Analog VLSI: Circuits and Principles

Analog VLSI: Circuits and Principles

Shih-Chii Liu, Jorg Kramer, Giacomo Indiveri, Tobias Delbriick, and

Rodney Douglas
with contributions from Albert Bergemont, Chris Diorio, Carver A.
Mead, Bradley A. Minch, Rahul Sarpeshkar, and Eric Vittoz.

A Bradford Book
T he MIT Press
Cambridge, Massachusetts
London, England
© 2002 Massachusetts Institute of Technology

All rights reserved. No part of this book may be reproduced in any form by any electronic
or mechanical means (including photocopying, recording, or information storage and retrieval)
without permission in writing from the publisher.

This book was set in Times Roman by the authors using the N}3X document preparation system.
Printed on recycled paper and bound in the United States of America.

Library of Congress Cataloging-in-Publication Data

Analog VLSI : circuits and principles / Shih-Chii Liu ... ret al.] with contributions from Albert
Bergemont ... ret al.].
p. cm.
Includes bibliographical references and index.
ISBN 0-262-12255-3 (hc. : alk. paper)
I. Integrated circuits, Very large scale integration. 2. Linear integrated circuits. I. Liu,
Shih-Chii.
TK7874.75 .A397 2002
621.39'5--dc2 I
2002021915
This book is dedicated to the memory of our creative colleague and friend,
Misha Mahowald, who was a pioneer and an inspiration in this field.
Contents

Authors and Contributors xiii

Acknowledgments xv
Preface xvii
Foreword xix

1 Introduction 1

I SILICON AND TRANSISTORS

2 Semiconductor Device Physics - Jorg Kramer 7

2.1 Crystal Structure 7

2.2 Energy Band Diagrams 9

2.3 Carrier Concentrations at Thermal Equilibrium 13

2.4 Impurity Doping 15

2.5 Current Densities 19

2.6 p-n Junction Diode 24

2.7 The Metal-Insulator-Semiconductor Structure 35

3 MOSFET Characteristics - Shih-Chii Liu and Bradley

A.Minch 47

3.1 MOSFET Structure 48

3.2 Current-Voltage Characteristics of an nFET 52

3.3 Current-Voltage Characteristics of a pFET 70

3.4 Small-Signal Model at Low Frequencies 71

3.5 Second-Order Effects 75

3.6 Noise and Transistor Matching 80

3.7 Appendices 81

4 Floating-Gate MOSFETs - Chris Diorio 93

4.1 Floating-Gate MOSFETs 93

4.2 Synapse Transistors 98

4.3 Silicon Leaming Arrays 107

4.4 Appendices 116

viii Contents

II STATICS

5 Basic Static Circuits - Jorg Kramer 123

5.1 Single-Transistor Circuits 124

5.2 Two-Transistor Circuits 127

5.3 Differential Pair and Transconductance Amplifier 133

5.4 Unity-Gain Follower 142

6 Current-Mode Circuits - Giacomo Indiveri and Tobias

Delbruck 145

6.1 The Current Conveyor 145

6.2 The Current Normalizer 148

6.3 Winner-Take-All Circuits 150

6.4 Resistive Networks 164

6.5 Current Correlator and Bump Circuit 168

7 Analysis and Synthesis of Static Translinear Circuits

- Bradley A.Minch 177

7.1 The Ideal Translinear Element 179

7.2 Translinear Signal Representations 181

7.3 The Translinear Principle 183

7.4 ABC's of Translinear-Loop--Circuit Synthesis 195

7.5 The Multiple-Input Translinear Element 202

7.6 Multiple-Input Translinear Element Networks 205

7.7 Analysis of MITE Networks 210

7.8 ABC's of MITE-Network Synthesis 216

III DYNAMICS

8 Linear Systems Theory - Giacomo Indiveri 231

8.1 Linear Shift-Invariant Systems 231

8.2 Convolution 234

Contents ix

8.3 Impulses 236

804 Impulse Response of a System 237

8.5 Resistor-Capacitor Circuits 240

8.6 Higher Order Equations 241

8.7 The Heaviside-Laplace Transform 243

8.8 Linear System's Transfer Function 244

8.9 The Resistor-Capacitor Circuit (A Second Look) 246

8.10 Low-Pass, High-Pass, and Band-Pass Filters 249

9 Integrator-Differentiator Circuits - Giacomo Indiveri

and Jorg Kramer 251

9.1 The Follower-Integrator 252

9.2 The Current-Mirror Integrator 256

9.3 The Capacitor 261

904 The Follower-Differentiator Circuit 263

9.5 The diffl and diff2 Circuits 264

9.6 Hysteretic Differentiators 270

10 Photosensors - Jorg Kramer and Tobias Delbruck 275

10.1 Photodiode 275

10.2 Phototransistor 283

10.3 Photogate 284

lOA Logarithmic Photosensors 286

10.5 Imaging Arrays 299

10.6 Limitations Imposed by Dark Current on Photosensing 307

IV SPECIAL TOPICS

11 Noise in MOS Transistors and Resistors· Rahul

Sarpeshk ar, Tobias Delbruck, Carver Mead, and
Shih-Chii Liu 313
x Contents

11.1 Noise Definition 313

11.2 Noise in Subthreshold MOSFETs 317

11.3 Shot Noise versus Thermal Noise 325

11.4 The Equipartition Theorem and Noise Calculations 328

11.5 Noise Examples 333

12 Layout Masks and Design Techniques. Eric Vittoz,

Shih-Chii Liu , and Jorg Kramer 341

12.1 Mask Layout for CMOS Fabrication 341

12.2 Layout Techniques for Better Performance 346

12.3 Short List of Matching Techniques 351

12.4 Parasitic Effects 353

12.5 Latchup 355

12.6 Substrate Coupling 356

12.7 Device Matching Measurements 359

13 A Millennium Silicon Process Technology. Albert

Bergemont, Tobias Delbruck, and Shih-Chii Liu 361

13.1 A typical 0.25 JLm CMOS Process Flow 361

13.2 Scaling Limits for Conventional Planar CMOS

Architectures 373

13.3 Conclusions and Guidelines for New Generations 382

14 Scaling of MOS Technology to Submicrometer

Feature Sizes· Carver Mead 385

14.1 Scaling Approach 386

14.2 Threshold Scaling 394

14.3 Device Characteristics 395

14.4 System Properties 402

14.5 Conclusions 402

Contents xi

Appendix A:
Units and symbols 407

References 415
Index 429
Authors and Contributors

Albert Bergemont Shih-Chii Liu

Maxim Integrated Products Institute of Neuroinformatics,
3725 North First Street, ETH/UNIZ
San Jose, CA 95134-1350 Winterthurerstrasse 190
U.S.A. 8057 Zurich, Switzerland

Tobias Delbriick Carver A. Mead

Institute of Neuroinformatics, Department of Computation and
ETH/UNIZ Neural Systems
Winterthurerstrasse 190 California Institute of Technology
8057 Zurich, Switzerland Pasadena, CA 91125
U.S.A.
Chris Diorio
Department of Computer Science Bradley A. Minch
and Engineering Department of Electrical Engineering
The University of Washington Cornell University
114 Sieg Hall, Box 352350 405 Phillips Hall
Seattle, WA 98195 Ithaca, NY 14853-5401
U.S.A. U.S.A.

Rodney Douglas Rahul Sarpesbkar

Institute of Neuroinformatics, Research Laboratory of Electronics
ETH/UNIZ Massachusetts Institute of
Winterthurerstrasse 190 Technology
8057 Zurich, Switzerland Cambridge, MA 02139
U.S.A.
Giacomo Indiveri
Institute of Neuroinformatics, Eric Vittoz
ETH/UNIZ Chief Scientist
Winterthurerstrasse 190 Advanced Microelectronics
8057 Zurich, Switzerland Center for Electronics and
Microtechnology
J6rgKramer Jaquet-Droz 1
Institute of Neuroinformatics, 2007 Neuchatel
ETH/UNIZ Switzerland
Winterthurerstrasse 190
8057 Zurich, Switzerland
Acknowledgments

This book was written by a small group of authors who represent the work of a
far larger community. We would like to acknowledge our colleagues who have
contributed to the advance of concepts and circuits in neuromorphic engineer­
ing; in particular, John Lazzaro, Massimo Silvilotti, John Tanner, Kwabena
Boahen, Paul Hasler, Steve Deweerth, Ron Benson, Andre van Schaik, John
Harris, Andreas Andreou, Ralph Etienne-Cummings, and many others. We es­
pecially wish to thank the following people for their help in the completion
of this book: Andre Van Schaik, Regina Mudra, Elisabetta Chicca, and Ralph
Etienne-Cummings for their constructive comments in earlier versions of the
book; Samuel Zahnd for putting together the material for the example circuits
on the website; Adrian Whatley for ensuring the integrity of the bibliography,
David Lawrence for dealing with computer mishaps, Mietta Loi for entering
some of the material in the book, Kathrin Aguilar-Ruiz for dealing with legal
details, Claudia Stenger for her endless patience with all sorts of requests, and
Donna Fox for always providing the answers for difficult requests. We also
thank Sarah K. Douglas for the cover design of this book. The work in this
fledging field has been supported by progressive funding organizations: Na­
tional Science Foundation, Office of Naval Research, Gatsby Charitable Foun­
dation, Swiss National Science Foundation, Whitaker Foundation, Department
of Advanced Research Projects Agency, and our various home institutions. We
also acknowledge Mike Rutter for his enthusiasm in starting this project, and
Bob Prior for seeing the project to its completion.
Preface

The aim of this book is to present the collective expertise of the neuromor­
phic engineering community. It presents the central concepts required for cre­
ative and successful design of analog very-Iarge-scale-integrated (VLSI) cir­
cuits. The book could support teaching courses, and provides an efficient intro­
duction to new practitioners who have some previous training in engineering,
physics, or computer science.
Neuromorphic engineers are striving to improve the performance of artifi­
cial systems by developing chips and systems that process information collec­
tively using predominantly analog circuits. Consequently, our book biases the
discussion of analog principles and design towards novel circuits that emulate
natural signal processing. These circuits have been used in implementations
of neural computational systems or neuromorphic systems and biologically­
inspired processing systems. Unlike most circuits in commercial or industrial
applications, our circuits are operated mainly in the subthreshold or weak in­
version region. Moreover, their functionality is not limited to linear operations,
but encompasses also many interesting nonlinear operations similar to those
occuring in natural systems.
Although digital circuits are the basis for a large fraction of circuts in cur­
rent VLSI systems, certain computations like addition, subtraction, expansion,
and compression are natural for analog circuits and can be implemented with
a small number of transistors. These types of computations are prevalent in the
natural system which has an architecture which is not that of a conventional
Turing machine. The mechanisms for signaling in the neural system which
are govemed by Boltzmann statistics can be captured by circuits comprising
metal-oxide-semiconductor field effect transistors (MOSFETs) that operate in
the subthreshold or weak inversion regime. Because the exponential depen­
dence of charges on the terminal voltages of a MOSFET is similar to those of
the bipolar junction transistor (BIT), current techniques for constructing a cir­
cuit which implements a given function using bipolar circuits, can be extended
to MOSFET circuits 1• Besides the advantage of the reduced power consump­
tion of MOSFET circuits that operate in the weak inversion regime, this new
circuit philosophy also translates to novel circuits and system architectures.
Local memory is an essential part of any artificial parallel distributed pro­
cessor or neural network system. In this book, we show circuits for analog
memory storage and for implementing local and global learning rules us­
ing floating-gate charge modulation techniques in conventional CMOS tech-

1 BITs are traditionally used for analog circuits in industry.

xviii Preface

nology. We also show how by using floating-gate circuits together with the
translinear principle, we can develop compact circuits which implement a large
class of nonlinear functions.
The first integrated aVLSI system that implemented a biological function
was a silicon retina by Mead and Mahowald. This system used analog circuits
that performed both linear and nonlinear functions in weak-inversion opera­
tion. Following this initial success, subsequent examples of simple computa­
tional systems and novel circuits have been developed by different labs. These
examples include photoreceptor circuits, silicon cochleas, conductance-based
neurons, and integrate-and-fire neurons. These different circuits form the foun­
dation for a physical computational system that models natural information
processing.
The material presented in this book has evolved from the pioneering series
of lectures on aVLSI and principles introduced by Carver Mead into the
Physics of Computation curriculum at the California Institute of Technology
in the mid 80s. Today, similar courses are taught at many institutions around
the world; and particularly, at the innovative annual Telluride Neuromorphic
Workshop (funded by the US National Science Foundation, and others). Many
of the people who teach these courses are colleagues who were trained at
Caltech, or who have worked together at Telluride.
We have been fortunate to obtain the enthusiastic participation of the many
authors who provided material and the primary text for this book. Their names
are associated with each chapter. However, the result is not simply an edited
collection of papers. Each of the authors named on the cover made substantial
contributions to the entire book. Liu and Douglas have edited the text to
provide a single voice.
We have attempted to make the material in this book accessible to readers
from any academic background by providing intuition for the functionality of
the circuits. We hope that the book will prove useful for insights into novel
circuits and that it will stimulate and educate researchers in both engineering
and interdisciplinary fields, such as computational neuroscience, and neuro­
morphic engineering.
Foreword

Carver Mead

In the beginning of any new technology, the applications of the technology

cannot be separated from the development of the technology itself. With vac­
uum tubes, discrete transistors, or integrated circuits, the first circuit topologies
were invented by those closest to the device physics and fabrication process.
As time passes, the knowledge of how to make the devices gradually separates
from the knowledge of how to use the devices. Abstractions are developed that
greatly simplify our conceptual models of the underlying technology. Canoni­
cal circuit forms are encapsulated into symbols that have meaning in the field
of use, rather than in the space of implementation. Familiar examples are logic
gates, operational amplifiers, and the like. As more time passes, the abstrac­
tions become entrenched in university courses and industrial job descriptions.
It is common to hear phrases like "we need to hire a system architect, two
logic designers, a circuit designer, and a layout specialist" or "who will we get
to teach the op-amp course next year?" This inexorable trend toward special­
ization fits well with the myth that putting more engineers on a design task will
make it happen faster. In my personal experience, I have never seen this myth
reflected in reality. In many cases, I have seen chips designed by large groups
fail to converge at all. Chip designs that converge rapidly and perform well
have always been done by small, cohesive groups. For the truly great chip de­
signs, there has always been a single person that could keep the entire design in
their head. The lesson from these experiences is unmistakable: leading contrib­
utors are able to lead because they move past the limitations of the prevailing
paradigm.
The second myth is that digital techniques are displacing analog techniques
throughout modern electronics. In fact, an explosion of new analog applica­
tions is occurring as this sentence is being written. Cellular telephones, fiberop­
tic transmitters and receivers, wideband wired and wireless data modems, elec­
tronic image capture devices, digital audio and video systems, "smart" power
control, and many others all have explicitly analog circuits. But, and perhaps
even more important, all circuits are analog, even if they are used as digital
building blocks. Rise times, delays, settling times, and metastable behaviors
are analog properties of nominally digital circuits. As the speed of operation
approaches the limits of the technology, analog considerations increasingly
dominate the process of digital design.
xx Foreword

