dividers can be

analyzed using the same model.

Both RFDs and ILFDs can be classified based on their core oscillator, injection method, and

injection phases. The core oscillator can be either an LC-based oscillator [21][23] or a ring-

16
oscillator [24][27]. LC-based oscillators can operate at relatively higher frequencies but they

suffer from a narrow locking range which makes them sensitive to PVT variations. In addition,

they consume large areas due to the use of inductors. Ring oscillators, on the other hand, have

lower frequencies but are more compact and have wider locking ranges reducing their sensitivity

to PVT variations.

Signal injection can be achieved by two different methods. One method is by varying the

transconductance of the amplifying transistors with the input signal. This is usually achieved by

driving the tail bias-current of these transistors with the input and thus is referred to as tail

injection [21][27]. Another method is by varying the load impedance, usually achieved by

modulating the conductance of a load that is directly connected to the divider output and thus is

called direct injection [23][24]. Fig. 2.8 shows an example of each of the 4 types of dividers

mentioned.

Dividers can also be classified based on the number of injection phases. Single-phase

injection means that only one phase is required at the input [25]. Similarly, the divider is said to

have a differential-injection when two opposite phases are used at the input [23][24]. In general,

multiphase injection can be used when M phases are used at the input (M=1, 2, 3, …) [26-28].

The idea of multiphase injection for frequency dividers has been introduced in [26]. The

authors use a frequency domain model to derive the optimum phase shift between different

injected signals to maximize the locking range. The analysis is applied to a 6-phase ring

oscillator composed of 3 CML differential buffers and with tail injection at two stages with 120o

phase shift. In [27], the authors explain the concept of multiphase injection using a time-domain

17
model. They apply it on a 4-phase divide-by-2 and a 12-phase divide-by-6 both composed of

CML cross-coupled latches and having 4 injection phases.

Fig. 2.8. Different types of injection-locked frequency dividers classified based on the core oscillator and
the injection method.

The term, superharmonic injection-locked oscillator (SHIL-O), has been used in [29] to

describe ILFDs. Superharmonic injection-locking is a phenomenon where an oscillating system

can be locked in phase at steady state to a periodic input signal whose frequency is a

superharmonic, or an integer multiple, of its own output frequency. In that sense, both RFDs and

ILFDs can be considered SHIL-Os. The only difference between them is that RFDs use a core

oscillator whose free-running frequency is equal to zero.

18
 We propose in Chapter 4 of this dissertation a general model for superharmonic injection-

locked multipath ring oscillators (SHIL-MPROs). The model is valid for both RFDs and ILFDs

with either tail-injection or direct-injection and with any number of input injection phases. In our

analysis, latch-based dividers are treated as a special case where the injection is large. We show

that SHIL-MPROs have different mechanisms that limit the locking range and derive the

mathematical expressions needed to optimize the locking range. We also derive a mathematical

condition that gives the possible clocking configurations for correct injection phases. With

multiple injection phases and multipath sizing factors, the design space of SHIL-MPROs can be

very large. A design procedure that uses the model is presented to allow designers to quickly

converge on a final design.

2.4. Serial Link Transmitters

A serializing transmitter is used as an application of the SHIL-MPRO design procedure and

verification of the analysis. Fig. 2.9 shows a typical block diagram of a serializing transmitter [5-

9]. The transmitter consists mainly of a serializer, a driver, and a clock generator. The serializer

is a chain of multiplexers that combine a number of low-speed input data channels and produce a

single high-speed output data stream to be transmitted over a certain communication channel.

The driver converts the binary data into a signal format suitable for the channel. For the serializer

to operate properly, clock signals at different frequencies are required to drive the different

multiplexing stages. These clock signals are provided by a voltage-controlled oscillator and a

chain of dividers that are controlled by a phase-locked loop.

19
 Fig. 2.9. Conceptual block-diagram of a serializing transmitter.

The range of output data rates where the transmitter can operate properly is limited by several

factors. These factors include the circuit bandwidths, the oscillator tuning range, the locking

range of the frequency dividers, and the timing constraints of the different serializer stages.

For multi-Gb/s transmitters, LC oscillators with a combination of coarse and fine tuning

capacitors are typically used to generate the required high-speed clock at low phase noise and

wide enough tuning range. To meet the speed requirements, designers usually use inductive

peaking to extend the circuits bandwidths at the cost of the large area consumed by the inductors

[42-44]. This bandwidth extension is usually applied to the final multiplexing stages in the data

path and to the dividers in the clock network.

To explain how the timing constraints of a certain multiplexing stage can limit the operation

range of the transmitter, Fig. 2.10(a) shows an N-to-1 multiplexer driven by an N-phase input

clock signal of frequency fo/N where fo is the bit rate at the multiplexer output. This clock signal

also samples, and if necessary, retimes, the N outputs of the preceding stage. Fig. 2.10(b) is a

timing diagram that describes the operation. The signal D<1> is an output of the preceding

multiplexer. Transitions in D<1> occur after a delay t(C-Q)MUX from the edge of the divided clock

20
signal CKDIV<1>. This divided clock signal in turn has transitions that are delayed by tdiv from

the edge of the input clock signal CKin<1>. The signal D<1> is sampled by CKin<i> which is

phase-shifted from CKin<1> by (2π)θ. In Fig. 2.10(a), θ can only take discrete values that are

multiples of 1/N. In general, if a buffering stage or variable delay is used between the input clock

and the N-to-1 MUX, θ can be any positive real number. Depending on the number of phases at

the output of the divider and the number of latches used in the retiming flip-flop, the same data

phase can be sampled by more than one input clock phase. We assume the phase difference

between the first and last clock edges sampling the same data phase is equal to (2π)φ.

(a)

(b)

Fig. 2.10. Timing requirements in an N-to-1 multiplexer: (a) Block diagram. (b) Timing diagram.

21
 For the setup time condition to be met, the delay between the triggering edge CKin1 and the

first sampling edge CKini has to be long enough to allow the data to switch before the edge of

CKini by at least one setup-time (tsu). This delay sets an upper limit on the output bit rate as

given by (2.3). On the other hand, for the hold time constraint to be met, the delay between the

triggering edge CKin1 and the last sampling edge CKinj has to be short enough to allow at

least one hold-time (th) before the next data transition. Assuming th<t(C−Q)MUX+tdiv, the hold time

condition sets a lower limit on the allowed output bit rate as given by (2.4).

N
f max  (2.3)
tdiv  tC Q MUX  tsu

N     1
f min  (2.4)
tdiv  tC Q  MUX  th

Fig. 2.11 shows the effect varying the phase shift θ and the order of the multiplexer on the

valid bit rate window. In the figure, we assume φ=0 which can be achieved by retiming the

sampled data before applying it to the multiplexer. As θ increases for a certain multiplexer, the

maximum allowed bit rate increases but the overall window of allowed rates becomes narrower.

The graph shows that a tradeoff exists between maximizing the output bit rate of the transmitter

and its robustness where the narrow window can shift with PVT variations. When the order of

the multiplexer (N) increases, this tradeoff between data rate and robustness becomes more

relaxed due to the increase in the input clock period.

22
 Fig. 2.11. Allowed bit rate window as a function of θ assuming φ = 0.

One technique for dealing with this delay skew problem is to insert a replica delay buffer

between the input clock and the N:1 MUX [45]. Ideally, if the buffer delay matches the total

delay of the divider and the preceding MUX, the timing conditions can be met independent on

the bit rate. However, the replica buffer consumes extra power and occupies an additional area.

Besides, it is usually hard to achieve good matching over all PVT corners [8].

Another technique for dealing with the delay skew problem is the insertion of a variable delay

or a phase interpolator between the divided clock and the preceding MUX [7-8]. A feedback loop

adjusts the value of θ so as to position the valid data rate window around the desired operation

data rate. The drawback of this technique is the added design complexity in addition to the power

consumption and area overhead. For example, the power overhead in [8] is 72mW which is

slightly less than the total power consumption of our whole transmitter in the same 65nm CMOS

technology. The area overhead is 0.88mm2 which is 73% of our total chip area.

Equations (2.3-2.4) and Fig. 2.11 show that as the order of the multiplexer (N) increases, the

timing constrains become more relaxed due to the increased bit period at the input. In addition,

as more phases are used for the input clock, the phase shift θ can be selected at design time with

23
a higher resolution allowing the optimum positioning of the nominal bit rate window such that

the desired bit rates can fall within this window for all PVT corners.

Since the timing conditions have to be satisfied in every stage in the serializer, it is

advantageous if the divided clocks also have multiple phases. This is because this allows more

flexibility in setting the maximum and minimum allowed data rates for all stages. In Chapter 5,

we show how multiple phases in the divided clock signals can also be useful in reducing the

retiming elements to a minimum which leads to significant power and area savings.

24
 CHAPTER 3

Multipath Ring Oscillators

This chapter presents the analysis of multipath ring oscillators. We first explain a general

model that can represent any multipath ring oscillator. Then, we derive expressions for the

oscillation frequency and amplitude and show how the frequency can be maximized taking mode

stability into consideration. After that, we present the phase noise analysis.

3.1. Oscillator Modeling

3.1.1. General Form of the Multipath Ring Oscillator

Fig. 3.1 illustrates the general form of an MPRO. The oscillator consists of N stages where

each stage consists of N inverting buffers connected together at their outputs. The input of each

buffer is connected to one of the N different nodes of the oscillator to cover all possible coupling

paths. The buffers of the main loop are all scaled by a sizing factor hN with respect to a unit

reference buffer of a given size. Buffers that are driven by signals skewed by one stage are sized

by a factor hN-1. Similarly, buffers driven by signals skewed by j stages are sized by hN-j.

Fig. 3.1. General form of an MPRO having an arbitrary coupling structure.

25
 This general form can be used to model any MPRO by selecting the appropriate values of the

sizing factors, hi. For example, an ordinary SPRO is a special case of the general form where

hi=0  iN. A dual-path ring oscillator has hi=0  i{N,N-j} where j is the skew of the second

input and 0<j<N. Cross coupling, usually used to achieve differential operation, can be modeled

for an even numbered N by h(N/2)+10. The coupled oscillator [10] can also be treated as a special

case. By using a 53 coupled oscillator as shown in Fig. 3.2(a) as an example, when numbering

the nodes diagonally as shown in the figure, the oscillator can be visualized as a 15-stage dual-

path oscillator with a skew of 2 as shown in Fig. 3.2(b).

(a)

(b)

Fig. 3.2. An example using a 5×3 coupled-oscillator: (a) Coupled oscillator with its nodes numbered
diagonally. The coupling inverters are scaled by h15 while the ring inverters are scaled by h13. (b) Equivalent
dual-path ring oscillator with a skew of 2.

26
 Fig. 3.3 shows the circuit schematic of the unit inverter used in all the simulations in this

chapter. All simulations are performed using a 65nm CMOS technology with a supply voltage of

1V. Transistors M1 and M4 are used as enable devices where their gates are always driven by

VDD for M1 and GND for M4. Transistors M2 and M3 are the main inverter devices. They have

dimensions of (4m/L) for M2 and (8m/L) for M3 as shown in the figure. In different

simulations, L takes values of 60nm, 100nm, 200nm, or 400nm. The driving strength of each

inverter is determined by the number of units used which is indicated by hi in Fig. 3.3. Unless

otherwise stated, all the simulation results shown in this chapter are obtained using the periodic

steady state (PSS) and the periodic noise (PNoise) simulators in Cadence-SpectreRF.

Fig. 3.3. Schematic diagram of the unit inverting buffer.

3.1.2. Linear Delay Stage Model

Fig. 3.4(a) shows the inverting buffer modeled as a linear transconductor. The input-output

relationship of a single buffer scaled by hi and driving the input capacitance of a similar buffer

can be expressed as

27
 dvout  t 
hi g m vin  t   hi g o vout  t   hi C g 0
dt
(3.1)
dvout  t 
 an vin  t   vout  t    0
dt

where gm is the transconductance, go is the output conductance, Cg is the buffer input capacitance

which also acts as the load capacitance for the driving buffer, an=gm/go is the linear dc gain of the

buffer and =Cg/go is a time constant. In practice, parasitic capacitance due to routing and any

external loading can be absorbed into the capacitance, Cg.

(a) (b)

Fig. 3.4. Model of a single unit inverting buffer (a) Linear model (b) Nonlinear model.

A single stage of an MPRO consists of N buffers and is loaded with N buffers of similar

sizing factors as shown in Fig. 3.1. Using the linear model, the output voltage of any stage can be

expressed in terms of the N input voltages using a linear differential equation. Equation (3.2)

expresses v1, the output of the first stage in Fig. 3.1, in terms of the N inputs of the first stage.

N N N
dv1  t 
 h g v t    h g v t    h C
i 1
i m i
i 1
i o 1
i 1
i g
dt
0 (3.2)

N
Defining the fractional sizing factors and the total sizing factor as xi=hi /H and H   hi ,
i 1

respectively, equation (3.2) can be simplified into (3.3) where an=gm/go and =Cg/go are the same

dc gain and time constant defined in (3.1).

28
 N
dv1  t 
an  xi vi  t   v1  t    0 (3.3)
i 1 dt

3.1.3. Nonlinear Delay Stage Model

To model the nonlinear behavior of the inverter, we assume the transconductance and the

output conductance to be nonlinear functions of the input and the output voltages, respectively.

This nonlinear model is shown in Fig. 3.4(b). In this model, we assume the inverter input

capacitance is linear. The output voltage of a stage can be expressed as a function of the N inputs

using a nonlinear differential equation. Equation (3.4) shows the output of the first stage of the

general MPRO in Fig. 3.1 as a function of the N inputs. When Fg(vi)=gmvi and Fo(v1)=gov1, this

equation reduces to the linear case (3.3).

N
dv1  t 
 x F v t   F v t   C
i 1
i g i o 1 g
dt
0 (3.4)

3.2. Oscillation Frequency and Amplitude

3.2.1. Oscillation Modes

In this section, we first use linear analysis to explain the different oscillation modes that can

exist in an MPRO and determine which is dominant. Linear analysis also helps in understanding

how we can size the different coupling paths of the MPRO and what limits its maximum

oscillation frequency. Next, we use a simplified nonlinear analysis to find an expression for the

oscillation amplitude and a more accurate expression for the oscillation frequency.

Due to the symmetry of the system in Fig. 3.1 and in the absence of mismatches, if the MPRO

is to oscillate at a certain frequency, n, the phase shift between each two successive nodes and

29
the amplitude of oscillation at each node in the oscillator will ideally be the same. Therefore, the

oscillation waveform at the ith node can be expressed as

vi  t   Vo cos nt   i  (3.5)

where n is the oscillation frequency and Δφ=2πn/N since the total phase shift around the loop

should be multiples of 2, N is the number of stages in the oscillator, and n can take values

between 0 and N-1 representing the different possible modes of oscillation.

Using (3.5), equation (3.3) can be written as

N
 2 n   2 n   2 n 
an  xi cos  n t  i   cos  n t    n sin  n t  0 (3.6)
i 1  N   N   N 

By equating the cos(nt) and the sin(nt) terms of (3.6), we get expressions for the oscillation

frequency of the nth mode (3.7) and the minimum dc gain required for this mode to exist (3.8).

We refer to this gain as the mode gain.

 2 n
 i  1 
N

 x  sin 
i
N 
n  i N1 (3.7)
 2 n
 xi  cos   i  1 
i 1  N 

1
an 
 2 n
 i  1 
N
(3.8)
 xi  cos 
i 1  N 

In practice, the oscillator starts first from a linear mode of operation where all the buffers are

indeed acting as linear transconductors. All oscillation modes that have mode gains, an, lower

than the actual dc gain of the inverter, ao, start to grow. As the oscillation amplitude grows, the

effective gain of the inverter drops due to nonlinearity. Consequently, modes with higher mode

30
gain die out and only the mode that requires the minimum gain continues to oscillate and hence

is the dominant mode. Therefore, for an MPRO with N stages, once we choose a certain value for

the relative sizing vector x =[x1 x2 x3 … xN]T N, the index, n*, of the dominant mode is given

by (3.9). Note that this definition for the dominant mode is dependent only on the relative sizing

vector, x , of the MPRO and is valid for any buffer design used to build the actual oscillator.

 
 1 
n*  arg min  N  0
 2 n
n 1,2,..., N 
 
 xi  cos   i  1   (3.9)
i 1  N  
 N  2 n 
 arg max   xi  cos   i  1  
n 1,2,..., N  i 1  N 

3.2.2. Oscillation Amplitude and Nonlinear Time Constant

The nonlinear nature of the buffer elements in the MPRO affects the oscillator in two ways.

First, the nonlinearity determines the oscillation amplitude. Second, it can change the oscillation

frequency since the effective output conductance is often dependent on the oscillation amplitude.

Equation (3.7) predicts the oscillation frequency normalized to a certain time constant.

Depending on the unit inverting buffer used to implement the MPRO, this time constant may not

be actually, constant. To explore the dependence of the time constant, , and the oscillation

amplitude ,Vo, on the sizing vector, x , we simplify equation (3.4) by solving it using a single

tone harmonic balance (3.10). This assumption is later verified to be a good approximation when

comparing the equations with simulation results.