Modern integrated circuit technology is extremely sophisticated; it is ca­

pable of realizing a vast array of useful device and circuit structures. Designers
who avail themselves of the richness of the technology have a powerful advan­
tage over those who are limited to logic gates and operational amplifiers. This
book contains precisely the information required by such renaissance design­
ers. Device physics, process technology, linear and nonlinear circuit forms,
photodetectors, floating-gate devices, and noise analysis are all described in
clear, no-nonsense terms. It stands in refreshing contrast to the litany of oper­
ational amplifier and switched capacitor techniques that are widely mistaken
for the whole of analog design.
1 Introduction

This book presents an integrated circuit design methodology that derives its
computational primitives directly from the physics of the used materials and
the topography of the circuitry. The complexity of the performed computations
does not reveal itself in a simple schematic diagram of the circuitry on the
transistor level, as in standard digital integrated circuits, but rather in the
implicit characteristics of each transistor and other device that is represented
by a single symbol in a circuit diagram. The main advantage of this circuit­
design approach is the possibility of very efficiently implementing certain
'natural' computations that may be cumbersome to implement on a symbolic
level with standard logic circuits. These computations can be implemented
with compact circuits with low power consumption permitting highly-parallel
architectures for collective data processing in real time. The same type of
approach to computation can be observed in biological neural structures, where
the way that processing, communication, and memory have evolved has largely
been determined by the material substrate and structural constraints. The data
processing strategies found in biology are similar to the ones that tum out to
be efficient within our circuit-design paradigm and biology is thus a source of
inspiration for the design of such circuits.
The material substrates that will be considered for the circuits in this
book are provided by standard integrated semiconductor circuit technology and
more specifically, by Complementary Metal Oxide Silicon (CMOS) technol­
ogy. The reason for this choice lies in the fact that integrated silicon technology
is by far the most widely used data processing technology and is consequently
commonly available, inexpensive, and well-understood CMOS technology has
the additional advantages of only moderate complexity, cost-effectiveness, and
low power consumption. Furthermore it provides basic structures suitable for
implementation of short-term and long-term memory, which is particularly im­
portant for adaptive and learning structures as found ubiquitously in biological
systems. Although we will specifically consider CMOS technology as a phys­
ical framework it turns out that various fundamental relationships are quite
similar in other frameworks, such as in bipolar silicon technology, in other
semiconductor technologies and to a certain extent also in biological neural
structures. The latter similarities form the basis of neuromorphic emulation of
biological circuits on an electrical level that led to such structures as silicon
neurons and silicon retinas.
2 Chapter 1

The book is divided into four sections: Silicon and Transistors; Statics;
Dynamics; and Special Topics. The first section (Silicon and Transistors) pro­
vides a short introduction into the underlying physics of the devices that are
discussed in the rest of the book; in particular the operation of the MOSFET
in the subthreshold region and a discussion of analog charge storage using
floating-gate technology. Chapter 2 discusses useful equations that can be de­
rived from modeling the physics of the basic devices in the silicon substrate.
These device models provide a foundation for the derivation of the equations
governing the operation of the MOSFET as described in Chapter 3. From re­
sults discussed in Chapters 2 and 3 we show in Chapter 4 how MOS technology
can be used to build analog charge storage elements. Readers who are more in­
terested in circuits at the transistor level description may omit Chapters 2 and
4 and continue to the Statics section.
The Statics section comprises three chapters. These chapters describe ex­
amples of linear and nonlinear static functions that can be implemented by
simple circuits. Chapter 5 presents some basic circuits which show the richness
of the processing that can be performed by the transistor. Chapter 6 introduces
an analog circuit design concept where currents represent the signal and state
variables in a circuit. As examples, some current-mode circuits are described
that implement nonlinear functions that are prevalent in natural systems, for
example, a winner-take-all circuit. Chapter 7 derives a methodology for im­
plementing a large class of linear and nonlinear functions using a particular
building block called a multiple-input translinear element.
The Dynamics section describes circuits which process time-varying sig­
nals. Chapter 8 reviews the basics of linear systems theory, which is a useful
tool for the small-signal analysis of circuits both in the time and space domain.
We apply this theory in Chapter 9 to selected examples of simple circuits for
first-order and second-order filters. Chapter 10 provides a brief introduction
into semiconductor photosensors and focuses on circuits that model prominent
properties of biological photoreceptors. It also gives an overview of common
image sensing principles.
The last section (Special Topics) contains chapters which expound fur­
ther on the basics of semiconductor technology. These chapters cover topics
on noise in transistors, the flow from design to layout to fabrication of an in­
tegrated circuit, and the issue of scaling semiconductor technology into the
future. Chapter 1 1 describes the different noise sources in a transistor and how
these sources can be measured. It also presents a novel way of demonstrat­
ing the equivalence of thermal noise and shot noise. The circuit layout masks
Introduction 3

needed to specify a layout to a fabrication house are listed in Chapter 12 along

with useful layout tips for good circuit perfol1llance. Chapter 13 describes the
processing steps executed in a 0.25 f..tm process and Chapter 14 projects how
transistors will scale in future technologies.
This book is directed towards students from a variety of backgrounds. The
students who are more interested in circuits can omit Section I and should
still be able to follow the chapters in Sections II and III. Most of the material
in this book has been taught in classes at the Institute for Neuroinfol1llatics,
University of Zurich! ETH Zurich, Switzerland, and also at the Telluride
Neuromorphic Engineering Workshop.
Examples of simulation and layout files for simple circuits are available
from the Institute of Neuroinfol1ll atics website. These circuits were fabricated
and used in our laboratory courses. They have been developed using the
integrated circuit design tools available from Tanner Research, Inc. Students
who are interested in simulating these circuits can also use the free public
domain software AnaLOG by John Lazzaro and David Gillespie.
I SILICON AND TRANSISTORS
2 Semiconductor Device Physics

The purpose of this chapter is to provide an introduction to the basics of the

semiconductor physics needed for the understanding of the devices described
in this book. Most of this introduction pertains to semiconductors in general.
Where general statements are not possible we focus on silicon. The values of
material constants, and the typical values of other parameters, are for silicon.
It is not intended to provide a detailed step-by-step derivation of the formulas
describing device behavior. Often we limit ourselves to stating the necessary
conditions for the derivation to hold and the important results without formal
derivation. More extended summaries of solid-state and semiconductor physics
can be found in standard texts (Grove, 1967; Sze, 1981; Kittel, 1996; Singh,
2001). Detailed analyses of the subject fill entire books (Dunlap, 1957; Smith,
1979; Moss, 1980).

2.1 Crystal Structure

In semiconductors and other materials the atoms are arranged in regular struc­
tures, known as crystals. These structures are defined and held together by
the way the valence (outermost) electrons of the atoms are distributed, given
that electrons tend to form pairs with antiparallel spin. Figure 2.1 shows the
crystal structures of some important semiconductors; silicon (Si) and gallium
arsenide (GaAs). A silicon atom, for example, has four unpaired valence elec­
trons that can form covalent bonds with a tetrahedral spatial characteristic
(Fig. 2.1(b)). Pure silicon naturally crystallizes in a diamond structure. The
diamond lattice is based on the face-centered cubic (fcc) arrangement shown
in Fig. 2.1(a), which means that the atoms are located at the corners and face
centers of cubes with a given side length a, called the lattice constant. The di­
amond structure consists of two interleaved fcc lattices that are displaced by
aj4 in each dimension. Silicon has a lattice constant of a = 5.43 A. Gallium
arsenide has the same structure as silicon, except that one of the interleaved
fcc lattices holds the gallium atoms and the other the arsenic atoms. This ar­
rangement is known as zincblende structure.
8 Chapter 2

[100]

S i mple cubic Face-centered cubic

(a)

Diamond Zi ncblende
(C, Ge, Si, etc) (GaAs, GaP, etc)

(b)

Figure 2.1
Various crystal lattice structures, with lattice constant a. (a) Simple cubic lattice with atoms at the
cube comers, and face-centered cubic (fcc) lattice with additional atoms on the cube faces. The
most important crystal directions [1 00], [I 1 0], and [III] of the simple cubic lattice are indicated.
(b) Structures with two interleaved fcc lattices: Diamond structure, consisting of one kind of atom,
and zincblende structure, consisting of two kinds of atoms. Figure adapted from S. M. Sze (1981),
Physics of Semiconductor Devices, 2nd Edition. © 1981 by John Wiley & Sons, Inc. Reprinted by
permission of John Wiley & Sons, Inc.
Semiconductor Device Physics 9

E E E

e e e e

Ins u l ator Semiconductor Conductor

Figure 2.2
Schematic representation of electron energy bands in a crystal for an insulator, a semiconductor,
and a conductor (or metal). The energy bands are represented as boxes. The hatched areas
symbolize the states in the energy bands that are occupied by electrons at zero temperature. At non­
zero temperatures some electrons (denoted by circled minus signs) occupy higher-energy states,
leaving "holes" (denoted by circled plus signs) in the unoccupied lower-energy states.

2.2 Energy Band Diagrams

Crystals and other solids are classified according to their electrical conduc­
tivity into insulators, semiconductors, and conductors or metals in the order
of increasing conductivity. Electrical processing structures, such as transistors,
are mostly fabricated from semiconductors because they operate at an inter­
mediate conductivity level, which can be modulated by varying the electrical
boundary conditions and by introducing atoms of foreign elements into the
crystal structure. This latter process is called impurity doping. Conductors and
insulators also play an important role in electrical circuits, because they con­
nect and separate respectively, the nodes of the processing structures. For ex­
ample, in integrated silicon technologies silicon dioxide (Si02) (often referred
to as oxide) is commonly used as an insulator, while polycrystalline silicon,
also known as polysilicon, and aluminum are used as conductors.
The physical basis for the above classification of the materials lies in the
properties of the atoms and their arrangement. Electrical currents in solids are
carried by the motion of valence electrons, which are attracted to the fixed
positively charged ion cores. A valence electron can thus either be bound
to a particular ion core by electromagnetic forces or it can be mobile and
10 Chapter 2

contribute to the current flow. In order to be mobile, an electron must acquire

a certain minimum energy to break free of its ion core. This energy is called
ionization energy. While free electrons can be in any energy state, the energies
of electrons in solids lie in certain ranges of values, which are separated by
forbidden zones due to the interaction of the valence electrons with the ion
cores. The allowed ranges of energy values are called energy bands and the
forbidden zones are called energy gaps or bandgaps. Each energy band in a
crystal consists of closely-spaced discrete levels. This discretization of energy
levels is a quantum-mechanical effect that is due to the spatial confinement
of bound electrons and to the spatial periodicity of the potential energy for
mobile electrons. An energy level supports a limited number of states, each of
which is either occupied by an electron or empty at any given time. Simplified
energy band diagrams of an insulator, a semiconductor, and a conductor are
schematically shown in Fig. 2.2. Under normal conditions, the energy bands of
insulators are either completely filled or completely empty, that is all electrons
are bound. A metal has a partly filled bands within which electrons can move.
Some metals have two or more overlapping energy bands that are partly filled.
They are often referred to as semimetals. In a semiconductor, one or more
bands are either almost filled or almost empty. Current flow is then influenced
by certain physical parameters and boundary conditions.
The energy bands in semiconductors are called valence bands and conduc­
tion bands according to whether they are almost filled or almost empty. In a
conduction band, electrons are essentially free to move, while in valence bands
the electrons are bound to the atom bodies and can only move from one of the
few unoccupied states to another one, if they cannot acquire enough energy
to bridge the bandgap to a conduction band. An elegant mathematical con­
cept, which is commonly used throughout the semiconductor literature, allows
a more symmetric view of this mechanism. This concept considers unoccupied
energy states in a valence band as holes. The term derives from the notion that
an electron is absent from a state that is usually occupied. An electron valence
band can thus be considered as a hole conduction band, since the holes are
quite free to move around, that is electrons in the valence band in the vicinity
of an unoccupied state may hop to that state and leave their initial state free for
another electron to hop in, and so on. In common terminology the definition
of valence and conduction bands is related to electrons and the holes are thus
said to move in the valence band. In mathematical expressions the holes can
be treated as positively charged particles that within a semiconductor acquire
as much physical reality as electrons. Consequently, they are also attributed
Semiconductor Device Physics 11

other physical parameters used to characterize particles, such as mass and mo­
bility. However, it is important to note that holes are not just positively charged
electrons, but have different characteristic parameter values. Furthermore, you
should keep in mind that the symmetry between electrons and holes breaks
down as soon as the charge carriers leave the semiconductor, as we shall see,
for example, in the chapters dealing with floating-gate structures.
Energy bands in a crystal have certain properties that are closely related
to the crystal structure and thus vary significantly between different types of
crystals. Graphic representations of the energy-band diagrams of Si and GaAs
are shown in Fig. 2.3. The allowed electron energies are plotted as a function of
the electron momentum for two sets of directions (cf. Fig. 2.1 (a)), namely the
[100] directions along the edges of the crystal lattice and the [111] directions
along the lattice diagonals. By convention, energy-band diagrams are drawn
such that electron energy increases in the upward direction. Energy values
are usually specified in units of electron volts (eV). This unit is convenient
for the conversion of an energy-band diagram into an electrostatic potential
distribution, which is obtained by dividing the energy values by the electron
charge, that is the negative value of the elementary charge q = 1.60218 X
10-19 C. An electron volt is the energy corresponding to a potential change of
an electron of one volt.
For simplicity, only the valence band and the conduction band with the
smallest energy separation are shown in Fig. 2.3. The lines represent the band
edges, that is the highest-energy states of the valence band and the lowest­
energy states of the conduction band. The band edges tell us the minimum
amount of energy an electron has to acquire or lose to bridge the bandgap for
a given change in momentum. The difference between the lowest conduction
band energy and the highest valence band energy is called bandgap energy E g.
The bandgap energy of silicon at room temperature is 1.12 eV. The valence
band edge appears at zero momentum and is degenerate, that is common to
several valence bands, for the most widely-used semiconductors. The momen­
tum associated with the conduction band edge may be zero or not, depending
on the semiconductor, and the conduction band edge is not degenerate. If the
minimum of the conduction band edge is at the same momentum as the maxi­
mum of the valence band edge we speak of a direct bandgap, otherwise of an
indirect bandgap. As we can see from Fig. 2.3, gallium arsenide has a direct
bandgap and silicon has an indirect bandgap.
Electron energy and momentum changes can be induced by different phys­
ical processes, the most important of which are interactions with lattice vibra-
12 Chapter 2