N
 2 n   2 n   2 n 
I g Vo   xi cos   n t  i   I o Vo  cos   n t     n C gVo sin   n t  0 (3.10)
i 1  N   N   N 

31
 In (3.10), Vo is the amplitude of the fundamental component of the voltage waveform, Ig(Vo)

and Io(Vo) are the fundamental components of the transconductor and the output conductance

currents, respectively. Similar to the linear analysis, equation (3.10) can be decomposed into the

following:

 2 n
 i  1   I o Vo   0
N
I g Vo   xi  cos  (3.11)
i 1  N 

 2 n
 i  1   n C gVo  0
N
I g Vo   xi  sin  (3.12)
i 1  N 

Using the frequency and mode gain definitions in (3.7) and (3.8), we can reduce (3.11) and

(3.12) for the dominant oscillation mode into:

I g Vo   an*  Io Vo  (3.13)

N
 2 n* 
 CV   x  sin 
i  i  1 
n*  g o   i N1  N 
(3.14)
 I o Vo    x  cos  2 n  i  1 
*

 i  
i 1  N 

Equation (3.13) shows that the oscillation amplitude of an MPRO depends only on the type of

nonlinearity that the buffer exhibits and the value of the dominant-mode gain. Fig. 3.5 is an

interpretation for (3.13). The figure shows arbitrary amplitude characteristics for an inverter. At

small input amplitudes, the inverter acts as a weakly nonlinear amplifier that can be

approximated by a third order nonlinearity coefficient. As the input amplitude increases, the

output resembles a square-wave whose fundamental component is (2/π)Vdd. As the figure shows,

the oscillation amplitude is determined by the amplitude that causes enough gain compression

such that the effective dc gain of the inverter is exactly equal to the gain required to sustain the

oscillation mode, i.e. the mode gain.

32
 Fig. 3.5. Example of the amplitude characteristics of an inverter.

To find an accurate expression for the oscillation amplitude, the exact shape of the buffer

nonlinearity must be known. In general, two limiting cases can be examined. The first case is

when the mode gain is close to unity where the oscillation waveform is approximately a square-

wave with a fundamental component (2/π)Vdd. In this case, the oscillation amplitude can be

approximately given by (Vo/Vdd)= (2/π)(1/an*). The other limiting case is when the mode gain is

close to the small signal dc gain of the inverter, ao, which is the maximum gain an inverter can

provide. This case causes the oscillation amplitude to be very small and the inverter nonlinearity

can be approximated using a third-order nonlinearity coefficient. This coefficient can be chosen

to match the simulation results at small oscillation amplitudes.

Equation (3.15) is an empirical expression for the oscillation amplitude. This expression is

derived by taking the average of the two limiting cases discussed above. The third order

nonlinearity coefficient is empirically chosen such that Vo is equal to (2/π)Vdd when an*=1.

Vo 1  1 a a 
    o n*  (3.15)
Vdd   an* ao  1 

33
 Fig. 3.6 shows the simulated oscillation amplitude as a function of the dominant mode gain

together with equation (3.15). The values of the dc gain, ao, used to plot equation (3.15) for

different channel lengths are shown in Table 3.1. These values are obtained using ac simulation

and will be used throughout the rest of this chapter. For each value of the mode gain, we use an

MPRO similar to that in Fig. 3.1 with the unit inverter shown in Fig. 3.3 and with sizing factors

as shown in Table 3.2. The way these sizing factors are derived is explained in the next

subsection. To verify the validity of the expression for different channel lengths, each simulation

is repeated using different channel lengths for the NMOS and the PMOS transistors of the

inverter. As the figure shows, very good agreement between the empirical equation and the

simulation results is observed.

600 Theory
Vo (mV)

400 Simulation

200

0
0 2 4 6 8 10
an*

(a)

600 Theory
Vo (mV)

400 Simulation

200

0
0 4 8 12 16 20
an*

(b)

Fig. 3.6. Fundamental component of the oscillation waveform as a function of the dominant mode gain.
(a) L=60nm (b) L=200nm.

34
 Table 3.1. Small-signal dc-gain and time-constant parameters of the unit inverter for different channel
lengths
Channel Length (L) 60nm 100nm 200nm 400nm

Small-signal dc gain (ao) 8 20 30 45

Time-constant (τo) 35ps 145ps 580ps 2800ps

Fanout-1 delay (τp) 10.2ps 18.4ps 50ps 166ps

To find an expression for the nonlinear time constant, we compare equations (3.14) and (3.7)

to derive (3.16), the effective time constant. Equation (3.16) shows that the effective time

constant is similar to the oscillation amplitude in that it only depends on the dominant mode gain

and the buffer nonlinearity. This result implies that two MPROs with the same dominant mode

gain but with different sizing factors and possibly, different number of stages would still have the

same oscillation amplitude and effective time constant.

CgVo
    an*  (3.16)
I o Vo 

As we did for the oscillation amplitude, we first examine the two limiting cases when an*1

and when an* ao, the small signal dc gain of the buffer. These two cases represent the

minimum and the maximum dominant mode gains that can exist. Obviously, a mode gain below

unity would not satisfy the oscillation unity gain condition and an oscillator with a dominant

mode gain greater than ao would not oscillate.

For the case when an*1, the equivalent time constant (an*) approaches p, the buffer’s

fanout-1 propagation delay. To explain this, we examine equations (3.7) and (3.8) with xN=1 and

N a large odd number. In this case, we see that an*1 and fn*1/(2N) which is the familiar

equation for the oscillation frequency of a long-chain SPRO having a delay per stage, p.
35
 When an*ao, the oscillation amplitude is infinitesimally small since any growth in amplitude

will cause the gain to drop due to nonlinearity. Accordingly, the linear expression for the

oscillation frequency becomes valid and the equivalent time constant is equal to the small signal

time constant of the buffer, o=Cg/go.

Fig. 3.7 shows the simulated values of the effective time constant as a function of the mode

gain for different channel lengths of the inverter. The plot is obtained by simulating different

MPROs having the sizing factors shown in Table 3.2. For each MPRO, we compare the

oscillation frequency obtained from simulation, fsim, to the expression in (3.7) and find the value

of the equivalent time constant, τsim, as shown in equation (3.17).

N
 2 n 
1  x  sin 
i
N
 i  1 

 sim   i N1 (3.17)
2 f sim  2 n
 xi  cos   i  1 
i 1  N 

40
30
 [ps]

20
Theory: Eq.(18)
10
Simulation: Eq.(17)
0
0 2 4 6 8 10
an*

(a)

400
300
 [ps]

200
100 Theory: Eq.(18)
Simulation: Eq.(17)
0
0 4 8 12 16 20
an*

(b)

Fig. 3.7. Effective time constant as a function of the dominant mode gain. (a) L=60nm (b) L=200nm.
36
 As the figure shows, the effective time constant can be very well approximated using a linear

interpolation between the two extreme cases of mode gain. Equation (3.18) shows the

approximate expression for the effective time constant as a function of the dominant mode gain.

The validity of this approximation is verified by repeating the simulation for different channel

lengths as shown in Fig. 3.7. In all cases, we found a very good agreement between the linear

interpolation and the simulated values of the time constant. The values for ao, τo and τp used in

Fig. 3.7 and throughout the rest of this chapter are shown in Table 3.1. These values were

obtained using ac simulation for ao. Values of τp are the fanout-1 propagation delay of the unit

cell obtained by simulating a 19-stage SPRO. Finally, values of τo are those that best fit the

curves in Fig. 3.7. These values were found to be very close to the actual small-signal time

constant obtained from ac simulation.

  o  p 
  a n*    p      an*  1 (3.18)
 ao  1 

3.2.3. Oscillation Frequency

Using equations (3.7), (3.8), and (3.18), an expression for the oscillation frequency can be

derived (3.19). Equation (3.19) gives the oscillation frequency of the dominant mode of any

MPRO having any arbitrary coupling structure and number of phases.

N  2 n*  i  1 
 ao  1   xi sin  
1 i 1  N 
f n*  (3.19)
2 N  2 n*  i  1 
 ao p   o     xi cos  N    o   p 
i 1  

To verify this general expression for the MPRO frequency, the frequency is calculated for

different oscillator configurations and compared against circuit simulation results. We start with

37
the degenerate case of an SPRO. In this case, xN=1 and all other fractional sizing factors are

equal to zero. Equation (3.9) shows that the dominant oscillation mode for SPROs is given by

n*=ceil(N/2). For even N, n*=N/2 and the oscillation frequency of the dominant mode is always

equal to zero indicating the fact that even numbered SPROs latch and do not oscillate. For odd N,

n*=(N+1)/2 showing that the phase shift between every two successive nodes is π+(π/N) which is

a well-known result for SPROs. The oscillation frequency can be found from (3.19) and is

expressed by (3.20) and compared to simulation in Fig. 3.8. The expression reduces to the

familiar form fSPRO=1/(2Nτp) when N is a large odd integer.

24
Theory
18
freq [GHz]

Simulation
12
6
0
3 7 11 15 19
N

(a)

6
Theory
4.5
freq [GHz]

Simulation
3
1.5
0
3 7 11 15 19
N

(b)

Fig. 3.8. Oscillation frequency of an SPRO for different numbers of stages. (a) L=60nm (b) L=200nm

Fig. 3.9 shows the comparison for the case of a 3-stage dual-path ring oscillator. Fig. 3.9(a)

shows the schematic diagram of the oscillator. In this special case, the mode gain is equal to 2

and is independent on the sizing coefficients. Accordingly, the effective time constant is also

independent on sizing. The oscillation frequency can be obtained from (3.19) and is plotted in
38
Fig. 3.9(b). The analytical results were found to match the simulation for all the simulated

channel lengths. The figure only shows the results for a channel length of 200nm.

(a)

4
3
freq [GHz]

2
Theory
1 Simulation
0
0 2 4 6 8 10
h3 / h2

(b)

Fig. 3.9. Oscillation frequency of a 3-stage dual-path ring oscillator.

(a) Schematic diagram of the oscillator (b) Simulation results and analytical expression for L=200nm.

Fig. 3.10 and Fig. 3.11 show the comparison between equation (3.19) and the simulation

results for two different 8-stage dual-path ring oscillators. The oscillator in Fig. 3.10(a) has a

skew of 5 stages meaning that v1 is driven by v8 and v3. Similarly, the oscillator in Fig. 3.11(a)

has a skew of 3 stages. Fig. 3.10(b) and Fig. 3.11(b) show the simulation results and the

analytical expression for the oscillation frequency. The figures also mark the different regions

where different modes are dominant. These regions are found using equation (3.9). The

boundaries between these regions are characterized by an abrupt change in frequency

accompanied by a change in the output phase order when the sizing factors are changed. For this

reason, if the MPRO is to be used as an oscillator that is desired to operate at a stable, well-

defined oscillation mode, these boundaries must be avoided.

39
 In both figures, a discrepancy between the simulation and the analytical prediction at the

boundary between the modes n*=5 and n*=4 is observed. This discrepancy is due to the very

small difference in mode gains at this region and the way these oscillators are simulated. This

small difference in mode gain causes the possibility of exciting a non-dominant mode if the

initial conditions of the oscillator are suitable for this. More about mode stability and dependence

on initial conditions is discussed in section 3.3.5.

(a)

(b)

Fig. 3.10. Oscillation frequency of an 8-stage dual-path ring oscillator with a skew of 5 stages. (a)
Schematic diagram of the oscillator (b) Simulation results (black dots) and analytical expression (gray curve)
for L=200nm.

40
 (a)

(b)

Fig. 3.11. Oscillation frequency of an 8-stage dual-path ring oscillator with a skew of 3 stages. (a)
Schematic diagram of the oscillator (b) Simulation results (black dots) and analytical expression (gray curve)
for L=200nm.

3.2.4. Maximum Oscillation Frequency

Using the derived analytical expression for oscillation frequency as a function of the

fractional sizing vector, x , we can find the value of x that maximizes the oscillation frequency

for a given number of stages and certain constraints on the mode gain. This maximization is

useful in determining whether a particular frequency is achievable. Furthermore, as shown in the

next section, the design for maximum frequency allows us to design oscillators with lower phase

noise.

We first define the two matrices A and B as given by (3.21) and (3.22) and then rewrite (3.7)

and (3.8) in matrix notation as follows:

41
   01  0N 
   2 n
A     ;  ni  sin   i  1  (3.21)
   N 
  N 11
   N 1 N 

 b01  b0 N 
   2 n
B     ; bni   cos   i  1  (3.22)
b   N 
  
N 1 1
 b  
N 1 N

1
an  (3.23)
BnT x

AnT x
n  (3.24)
BnT x

where x =[x1 x2 x3 … xN]T N is the fractional sizing vector and each of the row vectors AnT

and BnT is the nth row in the matrices A and B, respectively.

The goal is to maximize n* where the feasible region of the vector x in N can be

expressed by a group of linear equations and linear inequalities. This region is defined by

   x N : x  0, 1T x  1 (3.25)

where 1 T = [1 1 1 … 1]T. The first constraint requires x being positive while the second

constraint follows from the definition of x summing to unity.

The algorithm used to solve this optimization problem can be summarized in the following

steps:

1. For each mode n, where n and N do not have common factors (to guarantee N

distinctive phases at the output), find the region of space n where this mode is

dominant. This can be achieved by adding two additional constraints to the feasible

42
 region  as shown in (3.26). The first inequality in (3.26) constrains the mode gain to

be less than a certain maximum value, amax. This is important to guarantee certain

minimum oscillation amplitude. The second inequality guarantees the mode gain of the

mode n is less than that of all other modes. This allows us to maximize the frequency

of each mode in the region where this mode is dominant. The parameter m allows us to

set a certain margin between the dominant mode gain and other modes to guarantee

mode stability. This will be explained more in the next section.

 n   x   : BnT x  1 amax , BnT x  1   m  BmT x ; m  n (3.26)

2. Find the maximum normalized oscillation frequency of each mode and the

corresponding optimum sizing vector, x n * , by solving the following optimization

problem:

maximize AnT x BnT x

(3.27)
subject to: x  n

This problem is a linear fractional optimization problem. It can be solved accurately by

iteration using a binary search algorithm. Each of the iterations in this algorithm is a

linear programming feasibility problem that can be solved in MATLAB by methods such

as the Simplex algorithm or interior point methods [46][47].

3. Find the absolute maximum frequency over all modes and the corresponding optimum

sizing vector, x * .

4. Repeat the previous steps for MPRO’s with different numbers of stages.

43
 Fig. 3.12 shows the results obtained from running the above algorithm assuming amax is

infinite and m= 0. These results represent the absolute maximum normalized frequency that can

be obtained from an MPRO of a certain number of stages. The results show a linear increase in

the maximum possible normalized oscillation frequency as the number of stages increases

provided that the dc gain of the buffer sufficient to provide the required amplification. Table 3.2

shows optimum values of the sizing factors, hi, and the corresponding mode gain for varying

number of stages. Note that in this optimization and subsequent derivations, x1 (an accordingly,

h1) is set to 0 since a diode connected transconductor mainly reduces the effective value of the

time constant and thus can be absorbed in the parameter, .

12
(2/)*N
Optimization Output
(  n )max

0
0 4 8 12 16 20
N

Fig. 3.12. Maximum normalized oscillation frequency.

It is interesting to note that the results in Table 3.2 are considerably systematic and can be

easily deduced for higher values of N. Using (3.8) and the values in Table 3.2, the dominant

mode gain for maximum frequency is given by an*=N–1 as can be seen in Table 3.2. From (3.7),

it can be shown that n* approaches (2/)N as N increases.

Since the results in Fig. 3.12 and Table 3.2 were derived assuming m= 0, these results cause

the oscillator to oscillate at the boundary between different oscillation modes. In fact, it can be

easily shown by substituting the values in Table 3.2 into (3.8) that for each of these MPROs, the

44
value of the mode gain is the same for all modes n= 1, 2, …, N-1. This means that the maximum

oscillation frequency of the MPRO is achieved at the cost of the minimum possible mode

stability. Accordingly, if the MPRO is to be used as an oscillator that has a stable well-defined

oscillation mode, some back-off from this maximum frequency is needed to guarantee mode

stability. However, the results obtained for the maximum frequency and the corresponding sizing

factors are still important for two reasons. First, they serve as a benchmark to show the

maximum achievable frequency under all circumstances. And second, they can still be used for

driven MPROs like in injection-locked oscillators and frequency dividers where mode stability is

guaranteed by the external driving signals.