Si GaAs

(111) POOl x [111] r (100] x

Wave vector Wave vector

(a) (b)

Figure 2.3
Energy-band diagrams of (a) silicon (Si) and (b) gallium arsenide (GaAs). Only the edges of the
uppermost valence band and of the lowermost conduction band are shown as a function of the
wave vector for two sets of directions in the crystal. The r point corresponds to charge carriers
being at rest. The [III] set of directions is along the diagonals of the crystals, while the [1 00]
set is oriented along the edges of the crystals, as shown in Fig. 2.I(a). The L point stands for
wave vectors (1l,/a)(±1, ±1, ±1); and the X point stands for wave vectors (21l' ja)(±l, 0, 0),
(21l' ja)(O, ±1, 0) and (21l' ja)(O, 0, ±1). The momentum p of a charge carrier is computed from
its wave vector k, as p hk where h is the reduced Planck constant. The bandgap energy Eg
=

is the separation between the top of the topmost valence band and the bottom of the bottommost
conduction band. These extrema appear at different momenta for Si and at the same momentum for
GaAs. The bandgap of Si is thus called indirect and the bandgap of GaAs is called direct. Figure
adapted from 1. R. Chelikowsky and M. L. Cohen (1976), Nonlocal pseudopotential calculations
for the electronic structure of eleven diamond and zinc-blende semiconductors, Phys. Rev., 814,
556-582. @1976 by the American Physical Society.

tions, that is collisions with the ions in the crystal, and with electromagnetic
waves. The energies that are transferred during these interactions are quan­
tized. The energy quantum of a crystal lattice vibration is called a phonon and
the energy quantum of an electromagnetic wave is called a photon. Absorption
Semiconductor Device Physics 13

or emission of a phonon changes mainly the momentum of an electron while

the energy change is typically small (",0.01 eV to 0.03 e V) compared to the
bandgap energy. On the other hand, the momentum transfer by a photon with
an energy of the order of the bandgap energy is negligible. This means that in
the case of an indirect bandgap the transitions between valence and conduction
bands with the smallest energies typically involve phonons and photons, while
direct bandgaps can, for example, be bridged by photons alone. This is the
reason why materials with direct bandgaps can be used as efficient sources of
electromagnetic radiation, while those with indirect bandgaps are inefficient,
because the minimum energy transitions, which are the most probable ones,
depend on the availability of a phonon with the proper momentum.
Most semiconductor devices do not make use of the electromagnetic prop­
erties of the material and the circuits are therefore shielded from high-energy
electromagnetic radiation. Changes in electron momentum are then much more
easily induced than changes in energy that are large enough to make the elec­
tron bridge the bandgap. The majority of electrical properties of semiconduc­
tors can thus be sufficiently accurately described without considering the mo­
mentum space at all. In the energy band diagrams the electron energies of the
maximum of the valence band edge and the minimum of conduction band edge
is therefore usually plotted as a function of position in one- or two-dimensional
space, while the structure of the bands in momentum space does not appear
anymore. Such a simplified energy band diagram is shown in Fig. 2.4.

2.3 Carrier Concentrations at Thermal Equilibrium

Thermal energy expresses itself in vibrations of the crystal lattice. Energy

transfer from the lattice to the electrons is thus established through absorption
or emission of phonons. At zero temperature, all low-energy states are filled
and all high-energy states are empty. At higher temperatures some electrons
will leave their lower-energy states in favor of higher-energy states. The occu­
pancy of energy states is statistically described by a probability distribution.
This distribution is known as Fermi-Dirac distribution. The probability that an
energy state with value E is occupied is given by

F(E) = ( 1 + e(E-EF)/kT ) -l (2.3.1)

where E F denotes the energy at which the occupation probability is 0.5, called
Fermi level or chemical potential, k= 1.38 066 X 10- 2 3 JIK is the Boltzmann
14 Chapter 2

Electron
energ y

Conduction band

.. ----
EC 1-.."-
Bandgap

Ev

I-... .,.+�----
Valence band
Hole
energy

Position

Figure 2.4
Simplified semiconductor energy-band diagram. The energy is plotted as a function of position in
one dimension. Mobile charge carriers are symbolized by the signed circles.

constant, and T is the absolute temperature. For intrinsic (undoped) semicon­

ductor crystals E F is very close to the center of the bandgap. For typical im­
purity doping concentrations and bandgaps of commonly used semiconductors
EF is well-separated from the valence and conduction band edges, such that
IE - EFI > > kT for all allowed energy states. This simplifies the Fermi-Dirac
distribution in the conduction band to the Boltzmann distribution

1 F(E) = e-(E-EF)/kT .1 (2.3.2)

This probability distribution is the reason for the exponential characteristics of

diodes and transistors that will be described in this and the next chapter; and
these devices will determine the characteristics of most circuits in this book.
The electron density dn(E, dE) within an energy interval dE around an
energy E is given by

dn(E, dE) = N(E) F(E)dE (2.3.3)

where N(E) ex: viE - Ec near the bottom of the conduction band. In thermal
equilibrium, that is if no external voltage is applied to the semiconductor
and no net current flows, the total electron density in the conduction band is
Semiconductor Device Physics 15

obtained by integrating Eq. 2.3.3 with respect to energy from the conduction
band edge to infinity, resulting in

n = Nc e-(Ec-EF)/kT (2.3.4)
where Nc denotes the effective density of states in the conduction band near
its edge, and Ec is the energy of the conduction band edge. A corresponding
equation can be derived for the hole density near the top of the valence band:
p = Nv e-(EF-Ev)/kT . (2.3.5)
For intrinsic semiconductors n and p are equal. We define an intrinsic carrier

-
density ni as
_
n2i np (2.3.6)
It follows from Eqs. 2.3.4, 2.3.5, and 2.3.6 that

ni = .;np = y'NcNv e-Eg/2kT (2.3.7)

and that the Fenni level Ei for an intrinsic semiconductor is given by

Et. -
- Ec + Ev +
2
kT
2
I
og
( )
Nv
Nc
. (2.3.8)

For silicon, Nc = 2.80 X 1019 cm-3 , Nv = 1.04 X 1019 cm-3 , and ni =

1.45 x 1010 cm-3 at room temperature. The concentration of Si atoms in a
crystal is 5 x 1022 cm-3 , which means that only One out of 3 x 10 12 atoms is
ionized at room temperature and therefore conductivity is very low.

2.4 Impurity Doping

The conductivity of a semiconductor can be increased significantly by doping

it with impurities. In the doping process a small fraction of the semiconductor
atoms in the crystal structure are replaced by atoms of a different element.
As illustrated in Fig. 2.5, a donor impurity is an atom with a valence electron
more than the semiconductor atom and an acceptor impurity is an atom with
a valence electron less than the semiconductor atom. Note that impurity atoms
are electrically neutral: The difference in number of electrons is balanced by
an equal difference in number of protons in the nucleus. Since the additional
valence electron of a donor does not fit into the crystal bond structure it is
only loosely bound to its nucleus by electromagnetic forces. In the energy­
band diagram donors fonn an energy level in the bandgap that is typically
16 Chapter 2

, ''- A A.. A./ ' ''- AA.. A./

, / , / , / , /
, / , , , / ,
-
/
acceptor
( \ ( - \ donor ( \
- ( \

/' A.. ): ,,-imp urity /'

- '\ - - - /' "- impurity
/ , - / , e- , / ,
/ / , / , - / , -
( - \ e ( - \ ( ( - \ ( \ ( \
"
' ''- AA.. ,/'./ ' ''- /'A.. ,/'./
, / , / , / , /
, / , / , / , /
\ ( \ ( \ ( \
AA.. A"- AA.. A"-
/ , / , / , / ,
/ , / , / , / ,
\ ( \ ( \ ( \ (
, ./ , ./ , ./ , ./ , ./ , ./

(a) (b)

Figure 2.5
Illustration of a semiconductor with (a) donor and (b) acceptor impurity doping. The ion cores
(dashed circles) are bound into the crystal structure by covalent bonds (dashed lines). The excess
charge carrier (solid circle) that is introduced with the impurity does not fit into the covalent bond
structure. This excess charge carrier is only loosely bound to the ion core of the impurity atom by
electromagnetic forces and so is mobile. The hole introduced by an acceptor is a missing electron
in a covalent bond.

close to the conduction-band edge. As long as the surplus electron is bound to

its nucleus the corresponding state on that level is occupied and defined to be
neutral. Conversely, acceptor atoms have loosely bound holes that appear as
neutral states on an energy level in the bandgap that is typically close to the
valence-band edge.
Figure 2.6 shows the energy-band diagrams, the state densities, the Fermi­
Dirac distributions, and the carrier concentrations for differently-doped semi­
conductors. The introduction of donors moves the Fermi level from near the
center of the bandgap further towards the conduction-band edge, whereas the
introduction of acceptors moves it closer to the valence-band edge. If the
donor-doping density is larger than the acceptor-doping density there
are more electrons in the conduction band than holes in the the valence
band (n > p) under charge-neutrality conditions and the semiconductor is
said to be n-type. In the reverse case, p > n and the semiconductor is said
to be p-type. The doping strength is often indicated by plus or minus signs.
For example, a weak p-type doping is denoted by p - and a very strong n­
type doping by n++. If the semiconductor is so strongly doped that the Fermi
Semiconductor Device Physics 17

n = NceXp[-(Ec- Ep)lkTj
(= nil
---E ----
---L-- _ � p�( NveXP [-(EF- Ev)lkTj
n il
i1
=

(a)

I n
---1----- Ec
1
No EF ----EF
1
1
---1---
1
1
1

(b)

1
----,-----
1
1-00=-,,-1 -
---- EF - np= n;
---4--- -E
1
1 P
1
N(E) 0 0.5 1.0 F(E) n,p
(c)

Figure 2.6
Energy-band diagram, density of states, Fermi-Dirac distribution, and carrier concentrations for
(a) intrinsic, (b) n-type, and (c) p-type semiconductors at thermal equilibrium. The concentrations
of mobile electrons and holes are indicated by the hatched areas in the plots on the right. Figure
adapted from S. M. Sze (1981), Physics of Semiconductor Devices, 2nd Edition. © 1981 by John
Wiley & Sons. Inc. Reprinted by permission of John Wiley & Sons. Inc.
18 Chapter 2

level is within the conduction or valence band or very near the edge of one
of these bands, such that a large fraction of the states at the band edge are
occupied, its properties become similar to those of a metal and we speak of a
degenerate semiconductor. This happens for acceptor doping concentrations of
around Nc and donor doping concentrations of around N v. Commonly-used
doping elements for silicon are phosphorus (P) and arsenic (As) as donors and
boron (B) as an acceptor. The ionization energy of such an impurity atom,
that is the energy required to remove the loosely-bound charge carrier from
its ionic core, is on the order of 0.05 eV. This is only a small fraction of the
bandgap energy and most donors and acceptors are thermally ionized at room
temperature. The condition of charge neutrality in the crystal can then be stated
as
(2.4.1)

where NA denotes the acceptor impurity concentration and N D the donor im­
purity concentration. Furthermore, Eq. 2.3.7 is also valid for doped semicon­
ductors.
The mobile charge carriers that are more abundant in a semiconductor
in thermal equilibrium are called majority carriers, whereas the sparser ones
are called minority carriers. Using Eqs. 2.3.7 and 2.4.1 the concentration of
majority electrons in the conduction band of an n-type semiconductor can be
approximated by

(2.4.2)

and the concentration of majority holes in the valence band of a p-type semi­
conductor by

(2.4.3)

For strongly doped n-type material with N D >> NA and ND - NA > > ni

(2.4.4)

and for strongly doped p-type material with N A >> ND and NA ND - > > ni

(2.4.5)
Semiconductor Device Physics 19

The minority carrier concentrations can be computed from Eq. 2.3.6 as

n�t "'" n�t
P nO--""'
_
(2.4.6)
nn o ND
and
n�t "'"
n�t
npo -
- "'" (2.4.7)
Ppo NA •

Using Eqs. 2.3.4 and 2.4.4 we can approximate the Fermi level of a higbly­
doped n-type semiconductor by

EF = Ec - kTlog (Z�) . (2.4.8)

Correspondingly, Eqs. 2.3.5 and 2.4.5 yield an approximation of the Fermi

level of a strongly-doped p-type semiconductor:

EF = Ev + kTlog (Z:) . (2.4.9)

Hence, the Fermi level is near the conduction band edge for ND � Nc and
near the valence band edge for N A � Nv, as we noted before.

2.5 Current Densities

In the presence of external electric and magnetic fields the thermal equilibrium
in the semiconductor is disturbed. The behavior of charged particles in such
fields is described by the Maxwell equations. In normal semiconductor oper­
ation magnetic effects can be neglected. The most important consequence of
the Maxwell equations, for our purposes, relates the charge density p (charge
per volume) to the divergence of the electric field£:
P
\7. £=- (2.5.1)
cs
where \7 is the Nabla operator!, and Cs = coc is the electrical permittivity
of the semiconductor with co = 8.85418 X 10-12 F/m denoting the vacuum
permittivity, and c is the dielectric constant of the semiconductor. For silicon,
c = 11.9. This equation holds for homogeneous and isotropic materials under
quasi-static conditions and is called the Poisson equation. The gradient of the
20 Chapter 2

electrostatic potential V is given by

lV'v = -£ · 1 (2.5.2)

Hence, the Poisson equation can be rewritten as

I �V = - t l (2.5.3)

where � is the Laplacian operator2• Since the potential energy of an electron is

-q V , and the potential energy of a hole is q V , the differential spatial structure
of the energy band edges can be computed from the electric field as

V'Ee = V'Ev = -q V'V = q£ . (2.5.4)

The mobile charge carriers in a material that is not in thermal equilibrium

give rise to current flow. The total current density J (charge flowing through a
given cross-section during a given time interval) in a semiconductor is the sum
of an electron current density J n and a hole current density J po By historical
definition the electron is assigned the negative elementary charge -q , where
we define q to be positive. However, the direction of the current density is
defined as the direction of positive charge flow. Consequently, J nis antiparallel
to the electron flow and J p is parallel to the hole flow. The average electron
flow velocity can be expressed as

In
vn= - ­ (2.5.5)
qn
and the average hole flow velocity as

v p -
Jp
(2.5.6)
-
qp
For each carrier type the current flow is due to two basic mechanisms,
namely diffusion and drift. Diffusion is a term borrowed from gas dynamics. It
describes the process by which a net particle flow is directed from a region of
higher particle density to a region of lower particle density along the density
gradient. This phenomenon is a direct consequence of the assumption of statis­
tical isotropic motion of the particles. The electron and hole diffusion current
Semiconductor Device Physics 21

nr �
...
n
p

r �
...