Table 3.2. Optimum sizing factors for maximum oscillation frequency assuming unlimited dc gain and zero
mode gain margins
N an h2 h3 h4 h5 h6 h7 h8 h9 h10 h11 h12 h13 h14 h15 h16
3 2 0 1
4 3 0 1 2
5 4 0 0 1 1
6 5 0 0 1 2 2
7 6 0 0 0 1 1 1
8 7 0 0 0 1 2 2 2
9 8 0 0 0 0 1 1 1 1
10 9 0 0 0 0 1 2 2 2 2
11 10 0 0 0 0 0 1 1 1 1 1
12 11 0 0 0 0 0 1 2 2 2 2 2
13 12 0 0 0 0 0 0 1 1 1 1 1 1
14 13 0 0 0 0 0 0 1 2 2 2 2 2 2
15 14 0 0 0 0 0 0 0 1 1 1 1 1 1 1
16 15 0 0 0 0 0 0 0 1 2 2 2 2 2 2 2

When the mode gain is limited by a certain finite amax, the optimum values of hi will be

different from those shown in the Table 3.2 for N>(1+amax). The new values can be obtained

similarly by running the same algorithm as before and setting the appropriate value of the

maximum gain. In this case, the maximum normalized oscillation frequency saturates

approximately at (2/)(amax+1) where amax is the upper bound of the mode gain used in the
45
optimization problem. In general, if we use the upper bound shown in Fig. 3.12 for (nτ)max, that

is (nτ)max < (2/π)N, together with (3.18), an approximate maximum limit on the oscillation

frequency can be formulated as given by (3.28).

1 N  ao  1
 2 ; 3  N  amax  1
   ao  1  p   N  2   o   p 

f max  N    (3.28)

 1  amax  1 ao  1 ; N  amax  1
   ao  1 p   amax  1  o   p 
2


From this expression, we see that theoretically, the absolute maximum frequency of an MPRO

can be achieved when amax is set equal to ao, the maximum available dc gain of the inverter.

When this gain is high enough, the limit is equal to (1/π2)ao/τo= (1/π2)gm/Cg which is

proportional to the cut-off frequency of the unit inverter used to build the oscillator.

Fig. 3.13 shows the actual maximum oscillation frequency obtained from circuit simulation

together with the theoretical expression in (3.19) and the maximum limit in (3.28). The increase

in the actual frequency is only sub-linear as the figure shows due to the increase of the effective

time constant with mode gain as explained in the previous subsection. As the figure shows, as N

increases, both equations (3.19) and (3.28) converge and they match the simulation results very

closely. In these simulations, the initial state of the oscillator is adjusted such that the desired

oscillation mode is excited. It should be noticed that for N=3, the maximum frequency is that of a

3-stage SPRO as indicated in Table 3.2.

46
 15

freqmax [GHz]
13
Simulation
11 Equ. (19)
Equ. (28)
9
0 4 8 12 16 20
N

(a)

6
freqmax [GHz]

5
Simulation
4 Equ. (19)
Equ. (28)
3
0 4 8 12 16 20
N

(b)

Fig. 3.13. Simulated maximum oscillation frequency for different numbers of stages.
(a) L=100nm (b) L=200nm.

3.2.5. Mode Stability

A common problem in MPRO design is the stability of the dominant oscillation mode. Mode

stability refers to whether the MPRO always oscillates at the same mode regardless of the initial

conditions of the oscillator. This problem is especially pronounced for MPROs with a large

number of phases. This is due to the existence of many modes and the very small differences in

the value of the mode gain of adjacent modes if the MPRO is not well designed.

Analyzing the problem of mode stability accurately requires an accurate nonlinear state-space

analysis to identify the different regions of attraction for each of the oscillation modes in the N-

dimensional state space of the oscillator [30]. Such accurate analysis is out of the scope of this

work and cannot be made using the presented model. However, our model still provides
47
sufficient information for designing MPROs that are less sensitive to their initial conditions and

that have a stable, well-defined oscillation mode regardless of their initial state.

As an example, we use the two 8-stage dual-path ring oscillators shown in Fig. 3.10 and Fig.

3.11 with coupling skew of 5 and 3, respectively. Fig. 3.14 shows the mode gain of the different

possible modes of oscillation for the two oscillators as a function of the sizing factors. These

curves can be obtained using equation (3.8) for different values of n. Modes that are not plotted

in the figure are either degenerate or impossible (having negative mode gain). As the figure

shows, for both oscillators, the difference in the values of mode gain for n=4 and n=5 is very

small near the boundary between these two modes. In general, when the mode gain difference

between two modes is small, the oscillator can operate in either one depending on initial

conditions. Hence, designing an oscillator near this boundary is undesirable although the

maximum oscillation frequency exists at this region.

Another example is shown in Fig. 3.15. The figure shows the mode-gain profile of two 47-

stage MPROs. The first MPRO is designed using the design procedure explained in the previous

section for maximum oscillation frequency at a maximum mode gain amax=2. The oscillator has

sizing factors, (h47,h3)=(2,1), a dominant mode, n*=31, a dominant mode gain, an*=2, and a

normalized oscillation frequency, n*τ=1.73. Fig. 3.15(a) plots the mode gain profile of this

oscillator. The even symmetry in the figure directly follows from (3.8). Combined with the odd

symmetry in equation (3.7), every two mirror image modes are actually degenerate, i.e. they are

the same mode. As the figure shows, the mode gain margin between the dominant mode, n=31,

and the adjacent modes is very small. This causes the actual oscillation mode to be sensitive to

initial conditions.

48
 (a)

(b)

Fig. 3.14. Mode gain of the different possible oscillation modes of an 8-stage dual-path ring oscillator
having (a) Skew=5 (b) Skew=3. The vertical dashed lines represent boundaries between different modes.

Fig. 15(b) shows the mode gain profile of a second MPRO. The sizing factors of the second

oscillator are also obtained using the optimization algorithm with the mode gain margin, m, set

to 50% to allow for mode stability. After quantizing the sizing factors into integers, the second

MPRO has a coupling configuration, (h47,h25,h22,h14,h11)=(3,5,4,5,2), a dominant oscillation

mode, n*=30, a dominant mode gain, an*=1.84, and a normalized frequency, n*τ=1.48. As Fig

15(b) shows, all the modes other than the dominant mode have at least 50% higher mode gain

compared to the dominant mode.

49
 (a)

(b)

Fig. 3.15. Mode gain profile of a 47-stage MPRO having a coupling configuration.
(a) (h47,h3)=(2,1) (b) (h47,h25,h22,h14,h11)=(3,5,4,5,2).

To verify the two MPROs for mode stability, both oscillators are simulated using the test-

bench shown in Fig. 3.16. In this simulation, an exponentially decaying sinusoidal waveform is

coupled to one node of the MPRO using a small coupling capacitor of 1fF. The coupling

capacitor is small enough not to cause any significant loading or mismatch for the MPRO when

oscillating at steady state. The excitation frequency, fs, is varied and the steady-state oscillation

frequency and oscillation mode are measured. For the first oscillator in Fig. 3.15(a), the final

oscillation mode was found to be dependent on the excitation frequency. This is attributed to the

small mode gain margin as explained before. For the second oscillator, we found the final

oscillation mode to be always n*=30, independent on the excitation frequency.

50
 Fig. 3.16. MPRO simulation test-bench. The oscillator is excited using an exponentially decaying
sinusoidal waveform of frequency fs coupled to v1 using a 1-fF capacitor.

In general, the design methodology presented in this section can be used to design any MPRO

with any number of phases. The designer must verify the mode stability by simulating the

oscillator using the different expected initial conditions. Improving the mode stability can be

achieved by increasing the mode gain margin and this usually comes at the cost of reduced

maximum frequency.

3.3. Phase Noise Analysis

The approach for analyzing phase noise follows a linear analysis similar to that in [31]. The

MPRO is modeled as an unstable negative feedback system with the noise sources in the system

acting as inputs. Fig. 3.17(a) shows the small signal model of the MPRO. Every single

transconductor injects noise into the oscillator by an equivalent noise current source,  hi  in . To
2

find the system transfer function, the loop can be opened at node v1 as shown in Fig. 3.17(b)

where all the noise sources injecting noise current at v1 are lumped into one noise source, H  in
2

where H is the total sizing factor. The system is equivalent to the circuit in Fig. 3.17(c) where the

output current from the equivalent transadmittance across v1 is subtracted from the input noise

51
current. The error current is converted into a noise voltage by the impedance Z(j)/H seen at

node v1 to ground where Z(j)=1/(go+jCg)=r/(1+ jτ).

Based on this model, and using Fig. 3.17(d), the open–loop gain, G(j), and the closed–loop

transfer function, T(j), of the MPRO can be expressed in equations (3.29) and (3.30)

respectively.

1 
G  j     Z  j     H  Ym  j  
H  (3.29)
 Z  j   Ym  j 

vn 1 Z  j 
T  j    j    (3.30)
iin H 1  G  j 

At the oscillation frequency, n, linear analysis predicts an infinite closed-loop transfer

function and an open-loop gain, G(jn) = –1. Using the first order Taylor series approximation

for G(j) and assuming that <<n, we can approximate T(jn+j) as

1 Z  jn 
T  jn  j   
H d (3.31)
G  j      
d   n

The noise voltage at the node v1 due to the noise current injected at this node can be expressed

as

2
___
___
i2
Z  jn 
vn2  jn  j    n
(3.32)
H d
G  j      
d    n

52
 (a)

(b)

(c) (d)

Fig. 3.17. Linear model used for noise analysis. (a) Schematic diagram for the MPRO in closed loop. (b)
Schematic of the MPRO in open loop configuration. (c) Equivalent circuit of the MPRO. (d) Block diagram
of the equivalent negative feedback system.

53
 By using assumptions similar to [31]1, the expression for the single-sideband phase noise due

to the noise sources from the N nodes is

___
2
1 v
L      Nn
2
2 V rms
2
(3.33)
Z  jn  r
2
i r N  
2 2
 n
  n 
Vo2 H
n
d
G  j     
d   n

where the mean-squared voltage of the oscillation waveform is Vrms2=Vo2/2.

To find an expression for the oscillator figure of merit, the normalized power consumption,

Pu, is defined as power consumed in mW by a single stage in the MPRO having a total sizing

factor, H=1. With this definition, the total power consumption of the MPRO can be expressed as

Ptot , mW  N  H  Pu (3.34)

Combining (3.33) and (3.34), the figure of merit of the MPRO is expressed in equation (3.35)

where Io(Vo) is the fundamental component of the output conductance current as in (3.10) and the

power is in mW.
1
 L     Ptot , mW 
FOM   
 n  
2

 d
2

 n G  j   
 Vo 2
1  d   n 
    (3.35)
 n
i 2
 r 2 P
 
u  N  Z  j n  r 
 
 
I 2 1 
  o    FMPRO  x 
 in Pu 
2

1
Assume that half of the power is causing phase noise while the other half is causing amplitude noise that is rejected by the nonlinear
amplitude limiting mechanism of the oscillator. Also assume that the voltage gain between different nodes of the MPRO close to the oscillation
frequency is approximately unity.
54
 The first term of (3.35) is dependent on both the circuit design of the unit buffer and the

MPRO coupling configuration. The coupling configuration determines the value of the dominant

mode gain which in turn determines the value of the oscillation amplitude, Vo, the output
___
conductance current, Io(Vo), and the equivalent noise current, in2 [32][33]. The normalized power

dissipation, Pu, is also dependent on mode gain at low values of the mode gain where power is

dominated by the switching power, Pu,sw=Vdd(2Vo)Cgf. Power is impacted not only through Vo but

is also proportional to (nτ). At higher values of the dominant mode gain, the total power

remains roughly constant since the inverting buffers behave as class-A amplifiers as the

oscillation amplitude decreases.

The second term in (3.35), FMPRO( x ), is only dependent on the coupling structure and not on

the unit buffer design. This term represents the filtering or the frequency selectivity of the

MPRO. In other words, it represents the effective quality factor of the oscillator. By optimizing

this term, we can determine the best phase noise performance of an MPRO given a certain mode

gain.

To derive an expression for FMPRO( x ), we first derive an expression for the transadmittance,

Ym(j). From Fig. 3.17(b), the output current is given by

iout  j   Hg m x T v  j   HYm  j  v1 (3.36)

where ~x =[x2 x3 … xN]T and v =[v2 v3 … vN]T. We assume as before that x1=0. Therefore, the open-

loop gain of the MPRO can be expressed as

gm go T
G  j   x  v v1  (3.37)
1  j

55
 To find the vector ( v /v1), we apply KCL at the node vk with k1. By defining the matrix X as

shown in (3.38) and the vector ~xF =[xN xN-1 … x2]T, we can express the vector ( v /v1) as shown in

(3.39). The open-loop gain can then be expressed by (3.40).

1  j 
g g x2  xN 1 
 m o 
 1  j 
 xN  xN  2 
X  g m go  (3.38)
     
 
 x 1  j 
x4 
 3
g m g o 

v v1    X 1  x F (3.39)

 g g  T
G  j     m o 1
 x  X  xF (3.40)
 1  j 

By differentiating (3.40) with respect to j and using the fact that G(jn)≈-1, the expression

for FMPRO( x ) can be written as (3.41). In this step, we use the effective transconductance and

output conductance, not the small-signal ones. Thus gm/go is equal to an.

2
  
FMPRO  x    n  1  x T  X n 1  X n 1  x F
2
(3.41)
 N 

From (3.40) and using the values of the normalized frequency and the mode gain in (3.6) and

(3.7), (1+jωτ)/an can be expressed as a sum of complex exponentials as (3.42)

x T  X n 1  x F   x T  Cn  C FnT  x F (3.42)

2 n  2 1 2 n  3 1 2 n  N 1 2 n  N 1 2 n  N  2  2 n  2 1

where Cn  [e j N
e
j
N
 e
j
N
]T and CFn  [e
j
N
e
j
N
 e
j
N
]T . Since all

elements of x and xF and all the real and imaginary components of Xn-1 are greater than or equal

56
to 0 and with at least one element greater than 0, then, X n 1  xF  Cn and xT  X n1  CFnT

which means

2
  
FMPRO  x    n  1  CFnT  Cn  n 
2 2
(3.43)
 N 

Interestingly, the value of FMPRO( x ) is equal to (n)2 hence reducing the expression for phase

noise to (3.44) and the FOM to (3.45). Since the unit power consumption is proportional, at

most, to (n) while FMPRO( x )=(n)2, maximizing the normalized oscillation frequency using

the algorithm explained in the previous section, also maximizes the FOM given a certain value of

the dominant mode gain.

2
in2  r 2 1 1  
L        n  (3.44)
Vo 2
HN n 2   

I 2 1 
FOM   o    n 
2

 i 2 Pu  (3.45)
 n 

The numerical value for the FOM depends on how the different terms in (3.45), other than

(ωnτ), vary as a function of the dominant mode gain. The analysis has been done for SPROs [34-

37]. The resulting expressions for FOM from [34-37] vary depending on the assumptions but the

final numerical values in all derivations predict an FOM in the vicinity of 170dB when the mode

gain approaches unity2.

In this work, we do not derive expressions for the different terms in (3.45). Although this is

possible, we believe that these expressions are of limited practical importance. The more

important conclusion in this paper is that for a certain value of the mode gain, maximizing the

2
An SPRO with a large number of stages is the case when the mode gain approaches unity.
57
normalized frequency gives the best FOM and accordingly, the best phase noise versus power

tradeoff.

Fig. 3.18 shows the simulated maximum FOM measured at a 10MHz offset frequency as a

function of the dominant mode gain for different channel lengths. These results are obtained by

simulating the MPROs in Table 3.2. The result for an1 is obtained using a 19-stage SPRO as an

example of a long chain SPRO and it agrees with the predictions in [34-37]. The maximum FOM

for an=2 is equal to that of a 3-stage SPRO as indicated in Table 3.2. In general, as the figure

shows, as the dominant mode gain increases, the maximum FOM decreases due to the faster

reduction in the first term in (3.45) compared to the possible increase in (ωnτ). Initially this

reduction is relatively fast for an between 1 and 3, and then the FOM stays relatively constant

until the mode gain approaches the small-signal dc gain of the unit inverter. At this point, the

FOM decreases sharply due to the fast reduction in oscillation amplitude.

Fig. 3.18. Simulated maximum FOM as a function of the mode gain. Phase noise is measured at a 10-MHz
offset frequency.

For applications requiring multiple clock phases, it is important to note that equation (3.45)

does not directly depend on the number of phases of the MPRO. As long as the mode gain and

the oscillation frequency stay constant, increasing the number of phases does not change the

58
FOM, theoretically. When the MPRO is sized correctly, the additional phases come at zero

power cost.

Table 3.3 shows the coupling configurations for a number of MPROs having different number

of phases spanning a range from 8 to 32. These coupling configurations have been obtained

using the optimization algorithm in Section 3.3.4 where mode stability is also taken into

consideration. The coupling configurations shown in the table are those obtained after quantizing

the resulting sizing factors into integers to be suitable for practical implementation.

The simulation results shown in Table 3.3 are obtained using the same unit cell in Fig.3.3 with

a channel length 200nm. As we can see from the table, although the number of phases spans 2

octaves, the FOM of all MPROs differs only by less than 0.5dB and their oscillation frequency

differs only by 3%. The FOM stays roughly constant because all of the oscillators are designed

to have approximately the same value for the dominant mode gain, an*, and the normalized

oscillation frequency, (nτ). If these MPROs are scaled such that all of them have the same total

power consumption then all of them would have the same thermal phase noise. These results

indicate that the oscillation frequency can be decoupled from the desired number of phases and

the additional phases can theoretically be obtained at zero power cost.

In practice, however, for a given power consumption, the number of phases cannot be

increased indefinitely. As the number of stages increases at the same total power, the driving

capability of the MPRO is divided among a greater number of stages reducing the driving

capability of each single phase. In addition, as the number of stages increases, the layout routing

becomes more complex leading eventually to a need for increasing the total oscillator power in

order to drive this increased loading.