V V

� p •
Vn,diff Vp,diff

... •

In,diff Jp,diff

(a) (b)

Figure 2.7
Diffusion of (a) electrons and (b) holes. The directions of the carrier concentration gradients,
carrier motion, and electrical currents are shown.

densities are respectively given by

In,di// = q Dn\7n (2.5.7)

J p ,di// = -q Dp \7p (2.5.8)

where Dnand Dp are positive constants denoting the electron and hole diffu­
sion coefficient, respectively. The average diffusion velocities are

\7 n
Vn,di// = - Dn-;;;: (2.5.9)
\7p
Vp ,dif/ = - Dp p ' (2.5.10)

The relationships between the carrier concentrations, their gradients, the dif­
fusion velocities, and the diffusion current densities are shown schematically
in Fig. 2.7. As we shall see, diffusion determines the current flow in diodes
and, within the operating range mainly considered in this book, in transistors.
Diffusion also governs the ion flows in biological neurons.
22 Chapter 2

Drift currents are caused by electric fields. For low electric fields the
electron and hole drift current densities, respectively, are given by

In,drift = q JLnn£ (2.5.11)

J p ,drift = q JLp P£ (2.5.12)

where JLnand JLp are positive constants denoting the electron and hole mobility,
respectively. The mobilities are the proportionality constants that relate the
drift velocities of the charge carriers to the electric field according to

V n,drift = - JLn£ (2.5.13)

v p ,drift = JLp £ . (2.5.14)

The mobilities decrease with increasing temperature as JLoc T -n, where n =1.5
in theory, but empirically is found to be closer to n=2.5. The relationships
between the different parameters are illustrated in Fig. 2.8. At sufficiently large
electric fields the drift velocities saturate due to scattering effects and the term
JL£ in the above equations must be replaced by a constant term v s, which is
of the same order of magnitude as the thermal velocity. For intrinsic silicon at
room temperature, approximate values of the mobilities are JL n = 1500 cm2Ns
and JLp = 450cm2Ns, and the thermal velocity is 5 x 106 cm/s. Mobilities
decrease with increasing impurity doping concentrations.
For non-degenerate semiconductors, there is a simple relation between
diffusion constants and mobilities that was discovered by Einstein when he was
studying Brownian motion, and is therefore known as the Einstein relation:

Dn = UT JLn (2.5.15)
Dp = UT JLp (2.5.16)

where UT = kT/ q is the thermal voltage, and is the natural voltage scaling unit
in the diffusion regime. Its value at room temperature is approximately 25 mV.
From Eqs. 2.5.7-2.5.16 we then obtain the total electron and hole current
densities

In = In,drift + In,diff = q JLn( n£ + UT 'Vn) (2.5.17)

J p = J p ,drift + J p ,diff = q JLp (p£ - UT 'Vp) . (2.5.18)

In thermal equilibrium, diffusion and drift currents are balanced, that is J n =

Jp = 0 and the carrier concentration gradients can be computed by differen­
tiating Eqs. 2.3.4 and 2.3.5. If we further use Eq. 2.5.4 to express the electric
Semiconductor Device Physics 23

v v

r �
£
•
r �
£
•

... •

Vn,drift Vp,drift

• •

In,drift Jp,drift

(a) (b)

Figure 2.8
Drift of (a) electrons and (b) holes in an electrostatic potential V. The directions of the electric
field £, carrier motion, and electrical currents are shown.

field in tenns of the gradient of the energy-band edges, we obtain the important
result that in thennal equilibrium

'lEF = O. (2.5.19)

That is, the Fenni level is constant. This result is intuitively clear, because
otherwise a state of a given energy would more likely be occupied in one spatial
position than in another. More mobile charge carriers would then move to this
position than away from it, and so the energy states would be filled up until the
probabilities would be matched everywhere.
The temporal dynamics of the the carrier density distributions are de­
scribed by the continuity equations, which are a direct result of the Maxwell
equations:
on 1
at = Gn- Rn+ q'l In . (2.5.20)
oP
at = GP - RP -�'l.JP. (2.5.21)
q
where Gn and Gp denote the electron and hole generation rate and R n and
24 Chapter 2

Rp the electron and hole recombination rate, respectively. Generation and

recombination effects account for the creation and annihilation of electron­
hole pairs due to transitions between valence band and conduction band.
Generation requires a certain amount of energy that can be supplied by thermal
effects, optical excitation (discussed in Chapter 10) or impact ionization in
high electric fields. Recombination counterbalances generation and is driven
by the principle that a system tends towards a state of minimum energy.
For a recombination process to take place an electron and a hole have to be
present in close vicinity. Recombination is therefore limited by the availability
of minority carriers. Approximations of the recombination rates under low
injection conditions, where the majority carrier densities are much larger than
the minority carrier densities are given by

Rn _ n p - n po (2. 5. 22)
- Tn
Rp _P n- Pno (2. 5. 23)
- Tp
where n p and Pnare the minority carrier densities and n po and Pno their values
at thermal equilibrium. The minority carrier lifetimes T n and Tp are equal if
electrons and holes always recombine in pairs and no trapping effects occur.

2.6 p-n Junction Diode

The p-n junction diode is the fundamental semiconductor device. It is used as

a basis for every transistor type. Furthermore, it is the dominant light-sensing
device, and it will also become the most widely used sensor for electronic
imaging applications. Light-sensing applications of diodes will be discussed
in Chapter 10. The p-n junction has also found wide-spread use as the light­
emitting diode (LED). However, light-emitting diodes are very inefficient for
semiconductors with indirect bandgaps, such as silicon, and will not be treated
in this book.

Thermal Equilibrium

Consider what happens when an n-type semiconductor and a p-type semicon­

ductor are brought into physical contact 3• The diffusion processes described
in Section 2. 5give rise to a net electron flow from the n-type region to the
3 In practice. surface oxidation of the semiconductor materials would prevent this.
Semiconductor Device Physics 25

+
+ 8 8 8+ +

+
+ 8 8 +
+

8 8 8
+ + +
+ +
�------�--� �
x
p

I x

Eo

V l:::
�----- I
J:bi �
X

E -
];¢bi
- - - - - '-

-
I
� -
I
Ec

I -:-. I Ev
II •
•
x

p region Depletion region n region

Figure 2.9
Characteristics of an abrupt p-n junction in thermal equilibrium with space-charge distribution p.
electric field distribution E, potential distribution V. and energy-band diagram E.
26 Chapter 2

p-type region and a net hole flow from the p-type region to the n-type re­

dil n,dil p,dil

gion. The combination of these two effects results in a diffusion current den­
sity J I = J I +J I from the p-type to the n-type region, as given
by Eqs. 2. 5. 7 and 2. 5. 8. The diffusing minority carriers recombine with ma­

is largely devoid of mobile charge carriers and <<np n

jority carriers in the vicinity of the junction. As a result, this diffusion region
�. This region is
therefore called a depletion region. This situation is schematically shown in
the topmost graph of Fig. 2. 9. While the n-type and p-type semiconductor are
electrically neutral in isolation, this does not apply to the depletion region. On
the n-type side of the junction the donor electrons are absent and the donors
have a surplus proton; whereas on the p-type side the acceptor holes are filled
with the donor electrons from the n-type side, and so have a surplus electron.
Consequently, the depletion region is also called space-charge region. With a
positive net charge on the n-type side and a negative net charge on the p-type
side an electric field builds up in the depletion region that points from the n­
type side to the p-type side, according to Eq. 2. 5. 1. This electric field in turn

drilt n,drilt p,drilt

generates electron and hole drift currents with a combined current density of
J =J +J from the n-type to the p-type region, as given by
Eqs. 2. 5. 11and 2. 5. 12.In thermal equilibrium, diffusion and drift currents bal­
ance each other out and no net current flow is observed. Because of the electric
field the electrostatic potential, and thus the energy band edges, vary across the
space-charge region. Outside the space-charge region the energy bands flatten
out, leaving a constant offset of the band edges between the neutral n-type and

diffusion potential or built-in potential �bi.

p-type regions. The electrical potential corresponding to this offset is called
Using the fact that the Fermi level
is constant in thermal equilibrium (Eq. 2. 5. 19) and using Eqs. 2. 3. 4- 2. 3. 7 we
can compute the built-in potential to be

�bi = UT log (nnnoi;po) = UT log (nnnpoo) = UT log (PnpPoo) (2. 6. 1)

For highly-doped n-type and p-type regions Eqs. 2. 4. 4and 2. 4. 5
are valid and
the built-in voltage can also be expressed as

�bi = UT log (N�fA) (2. 6. 2)

The simplest p-n junction to analyze is the abrupt junction where the
n-type and p-type regions are each homogeneously doped and have a sharp
boundary. This case is illustrated in Fig. 2. 9. Assuming that the space-charge
Semiconductor Device Physics 27

region is fully depleted we obtain charge densities of

(2. 6. 3)
(2. 6. 4)
within the depletion regions of the n-type and p-type material, respectively. The
net charge density outside the depletion regions is zero, since the n-type and
p-type bulks are electrically neutral. According to Eq. 2. 5. 1the relationship
between charge density distribution and electric field is given by

aEx (x) p (x) (2. 6. 5)

ax Cs
in the one-dimensional case, where xis the coordinate along an axis perpendic­
ular to the junction plane with Ex pointing along that axis. Using the condition
that the electric field is zero at the boundaries x n and xp of the depletion re­
gion in the n-type and p-type material, respectively; and the charge neutrality
condition ND Xn = -NA Xp ; Eq. 2. 6. 5can be integrated to yield

(2. 6. 6)
in the n-type depletion region and

qNA (X- xp ) = Eo qNA x

_
(2. 6. 7 )
Cs Cs
in the p-type depletion region. Here
qND Xn _ qNA xp (2. 6. 8)
Cs Cs
is the electric field at the junction; where it reaches its largest magnitude. The
one-dimensional version of Eq. 2. 5. 2states that the electric field is the partial
derivative of the potential distribution along the xdirection according to

(2. 6. 9)
Since potentials are always measured with respect to a reference value the
offset of the V (x) curve is arbitrary. Choosing V (x = 0) = 0 we find

V (x) = - Eox+ 2Eo x 2

(2. 6. 10)
Xn -
28 Chapter 2

in the n-type depletion region and

Eo 2
V(x) = -Eox + x (2.6.11)
2 Xp
-

in the p-type depletion region. The built-in potential can then be expressed in
tenns of the depletion region width

(2.6.12)

as
(2.6.13)

Eliminating Eo from Eqs. 2.6.8 and 2.6.13 and solving for the depletion region
width we obtain
2 CB NA +ND
d= � bi · (2.6.14)
q NAND
The p-n junctions fabricated with typical silicon processes are not abrupt,
but have a more gradual profile. Their characteristics have to be detennined nu­
merically, but are qualitatively similar to those of the abrupt junction analyzed
above.

Forward and Reverse Bias

Having characterized the p-n junction under thennal equilibrium conditions we

now consider the cases where a net current flows through the diode. In practice,
the most common way of generating such a current flow is by changing the
boundary conditions for the n-type and p-type regions through the external
application of a potential difference. If a positive voltage is applied to the p­
type region relative to the n-type region the potential difference is called a
forward bias, if the voltage applied to the n-type region is higher we speak of
a reverse bias. Consequently, a current flowing from the p-type region to the
n-type region is calledforward current and a current flowing from the n-type
region to the p-type region is called reverse current.
In steady state, the total current density must be constant throughout the
diode. In the n-type and p-type bulk regions the current is made up of majority
carriers. The electron current in the n-type region is thus transfonned into a
hole current in the p-type region. This transfonnation happens in the vicinity
of the depletion region. The applied voltage V appears across the depletion
region as a change in the built-in voltage and thus modifies the width of the
Semiconductor Device Physics 29

depletion region and the minority carrier densities outside the depletion region
boundaries. The depletion region width can be computed from Eq. 2. 6. 14by
substituting c)bi with c)bi - V, if we define V to be positive for a forward bias:

2cs NA +ND
d= c)bt V) (2. 6. 15)
q NAND (
_
. •

According to the Boltzmann distribution (Eq. 2. 3. 2) the minority carrier

densities grow exponentially with decreasing electrostatic potential, so that
outside the depletion region boundaries they become

n p - n po eV/UT (2. 6. 16)

Pn= PnoeV/UT . (2. 6. 17 )
Since the majority carrier distributions are approximately constant throughout
the neutral regions, the np product is now given by

(2. 6. 18)
at the depletion region boundaries. The probability distributions for the occu­
pancy of a given energy state are now centered around the so-called quasi­
Fermi levels q c)nand q c)p , where

(2. 6. 19)
at the depletion region boundaries. The same argument that led to Eq. 2. 5. 19
for the thermal-equilibrium case now gives

In = -q JLnn \1c)n (2. 6. 20)

Jp = -q JLpp \1c)p . (2. 6. 21)
The current densities are therefore proportional to the gradients of the quasi­
Fermi levels. Outside the depletion region the electric field is small, as in the
thermal-equilibrium case, and the current flows mainly by diffusion. Within the
depletion region the concentration of mobile charge carriers is very low, and
therefore no significant recombination effects take place there. Consequently,
the electron and hole currents are almost constant throughout the depletion
region. The energy band diagrams for the two different biasing conditions
are shown in Fig. 2. 10and the carrier distributions and current densities in
Fig. 2. 11.
A forward bias diminishes the potential step across the junction. As a re­
sult, the minority carrier concentration and thus the np product on either side
30 Chapter 2

J
E
EC -----'--....
- � .. - -:- - (q(ct>bl-V)
j �' - - __ - ----- - - -
- -
- , - - -
--
-qct>p qV - -
T
-'"'\---- - - ,

�-
��I --'---
,•• --------------
d Ev

p reg ion Depletion region n regi on

(a)

---- : �
J
E
EC ...,.-""" - - ---:--[q(ct>bl-V)
____ L_..)___ ""--- '
-qct> -qV
P ' "'1z -
-' - - -
-__
- - - ...
�
,
---
-----
-

� --�
" ,.. �' � ------
d
------
----- Ev

p reg ion Depletion reg ion n region

(b)

Figure 2.10
Energy band diagram of a p-n junction diode for (a) forward bias, and (b) reverse bias.

of the depletion region are increased. This condition leads to an increase in

the recombination rate outside the depletion region boundaries and thus to a
minority carrier gradient that gives rise to a forward diffusion current. Since
the minority carrier densities at the depletion region boundaries increase expo­
nentially with applied forward bias (Eqs. 2.6.16 and 2.6.17) the recombination
rate, and therefore the forward current density, increase exponentially.
Semiconductor Device Physics 31

p reg i on p reg ion n reg ion

n,p n,p I
I
I
I
I
I
I
L ___ �no
Pno I
I
I
I

o�------��- o�--��--�----�
x x

J J

(a) (b)

Figure 2.11
Minority carrier distributions and current densities in the vicinity of a p-n junction for (a) forward
bias, and (b) reverse bias. Figure adapted from S. M. Sze (1981), Physics of Semiconductor
Devices, 2nd Edition. © 1981 by John Wiley & Sons, Inc. Reprinted by permission of John Wiley
& Sons, Inc.