59
 Table 3.3. Coupling Configurations and Simulation Results for Unloaded MPROs Designed for Dominant
Mode Gain  2 and Mode Gain Margin > 60% for Mode Stability.
Dominant Mode Mode Gain Normalized Freq Freq. Pwr FOM
N Coupling Configuration Mode (n*) Gain (an*) Margin (m) Freq (n*τ) (GHz) (GHz) (mW) (dB)
Eq (9) Eq (3.8) Eq (3.26) Eq (3.7) Eq (3.19) [Sim.] [Sim.] [Sim.]
8 (h8,h5,h3)=(4,1,3) 5 2.09 91% 1.52 3.46 3.31 9.55 165.4
9 (h9,h7,h5)=(2,4,1) 5 1.89 163% 1.39 3.34 3.12 8.94 165.8
10 (h10,h3)=(1,1) 7 1.79 123% 1.38 3.41 3.15 2.75 166.3
11 (h11, h8,h3)=(4,1,2) 7 1.81 122% 1.37 3.36 3.12 10.65 166.1
12 (h12,h10,h5)=(2,1,4) 7 1.88 86% 1.46 3.52 3.28 11.88 166.2
13 (h13,h8,h5,h3)=(5,2,1,2) 8 1.93 128% 1.43 3.40 3.20 18.71 165.8
14 (h14,h11,h3)=(3,1,2) 9 1.87 74% 1.47 3.55 3.29 11.83 166.2
15 (h15,h11,h7)=(2,3,4) 11 1.82 70% 1.40 3.43 3.19 18.78 166.2
16 (h16,h13,h6,h3)=(1,2,1,4) 11 1.93 107% 1.48 3.52 3.30 18.44 166.0
32 (h32,h20,h15,h10,h5)=(3,2,2,1,1) 19 1.83 64% 1.43 3.49 3.23 40.14 166.2

Based on the analysis we present in this paper, the design of a ring oscillator, given a certain

unit buffer design, should proceed in the following steps. First, the required number of phases,

the target oscillation frequency and the power budget should be specified. Second, the values of

the parameters ao, τo, and τp should be found by simulation for different channel lengths of the

devices. Third, using the optimization algorithm explained in Section 3.3, the optimum sizing

factors can be found. The constraint on maximum allowed mode gain must be decreased until the

minimum possible mode gain is found. This is done by running the algorithm and comparing the

oscillation frequency in (3.19) to the target frequency until the minimum mode gain is found. If

the required number of phases is odd and the target frequency is less than or equal the frequency

of an SPRO (3.20), the optimum design converges to an SPRO. Finally, the phase noise should

be simulated. If the phase noise performance is limited by flicker noise, then the overall

performance can be improved by using longer channel lengths for the devices at the cost of lower

speed. To compensate for the reduced frequency, the previous step should be repeated where the

optimum mode gain needs to be increased. In this way, the flicker noise performance is

improved at the cost of thermal noise.

60
 3.4. Conclusion

In this chapter, a general model for multipath ring oscillators having arbitrary coupling

structures is presented. With multiple modes of oscillation possible in MPRO structures, we

define the mode gain as the minimum buffer dc gain required to sustain a certain oscillation

mode. The value of the mode gain is only a function of the sizing factors of the different

coupling branches. The dominant mode is the oscillation mode with the lowest mode gain. We

show that several properties of the oscillator like the oscillation amplitude, the effective time

constant, the normalized power consumption and the effective noise factor are dependent on the

oscillator sizing only through the mode gain. An algorithm is introduced to calculate the

optimum sizing factors for the different coupling branches to achieve the maximum oscillation

frequency and the best phase noise taking into consideration the stability of the oscillation mode.

The results and methodology are verified by circuit simulation.

This chapter provides several guidelines for designing MPROs. The results of the

optimization show a simple heuristic for sizing an MPRO for maximizing the output frequency.

While frequency can be maximized by maximizing the mode gain of the dominant mode, a

designer should not choose a target mode gain near the maximum. Backing off from the

maximum avoids boundaries between modes, and improves an oscillator’s figure of merit.

Furthermore, the analysis indicates that for a given oscillation frequency and target mode gain,

the number of output phases can be increased without penalizing the figure of merit.

In the next chapter, the MPRO analysis is extended for optimizing the design of multiphase

frequency dividers. The divider is considered to be a further generalization of the multipath ring

oscillator, namely, a superharmonic injection-locked multipath ring oscillator.

61
 CHAPTER 4

Superharmonic Injection-Locked Multipath Ring Oscillators

In this chapter, we extend the analysis we presented in the previous chapter to include

frequency dividers. We show that any latch-based divider can be modeled and analyzed as an

injection-locked oscillator. The locking range of this oscillator is dependent on the oscillator

structure and the injection strength. By deriving accurate expressions for the locking range, a

design procedure can be deduced to maximize the maximum division frequency. If the divider is

not required to operate at the maximum frequency, high-speed can be traded-off for power

savings.

4.1. General Model

A superharmonic injection-locked MPRO (SHIL-MPRO) is an MPRO that is locked in phase

at steady state to an input periodic signal with a frequency that is an integer multiple (i.e. a

superharmonic) of the frequency of the MPRO itself [29]. Accordingly, a SHIL-MPRO is

typically used as a frequency divider. In this section, we propose a general model for a SHIL-

MPRO. This model can be used to find the optimal sizing factors and clocking configuration that

maximize the locking frequency of the divider.

To explain our model, we begin first by a simple example of the divide-by-2 shown in Fig.

4.1(a). The divider consists of two latches driven by a differential clock input. Fig. 4.1(b) shows

a CML implementation of the latch while Fig. 4.1(c) shows a possible CMOS implementation. In

both implementations, the clock signal modulates the transconductance of the amplifying

62
transistors by modulating their bias current. For the CMOS implementation, the clock signal also

modulates the output conductance of the inverters since it is dependent on the bias current.

(a)

(b) (c)

Fig. 4.1. Latch-based divide-by-2 (a) Divider schematic (b) CML latch implementation (c) CMOS latch
implementation

As a result of this mixing effect, the output current of any of the transconductors forming the

divider consists of many harmonic components. One component is a linear product due to the

input of the transconductor. Other components are the different intermodulation products of the

latch input with the clock signal. Fig. 4.2 shows an equivalent block diagram for the quadrature

divider. The divider can be considered an injection-locked quad-phase MPRO. The injection

currents are the result of mixing the MPRO outputs with the different input clock phases.

63
 Fig. 4.2. Equivalent model of the quadrature divider.

A general model for an M-phase input, N-phase output SHIL-MPRO is shown in Fig. 4.3. The

core oscillator is a general MPRO similar to that in Fig. 3.1. The parameters {hi} are sizing

parameters with respect to a reference unit inverter. The parameters {αi} take values 1 or 0

according to whether the ith branch of inverters is driven by the clock or not, respectively. We

assume in our analysis that a single-tone harmonic balance is accurate enough. Therefore,

assuming the MPRO is locked at a frequency ωo, the output voltages can be approximated by

(4.1). The transconductance of the inverter driven by the node vp and driving the node vq is

modulated by the clock signal u(q,p)(t) given by (4.2). We assume the input clock has a frequency

ωin and M distinct phases available. The clock phases distribution is determined by the

parameters (m1,m2) which can take integer values between 0 and M-1. By varying the values of

(m1,m2), all the possible uniform phase distributions can be modeled. Finally, the angle ψ is a

phase shift that depends on the injection frequency. In this analysis, we only consider the case of

tail injection. In general, direct injection can also be included by considering an additional set of

clock inputs modulating the output conductance of the inverters.

64
 Fig. 4.3. General model of an M-phase input N-phase output SHIL-MPRO.

 2 n  i 1 
vi  t   Vo cos  ot   (4.1)
 N 

 2  m1  q  1  m2  p  1   (4.2)
u( q , p )  t   U o cos  int   
 M 
 

The transconductance current of the inverter driven by vp and driving vq can be expressed by

(4.3). We assume the transconductance can be modeled as a memoryless nonlinear time varying

element. The first term in (4.3) represents the different components due to the transconductance

nonlinearity. The second component is the time variation function caused by the input clock

signal. This function is described by the coefficients αi,k. If the ith branch is clocked (αi=1), then

αi,k is equal to λk, the kth Fourier coefficient of the time modulation function. If the ith branch is

not clocked (αi=0), then αi,k is equal to unity for k=1 and is zero for k≠1.

   2 n  p  1  
ig ( q , p )  hp q 1   g m,kVo cos  kot  k 
 k 1  N  
(4.3)
   2  m1  q  1  m2  p  1   
   p q 1,k cos kint  k  k  
 M  
 k 0

65
 From all the frequency components given by (4.3), we are only interested in the component at

ωo. To calculate this component, we assume that at lock, the input frequency is an integer

multiple of the output frequency, that is ωin=dωo. By using this assumption, the

transconductance current component at ωo can be calculated as shown in (4.4). The first

component of this current is the linear gm product. The second and third components are the

injection current components due to upper side and lower side injection, respectively.

 2 n  p  1 
  hp  q 1   p  q 1,0  g m ,1Vo cos o t 
ig ( q , p ) 
o N
 
 2 n  p  1 2  m1  q  1  m2  p  1  
 

V
 o h p  q 1  g m , kd 1   p  q 1, k cos o t   kd  1 k  k 
2 k 1
 N M  (4.4)
 2 n  p  1 2  m1  q  1  m2  p  1  
 

Vo
 h p  q 1  g m , kd 1   p  q 1, k cos o t   kd  1 k  k 
2 k 1  N M 

At any node vq of the oscillator, the harmonic balance equation can be written as (4.5) where

go and Cg are the output conductance and input capacitance of the unit inverter, respectively. The
N
parameter H   hi is defined similar to the MPRO case. Substituting from (4.1) and (4.4) into
i 1

(4.5), the harmonic balance equation can be rewritten as shown in (4.6).

N dvq
i 
i 1
g q ,i  q 1
 Hg o vq  HC g
dt
0 (4.5)

N  2 n  q  1 2 n  i  1   2 n  q  1   2 n  q  1 
 h i i ,0 g m ,1Vo cos o t 
N

N
  Hg oVo cos o t 
N
  HC g oVo sin o t 
N

i 0      
Vo N 
 2 n  q  1 2 n  i  1   nd m2   nd m1  m2   

2
h  g i m ,  kd 1
 i , k cos o t 
N

N
 k  2 
N

M   i  1  2  N  M   q  1    
i 1 k 1        
Vo N 
 2 n  q  1 2 n  i  1   nd m2   nd m1  m2   

2
h  g i m ,  kd 1
 i , k cos o t 
N

N
 k  2 
N

M   i  1  2  N  M   q  1    
i 1 k 1        
0

(4.6)
66
 By comparing the first line in (4.6) to equation (3.2), we see that it represents the behavior of

the MPRO. The second and third lines of equation (4.6) represent the injection current. For the

output of the SHIL-MPRO to have uniform phases, the gain and phase conditions of the

oscillator at lock should be independent on the node where the harmonic balance is derived.

From (4.6), we can see that this can only happen if the condition given by (4.7) is satisfied. This

condition represents the allowed input clock phase distributions that can result in uniform output

phases. For the oscillator outputs to have N distinct phases, the mode index n and the number of

stages N should have no common factors. This condition implies that the minimum number of

input phases, M, required to have uniform phases at the output throughout the whole locking

range is equal to the number of output phases, N, divided by the division ratio, d, or M≥N/d.

Equation (3.7) thus serves as the mathematical formulation for the progressive injection phase

condition that is argued qualitatively in [27].

m1  m2 nd
 (4.7)
M N

To simplify the analysis, we consider only the lowest order intermodulation product and

assume that all the other components of the injection current can be neglected. Using this

assumption, (4.6) can be simplified to (4.8). The parameters hR,i HR, gR,o, and CR,g are equivalent

to the same parameters without the subscript ‘R’ in the free running MPRO. These are given by
N
hR,i=hiαi,0, H R   hii ,0 , gR,o=goH/HR, and CR,g=CgH/HR, respectively. The injection current
i 0

amplitude is given by (4.9) where αi,1= αiλ1 and λ1 is the first Fourier coefficient of the time-

domain modulation function.

67
 N
 2 n  q  1 2 n  i  1 
h g m ,1Vo cos o t 
R ,i
N

N

i 0  
 2 n  q  1   2 n  q  1 
 H R g R ,oVo cos o t    H R C R , g oVo sin o t   (4.8)
 N   N 
 2 n  q  1 
 I o cos o t   
 N 

1 N
 m n 
Io 
2
Hg m , d 1Vo  x
i 1
i i exp  j 2  1    i  1 
 M N 
(4.9)

Equation (4.8) describes the injection-locked MPRO (IL-MPRO) shown in Fig. 4.4. The unit

inverter of this IL-MPRO has a transconductance gm,1, an input capacitance CR,g and an output

conductance gR,o. The mode gain and the normalized free running frequency of the equivalent

MPRO without injection are given by (4.10) and (4.11), respectively where xR,i=hR,i/HR. The

injection current at the ith node of the IL-MPRO is given by (4.12) where Io is expressed by (4.9).

The substitution Ψ=tan-1(ωR,nτ)+θ is made for convenience since it makes the arbitrary phase

shift θ equal to zero when the locking frequency is equal to the free running frequency of the

equivalent core oscillator.

Fig. 4.4. Equivalent injection-locked MPRO model.

68
 1
aR , n  N
 2 n  i  1  (4.10)
  xR ,i cos  
i 1  N 

N
 2 n  i  1 
xsin 
R ,i
N

i 1
R ,n  N   (4.11)
 2 n  i  1 
 xR ,i cos  
i 1  N 

 2 n  i  1 
iinj ,i  t   I o cos  ot   tan 1 R,n     (4.12)
 N 

Equations (4.7)-(4.12) represent the mathematical model we use for the SHIL-MPRO

analysis. We use this model in the next section to calculate the locking range of a SHIL-MPRO.

4.2. Locking Ranges

Fig. 4.5 shows a phasor diagram for an injection-locked MPRO at lock. The current Ig is the

amplitude of the total transconductance current injected into a certain node vq. This current is

given by (4.13) where the negative sign in the first line indicates the current direction out of the

transconductors. The total current injected into the oscillator node IT can be decomposed into a

resistive component IT,R that is in phase with the node voltage and an orthogonal capacitive

component IT,C. The total transconductance current Ig has a phase shift of tan-1(ωR,nτ) compared

to the resistive current. For the oscillator to be locked at a frequency ωo, the total current IT must

have a phase shift tan-1(ωoτ) with respect to the resistive current. The extra phase shift, φ =

tan−1(ωoτ) − tan−1(ωR,nτ), must be provided by the injection current. The angle θ is greater than or

less than zero depending on whether the injection frequency is greater than or less than the free

running frequency of the equivalent MPRO, respectively.

69
 N
 2 n  q  1 2 n  i  1 
ig ,q  t    g m,1Vo hR ,i cos  ot   
i 1  N N 
H g V  2 n  q  1 
 R m,1 o 1  R ,n  cos  ot   tan 1 R ,n  
2
(4.13)
aR , n  N 
 2 n  q  1 
 I g cos  ot   tan 1 R ,n  
 N 

Fig. 4.5. Phasor diagram for an injection-locked MPRO at lock

The gain and phase conditions for locking can be derived from (4.8) by separating the

cos[ωot−2πn(q−1)/N] and sin[ωot−2πn(q−1)/N] coefficients. Equation (4.14) gives the equivalent

dc gain of the unit inverter at lock and (4.15) gives the phase condition from which the value of

the angle θ can be calculated. The injection coefficient η is defined in (4.16) where the value of

the injection current derived in (4.9) is used.

aR , n
an  (4.14)
1   cos  tan 1  R , n    

     sin  tan 1  R , n    
o  
R ,n
(4.15)
1   cos  tan 1  R , n    

70
 Io
1   R , n 
2

Ig
aR , n I o
 (4.16)
H R g m ,1Vo
1  H   g m , d 1  N
 m n 

2
 aR , n         x i i exp  j 2  1    i  1 
 H R   g m ,1  i 1  M N 

In the remainder of this section, we derive expressions for the different possible limitations on

the locking range.

4.2.1. Phase-Limited Locking Range

The first limitation on the locking range is due to the maximum phase shift φmax that can be

caused by the injection current. This limitation defines the phase-limited locking range (PLLR)

[22]. Fig. 4.6 shows the phasor diagram of the IL-MPRO at the edge of the PLLR. An expression

of the maximum and minimum locking frequencies due to the PLLR can be derived from the

figure trigonometry or by differentiating equation (4.15) with respect to θ. The PLLR expression

is given by (4.17) where η is the injection coefficient defined in (4.16).

Fig. 4.6. Phasor diagram for an IL-MPRO at the edge of the phase-limited locking range.

     1   2   R ,n 
2

o  PLLR max min  

R ,n (4.17)
1  2

Fig. 4.7 is a plot of the PLLR expression in (4.17) as a function of the injection coefficient η

for different values of the normalized free-running frequency (ωR,nτ). As the figure shows, for
71
small values of η, the locking range is a narrow range around the oscillator free running

frequency. As the injection strength increases, the locking range starts to increase. This behavior

is characteristic for weak injection and is explained in various previous publications as in [22].

However, as Fig. 4.7 shows, as the injection strength increases, the lower end of the PLLR

approaches zero while the upper end increases significantly until it tends to infinity when η

approaches unity. In practice, no matter how strong the injection level is, the locking range can

never tend to infinity which suggests the existence of additional limitations on the locking range

of a divider.

8
( R,n )=1
7
( R,n )=2
6 ( R,n )=3
( o )PLLR,(max/min)

4
3

2
1

0
0 0.2 0.4 0.6 0.8 1


Fig. 4.7. Phase-limited locking range.

4.2.2. Gain-Limited Locking Range

The second limitation on the locking range arises from the gain condition in (4.14). For

locking to occur, the available dc gain of the inverters of the equivalent MPRO must be greater

than the value given by (4.14). Otherwise, the oscillator cannot be locked. This limitation defines

the gain-limited locking range (GLLR). An expression for the GLLR can be derived from (4.14)

72
and (4.15) and is given by (4.18). The parameter amax is the maximum allowed mode gain for the

IL-MPRO at lock (4.14) while aR,n is the mode gain of the equivalent free running MPRO (4.10).

For our argument in this subsection, amax is the small signal dc gain of the inverters of the

equivalent MPRO and is equal to (HR/H)ao where ao is the small signal dc gain of the actual

inverters without switching.

 aR ,n  
2
a 
o GLLR MLLR max    max   R , n      1 
 2
  (4.18)
 aR ,n    amax  

For a certain value of (amax/aR,n), the upper end of the locking range becomes gain-limited

only when the injection coefficient, η, exceeds a certain threshold, ηg, given by (4.19). At this

value, both the PLLR and the GLLR have the same upper bound. When η exceeds this value, the

GLLR becomes the limit for the maximum locking frequency. The lower end of the locking

range is always phase-limited as long as amax≥aR,n which is the practical case of interest.

 R ,n    
2 2
a  a    a 
 R ,n  (4.19)
R,n
 g  1   R,n    R ,n  1   R ,n  
 amax  2 4  amax    amax  

Fig. 4.8 shows a plot of the GLLR given by (4.18) for (ωR,nτ)=2 and (amax/aR,n)= 1, 2, and 3.

As the figure shows, for small values of η, the locking range is phase limited. When η exceeds

the value of ηg, the maximum locking frequency becomes gain-limited. The limit increases when

the maximum available gain increases.

73
 Fig. 4.8. Gain-limited locking range for an IL-MPRO with (ωR,nτ)=2 and (amax/aR,n) = 1, 2, and 3

4.2.3. Mode-Limited Locking Range

The third limitation on the locking range arises for the existence of other possible modes of

oscillation for the MPRO. If the mode gain of the IL-MPRO (4.14) is higher than any of the

mode gains of the equivalent free running oscillator (4.10), this oscillation mode is able to build

up and the lock is lost. This limitation defines the mode-limited locking range (MLLR).

An expression for the MLLR is identical to that for GLLR if we define the maximum possible

mode gain amax as the minimum non-dominant mode gain. In general, equation (4.18) expresses

both the GLLR and the MLLR when amax is defined as given by (4.20).

 H g  
amax  min   a , min  a
 o l  0,1,..., N 1  R ,l  0 
 (4.20)
 H  ln 

74
 4.2.4. Quadrature-Phase Divider Example

In this subsection, we compare our analysis with simulation. A simple quadrature-phase

divide-by-2 is used in this example. In the simulations, we use the unit cell shown in Fig. 4.9. It

consists of differential CMOS inverters switched at the common source nodes of both the NMOS

and the PMOS transistors. The clock signal is assumed to be strong enough to turn off the current

path when it is in low state. Accordingly, the time modulation function is assumed to be a 50%

duty cycle square wave with an average λ0=0.5 and a first harmonic component λ1=2/π. When a

certain cell is not clocked, this means that CK=VDD at all times. The dc gain of the unswitched

inverter for VDD=1V in 65nm CMOS technology is equal to 8. The clock signal is a rail-to-rail

square-wave with 10% rise and fall time. The simulated maximum division frequency shown in

this section is for the unloaded divider which represents the absolute maximum frequency a

divider can work at. The effect of loading is discussed in the next section.

Fig. 4.9. Unit switched inverters used in simulation.

Fig. 4.10 shows the schematic of a quadrature phase (N=4) SHIL-MPRO. The sizing factors

of the feed-forward and cross-coupling inverters are (h4,h3)=(h,1). The equivalent MPRO has

two distinct oscillation modes, n=3 is the desired mode and n=2 is an undesired mode. Assuming

75
the maximum number of distinct input phases is M=4, the input clock phases must satisfy the

condition (m1+m2)=6 (4.7).

Table 4.1 shows all the allowed clocking configurations that satisfy equation (4.7). The table

also shows all the quantities needed to calculate the three locking ranges as a function of the

sizing ratio, h.

Fig. 4.10. Quadrature output divide-by-2.

Table 4.1. Possible clocking configurations for a quadrature-phase divide-by-2 shil-mpro

Index (α4,α3) m1 CK1 CK2 CK3 CK4 aR,3 aR,2 ωR,3τ η hGM
h h  0 h 1 0 a0  1
1 (0,1) X VDD 0o VDD 180o 1  0.71
0 h  0 0 2 0 a0  1

h 0  1 a0  1
1 3
2 (1,0) X 0o VDD 180o VDD 1  0h h 0  1 0h h 0 a 0  1
2

h1 1 0 a0  1
3 (1,1) 0 0o 180o 180o 0o 1 h h  h2  1  1.67
h1 20 0 a0  1
h1 1 0 a0  1
4 (1,1) 1 0o 90o 180o -90o 1 h h  h1  1.67
h1 20 0 a0  1
h1 1 0 a0  1
5 (1,1) 2 0o 0o 180o 180o 1 h h  h2  1  1.67
h1 20 0 a0  1
h1 1 0 a0  1
6 (1,1) 3 0o -90o 180o 90o 1 h h   h  1  1.67
h1 20 0 a0  1

Fig. 4.11 shows the normalized maximum locking frequency (ωoτ)max for the 6 clocking

configurations, calculated by substituting the values in Table 4.1 into (4.17–4.20), and the

maximum locking input frequency obtained by simulation. For the six configurations, at small
76
values of h, the free running frequency, (ωR,3τ), is small and thus the MPRO mode gain, aR,n, and

the IL-MPRO mode gain, an, are both smaller than the maximum available dc gain. This causes

the locking range to be phase limited. As the value of h increases, the value of an at the edge of

the PLLR exceeds the available dc gain. Accordingly, the locking range becomes gain limited.

As h increases further, the value of the non-dominant mode gain, aR,2, decreases below the

available dc gain. This decrease in the non-dominant mode-gain casuses locking range to be

mode limited. For a 4-phase SHIL-MPRO using this particular unit cell (ao=8), the maximum

locking frequency occurs at the intersection of the GLLR and the MLLR. The value of the sizing

ratio that satisfies this condition, hGM, is calculated for each clocking configuration in the last

column in Table 4.1.

Fig. 4.11 shows that the optimum sizing ratio for maximizing the upper limit of the locking

range can be predicted with good accuracy. Note that while the sizing factor of the maximum is

accurate, the plot shows some discrepancy in the magnitude or shape of the curves. The

discrepancy is expected because the normalizing time constant, τ, for an MPRO is a function of

the mode-gain. Accordingly for a SHIL-MPRO, τ is a function of the sizing parameters and

injection strength (4.14). Meanwhile, the theoretical calculation uses a constant τ for

normalization.

When the SHIL-MPRO is sized to maximize the upper end of the locking range, (4.9)

indicates that the equivalent injection current is large. Coupled with the low quality factor of ring

oscillators, the oscillator has a locking range that extends down to low frequencies and is

sufficient for most practical applications. A more detailed application of this model is presented

in Section 4.4.

77
 (a) (b) (c)

(d) (e) (f)

Fig. 4.11. Theoretical normalized maximum locking frequency and simulated maximum locking frequency
for an unloaded quadrature divide-by-2. (a) to (f) refer to the six clocking configurations in Table 4.1
respectively

4.3. Power Consumption and Effect of Loading

To study the effect of loading, we first consider an arbitrary unloaded SHIL-MPRO with a
N
locking range between fL and fH and with sizing factors {hi} where H   hi . For a fixed supply
i 1

voltage, the power consumption of this unloaded divider at lock is Pd(f). The power is only a

function of the injection frequency, fL ≤ f ≤ fH. The self-loading capacitance of this divider at each

of its outputs is equal to HCg where Cg is the equivalent input capacitance of the unit inverter as

defined before. The exact form of the function Pd(f) depends on the design of the unit inverter

78
and the divider architecture and is in general due to both the switching and the direct-path

currents.

If the whole unloaded divider is scaled with a factor Kd, the power consumption of the scaled

divider becomes KdPd(f) where fL ≤ f ≤ fH. If the scaled divider outputs are uniformly loaded with

a capacitance CL at each node, the divider behavior as a function of frequency scales in the

frequency axis by a factor (KdHCg+CL)/(KdHCg). Note that the RC ring-type ILO behavior is

always dependent on the normalized variable (ωoτ) rather than the absolute frequency. The

power consumption of the loaded scaled divider can be expressed by (4.21). Fig. 4.12(a) shows

an example for the frequency scaling due to loading. The figure shows the simulated power

consumption of the quadrature divider of Fig. 4.10 as a function of the injection frequency for

topology 2 in Table 4.1 which corresponds to Fig. 4.11(b). The divider uses the unit cell of Fig.

4.9 and has a sizing vector (h4,h3) = (3,1).

 K d HC g  C L  K d HC g K d HC g
Pdiv  f   K d  Pd  f  ; where fL  f  fH (4.21)
 K HC  K d HC g  C L K d HC g  C L
 d g 

To minimize power consumption, a divider must be scaled with the minimum scaling factor

Kd that still achieves the desired maximum operation frequency fmax<fH. This optimal scaling

factor and the corresponding power consumption can be expressed as (4.22) and (4.23)

respectively. In (4.23), Pu(fH) is the power consumption of a unit unloaded divider with a total

sizing factor H=1 at its maximum locking frequency fH and is equal to Pd(fH)/H. In practice, the

frequency fH should be chosen less than the actual maximum locking frequency by a factor large

enough to guarantee the divider achieves the desired frequency over all PVT variations.

79
 2

1.5

Pdiv(fin )/H [mW] 1 Simulated Pu(fin ): CL=0

[ equ.(24) ]: CL=4Cg
Simulation: CL=4Cg
0.5
[ equ.(24) ]: CL=12Cg
Simulation: CL=12Cg
0
0 10 20 30 40 50 60
fin [GHz]

(a)

100
Pdiv,opt(fmax ) / CL [ W / fF ]

80

60

40

20

0
0 10 20 30 40 50 60
fmax [GHz]

(b)

Fig. 4.12. Power consumption of a quadrature divider with sizing factors (h4,h3)=(3,1) and clocking
configuration (α4,α3)=(1,0) (a) As a function of input frequency for different load capacitances (b) As a
function of the maximum desired division frequency when the divider is scaled optimally.

CL f max
K d ,min  ; f max  f H (4.22)
HC g f H  f max

80
  f max  Pd  f H 
Pdiv ,opt  f max      CL
 f H  f max  HC g
(4.23)
 f max  Pu  f H 
   CL
 f H  f max  Cg

Fig. 4.12(b) is a plot of equation (4.23) where Cg=160fF for our unit cell in Fig. 4.9. If the

divider is always scaled for minimum power, the power consumption is always proportional to

the load capacitance. At frequencies that are small compared to fH, the power is proportional to

the maximum desired input frequency. As the input frequency approaches the maximum

injection frequency of the unloaded divider, fH, the required power increases rapidly with

frequency since the divider has to be scaled up in size to the limit where the external load

becomes small enough compared to its self-loading capacitance.

The analysis, in this section, shows that maximizing the maximum locking frequency of the

unloaded divider is useful even when a divider does not need to operate at such high locking

frequency since the divider can then be scaled down in size and save power compared to an un-

optimized design.

4.4. Design Procedure

The mathematical model we derived in equations (4.7) through (4.20) can be used in an

optimization algorithm to find the optimum design. Fig. 4.13 shows a flow chart indicating the

proposed design procedure. In step A, we set the required initial parameters like the number of

input and output phases, the inverter dc gain, the range of sizing factors to search, and the

characteristics of the time-domain modulation function. We use these parameters in step B to

form the design space grid. This grid consists of all the possible sizing factors, oscillation modes,

and clocking configurations. We limit the number of signal paths per stage to a maximum of 3

81
paths. In general, if the maximum frequency required cannot be achieved using 3-paths, then the

maximum number of paths should be increased and the search should be repeated. However, it

should be noted that in practice, as the number of paths increases, the layout design of the divider

becomes more complicated and the benefit obtained from using more paths starts to diminish.

In step C, we exclude the undesired solutions from the grid before the actual computation of

the maximum locking frequency. These solutions include those that contain any two branches, i

and i+N/2 since they have opposite phases. We also exclude all the modes that do not lead to

distinctive output phases. These are the modes that do not satisfy the condition gcd(n,N)=1 where

gcd() is the greatest common divisor of two integers. We finally exclude all the solutions that

contain ambiguity in the output phases.

In step D, we use (4.16-4.20) to calculate the maximum locking frequency for all the

remaining solutions in the design space. We sort these solutions in step E into three groups

according to the number of clocked branches. We then sort the solutions in each group in

descending order according to the maximum locking frequency. Finally in step F, we select the

best solution and plot its maximum locking frequency as a function of the sizing factors and

compare that to simulation. Since we do not have to examine all the possible combinations, the

design time is reduced significantly.

82
Fig. 4.13. Flow chart indicating the design procedure of a SHIL-MPRO divide-by-2.

83
 4.5. Octal-Phase Divider

In this section, we apply our procedure to design a quadrature-phase input, octal-phase output

divide-by-2. This divider is the first stage of a divider chain used to generate the clock signals

required by a high speed serializing transmitter which is explained in the next chapter. The initial

parameters for this design are N=8, M=4, in addition to the other parameters given in step A in

Fig. 4.13. The design space formed in step B includes three different kinds of architectures as

shown in Fig. 4.14. Each of these architectures can have different clocking configurations where

we only show the one where one branch is clocked. The first architecture in Fig. 4.14(a) uses two

separate differential-input quadrature-output dividers with a 90o phase shift between their input

clock phases. The advantage of this realization is that the maximum division frequency is equal

to that of a quadrature divider. However, due to the symmetry of the divider, this architecture

suffers from 180o ambiguity in the phase relationship between the outputs of the two dividers. As

part of step C, this ambiguity cannot be tolerated in most applications where the sampling phases

have to be predictable and accordingly it is eliminated from consideration. In general, this

condition can be identified whenever there exist isolated node sets in the divider.

Fig. 4.14(b) shows a second architecture where 4 cross-coupled latches are clocked with the

quadrature input clock to produce 8 phases at half the input frequency. This approach does not

suffer any phase ambiguity but our analysis, in addition to common practice, indicates that the

maximum division frequency is roughly half that of a 2-latch-based quadrature divider.

84
 (a)

(b)

(c)

Fig. 4.14. Quadrature-phase input, octal-phase output divide-by-2 with three possible implementations as (a)
Two separate differential-input quadrature-output dividers (b) Conventional 4 latches in a single loop and
(c) Triple-path SHIL-MPRO.

Fig. 4.14(c) shows a third realization as a triple-path SHIL-MPRO. If the coupling structure

and clocking configuration of this realization are optimized, the design can achieve a maximum

division frequency that is equal to that of a quadrature divider. At the same time, since the design

has no isolated node sets, it does not suffer any phase ambiguity. In fact, one can see that the

85
third realization is a general case that reduces to the first one if h8=0 and to the second one if

h3=0. In general, a quad-path SHIL-MPRO can also be used and the maximum division

frequency in that case can theoretically be even higher than a quadrature-divider. However, the

increased layout complexity practically limits the performance improvement. Accordingly, a

three-paths-per-stage architecture is chosen.

We used our model in the search algorithm to examine all the possible triple-path

implementations and find the best coupling structure and clocking configuration (steps D and E).

Fig. 4.15 shows the simulation results for the triple-path octal-phase SHIL-MPRO in Fig. 4.14(c)

where only the branch, h3, is clocked (α8,α5,α3) = (0,0,1). The figure shows the simulated and

theoretical maximum locking frequencies as a function of the sizing factor, h3, for different

values of h8. The value of h5 is given by h5=H-h8-h3 where H is arbitrarily taken equal to 10. The

figure shows that our model can predict the optimum design with very good accuracy. By

comparing Fig. 4.15 and Fig. 4.11, we see that the maximum locking frequency for the octal-

phase divider is approximately the same as that for the quadrature one. This shows that using

SHIL-MPROs, the maximum locking frequency and the number of output phases can be

decoupled which is the same conclusion we reached for MPROs.