A reverse bias increases the potential step across the junction. The minority
carrier concentrations, and the np products on both sides of the depletion
region are decreased and therefore the recombination rate is decreased. The
thermal generation rate now exceeds the recombination rate near the depletion
region boundaries. This condition results in a small minority carrier gradient
pointing away from the junction, and thus a small reverse diffusion current
density occurs.
32 Chapter 2

An approximation of the diode current-voltage relationship based on the

above considerations is given by the Shockley equation:

IJ = In+ Jp = Js (eV/UT 1) I
-
(2.6.22)

with
Js -_ q Dnn
Ln
po +
q Dp Pno
Lp (2.6.23)

where L n = ..j DnTnis called the electron diffusion length and L p = J Dp Tp

the hole diffusion length. This current-voltage relationship is illustrated in
Fig. 2.12(a). Since the reverse current density is limited by Js ' which is much
smaller than forward current densities at forward biases larger than about 4U T,
diodes are often used as rectifiers with a large conductivity in the forward
direction and a small conductivity in the reverse direction. This application is
reflected in the circuit symbol for the diode (Fig. 2.12(b», which is resembles
an arrow pointing in the forward current direction.

p n

(a) (b)

Figure 2.12
The p-n junction diode. (a) Current-voltage characteristic of an ideal diode according to the
Shockley approximation. (b) Diode symbol; the arrow indicates the direction of a forward current
density JF.

The Shockley equation is derived from the diffusion current density equa­
tions 2.5.7 and 2.5.8, the continuity equations 2.5.20 and 2.5.21, as well as
Eqs. 2.5.22 and 2.5.23 for the recombination rates. The underlying assump-
Semiconductor Device Physics 33

tions (Shockley approximation) are: Abrupt depletion layer boundaries; the

validity of the Boltzmann approximation given by Eq. 2. 3. 2, and of the low­
injection condition; negligible generation current in the depletion layer; and
constant electron and hole currents within the depletion layer.

Junction
breakdown

[ Reverse
--
- -- (e)-- ---
- --
'{

Ideal forward

Ideal reverse

1 0- 1 L---��--��---L�---L-----L----�lr-
o 5 10 15 20 25 30
qj VllkT

Figure 2.13
Comparison of the current-voltage characteristics of an ideal and a practical diode. (a) Generation­
recombination current domain. (b) Diffusion current domain. (c) High-injection domain.
(d) Series-resistance effect. (e) Reverse leakage current due to generation-recombination and sur­
face effects. Figure adapted from 1. L. Moll (1958). The evolution of the theory of the current­
voltage characteristics of p·n junctions. Proc. IRE. 46. 1 076. © 1958 IRE now IEEE.

For silicon p-n junctions there is only a qualitative agreement between

the observed behavior and the Shockley equation 2. 6. 22, because the above
approximations are not completely justified; and because of surface effects
at the semiconductor boundaries. Figure 2. 13compares the current-voltage
34 Chapter 2

characteristics of an ideal diode and a real diode. In semiconductors with small

intrinsic carrier concentrations ni, such as silicon, the reverse diffusion current
density (given by Js for reverse biases larger than approximately 4UT ) may
be dominated by a superimposed reverse generation current density J g e n . The
generation current is mainly due to trapping centers in the depletion region.
Trapping centers are imperfections in the crystal, which capture and release
mobile charge carriers. The generation current density due to trapping is given
by
_ qnid
Jg e n - (2.6.24)
Te

where Te is the effective lifetime of the trapping. Similarly, under forward bias
conditions there is a recombination current density component due to carrier
capture processes mainly in the depletion region that exhibits an exponential
behavior
(2.6.25)
Empirically, the total forward current density can be fit with the function

(2.6.26)
where n is a number between I and 2, depending on which current density
component dominates.
For large forward biases, where the minority carrier concentrations ap­
proach the majority carrier concentrations near the depletion region bound­
aries, part of the applied voltage appears as linear potential drops outside the
depletion region, which with increasing forward bias start to extend more and
more into the semiconductor between the diode terminals. In this domain, the
forward current-voltage characteristic is subexponential and finally asymptotes
to a linear behavior given by the series resistance of the bulk: regions.
For large reverse biases, a phenomenon called junction breakdown occurs
that expresses itself in a sudden increase of reverse current at a certain reverse
voltage. For silicon with typical impurity doping concentrations this effect is
due to impact ionization: The generation of electron-hole pairs by collision
with an electron or hole that has acquired sufficient kinetic energy in the elec­
tric field of the depletion region. A charge carrier may create multiple electron­
hole pairs during its transition through the depletion region. The generated car­
riers can in tum create electron-hole pairs if they acquire sufficient energy, and
so on. This effect is known as avalanche multiplication. It is characterized by
a sharp onset and a high gain with respect to a reverse voltage change.
Semiconductor Device Physics 35

2.7 The Metal-Insulator-Semiconductor Structure

As its name implies, the Metal-Insulator-Semiconductor (MIS) structure con­

sists of a conductor and a semiconductor that are separated by a thin insu­
lator layer. The most common version of the MIS structure is the Metal­
Oxide-Silicon (MOS) structure, where the 'oxide' is in most cases silicon
dioxide (Si0 2 ). The MOS structure and the p-n junction diode are the build­
ing blocks of today's most widely used type of transistor, commonly known
as Metal-Oxide-Semiconductor Field-Effect Transistor (MOSFET) . In addi­
tion, the MOS structure is the basic building block of the Charge-Coupled
Device (CCD), which is currently the most widely used device for electronic
imaging applications, and which is presented in Section 10. 5.
Operation Domains

In a typical MIS structure the insulator layer is sufficiently thick that it can­
not be crossed by charge carriers under normal operating conditions and suffi­
ciently thin that the charge on the conductor can influence the charge distribu­
tion in the semiconductor via the electrostatic potential it induces. A positive
charge on the conductor attracts mobile electrons from the semiconductor to
the semiconductor-insulator interface and repels mobile holes away from the
interface. Conversely, a negative charge on the conductor attracts holes and
repels electrons. If the semiconductor is n-type, a positive charge on the con­
ductor increases the majority carrier density near the semiconductor surface,
an effect known as accumulation, while a negative charge on the conductor
reduces the majority carrier density. With increasing negative charge on the
conductor, most majority carriers are driven from the region near the surface,
resulting in depletion, and eventually minority carriers start to accumulate at
the semiconductor surface, an effect called inversion. The same effects are ob­
served in p-type semiconductors, if the sign of the charge on the conductor is
reversed.
The energy-band diagram is a helpful tool to visualize these effects. In
order to be able to compare the energy levels and potentials in the conductor
and the semiconductor it is helpful to define a few more parameters, as shown
in the band diagram of Fig. 2. 14for the case of a p-type semiconductor. The
basic concept is that of the work junction, which is defined as the energy
difference of an electron between the Fermi level in the material and the
vacuum level in free space. The work function is denoted by q ¢ m, where ¢m is
the electrostatic potential difference corresponding to the work function. In the
36 Chapter 2

Vacu u m
level q Xj
- - - -

Ec
Egl2

Ej
q ljl B
- - - - - - - - -

Ev
--
dj

M etal Insu lator Sem iconductor

Figure 2.14
Energy-band diagram of an ideal MIS diode with no applied bias between the semiconductor and
metal for a p-type semiconductor.

same context the electron affinity is defined as the energy difference between
the bottom of the conduction band in the semiconductor or the insulator and
the vacuum level. For the semiconductor we denote the electron affinity by
QX, for the insulator by QX i . Furthermore, the potential difference between the
Fermi level in the metal and the insulator conduction-band edge is denoted by
¢B . The potential difference separating the Fermi level E F and the intrinsic
Fermi level Ei of the semiconductor can be computed from Eqs. 2. 3. 7 , 2. 3. 8,
and 2. 4. 9as

(2. 7. 1)

An MIS diode is called ideal if it has the following properties: Firstly, when
there is no applied bias, the work functions of the semiconductor and the metal
are equal: the Fermi levels line up and the energy bands in the semiconductor
are flat (flat-band condition). Secondly, the charge on the conductor plate is
equal to the total charge in the semiconductor with opposite sign. Finally, the
insulator is neither charged nor permeable to charge carriers. Note that for
certain applications deviations from this ideal behavior may be desirable, as
we will see in later chapters.
Semiconductor Device Physics 37

Steady-State Analysis

With the above terminology and assumptions we can now explain the effects of
accumulation, depletion, and inversion with the bending of the semiconductor
energy bands near the semiconductor-insulator interface, as shown in Fig. 2. 15
for a p-type semiconductor. In thermal equilibrium, the semiconductor Fermi
level is constant and separated from the conductor Fermi level by the energy
corresponding to the applied potential difference. Inversion occurs when the
intrinsic Fermi level crosses the Fermi level near the surface, corresponding
to the situation where the minority carrier density is larger than the majority
carrier density at the surface.

CL.>"-- Ec

M etal Sem iconductor Metal Semiconductor M etal Sem iconductor

I n s u l ator I n s u l ator I n s u l ator

(a) (b) (c)

Figure 2.15
Energy-band diagrams of an ideal MIS diode with applied bias for a p-type semiconductor in
(a) accumulation, (b) depletion, and (c) inversion.

For further analysis we define an electrostatic potential 'Ij; corresponding

to the difference between the local intrinsic Fermi level and the intrinsic Fermi
level in the bulk. The value of this potential at the semiconductor surface is
called surface potential 'lj;s. For a p-type semiconductor, 'lj; s < 0 corresponds
to accumulation; 'Ij; s = 0 to the flat-band condition; 0 < 'Ij; s < 'Ij;B to depletion;
and 'Ij; B < 'Ij; s to inversion. The potential distribution and therefore the bending
of the energy bands can be computed from the space-charge distribution using
38 Chapter 2

the Poisson equation

p(x) (2. 7. 2)
where the x axis is perpendicular to the semiconductor surface. The following
analysis will only be made for a p-type semiconductor. For an n-type semi­
conductor the results are the same if the symbols p and n are interchanged,
and the signs are appropriately changed. Using the flat-band charge-neutrality
condition the space-charge density can be expressed as

q(Nt - NA + Pp - n p ) = q( n po - Ppo + Pp - n p )
p(x) = (2. 7. 3)
where Nt and NA are the densities of ionized donors and acceptors, respec­
tively. The carrier concentrations are given by

n p - n po e 1/J/ UT (2. 7. 4)
Pp - Ppo e- 1/J/ UT . (2. 7. 5)
The Poisson equation can then be rewritten as

�:� = - :s (p po ( e- 1/J/ UT )
- 1 - n po ( e 1/J/ UT -1 )) (2. 7. 6)
Integration of this equation leads to

(2. 7. 7 )

e - 1/J/ UT + .!P....
UT
_ 1 +
n po
Ppo
( e 1/J/ UT - .!P...
UT -
. 1 )
(2. 7. 8)
where L D = VcsUT / qp po , and the electric field has the same sign as the
potential. Integrating Eq. 2. 6. 5from the bulk to the surface we can now define
an area charge Q s as the charge underneath a unit area of semiconductor
surface and relate it to the surface potential using Eq. 2. 7 . 7 :

Qs = -cs£s (2. 7. 9)
V2csUT
= =f --­
LD
e - 1/J . / UT +
1/Js
UT
_ 1 +
n po
Ppo
( e 1/J. / UT _
1/Js
UT
_ 1 )
(2. 7. 10)
where £s is the electric field at the surface. This relationship between area
Semiconductor Device Physics 39

1 0 -4

P - type Si (300K)
NA = 4 X 1 0 1 5 cm-3

1 0 -5 -exp (q\flsI2kT)
(Strong i nversion)

-exp (q/lflsf/2kT)
N' 1 0 --;; (Accu m u l ation)
E
�
(g>
1 0 -7
I"

1 0 -8 Weak
i nversion

1 0 -9 L __�____�__�L-��____-L____�-L�____
-
-0.4 -0.2 0 0.2 0.4 0.6 0.8 1 .0
\fIs (V)

Figure 2.16
Dependence of the area charge Q. on the surface potential for p-type silicon with acceptor density
NA = 4 X 1015 cm - 3 at room temperature. Figure adapted from S. M. Sze (1981), Physics of
Semiconductor Devices, 2nd Edition. © 1981 by John Wiley & Sons, Inc. Reprinted by permission
of John Wiley & Sons, Inc.

charge in the semiconductor and surface potential is plotted in Fig. 2. 16. The
different domains shown in Fig. 2. 15can be distinguished by the different
dependencies of Q s on 'lj; s. In accumulation ('Ij;s < 0) the first term under
the square root dominates, and we obtain Q s e-1/ls /2UT . In the flat-band
'"

situation ('Ij;s = 0), Q s = O. In depletion (0 < 'lj;s < 'lj;B ) the second
term dominates, and Q s - J'Ij; s / UT. According to these characteristics the
'"

inversion domain is separated into two distinct sub-domains; weak inversion

('Ij;B < 'lj; s < 2'1j;B) which exhibits the same relationship between Q s and 'lj; s
as depletion; and strong inversion ('Ij; s » 2'1j;B ) with Q s _ e1/ls /2UT . The
'"

transition region between weak and strong inversion ('Ij; s � 2'1j;B ) is called
40 Chapter 2

moderate inversion. Strong inversion shows a relatively abrupt onset at

(2. 7. 11)
In the case of inversion, the area charge consists of a contribution by mobile
electrons close to the surface, denoted by Q i , and a contribution by ionized
acceptors in the depletion region, Q d, and is equal with opposite sign to the
charge per unit area on the conductor plate, Q g :

(2. 7. 12)
where

(2. 7. 13)
and d is the depletion region width. Figure 2. 17 shows the distributions of
area charge, electric field and potential in the ideal MIS diode in inversion
with an externally applied potential difference V, under the assumption that
all mobile charge accumulates at the surface. Since the insulator is assumed
to be neutral, the electric field is constant and the potential decreases linearly
within the insulator. The total potential drop through the insulator is given by
Q sdi Qs (2. 7. 14)
Vi = £x i di = Ci
_
=
_

Ci
where Ex i denotes the electric field in the insulator, di the thickness of the
insulator, C i the permittivity of the insulator, and

(2. 7. 15)
is the insulator capacitance per unit area. The applied voltage is the sum of the
voltage drop across the insulator and the surface potential:

V = Vi + 'l/Js · (2. 7. 16)

In the depletion and weak: inversion case, the mobile charge can be neglected
and the space-charge density is given by

p = -qNA . (2. 7. 17 )
The potential distribution can be obtained as a function of the depletion region
width using Eq. 2. 7 . 2:
(2. 7. 18)
Semiconductor Device Physics 41

Metal , I ns u l ator , S e m i cond uctor

�_-----
Ec

�----.---
-
EF
Ev

De p letion

p
, � I nversion

-d a d x
j

Figure 2.17
Ideal MIS diode with applied bias for a p-type semiconductor in inversion: Energy-band diagram,
area charge distribution p. electric field distribution [;. and potential distribution 'I/J.

with
qNA rP .