For the first stage in our divider chain, we use a triple-path SHIL-MPRO with

(h8,h5,h3)=(2,2,6) and (α8,α5,α3)=(0,0,1). We scaled the design so as to minimize the power

consumption while still meeting the maximum frequency target at the worst PVT case.

86
 Fig. 4.15. Theoretical normalized (solid lines) and simulated (circles) maximum locking frequency for an
unloaded triple-path octal-phase divide-by-2 for (α8,α5,α3)=(0,0,1) and h5=10-h8-h3

4.6. Hexadecimal-Phase Divider

The second stage in the divider chain used in our transmitter is an octal-phase-input,

hexadecimal-phase output divide-by-2. Similar to the previous design, different implementations

are possible. Using the same design procedure, we chose a triple-path SHIL-MPRO with

(h16,h9,h7)=(2,2,6). Fig. 4.16 shows the clocking configuration and coupling structure.

Fig. 4.16. Octal-phase input, hexadecimal-phase output divide-by-2 as a triple-path SHIL-MPRO

87
 Fig. 4.17. Theoretical normalized (solid lines) and simulated (circles) maximum locking frequency for an
unloaded triple-path hexadecimal-phase divide-by-2 for (α16,α9,α7)=(0,0,1) and h9=10-h16-h7

Fig. 4.17 shows that the simulation results for the selected architecture have good agreement

with the analytical model. The maximum frequency of the 16-phase triple-path SHIL-MPRO is

approximately half that of the 8-phase design. This is acceptable for our case since it does

operate at half the frequency of the octal-phase divider. However, if a higher maximum speed is

needed, 4 or more paths per stage should be used at the cost of a more complicated layout.

4.7. Conclusion

We propose in this chapter a comprehensive analysis and model for superharmonic injection-

locked multipath ring oscillators. The analysis shows that SHIL-MPROs inherit the property of

88
MPROs such that one can increase the number of output phases without penalizing the maximum

division frequency. We explain three different limitations that determine the maximum locking

range of the frequency divider: phase (PLLR), gain (GLLR), and oscillation mode (MLLR). The

PLLR corresponds to the maximum phase shift that an equivalent injection current can cause.

The GLLR corresponds to the finite dc gain of the inverters forming the divider. And the MLLR

corresponds to the existence of other non-dominant oscillation modes that at some point can

become dominant and prevent the SHIL-MPRO from locking in the desired mode. We derive

expressions for each of these three limitations on locking ranges as functions of the sizing factors

and input clock distribution. These expressions can be used in a design procedure to find the

optimum coupling structure and clocking configuration that maximizes the locking frequency of

the divider. We also explain the effect of loading the divider with an external capacitive load and

show that maximizing the locking frequency can be used to save power if the divider is then

operated at a lower frequency. The design procedure can be used for a wide range of applications

for any ring-oscillator based divider.

The next chapter describes the detail implementation of a serial link transmitter that uses the

dividers presented in Section 4.5 and 4.6 to achieve low area and power for the data serialization.

89
 CHAPTER 5

A 32-48 Gb/s Serial Link Transmitter

The ability to achieve frequency division with multiple phases at relatively high frequency is

used to relax the timing requirements of the serializer hence allowing us to eliminate the need for

any delay calibration loops. This chapter applies this idea to a serial-link transmitter. We first

explain the system design of the transmitter in the first section and then give the details of the

circuit design in Section 5.2.

5.1. Transmitter System Design

Fig. 5.1 shows a block diagram of a serializing transmitter. The data path consists of a 64-to-1

serializer followed by a two-stage driver. The sampling clock signals required for the different

serializer stages are generated from a charge-pump phase-locked loop. An LC-VCO drives the

final N:1 multiplexer with an N-phase clock at a frequency fo/N where fo is the final bit rate. The

VCO clock is then divided by a chain of frequency-dividers to generate the lower frequency

clock signals required by the 64:N multiplexing stages.

To minimize the chip area of the transmitter, the use of inductors has been restricted only to

the VCO, the final multiplexing stage, and the following pre-driver and driver. All the other

multiplexing stages in addition to the whole frequency-divider chain are to be designed in

CMOS. In addition, to avoid the overhead added by delay-calibration loops or delay matching

buffers, the serializer must have a wide enough operation range to guarantee the timing

constraints are met for the desired data rate over all PVT corners.

90
 Fig. 5.1. Block-diagram of a 64-to-1 serializing transmitter.

In the following subsections, we explain the different architectural choices that we made to

arrive at the proposed architecture.

5.1.1. Final Multiplexing Stage Order

As explained in Chapter 2, as the order of the final multiplexer (N) increases, the timing

constraints at the final multiplexing stage become more relaxed. However, this has two

drawbacks. First, higher order multiplexers suffer lower bandwidth which means higher power

consumption to meet a certain target data rate. On the other hand, as N increases, the VCO output

frequency decreases which lowers the power consumption requirements of the divider chain as

explained in Chapter 4. Accordingly, to select the value of N that minimizes the total power

consumption, this trade-off has to be analyzed. The second problem is that when N increases, the

output signal becomes more sensitive to phase mismatches in the multiphase VCO. In addition, if

an LC-VCO is used, the number of inductors in the VCO increases increasing the chip area.

91
 By investigating the power trade-off between the MUX and the clock dividers, we found that

for the target data rate of ~40 Gb/s using 65-nm CMOS technology, a 4-to-1 MUX at the final

stage is the best compromise. If a 2-to-1 MUX is used, the first divider has to be clocked by a

~20-GHz input clock which requires burning a lot of power if inductive peaking is to be avoided.

On the other hand, if an 8-to-1 MUX is used, the power consumption required for the MUX to

meet the bandwidth requirement becomes excessive. In addition, a 4-to-1 multiplexer was found

to be a good compromise between relaxed timing constraints and phase-mismatch sensitivity.

5.1.2. Multiphase Sampling

Another design trade-off is the number of phases in the divided clock signals. Fig. 5.2 shows

the number of latches required to correctly retime the data input to the final 4:1 MUX. If the first

divided clock has a differential phase, the 4 outputs of the preceding 8:4 stage will have

transitions all at the same phase. Accordingly, the data has to be retimed before being applied to

the 4:1 MUX. Fig. 5.2(a) shows a possible implementation where a minimum of 7 latches is

required before the multiplexer. If the divided clock has a quadrature phase, the 4 data inputs will

have 2 transition phases which relaxes the number of latches to a total of 6 latches as shown in

Fig. 5.2(b). Finally, if the divided clock has 8 phases, the input data is already retimed and the

number of latches required is reduces to 4 latches as shown in Fig. 5.2(c).

Fig. 5.2 shows that if we can generate 8 phases for the divided clock, power and area savings

can be achieved by minimizing the required sampling elements at the input of the multiplexer.

This approach can be applied to all stages of the serializer. In general, for any multiplexing stage,

if the time resolution of the divided clock is equal to that of the input clock, the number of

92
retiming elements can be minimized. Time resolution can be maintained if the number of phases

of the divided clock is double that of the input clock.

(a)

(b)

(c)

Fig. 5.2. Required retiming elements at the input of the final 4-to-1 multiplexer when the divided clock has
(a) Two phases. (b) Four phases. (c) Eight phases.

93
 Based on the analysis in Chapter 4, multiphase dividers can be designed as superharmonic

injection-locked multipath ring oscillators at no additional power cost.

5.1.3. Transmitter Architecture

Fig. 5.3 shows the proposed transmitter architecture. The transmitter consists of a 64:1

serializer followed by a two-stage driver. The final stage in the serializer is a 4:1 MUX that

includes pulse generation and latching functionality. The required sampling clock signals are

generated from a phase-locked loop. A quadrature phase LC-VCO generates 4 phases at 12 GHz

that directly drive the final MUX and the first frequency divider. High speed, low power

frequency dividers are realized as superharmonic injection-locked multipath ring oscillators. The

dividers generate 8 phases at 6 GHz, 16 phases at 3 GHz, 8 phases at 1.5 GHz, and 16 phases at

750 MHz.

Since the number of phases at the output of each of the first and the second dividers is double

the number of phases at their inputs, the number of latches at the input of the 8:4 MUX and the

4:1 MUX can be minimized to a single latch per input. This saves a considerable amount of

power. In principle, it is possible to maintain the timing resolution for the whole divider chain.

However, due to the increased layout complexity of multiphase dividers and the lower power of

the low speed stages, the third divider is designed to have only 8 phases instead of 32. This

increases the number of latches required at the input of the 16:8 MUX to 1.5 latches per input

where the MUX itself is used as a latch. The final divider generates 16 phases at 750 MHz

maintaining the timing resolution of the preceding clock and thus allowing the use of a single

latch per stage for the 32:16 MUX. Since the timing constrains are relaxed at this low frequency,

94
we eliminate this latch. In effect, this means that the 32:16 MUX has an effective φ=0.5 since it

is transparent to the input for half the clock period.

A 64-channel PRBS generator generates a pseudorandom sequence of length (27-1), (215-1),

or (231-1) that is used as an input data source for the TX [48]. The PRBS generator can also be

bypassed and a 210-bit arbitrary sequence can be used as the input data. The data source is driven

by one of the 750 MHz clock phases and its output is retimed and applied to the input of the

serializer.

Fig. 5.3. Proposed transmitter architecture.

5.1.4. On-Chip Eye Monitor

To be able to measure the actual waveforms on the chip, we included an on-chip eye

monitoring system. A 6-bit phase interpolator [60] and a simple 8-bit R-2R DAC provide

variable time and voltage references for a high bandwidth comparator. The comparator

subsamples the output of the transmitter at 1/64 of the final data rate. The output of the

comparator is then compared in the control unit to one of the input channels of the transmitter

and bit errors are counted in a 32-bit counter. The measurement is repeated for every input clock
95
phase and reference voltage and the results are stored in a scan chain that is read in real time

using a logic analyzer.

Although the output is subsampled, the sampling aperture of the comparator has to be narrow

enough to avoid filtering the measured signal. In fact, the bandwidth requirements of the

comparator used here are the same as that that would be used in the receiver. The design of the

comparator is discussed in Section 5.2.

Fig. 5.4. On-chip eye monitoring system.

96
 5.2. Circuit Details

This section explains the details of the different circuit blocks used in the transmitter. We

begin with the building blocks of the phase-locked loop then we explain the design of the

multiplexer and finally, the design of the comparator is discussed.

5.2.1. Quadrature LC-VCO

Fig. 5.5 shows the schematic of the 12-GHz quadrature LC-VCO. Using the same naming

convention used for MPROs in Chapter 3, the cross-coupled inverters are sized by a scaling

factor h3 while the coupling inverters are sized by h4. An expression for the oscillation frequency

of the quadrature oscillator can be derived using a similar analysis to that of MPROs. In the

Appendix, we derive an expression for the oscillation frequency of a general multiphase,

multipath LC oscillator (MPLCO). This expression can be simplified as given by (5.1).

(a) (b)

Fig. 5.5. Schematic diagram of the quadrature LC-VCO: (a) VCO. (b) Unit capacitor used for coarse tuning.

2
R  
osc    R   o 2 (5.1)
2  2 

In (5.1), ωo is the resonance frequency of the equivalent tank circuit while ωR is the

oscillation frequency of the core ring oscillator that results if the equivalent inductance of the

97
tank circuit is eliminated. From the analysis in Chapter 3, As the coupling ratio (h4/h3) is

increased, the value of ωR increases causing the oscillation frequency to deviate more from

resonance. This increase in the coupling ratio improves the quadrature phase accuracy at the cost

of random phase noise degradation [49].

In practice, supply noise is a more serious source of jitter compared to the device noise. As

the supply voltage increases, both components that determine the oscillator frequency, ωR and

ωo, change in opposite directions. On one hand, the frequency of the core ring oscillator

increases with the supply voltage due to the increased current drive of the inverters. On the other

hand, the resonance frequency of the tank circuit decreases with supply voltage due to the

increase in the value equivalent tank capacitance [39]. The variation in oscillation frequency due

to supply variation can be calculated by differentiating (5.1) as given by (5.2). Since the first

term due to the core ring oscillator is always positive while the second term that represents the

tank resonance is always negative, a value of h4/h3 exists that causes the supply sensitivity to be

nulled.

osc 1  osc   R   o   o 

            (5.2)
Vdd 2  osc  R 2    Vdd   osc  R 2    Vdd 

Fig. 5.6 shows simulation results for the supply sensitivity defined by (5.3) and the oscillator

figure of merit (2.2). In addition to improving the phase accuracy, increasing the coupling ratio

can improve the supply rejection by more than 20dB at the cost of ~4dB degradation in random

phase noise for the same power consumption. In a typical mixed-signal chip environment, this

leads to an overall improvement in the overall jitter. Accordingly, in our design, we use a

coupling ratio h4/h3=1.

98
   osc osc  
SVDD  20 log   (5.3)
  Vdd Vdd  

Fig. 5.6. Effect of varying the coupling ratio (h4/h3) on the supply voltage sensitivity and the oscillator
figure of merit

5.2.2. Frequency-Divider Chain

The divider-chain consists of four stages. The first and second stages are explained in

Sections 4.5 and 4.6, respectively. They are realized as octal-phase and hexadecimal-phase

superharmonic injection locked triple-path ring oscillators. The first stage has a coupling

configuration (h8,h5,h3)=(2,2,6) and a clocking configuration (α8,α5,α3)=(0,0,1) while the second

stage has (h16,h9,h7)=(2,2,6) and (α16,α9,α7)=(0,0,1). These sizing ratios maximize the maximum

locking frequency as explained in the previous chapter with only one clocked branch to simplify

the layout design.

For the octal-phase divider in the third stage, we only used a conventional dual-path design as

the one in Fig. 4.15(b). The reason for this choice is because at the third stage, the frequency is 4

times less compared to that at the first stage and the loading capacitance is also less due to the
99
fanout in the transmitter data path. Accordingly, the optimally scaled triple-path design would

reach the limit of the minimum device size that we set and the dual-path design becomes more

power efficient due to reduced self-loading. Similarly, for the final hexadecimal-phase divider,

we used a dual-path design using 8 differential cross-coupled latches. The sizing ratios for the

third and fourth stages are (h8,h5)=(3,1) and (h16,h9)=(3,1), respectively. In both stages, only the

feedforward inverters are clocked. That is (α8,α5)=(1,0) for the octal-phase divider and

(α16,α9)=(1,0) for the hexadecimal-phase one.

5.2.3. Phase/Frequency Detector

Fig. 5.7(a) shows the block diagram of a conventional tri-state phase/frequency detector

(PFD) [51]. The flip-flop can be implemented as a TSPC-FF with an asynchronous reset signal

as shown in Fig. 5.7(b). Since the D-input of the flip-flop is always equal to ‘1’, node ‘X’ in Fig.

5.7(b) is always pulled down. Accordingly, both the first stage and the pull-down network of the

second stage can be removed without affecting the functionality of the flip-flop when used in a

PFD. Since the output has 0-to-1 transitions only when the reset input is activated, M6 in the

third stage can also be removed. The flip-flop thus reduces tot eh simplified form in Fig. 5.7(c).

Fig. 5.7(d) shows the schematic of the proposed PFD. It consists of two simplified dynamic

latches that require only a single phase for both the reference and divided clock inputs.

Minimizing the number of stacked devices allows the PFD to have sharp output transitions.

Simulations show fast rising/falling edges for frequencies up to more than 2 GHz.

100
 (a)

(b) (c)

(d)

Fig. 5.7. Evolution of the proposed phase/frequency detector (a) Block diagram of a conventional PFD
(b) TSPC flip-flop with an asynchronous reset (c) Simplified TSPC-FF assuming D=1 (d) Proposed PFD

101
 5.2.4. Charge-Pump

Fig. 5.8 shows the schematic diagram of the charge pump. Dual current compensation [52] is

used to achieve constant and matched UP/DN currents. In addition, current steering and OpAmp

biasing of node X reduces the dynamic mismatch [53]. Since the loop filter capacitors are

implemented using metal fringing capacitors, no leakage compensation is needed in the charge

pump. Due to the small static and dynamic mismatch, the measured spur level is less than −74

dBc without any calibration.

Fig. 5.8. Charge pump schematic

5.2.5. Multiplexer Design

The final stage in the serializer is a 4-to-1 multiplexer that operates up to 48 Gb/s. The MUX

is clocked with a quadrature-phase 12 GHz clock generated directly from an LC-VCO. Fig.

5.9(a) shows a block diagram of the multiplexer. Each unit cell generates a 25% duty-cycle pulse

by ANDing two adjacent clock phases. This pulse is used to connect one data input to the output

as shown in Fig. 5.9(b). The outputs of the 4 unit cells are added in the current-domain and

converted into a voltage by the output impedance of the multiplexer. Inductive shunt peaking is

used to extend the bandwidth of the circuit.

102
 Fig. 5.10 shows different possible methods to implement the multiplexer unit cell. In Fig.

5.10(a), a stack of 3 NMOS devices is used to perform the ANDing and the sampling operations

in a single stage. This technique is used in [54] to implement a 10 Gb/s 4:1 MUX in 0.13μm

CMOS technology. Although the number of internal nodes is minimal, the devices have to be

large enough to provide the current drive needed to produce the desired swing. Scaling up the

devices increases the capacitive loading of the VCO and the preceding latch which increases the

overall power consumption.

(a)

(b)

Fig. 5.9. Block diagram (a) and timing diagram (b) of a 4:1 multiplexer.

103
 (a) (b)

(c)

Fig. 5.10. Different possible implementations of the 4:1 MUX unit cell

To improve the current drive, a two-stage implementation can be used where pulse generation

is achieved using a separate AND gate as shown in Fig. 5.10(b). Since the AND gate is directly

driven by the sinusoidal VCO outputs, the output of the first NAND gate is not rail-to-rail and

the inverter burns an excessive amount of power due to both the switching and the direct-path

currents.

A modified two-stage implementation is shown in Fig. 5.10(c) [55]. Since the AND gate

output duty-cycle is increased to 50% and its switching activity is reduced by a factor of 2, the

power consumption can be reduced. In addition, the first stage now acts as a dynamic latch

where node X is pre-charged every clock cycle and then conditionally discharged by the input

data. Accordingly, no additional latches are required before the multiplexer.

The proposed implementation of the multiplexer unit cell is shown in Fig. 5.11(a). It is a

modified version of the design in [55] where the input latch is realized in PMOS. At the rising

104
edge of CK1, L1 holds the sampled data input at the internal capacitance of node X which then

propagates to the output as shown in Fig. 5.11(b). At the rising edge of CK2, node X is

discharged causing the output current to turn off. Since node X is pre-discharged rather than pre-

charged, the inverter is not needed and can be removed. The proposed design minimizes the

number of internal nodes and the direct-path current.

Fig. 5.11(c) shows the schematic diagram of the proposed 4-to-1 multiplexer. It uses a

differential version of the proposed unit cell. The transistor M1 is shared between the 4 branches

to provide a higher current drive for the discharging branch. Since each clock phase drives both

clock inputs in two different unit cells and each cell has a differential data input, the clock load is

symmetric for all phases and no phase mismatch can result from load imbalance.

(a) (b)

(c)

Fig. 5.11. Proposed 4-to-1 multiplexer (a) Unit cell (b) Timing diagram (c) Complete MUX
105
 5.2.6. Comparator Design

The comparator is used in the on-chip eye-monitoring system to sample the output waveform

of the transmitter. Accordingly, the sampling aperture of the comparator has to be narrow enough

such that it does not introduce significant filtering for the measured waveform.

Fig. 5.12(a) shows the schematic of the proposed comparator design. The comparator uses

both edges of the input differential clock to operate at double the input clock rate. During one of

the two input clock phases, one channel uses negative feedback to extend the tracking bandwidth

at the cost of the tracking gain. At the same time, the other channel uses positive feedback to

achieve regeneration. The switching between tracking and regeneration is achieved by reversing

the output polarity of the feedback stage as shown in Fig. 5.12(b). The design uses drain

switching rather than source switching to minimize the transition time from tracking to

regeneration which reduces the sampling aperture.

Fig. 5.13 shows the simulated sampling function of the comparator. The simulation is

performed using the periodic steady-state (PSS) and the periodic ac (PAC) analyses in spectreRF

as explained in [61]. The sampling instant is taken at the end of the regeneration period just

before the clock transition. The simulation is done using a 10 GHz input clock which

corresponds to a sampling rate of 20 GS/s. Fig. 5.13 shows that the sampling aperture is

approximately 14 ps which is less the minimum desired bit period of 20 ps.

106
 (a)

(b)

Fig. 5.12. Proposed drain-switched comparator: (a) Schematic diagram. (b) Equivalent block diagram.

Fig. 5.13. Simulated sampling aperture of the comparator.

107
 Fig. 5.14 shows the simulated sensitivity of the comparator. A ‘00110011’ sequence is

applied to the comparator input at double the clock rate. This causes a ‘0101’ output at each of

the comparator outputs. The time shift between the data transition and the clock edge is varied

and each time, the minimum signal amplitude needed to obtain correct outputs is found. The

minimum sensitivity occurs for a data-to-clock delay of ~30 ps which is equal to the delay of the

comparator, as can be seen from Fig. 5.13, and is less than 3-mV.

Fig. 5.13. Simulated sampling aperture of the comparator..

5.3. Conclusion

Serial link transmitters are systems that are inherently parallel and thus ones that can greatly

benefit from multiphase clocking. In this chapter we showed how multiphase sampling can be

used to optimize the operation bit-rate window such that the transmitter can operate for a wide

range of bit rates. This leads to designing transmitters that are robust to PVT variations and

eliminate the need for any delay matching buffers of delay calibration loops leading to a simpler,

more compact, and lower power design.

108
 Multiphase sampling can also allow the designer to minimize the number of sampling

elements between the different serializer stages which are needed for resynchronization. If the

timing resolution of the sampling clock is maintained between different stages, the required

number of latches can be reduced by 43% which leads to significant improvement in power and

chip area provided that phase generation can be achieved at no extra overhead. This is

accordingly a direct application for the results of our analysis in Chapter 4.

Various circuit techniques are proposed to improve the power efficiency and performance of

the transmitter. An accurate quadrature-phase low supply sensitivity VCO design has been

explained to minimize the deterministic jitter at the output of the transmitter. A novel PFD and

an accurate charge pump reduce the reference spur level to less than -74 dBc which leads to

negligible sinusoidal jitter at the output. A new 4-to-1 multiplexer design, that achieves the

desired bandwidth requirements at low power, is proposed. And finally, a low sampling aperture

comparator using drain-switching that is used in an on-chip eye monitoring system is explained.

In the next chapter, we show the measurement results of the fabricated transmitter chip.

109
 CHAPTER 6

Experimental Results

The transmitter explained in the previous chapter is fabricated in 65-nm 1P9M CMOS

technology and tested in a chip-on-board assembly. Fig. 6.1 shows the chip and the printed

circuit board photographs. The chip occupies a total area of 1.2 mm2. In this chapter, we explain

the details of the testing setup and discuss the measurements results.

(a) (b)

Fig. 6.1. Chip photograph (a) and testing board (b)

6.1. Testing Setup

Fig. 6.2 shows a block diagram of the testing setup. The transmitter output is probed using a

50-GHz bandwidth GSSG cascade probe. The probe is either connected to a 20-GHz bandwidth

subsampling scope or a 50-GHz bandwidth spectrum analyzer. An external signal generator

generates a 750-MHz reference clock that is used both as a reference to the PLL and to trigger

the scope. Different control bits are written into an on-chip scan chain using a pattern generator

110
and can be read out using a logic analyzer. The interface between the pattern generator, the logic

analyzer, the chip and the PC is controlled by a National Instruments data acquisition (NI-DAQ)

interface.

Fig. 6.2. Testing setup

6.2. Measurement Results

Fig. 6.3 shows the measured PLL output spectrum at 10.56-GHz output. To test the PLL, a

‘0011’ sequence is applied to the serializer causing the final 4:1 MUX, the pre-driver, and the

driver to act as a 3-stage clock buffer for the VCO output. The locking range of the PLL extends

111
from 7.9 GHz to 12.1 GHz limited only by the VCO tuning range. The worst reference spur level

over the locking range of the PLL is −74 dBc for a carrier frequency of 10.56 GHz.

-10

-20

-30
SPECTRUM (dBm)

-40

-50

-60

-70

-80

-90

9.81 10.06 10.31 10.56 10.81 11.06 11.31

FREQUENCY (GHz)

Fig. 6.3. Measured spectrum of the PLL output for a 10.56 GHz output carrier frequency

Fig. 6.4 shows the measured single side-band phase noise for a carrier frequency of 12.08

GHz near the upper end of the locking range. The out of band phase noise at 10 MHz offset is

−127.5 dBc/Hz. The small bump around 1 MHz is due to some under-damping in the loop

dynamics. The integrated phase noise from 100 Hz to 10 MHz is equal to 1.09 degree-rms which

corresponds to 251 fsecrms random jitter. As one would expect from the very low spur level, the

periodic jitter is very low.

112
 -60

PHASE NOISE (dBc/Hz)

-80

-100

-120

-140

100 1K 10K 100K 1M 10M

FREQUENCY OFFSET

Fig. 6.4. Measured phase noise at 12.08 GHz carrier frequency

Such low jitter numbers are not measurable using our equipment. Fig. 6.5 shows the time

domain measured jitter histogram. The measured jitter is less than 0.66 psecrms limited by the

triggering jitter of the scope.

Fig. 6.5. Measured jitter for a 12.08 GHz carrier frequency.

113
 The serializer operates properly over the whole locking range of the PLL. This range

corresponds to data rates of 32-48 Gb/s. This wide operation range is due to the optimum choice

of the sampling phases at all the multiplexing stages. Fig. 6.6 shows the measured single-ended

eye-diagram at the output of the transmitter at 48 Gb/s while Fig. 6.7 shows the measured eye

using the on-chip eye monitoring. The vertical eye opening measured on chip for the single-

ended output is ~270 mV while that measured by the scope is ~150 mV. This difference is due to

the attenuation caused by the probes, the 48” cables and the sampling front-end of the digital

scope. This attenuation is estimated as ~7 dB at 24 GHz.

Fig. 6.6. Measured eye-diagram at 48 Gb/s using a (231−1) PRBS

114
 1.2

1.1

1.0
[ Volt ]

0.9

0.8

0.7

0.6

-0.5 0 0.5
[ UI ]
Fig. 6.7. Eye diagram measured using on-chip eye monitor.

The chip operates at data rates 31.68−48.4 Gb/s. It consumes a total power of 88 mW from a

1.2 V supply. Table 6.1 shows the power breakdown of the transmitter while Table 6.2 compares

the achieved performance with other serializing transmitters operating at similar data rates. The

use of high-speed power-optimized multiphase frequency dividers allows us to remove all forms

of delay adjustment overhead that is used in other designs. In addition, it allows us to minimize

the number of retiming latches in the data path and to realize the whole divider chain in CMOS

without using any inductors. All of which leads to the significant power and area savings

compared to previous work. These savings are achieved with a phase noise and jitter

performance that is comparable with other approaches.

115
 Table 6.1. Power breakdown for the transmitter chip.

Block Power (mW) Fraction (%)

VCO 26.6 30.2

Divider Chain 18 20.5

Buffer/PFD/CP 2 2.3

Predriver/Driver 26.4 30

Serializer 15 17

Total 88 100

Table 6.2. Performance Summary

Ref [5] [7] [8] This Work

Technology 130nm CMOS 65nm CMOS 65nm CMOS 65nm CMOS

Supply Voltage [V] 1.45 1.1 1.2 1.2

Data Rates [Gb/s] 35.2−38.8 39.8−44.6 39.8−44.6 31.68−48.4

PLL Freq [GHz] 17.6−19.4 39.8−44.6 19.9−22.3 7.92−12.1

PN@10MHz (dBc/Hz) -113.5 -108 -120 -127.5

Jitter [fsecrms] 650 238 640 251

TX Area [mm2] 3.74 5 4.6 1.2

TX Power [W] 1.56 0.76 0.86 <0.1

116
 CHAPTER 7

Conclusion

We present in this dissertaion an accurate model and a comprehensive analysis that explain

the operation of multipath ring oscillators [57] and frequency dividers [58]. We showed that

using multipath coupling, the number of phases of a ring oscillator or a frequency divider can be

increased arbitrarily while maintaining the oscillation or division frequency. In fact, by

increasing the number of phases, it is possible to push the maximum operation frequency to

higher values at the cost amplitude reduction or equivalently degraded phase noise. We show

that the only limitation on the maximum operation frequency of a ring oscillator or a frequency

divider is the cut-off frequency of the technology and not the number of phases needed at its

output.

The mode stability problem which is characteristic of multipath ring structures is part of the

analytical study. To help designers traverse the large design space, a design procedure based on a

linear optimization algorithm is provided to allow a designer to quickly arrive at the optimum

coupling structure that maximizes the oscillation frequency given a certain desired number of

phases and phase noise performance. This optimum structure is chosen while considering mode

stability.

The oscillator analysis is further extended to apply to frequency dividers. The locking range

of superharmonic injection-locked multipath ring oscillators is studied under the condition of

strong signal injection. We show that, unlike the previously studied weak injection scenario, the

117
actual locking rang of the divider is not actually phase limited. Instead, it is either limited by the

maximum gain available in the circuit or by the existence of other non-dominant modes that can

prevent the desired mode from being sustained.

Similar to our work on oscillator design, a design procedure that uses the derived

mathematical model is proposed to help a designer to quickly arrive at the best divider design

that maximizes the division frequency for a certain power budget.

The ability to generate multiple phases at high frequencies and with low power is useful in

numerous applications. As an example, we showed how this can help design a high speed serial

link transmitter allowing it to operate at wide range of data rates without the need for any delay

adjustment overhead. As shown in out measurement results, the architecture also minimizes the

number of high-speed latches which results in significant power savings in comparison with

previously published results.

The analyses presented in this work can be extended to many of the applications mentioned in

Chapter 1. One example is to design a complete clock network as one big coupled oscillator. The

effect of different loads and routing parasitics can be taken into account and compensated for by

proper adjustment of coupling strengths. Since the figure of merit an optimally designed

multipath ring oscillator can be maintained regardless of the number of phases it has, this

suggests that as the clock network grows in size, the resulting phase noise or jitter at any of its

outputs can actually decrease. This conclusion is opposite to common wisdom which suggests

that as more buffering elements are added, the jitter performance should degrade.

Another extension is to perform this coupling in clock networks and embed the frequency

dividers. This extension is not unlike the clock divider approach described in this dissertation
118
except the feedback divider can be considered as part of the oscillator itself as opposed to an

explicit element downstream from the oscillator. In addition to reducing jitter, inter-stage

coupling allows the design of a clock network with accurate and well defined phase relationships

between the different outputs.

119
 Appendix

Oscillation Modes in Multipath LC Oscillators

Fig. A.1 shows the general form of a multipath LC oscillator (MPLCO). Similar to the

MPRO, we assume all coupling paths exist. Each row of inverters is sized by a sizing factor, hi,

with respect to a certain reference inverter where i=1,2,3,…,N. The unit inverter has an

equivalent input capacitance, Cg, an output conductance, go, and an equivalent transconductance,

gm. Unlike the MPRO, each of the output nodes of the MPLCO is loaded with an LC tank circuit

having a capacitance, Ct, and an inductance, L. The finite quality factor of the inductance is

modeled with a parallel conductance gt=1/(QωL).

Fig. A.1. General form of a multipath LC oscillator.

The output voltage vi can be expressed by (A.1) where Vo is the oscillation amplitude, ωn is

the oscillation frequency, and n=0,1,2,…,N-1 is the oscillation mode index. The gain and phase

120
oscillation conditions can be derived by applying KCL at any of the oscillator nodes similar to

the MPRO analysis. Equation (A.2) shows the harmonic balance equation at node v1.

 2 n  i  1 
vi  t   Vo cos  nt   (A.1)
 N 

N
 2 n  i  1   gt N 
 h g V
i m o cos 
 n t 
N
 o    hi  Vo cos n t 
 g
i 1    g o i 1 
(A.2)
 Ct N  1
n Cg    hi  Vo sin nt   V sin n t   0
 Cg i 1  n L o

By separating the coefficient of cos(ωnt), the gain condition can be derived and the mode gain

can be expressed by (A.3) where aR,n is the mode gain of the core MPRO that results if the LC

N
tanks are disconnected (3.8) and H   hi is defined similar to the previous MPRO and SHIL-
i 1

MPRO analyses. Equation (A.3) shows that the dc gain required from the inverters in the case of

an MPLCO is slightly higher than that for an MRPO due to the loading caused by the finite

quality factor of the tank circuit.

1   gt Hg o 
an 
 2 n
 i  1 
N
 xi  cos  (A.3)
i 1  N 
 1   gt Hg o   aR ,n

The phase condition can be derived by separating the coefficient of sin(ωnt) in (A.2). Due to

the presence of the inductor, the phase condition is a quadratic equation of the oscillation

frequency that has two solutions rather than just one as for an MPRO. The oscillation frequencies

of the nth mode can be expressed by (A.4) where ωR,n is the oscillation frequency of the core

MPRO if the tank circuits are disconnected, expressed by (3.7), and ωo=[L(Ct+HCg)]−1/2 is the

121
resonance frequency of the tank circuit loaded with the inverters. Equation (A.4) shows that if

the core MPRO does not have a zero frequency, the oscillation frequency of the MPLCO can

take one of two values that is either higher than or less than the resonance frequency. This each

value of the mode index, n, corresponds to two oscillation modes. The first mode has a higher

oscillation frequency than the resonance frequency where the tank load appears capacitive and

thus we refer to it as the capacitive mode and we denote it with the subscript ‘C’. The second

mode has a lower frequency than resonance where the tank circuit appears inductive and thus we

refer to it as the inductive mode and we denote it with the subscript ‘L’.