'l/Js =

2cs
(2.7.19)

The depletion region width reaches its maximum at the onset of strong inver-
42 Chapter 2

sion, which can be computed from Eqs. 2. 7. 11and 2. 7. 19as

(2. 7. 20)

The minimum voltage that has to be applied to the MIS structure to obtain
strong inversion is called threshold voltage. With the approximation that Q i «
Qd at the onset of strong inversion and using Eqs. 2. 7. 11, 2. 7. 13, 2. 7 . 14, 2. 7. 16,
and 2. 7 . 20the threshold voltage can be estimated to be

VT � - �: + 2'1j;B v'4cSi!:A'Ij;B + 2'1j;B . =

(2. 7. 21)

Conductor

I n s u l ator
. . . . . . . . . . . . . I n version l ayer
. .. . .. . .. . .. . .. . .. . .. . .. . .. . .. . .. . .. . .. . .. . . . .
'- " - " - " - '
Depletion reg ion

14--- Semiconductor

Figure 2.18
MIS diode in inversion with equivalent capacitive-divider circuit.

MIS Capacitance

Because the depletion layer does not contain any mobile charge, it can be
regarded as a capacitor and assigned an incremental capacitance per unit area
which is obtained by differentiating Eq. 2. 7 . 9:

Cd 8Qs (2. 7. 22)

e-W./UT + (eW./UT 1)
= _

8'1j;s
1 - � -

V2L D e-w. /UT

V + UT 1 + (ew./UT UT
Ppo
±�
1)
=
•

.:i!..L _ � - .:i!..L -
Ppo

(2. 7. 23)
Semiconductor Device Physics 43

In the case of depletion and weak: inversion this equation reduces to

(2. 7. 24)
and in the flat-band case to

(2. 7. 25)
The capacitances Ci and Cd are connected in series, as shown in Fig. 2. 18, so
that the MIS diode has a total incremental capacitance of

C Ci Cd (2. 7. 26)
Ci + Cd .
=

Since Ci is constant, it can be used to normalize C in terms of the capacitive

divider ratio
C Cd (2. 7. 27 )
Ci + Cd '
=

Ci
The dependence of the incremental capacitive divider ratio on the surface
potential, as given by Eq. 2. 7 . 22is shown in Fig. 2. 19(curve (a». The flat­
band capacitance is obtained from Eqs. 2. 7 . 15, 2. 7 . 25, and 2. 7 . 26as

C('ljJs 0)
ei (2. 7. 28)
= =
d
g' L
i+� D

In accumulation and in strong inversion Q s depends exponentially on 'ljJ s

and Cd » Ci , while in depletion and weak: inversion Cd is comparable to Ci .
Eq. 2. 7. 22and curve (a) of Fig. 2. 19describe the system in thermal equilib­
rium. If the applied voltage is modulated in time, however, the characteristic of
the capacitance Cd depends on the ability of the charge distribution to follow
the signal. In the case of inversion, this ability is already impaired for rela­
tively slow signals, because the inversion charge at the semiconductor surface
is electrically almost insulated and its concentration can only be changed via
the normal generation-recombination process. At large enough frequencies,
the inversion charge thus remains practically constant and only the majority
carriers in the substrate follow the signal. The incremental depletion capaci­
tance Cd then stays at a value comparable to C i and continues to determine the
capacitance-voltage characteristic into the range beyond the threshold voltage.
For typical silicon MIS structures, the change between the two characteristics
occurs at frequencies between 5Hz and 100 Hz. Around the threshold of in­
version, the depletion region width reaches a maximum value and becomes
44 Chapter 2

1 .0

CFB ( F l at band
0.8
capacitance)

�-
(.j
0.6

0.4
I
I
0.2 I
Semicond uctor !
breakdown I
I
0

Figure 2.19
Capacitance· voltage characteristics of an ideal MIS diode. (a) Low frequency. (b) High frequency.
(c) Deep depletion. Figure adapted from A. S. Grove et aI. (1 965), Investigation of thermally
oxidised silicon surfaces using metal-oxide-semiconductor structures, Solid-State Electronics, 8,
145-163. © 1 965, with permission from Elsevier Science.

independent of the applied voltage for larger voltages. The depletion capaci­
tance Cd is therefore also independent of the voltage (curve (b) in Fig. 2. 19).
If the voltage is switched rapidly beyond threshold from the accumulation do­
main, such that there is no time for inversion charge to build up at the surface,
the depletion region width grows beyond the limits normally set by the inver­
sion threshold. The depletion region width as a function of the applied voltage
then resembles that of a reverse-biased p-n junction diode. This condition is
called deep depletion and is shown by curve (c) in Fig. 2.19. The deep deple­
tion domain is important for the operation of charge-coupled devices, which
will be presented in Section 10. 5.
The relationship between the applied potential and the surface potential is
given by the coupling factor

(2. 7. 29)
Semiconductor Device Physics 45

With the help of Eqs. 2. 7 . 14, 2. 7. 16, 2. 7. 22,and 2. 7 . 26", can be expressed as
I '"=
Ci
Ci + Cd Cd
=
£ 1 !!...
= -
Ci
, I (2. 7. 30)
which is the incremental capacitive divider ratio as seen from the conductor
side. The coupling factor '" appears as a parameter in the basic equations
describing the operation of MOSFETs, where it will be called the subthreshold
slope factor, and will be used extensively in this book. In the small-signal
analysis of MOSFET circuits ", is assumed to be constant, but it must be kept
in mind that '" varies quite strongly with applied voltage, as can be seen from
Fig. 2. 19.
Parasitic Charges

We conclude this chapter by noting that in MOS processing technologies

the thin silicon dioxide layer is obtained by thermal oxidation of the silicon
surface. A consequence of this fabrication process is the creation of traps and
charges that must be taken into account in models of device behavior. There
are several types of such traps and charges that can be distinguished according
to their location either at the interface or within the oxide and according to
whether they are fixed or mobile. The most important effect of these charges is
a shift in the capacitance-voltage characteristics.
Typically, fixed positive charges are found close to the interface between
the silicon substrate and the oxide. They are not changed by variation of 1/J s .
Electrically neutral defects in the silicon dioxide, called oxide traps, can be
charged or discharged by electrons and holes introduced into the oxide. In
the presence of such charges an external voltage must be applied to the MOS
diode to establish the flat-band condition. This voltage is called the flat-band
voltage, Vfb . The relationship between applied voltage and surface potential is
then modified from Eq. 2. 7 . 16to become
V =
Vi + 1/Js + Vfb (2. 7. 31)
and the expression for the threshold voltage (Eq. 2. 7 . 21) is changed to
VT � Qd + 21/JB + Vfb
- =
J4csqNA1/JB + 21/JB + Vfb . (2. 7. 32)
Ci Ci
3 MOSFET Characteristics

The two most common devices used in today's integrated circuit technology
are the Metal-Oxide-Silicon Field Effect Transistor (MOSFET) 1 and the bipo­
lar junction transistor (BIT) 2 . The currents in these devices comprise either
positively-charged holes, negatively-charged electrons, or both holes and elec­
trons. The BJT is called a bipolar device because the current in the transistor
consists of both types of carriers, electrons and holes. The MOSFET 3 is called
a unipolar device because the current has only one type of carrier, either holes
or electrons.
In this book, we concentrate on MOSFETs and their current-voltage char­
acteristics in the subthreshold (also known as weak inversion) domain. We use
the transistor in this domain because the current here is exponentially depen­
dent on the control voltages of the MOSFET just as the ionic conductances of
a neuron are exponentially dependent on its membrane potential. Although the
current in a BIT is also exponentially dependent on its control voltages, we
only use BITs when there is a requirement for higher current drive, and for
lower offsets between transistors. Furthennore when MOSFETs are operated
in the subthreshold domain, they draw small currents so power consumption is
reduced.
We also describe to some extent the current-voltage characteristics of the
MOSFET in the above threshold domain. There are numerous texts and papers
that cover the transistor's characteristics in this domain (Weste and Eshraghian,
1 994; Ismail and Fiez, 1 994; Tsividis, 1 996; Johns and Martin, 1 997; Tsividis,
1 998; Gray et aI., 2(0 1 ). There are a few sources in which the operation
of the transistor in the subthreshold domain is described in detail (Mead,
1 989; Maher, 1 989; Andreou et aI., 1 99 1 ; Andreou and Boahen, 1 994; Enz
et aI., 1 995; Enz and Vittoz, 1 997; Tsividis, 1 998). We start by describing
the MOSFET structure and the biasing necessary for the different modes of

1 The field-effect transistor structure was first described in a series of patents by J. Lilienfeld that
were granted in the early 1930s. The MOSFET is the field-effect transistor type that is almost
exclusively used today. Historically, other field-effect transistor types were invented including
the junction field-effect transistor (JFET), and the metal-semiconductor field-effect transistor
(MESFET).
2 The pn junction field-effect and the bipolar transistor were invented by Bardeen, Brattain, and
Shockley and their colleagues at Bell Telephone Laboratories during 1947- 1 952. Even though the
FET was conceived earlier than the BIT, the latter was the first to be mass produced.
3 This device is also called a MOST (MOS transistor) or an IGFET (insulated gate field-effect
transistor).
48 Chapter 3

operation. We then derive the current-voltage characteristics of the MOSFET

in these different regimes using the equations derived for the pn junction in
Chapter 2. We also look at a small-signal MOSFET model and some second­
order effects that affect the operation of the transistor. Other issues such as
noise and transistor mismatch are discussed in Section 3.6, and in great detail
in Chapter 1 1 and Chapter 1 3 respectively.

3.1 MOSFET Structure

Gate (G) Drain (D)

(a)
Substrate (8)

Polysilicon gate
Metal Metal

Gate oxide
Field oxide
Field oxide

p' substrate
(b)

Figure 3.1
Structure of an n-type MOSFET in a p- body. The MOSFET has four terminals; the drain (0), the
source (S), the gate (G), and the bulk (8). (a) Pictorial view of the MOSFET. (b) A more realistic
picture of a cross-section of a fabricated MOSFET. Note that the gate oxide is much thinner than
the field oxide.
MOSFET Characteristics 49

The MOSFET is composed of a MOS structure and two pn junction diodes

that were both discussed in Chapter 2. The structure of an n-channel MOSFET
in a p- silicon body or substrate (lightly doped p-type) is shown in Fig. 3 . 1 .
Because the body i s p-type (that is, doped with acceptors), the majority carriers
in this region are holes. The n+ regions (heavily doped n-type) are the source
and drain regions of the transistor. These regions are symmetric to each other
and as we will see in the next subsection, they are defined only by the voltages
applied to these regions. Because the regions are n-type (doped with donors),
the majority carriers are electrons. These regions have a low resistivity because
of the heavy doping of donors.
The region underneath the gate and between the source and drain regions is
called the channel. The channel has a width W, and a length L. The minimum
transistor length in modern-day MOSFET processes is approaching 0. 1 j,tm.
(The issues in process scaling are discussed further in Chapters 1 3 and 14.) The
channel is insulated from the gate above by a layer of silicon dioxide (or gate
oxide). This oxide is thinner than the other oxide (field oxide), which covers the
rest of the substrate. The gate is made of heavily doped polycrystalline silicon

�l.�y
=+==�
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

:: . . : :: : : : :: : : ::: . : . : n: . . . . : : : . : . : : . :: . :
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

· ··· · . . · · · · · · · · · · · . . · · . · . . . · · · . . . . . . . .
. .

(a) (b)

Figure 3.2
Physical structure of (a) an nFET and (b) a pFET in a common p- substrate. The pFET rests in a
n-well within the substrate.

(also called polysilicon or poly), which has low resistivity. We can view the
transistor as having four terminals; the gate (G), the source (S), the drain (D),
and the bulk (B). Because the n+ source and drain regions can supply a lot of
50 Chapter 3

electrons to the channel, this device is called an n-channel MOSFET (nFET)4 .

The complementary type of MOSFET is the p-channel MOSFET (pFET) 5, in
which the charge in the channel is carried by holes supplied from the source
and drain regions.

s s s

B B B

D D D

(a)

s s s

B B G0---4 B

D D D

(b)

Figure 3.3
MOSFET symbols for (a) an nFET and (b) a pFET.

The fabrication steps needed to fabricate a MOSFET are described in

Chapter 1 3 . Almost all modem MOS processes are CMOS (Complementary
Metal Oxide Semiconductor) In a CMOS process, both nFETs and pFETs are
fabricated on the same substrate. In this chapter, we will assume a single well
CMOS process in which the nFET rests in the common p - substrate, and

4 It is also called an n-type MOSFET or NMOS.

5 It is also called p-type MOSFET or PMOS.
MOSFET Characteristics 51

the pFET rests in an n-well within the substrate as shown in Fig. 3.2. In an
alternative CMOS process, the common substrate could be n-type, with the
nFET resting in a p-well within the substrate; while the pFET rests directly
within the substrate. However, most CMOS processes now use a p-type starting
substrate. There are also processes in which both types of transistors rest in
individual bulks, within a common substrate. This arrangement is called a
twin-tub CMOS process. In circuit diagrams, the two types of MOSFETs are
indicated by the symbols shown in Fig. 3.3. There are three different sets of
symbols in common use. The bulk terminal (B) is usually not drawn when the
bulks of the transistors are connected to the appropriate power supply. In this
book, we use the symbols (without the bulk terminal, unless necessary) shown
in the last column.