2
   1   gt Hg o    R , n  1   gt Hg o  
n  C L    R , n        o 2 (A.4)
 2  1   Ct HCg    2  1   Ct HCg  

The analysis in its current form does not show whether the capacitive or the inductive mode is

dominant. This is due to the symmetry of the impedance of the tank load as a function of

frequency. In practice, the finite quality factor of the inductor is more precisely modeled as a

series resistance rather than a parallel conductance [56]. In this case, the equivalent parallel

conductance of the tank circuit is lower at higher frequencies causing the mode gain of the

capacitive mode to be lower than that of the inductive one and thus becomes more dominant.

If the quality factor of the tank circuit is limited by that of the capacitance rather than the

inductance, the situation is reversed where the inductive mode becomes the dominant mode. This

situation, however, is rare in practice and it can easily be avoided by good layout design.

In the special case of a cross coupled oscillator, the oscillation frequency of the core MPRO is

equal to zero and (A.4) reduces to the resonance frequency.

122
 References

[1] Anantha P. Chandrakasan, Samuel Sheng, and Robert Brodersen, “Low-Power CMOS
Digital Design,” IEEE Journal of Solid-State Circuits, vol. 27, no.4, pp. 473–484, Apr. 1992.

[2] Zhiyu Ru, Niels A. Moseley, Eric A. M. Klumperink, and Bram Nauta,“Digitally
Enhanced Software-Defined Radio Receiver Robust to Out-of-Band Interference,” IEEE Journal
of Solid-State Circuits, vol. 44, no. 12, pp. 3359–3375, Dec. 2009.

[3] David Murphy, Amr Hafez, Ahmad Mirzaei, Mohyee Mikhemar, Hooman Darabi, Mau-
Chung Frank Chang, Asad Abidi, “A Blocker-Tolerant Wideband Noise-Cancelling Receiver
with a 2dB Noise Figure,” International Solid-State Circuits Conference (ISSCC), pp. 74–75,
Feb. 2012.

[4] Xiang Gao, Eric A. M. Klumperink, and Bram Nauta, “Advantages of Shift Registers
Over DLLs for Flexible Low Jitter Multiphase Clock Generation,” IEEE Transactions on
Circuits And Systems – II: Express Briefs, vol. 55, no. 3, pp. 244–248, Mar. 2008.

[5] Jeong-Kyoum Kim, Jaeha Kim, Gyudong Kim, and Deog-Kyoon Jeong, “A Fully
Integrated 0.13-μm CMOS 40-Gb/s Serial Link Transceiver,” IEEE Journal of Solid-State
Circuits, vol. 44, no. 5, pp. 1510–1521, May 2009.

[6] Jaeha Kim, et al.,“A 20-GHz Phase-Locked Loop for 40-Gb/s Serializing Transmitter in
0.13-μm CMOS,” IEEE Journal of Solid-State Circuits, vol. 41, no. 4, pp. 899–908, Apr. 2006.

[7] Shunichi Kaeriyama et al., “A 40 Gb/s Multi-Data-Rate CMOS Transmitter and Reciever
Chipset With SFI-5 Interface for Optical Transmission Systems,” IEEE Journal of Solid State
Circuits, vol. 44, no. 12, pp. 3568–3579, Dec. 2009.

[8] Kouichi Kanda, et al., “A Single-40 Gb/s Dual-20 Gb/s Serializer IC With SFI-5.2
Interface in 65 nm CMOS,” IEEE Journal of Solid State Circuits, vol. 44, no. 12, pp. 3580–3588,
Dec. 2009.

123
[9] Nikola Nedovic, et al., “A 3 Watt 39.8-44.6 Gb/s Dual Mode SFI5.2 SerDes Chip Set in
65 nm CMOS,” IEEE Journal of Solid State Circuits, vol. 45, no. 10, pp. 2016–2029, Oct. 2010.

[10] John G. Maneatis and Mark A. Horowitz, “Precise delay generation using coupled
oscillators,” IEEE Journal of Solid-State Circuits, vol. 28, no. 12, pp. 1273–1282, Dec. 1993.

[11] S. J. Lee, B. Kim, and K. Lee, “A novel high-speed ring oscillator for multiphase clock
generation using negative skewed delay scheme,” IEEE Journal of Solid-State Circuits, vol. 32,
no. 2, pp. 289–291, Feb. 1997.

[12] Lizhong Sun and Tadeusz A. Kwasniewski, “A 1.25-GHz 0.35-m Monolithic CMOS
PLL Based on a Multiphase Ring Oscillator,” IEEE Journal of Solid-State Circuits, vol. 36, no.
6, pp. 910–916, Jun. 2001.

[13] W. Yan and H. Luong, “A 900-MHz CMOS Low-Phase-Noise Voltage-Controlled Ring

Oscillator,” IEEE Transactions on Circuits and Systems – II: Analog and Digital Signal
Processing, vol. 48, no. 2, pp. 216–221, Feb. 2001.

[14] Chan-Hong Park and Beomsup Kim, “A Low-Noise, 900-MHz VCO in 0.6-um CMOS,”
IEEE Journal of Solid-State Circuits, vol. 34, no. 5, pp. 586–591, May 1999.

[15] T. Kan, G. Leung, and H. Luong, “A 2-V 1.8-GHz Fully Integrated CMOS Dual-Loop
Frequency Synthesizer,” IEEE Journal of Solid-State Circuits, vol. 37, no. 8, pp. 1012–1020,
Aug. 2002.

[16] Yalcin Alper Eken, and John P. Uyemura, “A 5.9-GHz Voltage-Controlled Ring
Oscillator in 0.18-m CMOS,” IEEE Journal of Solid-State Circuits, vol. 36, no.1, pp. 230–233,
Jan. 2004.

[17] Matthew Z. Straayer, and Michael H. Perrott, “A Multi-Path Gated Ring Oscillator TDC
with First-Order Noise Shaping,” IEEE Journal of Solid-State Circuits, vol. 44, no. 4, pp. 230–
233, Apr. 2009.

124
[18] Junfeng Xu, Shwetabh Verma, Thomas H. Lee, “Coupled Inverter Ring I/Q Oscillator for
Low Power Frequency Synthesis,” in Proc. IEEE Symp. VLSI Circuits, 2006, pp. 172–173.

[19] S. Mohan, W. Chan, D. Colleran, S. Greenwood, J. Gamble, and I. Kouznetsov,

“Differential Ring Oscillators with Multipath Delay Stages,” in Proc. IEEE 2005 Custom
Integrated Circuit Conf., 2005, pp. 503–506.

[20] R. L. Miller, “Fractional-Frequency Generators Utilizing Regenerative Modulation,”

Proc. I.R.E., pp. 446–457, Jul. 1939.

[21] Aminghasem Safarian, Seema Anand, and Payam Heydari, “On the Dynamics of
Regenerative Frequency Dividers,” IEEE Transactions on Circuits and Systems—II: Express
Briefs, vol. 53, No. 12, pp. 1413–1417, Dec. 2006.

[22] Shwetabh Verma, Hamid R. Rategh, and Thomas H. Lee, “A Unified Model for
Injection-Locked Frequency Dividers,” IEEE Journal of Solid-State Circuits, vol. 38, no. 6, pp.
1015–1027, Jun. 2003.

[23] Marc Tiebout, “A CMOS Direct Injection-Locked Oscillator Topology as High-

Frequency Low Power Frequency Divider,” IEEE Journal of Solid-State Circuits, vol. 39, no. 7,
pp. 1170–1174, Jul. 2004.

[24] Stefano Dal Toso, Andrea Bevilacqua, Marc Tiebout, Nicola Da Dalt, Andrea Gerosa,
and Andrea Neviani, “An Integrated Divide-by-Two Direct Injection-Locking Frequency Divider
for Bands S Through Ku,” IEEE Transactions on Microwave Theory and Techniques, vol. 58,
no. 7, pp. 1686–1695, July 2010.

[25] Ken Yamamoto, Minoru Fujishima, “A 44-μW 4.3-GHz Injection-Locked Frequency

Divider With 2.3-GHz Locking Range,” IEEE Journal of Solid-State Circuits, vol. 40, no. 3, pp.
671–677, Mar. 2005.

125
[26] Jun-Chau Chien, and Liang-Hung Lu, “Analysis and Design of Wideband Injection-
Locked Ring Oscillators With Multiple-Input Injection,” IEEE Journal of Solid-State Circuits,
vol. 42, no. 9, pp. 1906–1915, Sep. 2007.

[27] Ahmad Mirzaei, Mohammad E. Heidari, Rahim Bagheri, and Asad A. Abidi, “Multi-
Phase Injection Widens Lock Range of Ring-Oscillator-Based Frequency Dividers,” IEEE
Journal of Solid-State Circuits, vol. 43, no. 3, pp. 656–671, Mar. 2008.

[28] Hui Zheng and Howard C. Luong, “A Double-Balanced Quadrature-Input Quadrature-

Output Regenerative Frequency Divider for UWB Synthesizer Applications,” IEEE Transactions
on Circuits and Systems–I: Regular Papers, vol. 55, no. 9, pp. 2944–2951, Oct. 2008.

[29] Hamid R. Rategh, and Thomas H. Lee, “Superharmonic Injection-Locked Frequency

Dividers,” IEEE Journal of Solid-State Circuits, vol. 34, no. 6, pp. 813–821, Jun. 1999.

[30] D. W. Jordan, and P. Smith, Nonlinear Ordinary Differential Equations, Oxford

University Press, 2007.

[31] B. Razavi, “A study of phase noise in CMOS oscillators,” IEEE Journal of Solid-State
Circuits, vol. 31, no. 3, pp. 331–343, Mar. 1996.

[32] A. Hajimiri, S. Limotyrakis, and T. H. Lee, “Jitter and phase noise in ring oscillators,”
IEEE Journal of Solid-State Circuits, vol. 34, no. 6, pp. 790–804, Jun. 1999.

[33] Alper Demir, A. Mehrotra, and J. Roychowdhury, “Phase noise in oscillators: a unifying
theory and numerical methods for characterization,” IEEE Transactions on Circuits and Systems
– I: Fundamental Theory and Applications, vol. 47, no. 5, pp. 655–673, May 2000.

[34] J. A. McNeill, “Jitter in ring oscillators,” IEEE Journal of Solid-State Circuits, vol. 32,
no. 6, pp. 870–879, Jun. 1997.

126
[35] L. Dai, and R. Harjani, “Design of low-phase-noise CMOS ring oscillators,” IEEE
Transactions on Circuits and Systems – II: Analog and Digital Signal Processing, vol. 49, no. 5,
pp. 328–338, May 2002.

[36] R. Navid, T. H. Lee, and R. W. Dutton, “Minimum achievable phase noise of RC

oscillators,” IEEE Journal of Solid-State Circuits, vol. 40, no. 3, pp. 630–636, Mar. 2005.

[37] Asad. A. Abidi, “Phase noise and jitter in CMOS ring oscillators,” IEEE Journal of Solid-
State Circuits, vol. 41, no. 8, pp. 1803–1816, Aug. 2006.

[38] Zhiming Deng, and Ali M. Niknejad, “The Speed–Power Trade-Off in the Design of
CMOS True-Single-Phase-Clock Dividers,” IEEE Journal of Solid-State Circuits, vol. 45, no.
11, pp. 2457–2465, Nov. 2010.

[39] Payam Heydari, and Ravindran Mohanavelu, “Design of Ultrahigh-Speed Low-Voltage

CMOS CML Buffers and Latches,” IEEE Transactions on Very Large Scale Integration Systems,
vol. 12, no. 10, pp. 1081–1093, Oct. 2004.

[40] Payam Heydari, and Ravindran Mohanavelu, “A 40-GHz Flip-Flop Based Frequency
Dividers,” IEEE Transactions on Circuits and Systems–II: Express Briefs, vol. 53, no. 12, pp.
1358–1362, Dec. 2006.

[41] Christian Kromer, George von Büren, Gion Sialm, Thomas Morf, Frank Ellinger, and
Heinz Jäckel, “A 40-GHz Static Frequency Divider With Quadrature Outputs in 80-nm CMOS,”
IEEE Microwave and Wireless Components Letters, vol. 16, no. 10, pp. 564–566, Oct. 2006.

[42] Jaeha Kim, et al.,“Design Optimization of On-Chip Inductive Peaking Structures for
0.13-μm CMOS 40-Gb/s Transmitter Circuits,” IEEE Transactions on Circuits and Systems – I:
Regular Papers, vol. 56, no. 12, pp. 2544–2555, Dec. 2009.

[43] Sudip Shekhar, Jeffrey S. Walling, and David Allstot, “Bandwidth Extention Techniques
for CMOS Amplifiers,” IEEE Journal of Solid-State Circuits, vol. 41, no. 11, pp. 2424–2439,
Nov. 2006.

127
[44] Jeffrey S. Walling, Sudip Shekhar, and David Allstot, “Wideband CMOS Amplifier
Design: Time-Domain Considerations,” IEEE Transactions on Circuits and Systems – I: Regular
Papers, vol. 55, no. 7, pp. 1781–1793, Aug. 2008.

[45] Hai Tao et al., “40–43-Gb/s OC-768 16:1 MUX/CMU Chipset With SFI-5 Compliance,”
IEEE Journal of Solid-State Circuits, vol. 38, no. 12, pp. 2169–2180, Dec. 2003.

[46] Dimitris Bertsimas, John N. Tsitsiklis, Introduction to Linear Optimization. Athena

Scientific, 1997.
[47] Stephen Boyd, and Lieven Vandenburghe, Convex Optimization. Cambridge University
Press, 2004.

[48] Ming-Shuan Chen, C.-K. Ken Yang, “A Low-Power Highly Multiplexed Parallel PRBS
Generator”, accepted for publication at IEEE Custom Integrated Circuits Conference (CICC),
Sept. 2012.

[49] Ahmad Mirzaei, Mohammad E. Heidari, Rahim Bagheri, Saeed Cherazi, and Asad Abidi,
“The Quadrature LC Oscillator: A Complete Portrait Based on Injection Locking,” IEEE Journal
of Solid-State Circuits, vol. 42, no. 9, pp. 1916–1932, Sep. 2007.

[50] Emad Hegazi, and Asad A. Abidi, “Varactor Characteristics, Oscillator Tuning Curves,
and AM-FM Conversion,” IEEE Journal of Solid State Circuits, vol. 38, no. 6, pp. 1033-1039,
Jun. 2003.

[51] Behzad Razavi, Design of Analog CMOS Integrated Circuits, McGraw Hill, 2001.

[52] Dong-Keon Lee, Jeong-Kwang Lee, and Hang-Geun Jeong, “A dual Compensated
Charge Pump With Reduced Current Mismatch,” Proceedings of the 4th WSEAS international
conference on Circuits, systems, signal and telecommunications, pp. 109–112, 2010.

[53] Woogeun Rhee, “Design of High Performance CMOS Charge Pumps in Phase-Locked
Loops,” IEEE International Symposium on Circuits and Systems, pp. 545–548, Jul. 1999.
128
[54] Patrick Chiang, William J. Dally, Ming-Ju Edward Lee, Ramesh Senthinathan, Yangjin
Oh, and Mark A. Horowitz, “A 20-Gb/s 0.13-μm CMOS Serial Link Transmitter Using an LC-
PLL to Directly Drive the Output Multiplexer,” IEEE Journal of Solid-State Circuits, vol. 40, no.
4, pp. 1004–1011, Apr. 2005.

[55] C.-K. Ken Yang and Mark A. Horowitz, “A 0.8-pm CMOS 2.5 Gb/s Oversampling
Receiver and Transmitter for Serial Links,” IEEE Journal of Solid-State Circuits, vol. 31, no. 12,
pp. 2015–2023, Dec. 1996.

[56] Sudip Shekhar, et al., “Strong Injection Locking in Low-Q LC Oscillators: Modeling and
Application in a Forwarded-Clock I/O Receiver,” IEEE Transactions on Circuits and Systems-I:
Regular Papers, vol. 56, no. 8, pp. 1818–1829, Aug. 2009.

[57] Amr A. Hafez, C.-K. Ken Yang, “Design and Optimization of Multipath Ring
Oscillators,” IEEE Transactions on Circuits and Systems—I: Regular Papers, vol. 58, no. 10,
Oct. 2011, pp. 2332-2345.

[58] Amr A. Hafez, C.-K. Ken Yang, “Analysis and Design of Superharmonic Injection-
Locked Multipath Ring Oscillators,” Submitted to IEEE Transactions on Circuits and Systems—
I: Regular Papers.

[59] Amr A. Hafez, Ming-Shuan Chen, and C.-K. Ken Yang, “A Multi-Phase Multi-frequency
Clock Generator Using Superharmonic Injection Locked Multipath Ring Oscillators as
Frequency Dividers,” Accepted in IEEE Asian Solid State Circuits Conference (ASSCC), Nov.
2012.

[60] Ming-Shuan Chen, Amr A. Hafez, and C.-K. Ken Yang, “A 0.1-1.5 GHz 8-Bit Inverter-
Based Digital-to-Phase Converter Using Harmonic Rejection,” Accepted in IEEE Asian Solid
State Circuits Conference (ASSCC), Nov. 2012.

[61] Jaeha Kim, Brian S. Leibowitz, Metha Jeeradit, “Impulse Sensitivity Function Analysis
of Periodic Circuits,” IEEE International Conference on Computer-Aided Design (ICCAD), pp.
386–391, Nov 2008.

129