Biasing the MOSFETs

First, we look at how an nFET should be biased. The drain voltage, Vd, and the
source voltage, Vs, of an nFET (see Fig. 3.2(a)) should be greater than or equal
to the bulk voltage, Vb, so that the pn junctions between the highly doped n+
regions and the substrate will be reverse biased. That is, Vsb = Vs - Vb � 0
and Vdb = Vd - Vb � O. These bias conditions guarantee that there will
only be a small reverse leakage current at these junctions and that most of
the transistor's current will flow in the channel. In an nFET, the n + region
biased at the higher voltage is called the drain, and the other n + region is
called the source6 • Because electrons are negatively charged, the direction of
positive current flow, I, is from drain to source, as shown in Fig. 3.2(a) and
Fig. 3.4(a), even though the carriers flow from source to drain. The currents
measured at the source and at the drain are approximately the same, that is,
there is very little loss of carriers along the channel.
For a pFET (shown in Fig. 3.2(b)), the p+ regions should be biased nega­
tive relative to its bulk, that is, Vsb ::; 0 and Vdb ::; 0 so that the pn junctions
are again reverse-biased. The n-type bulk (or n-well) of the pFET should be
biased higher than the p - substrate. For a pFET, the p+ region which is biased
at the higher voltage is called the source and the other p+ region is called the
drain7• Because holes are positively charged, positive channel current, I, flows
from the source to the drain as shown in Fig. 3.2(b) and Fig. 3.4(b). The bulk
of the pFET is usually connected to the highest voltage (Vdd) supplied to the

6 Electrons are supplied to the channel by the source, and removed by the drain.
7 Holes are supplied to the channel by the source, and removed by the drain.
52 Chapter 3

chip while the bulk of the nPET is tied to the lowest voltage (Vss). In a p­
substrate, where the pFET rests in an n-well, the substrate is connected to Vss,
and the well to Vdd.

3.2 Current-Voltage Characteristics of an nFET

We start by deriving the current-voltage characteristics of the nFET. The

characteristics of the MOSFET depend on the relative voltages between the
four terminals. The terminal voltages are usually referenced to the source or to
the bulk voltage.

(a) (b)

Figure 3.4
Biased MOSFETs showing the direction of conventional current flow. (a) For proper nFET
operation, we should ensure that Vg 2: Vb, Vs 2: Vb, and Vd 2: Vb. If Vd 2: Vs, the channel
current I is positive, flowing from drain to source as shown. If \I,i :::; Vs, then I is negative and
flows in the opposite direction. (b) For proper pFET operation, we should ensure that � :::; Vb,
Vs :::; Vb, and Vd :::; Vb. If Vd :::; Vs, the channel current I is positive, flowing from source to
drain as shown. If Vd 2: Vs, then I is negative and flows in the opposite direction.

We first consider the current flow in an nFET as a function of the gate-to­

source voltage, Vgs . Depending on the relative values of the four terminals of
the transistor, the nPET operates in one of the following regimes; accumula­
tion, depletion, weak inversion, moderate inversion, or strong inversion. These
regimes are equivalent to those of the MIS structure described in Chapter 2.
We say the MOSFET is in cutoff, when it is in either the accumulation or de-
MOSFET Characteristics 53

pletion regime; in subthreshold, when it is in weak: inversion ; and to be above

threshold, when the MOSFET operates in the strong inversion regime 8. In the
cutoff region, the current in the MOSFET is close to zero. For both the sub­
threshold and above-threshold regimes we can further divide the operation of
the MOSFET into the triode and saturation modes depending on the value of
Vds.
Subthreshold Region

Increasing the gate voltage increases the positive charge on the gate. This
charge repels the holes in the substrate and leaves behind negatively-charged
ions, that balance out the gate charge. The MOSFET operates in the sub­
threshold regime when the positive charge on the gate is almost balanced by
the negatively-charged depletion region underneath the gate (see Fig. 3.5(a)).
There is also a very thin layer of electrons beneath the gate (the inversion
layer). In subthreshold, we ignore the charge from the inversion layer because
it is almost negligible compared with the depletion charge. The energy band
diagrams in Figs. 3.5(b) and (c) are similar to the band diagrams for the pn
junction in Chapter 29. In these diagrams, the axis for the electron energy is
directed upwards, while the positive voltage axis is downwards. That is, the
quasi-Fermi level of the region that has a higher applied voltage (like the drain
of the transistor) is lower than the quasi-Fermi level of the source region.
We now compute the electron concentrations at the source and drain ends
of the channel. Assume for the moment that the potential in the channel (or the
surface potential), 1/Js, is constant. This constant potential corresponds to zero
drift current. The electron concentrations at the two ends of the channel depend
on the energy barrier that the electrons encounter. This barrier is determined by
the voltage difference between the surface potential 1/J s and the applied voltages
Vs and Vd at the source and drain respectively. The barrier heights at the source
end of the channel, 8 s, and at the drain end of the channel, 8 d, can be expressed
as

8 s=80 - q(1/Js - Vs) (3.2. 1 )

8 d=80 - q(1/Js - Vd) (3.2.2)

8 The below-threshold regime includes the subthreshold and cutoff regimes.

9 To assist non-physicists in making qualitative band diagrams, we assume that the quasi-Fermi
level, EF, of each region is connected to the external bias voltage that is applied to the region
(except for the 'l/Js region). This assumption is not tme in reality. Remember that the positive
voltage axis points downwards.
54 Chapter 3

p
Depletion
region
(lift ' (lift

E
(b)

Be ..�'.:- � Be
GG8r .GG8
EEcF
\
__

8888. \..8888

�----
_�r EV

Figure 3.5
An nFET in subthreshold. (a) Cross-section. (b) Energy band diagram in the linear regime. (c)
Energy band diagram in the saturation regime.

where 80 = q Vb i is the built-in energy barrier between the n + drain/source and

the p- substrate; Vb i is the built-in potential at this pn junction; and q is the
positive elementary charge.
The electron density (the number of electrons per unit volume) at the
source end of the channel, Ns, is given by

(3.2.3)

where No is the effective number of states per unit area in the channel. A
corresponding relationship holds for the electron density at the drain end of
MOSFET Characteristics 55

the channel:

(3.2.4)

These equations are obtained by solving the boundary conditions of the energy
states at the boundaries where the channel meets the source and drain depletion
regions10 11 . Because the barrier at the drain is higher than that at the source,
the electron concentration at the drain end of the channel is lower than at the
source end. The concentration gradient leads to the diffusion of electrons from
the source to the drain12. The electrons that diffuse to the drain end of the
channel are swept into the drain by the electric field in the depletion region
around the drain. The electrons in the channel form an inversion layer.
We can compute the current in the transistor using the electron diffusion
current density equation, Eq. 2.5.7 in Chapter 2:

1= In,diff Wt
dN
= -qWtDndi (3.2.5)

where In, dif f is the electron diffusion current density, W is the width of
the channel, t is the depth of the channel, D n is the diffusion coefficient
of the electrons, and �� is the concentration gradient across the channel.
In computing ��, we assume that the conduction is lossless. Therefore, the
current is constant as a function of position z in the channel; where z=O at the
source end of the channel. The concentration gradient can be expressed as

(3.2.6)

where L is the channel length, UT = kT/q is the thermal voltage13, and

10 The carrier density at the boundary is given by the integral with respect to the energy E, of
the product of the two-dimensional density of states, N(E), in the channel and the Fermi-Dirac
distribution of energy states in the source or drain region (see Eq. 2.3.3). This integral is as follows:

L� ( N(E)
�
e(E-E/ /kT + J dE

whereEc is the energy at the conduction band and Ef is the quasi-Fermi level.
11 This derivation, or a similar derivation can be found in Maher ( 1989); Tsividis ( 1998).
12 Similarly, the current flow in a neuron is due to the diffusion of ions between the intracellular
space and extracellular space of the neuron.
13 UT � 25.8 m V at room temperature.
56 Chapter 3

Nl = No e - 90 / kT• Substituting Eq. 3.2.6 in Eq. 3.2.5 , we get

1= _q W tDnNl ew. /uT ( e -vd /uT _ e - V. /UT ) (3.2.7)

L
= Ito eW. /UT ( e - V. /UT _ e - Vd / UT ) (3.2.8)

where Ito = q! tDnNl is the accumulated pre-exponential.

The surface potential 'l/Js is modulated by the gate voltage Vg' In subthresh­
old, the gate charge per unit area Q 9 is balanced primarily by the depletion
charge per unit area Q d underneath the gate as described in Section 2.7. As the
gate voltage increases, the depletion region also increases so that the negative
charge in this region balances the increased positive charge on the gate. The
surface potential relative to the bulk also increases because of the wider deple­
tion region (see Eq. 2.7 . 1 9). By assuming a small change around an operating
point, 'l/Jo , we can express 'l/Js as

(3.2.9)

where K, is the capacitive coupling ratio from gate to channel and is defined by
Eq. 2.7.29 in Chapter 2. Substituting Eq. 3.2.9 into Eq. 3.2.8, we can write the
current-voltage (I-V) characteristics for an nFET in terms of the gate, source,
and drain voltages which are referenced to the bulk voltage as

I 1= 10 en Vg /UT ( e - V. /UT - e - Vd / UT ) ·
I (3.2. 1 0)

Equation 3.2. 1 0 can be separated into the forward current, I f ; and the reverse
current, Ir; so that it can rewritten as

(3.2. 1 1 )

The directions of the currents are shown in Fig. 3.5. These component currents
are

If = 10 e (n Vg - V. ) /UT (3.2. 1 2)
Ir = 10 e (n Vg - Vd ) / UT . (3.2. 1 3)

The forward component I f depends only on Vg and Vs , while the reverse

component Ir depends only on Vg and Vd . This form of the subthreshold I-V
equation will be useful when we analyze circuits in later chapters.
MOSFET Characteristics 57

Measuring'" How can we detennine "'? The derivation for", was given by
Eq. 2.7.30 in Chapter 2:
Cox
"'= ---- (3.2. 14)
Cox + Cd
where Cox is the capacitance of the gate oxide per unit area, and Cd is the in­
cremental capacitance of the depletion layer per unit area. The depletion capac­
itance is a non-linear function of the gate-to-bulk: voltage, Vg b (see Eq. 2.7.24).
As vgb increases, the depletion width increases slowly; therefore Cd decreases,
which in tum leads to an increase in "'.
One way of computing'" is to measure the slope of the log-linear plot of
I versus Vg when the nFET operates in the subthreshold regime. An example
of this measurement is shown in Fig. 3.6. The data was measured from an
nFET fabricated in a 0.8 JLm CMOS process. The dependence of '" on Vg
can be measured from the local slopes of this curve. Because", varies slowly
with Vg , we can usually assume that'" is constant in an approximate analysis
of a subthreshold circuit. Since '" determines the slope of the MOSFET in
subthreshold, we also refer to'" as the subthreshold slope factor. The value of
'" ranges from 0.5 to 0.9 depending on the process in which the transistors are
fabricated.
Another way of computing '" is to source a constant current through the
MOSFET, and then to measure the source voltage of the transistor as a function
of the gate voltage. The slope of this curve also gives an approximate value for
"'. A curve of'" plotted against Vg computed in this way is shown in Fig. 5.5
in Chapter 5.

Drain-to-Source Voltage Dependence in the Subthreshold Region

The subthreshold equation (Eq. 3.2. 10) of the nFET encompasses two regions
of operation; the triode region and the saturation region. Which of these
regions the transistor operates in depends on the drain-to-source voltage (see
Fig. 3.7). In the triode region, the current depends on Vds while in the saturation
region, the current is almost independent of Vds.

Triode Region

The triode region describes the operation of the transistor for small Vds. It
is also called the linear region 14 . Eq. 3.2. 1 0 describes the current-voltage

14 It is also called the non-saturation region, the conduction region, or the ohmic region. The term
linear region comes from the above-threshold characteristics of the MOSFET where the current is
58 Chapter 3

\
Above threshold

Subthreshold

1O-1 2 '--__�___�__�___�__�
o
2 V (V) 3 4 5
gs

Figure 3.6
Current as a function of the gate-to-source voltage 'lgs, as measured from an nFET
(W/L= 1 2.8/ 1.6) in a 0.8 J,tm CMOS process for a fixed "ds and Ysb. The log-linear plot shows
the different regimes of operation. In the subthreshold regime, the current .{js is exponentially
dependent on Vgs. For this device, the current changes bye-fold for every 37 m V with a measured
K, of 0.576. In the above-threshold regime, the current has a quadratic dependence on 'lgs. The
moderate inversion regime lies in between the subthreshold and above-threshold regimes.

characteristic of the nFET operating in the linear region. This equation can
be rewritten as

(3.2. 15)

By taking a Taylor's series expansion of Eq. 3.2. 1 5 , we can show that I is

approximately linear with Vds for Vds ::; UT (see Fig. 3.7).

Saturation Region

As Vds increases beyond 4 UT, the concentration of electrons at the drain end
of the channel can be neglected with respect to the concentration at the source
end because of the larger barrier height as shown in Fig. 3.5( c). Any electrons
in the channel that diffuse close to the drain are immediately swept into the
drain by the electric field in this region. Because the diffusion current is no
longer dependent on the electron concentration at the drain, the current in
the transistor depends only on Vs and is approximately equal to I f . The 1-

linear with Vds.

MOSFET Characteristics 59

X10
-8

7 Vgs=0.7 V
6

5
/
Saturation region
� 4
.gj
- 3

2
I

�
Linear region
I
I

-1
0 4 0.2 0.4 0.6 O.S
Ur Vds (V)
Figure 3.7
The current, Ids as a function of "4s for an nFET in the subthreshold region. The current was
measured from the same device in Fig. 3.6. The current is approximately linear in v.is for very
small values of "4s and is approximately constant for "4s > 4UT .

V relationship in this region is described by

1I = If = 10 e ( I<Vg - Vs ) / UT ·1 (3.2. 1 6)

This region of operation is called the saturation region. As seen in Fig. 3.7, the
current is approximately constant in this region 15. Figure 3.8 shows a family
of curves measured from the same nFET in the subthreshold region with gate
voltages ranging from 0.3 V to 0.7 V. In these curves, the transition point from
the linear region to the saturation region occurs around Vds lOOmV and is =

independent of the gate voltage.

Above Threshold Region

In the subthreshold region, the transistor current flows primarily by diffusion.

As the gate voltage increases, the surface potential 'Ij; s also increases, caus­
ing more electrons to enter the channel from the source and drain regions. An
increasingly larger portion of the positive gate charge is balanced by the inver-

15 We will see in Section 3.5 that the current in this region is not actually constant because the
drain voltage modulates the effective channel length of the transistor, causing I to depend on \1'.
60 Chapter 3

7
10-
Vgs=0.7 V
/,,
10-
8
,
V
{:,, 0.6

10-
9
,
�'" V
- (:,, 0.5

-1
0
,
'"

10
V
11
r:,, 0.4

10-

10-
12
r� '
0.3 V

o 4 U 0.2 0.4 0.6 0.8

T Vds (V)
Figure 3.8

:::::
A family of curves showing I versus "ds, as measured from a subthreshold nFET for Vgs between
0.3 V and 0.7 V in increments of 0. 1 V. All curves start saturating around las 4UT.

sion charge fonned by the electrons in the channel. In this intennediate region
(or moderate inversion region), the current consists of both drift and diffusion
currents. As the gate voltage increases further, the transistor begins to operate
in the strong inversion region (or above threshold region). Here, the current
consists predominantly of drift current and the charge on the gate is balanced
primarily by the inversion charge. Further increase in the gate charge is bal­
anced by an increase in the inversion charge. Because of the larger electron
concentrations at the source and drain ends of the channel (as compared to the
concentrations in the subthreshold region), the now finite horizontal electric
field, pointing from the drain to the source, creates a potential drop across the
channel. In contrast to the subthreshold region, the surface potential along the
channel is no longer constant but varies from tP s ( O ) at the source end to tPs (1)
at the drain end, where tPs ( l ) > tPs ( O ) .
In above-threshold, the surface potential of the transistor no longer de­
pends on Vg through K,. The electron density at the source end of the channel
is exponentially dependent on tPs. Therefore, only a small change in tPs will
supply sufficient electrons to offset the additional positive charge on the gate
caused by the increase in Vg • The surface potential can be treated as if it is
clamped in the above threshold regime. Figure 3.9 shows the dependence of
MOSFET Characteristics 61

OJ Below Above threshold Ij!s

threshold
Ij!s=lj/o+ K Vg
I

\ I
I
I
I
I
,
,
,
,

I
I ,

I ," Q.f
I
I
----- ---- --,.-

Vr Vg

Figure 3.9
Qualitative plots showing the dependence of the inversion charge Q (dashed line) and the surface
potential 'l/Js (solid line) on the gate voltage in subthreshold. and above-threshold regimes of an
nFET. The threshold voltage VT delineates the two operating regions.

the inversion charge per unit area, Q i, and the surface potential 'IjJ s, on the gate
voltage in both the subthreshold and above threshold regions.

Threshold Voltage As shown in Fig. 3.9, the threshold voltage, VT , delin­

eates the two operating regimes of the transistor. This voltage is the minimum
gate voltage required to obtain strong inversion in the MOS structure. The
threshold voltage of the nFET at Vs = 0 (from Eq. 2.7.32 in Chapter 2) is
Qd
VTO = Vf b + 2'IjJB - C
- (3.2. 17)
ox

= Vf b + 2'IjJB +'Y V2'IjJB (3.2. 1 8)

where Vf b is the flat-band voltage, Q d is the incremental capacitance of the

depletion region per unit area, 2'IjJ B is the surface potential of the channel when
the concentration of electrons in this region is equal to the concentration of
acceptor ions, and'Y = � O�
is the body factor coefficient. The body factor
coefficient'Y is described in Section 3.2.
What happens when the source voltage increases? An increase in Vs leads
to a decrease in the current of the transistor. To return to the original current, 'IjJ s
has to increase by the same amount as the increase in Vs' Since o'IjJs = K, oVg ,
the gate voltage has to increase by a factor of (i > 1 ) more than Vs' The larger
62 Chapter 3

increase required for Vg can be treated as an increase in the effective threshold

voltage VT of the transistor:

Vs
VT = VTO + -. (3.2. 1 9)
'"
As mentioned previously, when the transistor is abvoe threshold, increases
in the gate charge are balanced primarily by increases in the inversion charge
so we can write the relationship between the inversion charge and the threshold
voltage as

Qi = - Cox (Vg - VT) (3.2.20)

VS
= - Cox (Vg - VTO _
)
'" .
(3.2.2 1 )

This equation will be used i n the derivation o f the I-V characteristics o f the
nFET in above-threshold.

Drain-to-Source Voltage Dependence in the Above Threshold Region

As in subthreshold, above-threshold also encompasses two regions of opera­

tion: the triode regime, and the saturation regime. These regimes can be seen
in Fig. 3 . 1 0, which shows a family of I versus Vds curves taken from an nFET
in the above-threshold region. Notice that the current do not longer saturates
around Vds � 1 00mV (compare with Fig. 3.8 for the subthreshold curve).
The drain-to-source voltage at which the current saturates depends on the gate
voltage.

Triode Region

We first compute the I-V characteristic in in the deep triode or ohmic regime,
where Vds is very small. Here, the drift current depends linearly on Vds. The
energy band diagram of the nFET is shown in Fig. 3. 1 1 (b). To compute the drift
current, we begin from the electron drift current density equation (Eq. 2.5 . 1 1).
The drift current density, In,drijt. (the drift charge flowing through a given
cross-section during a given time interval) is given by

(3.2.22)

where n is the carrier concentration, f..t n is the mobility of the electrons, and t:
is the horizontal field. The charge concentration can be re-expressed in tenns
MOSFET Characteristics 63

V =2.4 V

V =1.8
gs
V

----�--�--�
_1 L-
o 0. 5 2 2. 5 3 v;JV)
Figure 3.10
A family of curves showing the dependence of I on v.t. as measured from an nFET in above­
threshold. The curves were taken for ltg. between 1 .8 V and 2.4 V in increments of 0. 1 V.

of the inversion charge per unit area, Q i:

qn= - = --
QiWL Qi
--- (3.2.23)
WLt t
where W, L, and t are the width, length, and depth of the channel respectively.
We can then compute the drift current as

(3.2.24)

We now substitute Eq. 3.2.20 into Eq. 3.2.24 and, assuming a constant field
across the channel for small Vds, we getl6

1= ILn Cox (Vg - VT) W

Vd - Vs
L
= ILn Cox y (Vg - VT ) (Vd - Vs) .
W
(3.2.25)

We can rewrite Eq. 3.2.25 as

(3.2.26)

16 We also assume that the electron mobility is constant because of the small lateral field.
64 Chapter 3

where /3 = f.1n Cox r. From Eq. 3.2.26, we see immediately that the current
through the MOSFET is linearly proportional to Vds, and hence the transistor
acts like a linear resistor in this region. The conductance of the resistor can be
set by the gate voltage.

(a)

Inversion P :
Depletion
layer
region

(b) frift :
E
,

888e
•

�88�
8
- 8ee8
8888 eee
eeee

Figure 3.11
An nFET in the linear region in the above threshold domain. (a) Cross-section. (b) Energy band
diagram.

In the derivation of Eq. 3.2.26, we assumed that Q i is uniformly distributed

through the channel. In the triode and saturation regimes, (excluding the deep
triode regime), the distribution of Q i is non-uniform because the voltage drop
across the gate oxide, Vi (z ) = Vg 'Ij; s ( z ) , varies with position along the channel
-

length z17. Because the voltage drop Vi ( z ) decreases towards the drain end of
the channel, Q i also decreases towards the drain end as shown in Fig. 3. 1 1 .

17 Remember that the gate charge per unit length is determined by Cox Vi (z ).
MOSFET Characteristics 65

However, the current is constant along the channel if we assume no loss of

carriers along the channel.
Again, we use Eq. 3.2.24 to compute the drift current in the triode regime.
The inversion charge Q i can be computed indirectly from the incremental
channel capacitance per unit area, C, which consists of the oxide capacitance,
Cox , and the depletion capacitance, Cd :
8Q
C= i =Cox + Cd . (3.2.27)
81/Js
We no longer assume that the field along the channel is constant. Instead, it
depends on the gradient of the surface potential 1/J s along the length of the
channel:
d1/Js
£= _
(3.2.28)
dz .
If we substitute Eq. 3.2.27 into Eq. 3.2.28, we can rewrite £ as

" = ..!. i .
c dQ
(3.2.29)
_

C dz
Substituting Eq. 3.2.29 back into the Eq. 3.2.24, we solve

/Ln WQ dQ i 1 /Ln W d Q
1= - /Ln Q i W£=C 2
(3.2.30)
i dz = 2C dz i '
Integrating both sides of Eq. 3.2.30, we rewrite I as

I L=W :C ( Q� - Q;) . (3.2.3 1 )

where Q s and Q d are the inversion charges at the source end and the drain end
of the channel respectively. In this equation, I is negative because the z-axis is
directed from the source towards the drain. Our definition of positive current
flow is from the drain towards the source. So we rewrite I to accommodate this
convention:
(3.2.32)

We use Eq. 3.2.21 to compute Q sand Qd. The threshold at the source end
of the channel is given by VT( S ) = VTO + �. Therefore, the inversion charge
at the source end of the channel is
Vs
Qs= Cox ( Vg - VTO - - ). (3.2.33)
'"

The threshold at the drain end of the channel is given by VT (d) = VTO + �.
66 Chapter 3

P
---
I I

Inversion : : Depletion
region : Pin choff region

,drift:
+-----'

(b)
,

88Ge
888
GBB
8888
ee
E
eeG
eee
eeee
Ec

''---- Ev
Figure 3.12
An nFET in the above-threshold saturation region. (a) Cross-section. The pinchoff region lies
between the pinchoff point and the drain, where the inversion charge is negligible. (b) Energy
band diagram.

Therefore, the inversion charge at the drain end is

(3.2.34)

Replacing the definitions for Q sand Q d into Eq. 3.2.32, and using the defini­
tion K, = Cox /C, we get

(3.2.35)

where /3 = f.l Cox �. As in the subthreshold case, this current is the sum of a
MOSFET Characteristics 67

Table 3.1
Current-voltage relationship of nFET in subthreshold and above-threshold.

Mode Subthreshold Above threshold

Triode Ioe",VgjUT (e V.jUT - e VdjUT) ,f[Qs2 - Qd2]
Saturation IoeC",vg V.)jUT -£- [(�(Vg - VTO) - Vs)2]
Cutoff 0 0

forward component I f and a reverse component Ir:

If= ! [( �(vg - VTO) - Vs) 2 ] (3.2.36)

Ir= ! [( �(vg - VTO) - Vd) 2 ] . (3.2.37)

We can further reduce Eq. 3.2.35 to

�
1=,8[( Vg - VT ) (Vd - Vs) - 2 «Vd - Vs) 2 ) ] . (3.2.38)

By assuming that Vds is very small, we can neglect the second term in
Eq. 3.2.38 and obtain the same equation as Eq. 3.2.26 for the deep triode re­
gion.

Saturation Region

As we increase Vd further, the threshold towards the drain end of the channel
increases. Consequently, the inversion charge disappears at some point along
the channel because the gate charge needs only to be balanced by the depletion
charge. The point in the channel where the inversion charge first disappears is
called the pinchoffpoint. The region between the pinchoff point and the drain
is called the pinchoffregion and is almost depleted of electrons (see Fig. 3.12).
In fact, this pinchoff region is in subthreshold. The current in the pinchoff
still flows by drift and the electrons are swept into the drain region by the
electric field resulting from the potential difference between the channel and
the drain. The current in the transistor only depends on the source voltage, and
is independent of the drain voltage. Accordingly, this region of operation is
called the saturation region.
We derive the I-V relationship of the transistor in the above-threshold
saturation regime by setting Q d=0 in Eq. 3.2.35:
68 Chapter 3

which reduces to

(3.2.39)

The triode and saturation regimes of an nPET operating above threshold are
shown in Fig. 3.10 for different values of Vgs . The value of Vd where the
transition between the triode and saturation regimes occurs, depends on Vg •
This dependence on Vgs is unlike the subthreshold case where the transition
is independent of the gate voltage. The transition value is computed by setting
Qd = 0 in Eq. 3.2.35:
(3.2.40)

The I-V characteristics of the nFET in the subthreshold and above-threshold

regions are summarised in Table 3.1. In above-threshold, the definition of K, is
invalid because the incremental increase in charge underneath the gate due
to an increase in the gate charge arises from the inversion charge not the
depletion charge. We set K, = 1 in the above-threshold equations.

Body Effect

In the I-V equations that we have derived so far, the terminal voltages of
the transistor are referenced to the bulk. However, the bulk is also an input to
the transistor. In the subthreshold region, we can describe the influence of Vb
through the series of capacitors C ox and Cd (as we saw also for the gate input).
The effect on the surface potential can be written as

(3.2.41)

The influence of the bulk on the transistor18 is called the body effect or sub­
strate effect. The bulk is sometimes called the back gate; hence the body effect
is also called the backgate effect.
In a later chapter, we show a circuit in which the bulk is used as the input
to a transistor instead of the gate.
In the strong inversion, or above-threshold, regime the influence of the
bulk voltage is usually treated as an increase in the threshold voltage of the
transistor. If Vb decreases, then there is practically no change in the gate charge
because the voltage across the gate oxide is essentially unchanged (the surface

18 Remember that since the surface potential is referenced to the bulk, the actual surface potential
change will be -KO\1,.
MOSFET Characteristics 69

n
Depletion
(b)
region

�=o ====::J:::

/dill
-­
s

Figure 3.13
A pFET in subthreshold. (a) Cross-section. (b) Energy band diagram in the linear region. (c)
Energy band diagram in the saturation region.

potential remains approximately the same). However, the depletion region

underneath the gate increases 19. Since the negative charge from the depletion
region is now larger, less charge is required in the inversion region to balance
the gate charge. The inversion region becomes smaller, so leading to a smaller
I. To restore I to its original value, we increase the gate voltage Vgs . The
effective threshold is now

(3.2.42)

19 Remember from Chapter 2 that increasing the reverse bias across a pn junction causes the
depletion width to increase.
70 Chapter 3

where "I is the body effect coefficient described in Eq. 3.2.18, and Vbs is the
bulk -to-source voltage20•
Assuming that we do not forward the pn junctions between the drain/source
regions and the bulk: What happens if the bulk voltage Vb is increased by Ll V?
This scenario is the same as decreasing Vg , Vs, and Vd by the same Ll V. In the
subthreshold region, the change in 'ljJ s will now be '" Ll V. The barrier height
-

is decreased at both ends of the channel, and the current increases. Hence, the
bulk acts like the gate, but it has a weaker influence on the transistor current.

3.3 Current-Voltage Characteristics of a pFET

The current in a pFET arises from the transport of holes across the channel
from the source to the drain 21• In subthreshold, the current is primarily due to
diffusion. The structure of the pFET and its energy band diagrams in the linear
and saturation regions of the subthreshold domain are shown in Fig. 3.13(a),
(b), and (c) respectively. Since the diffusion process in a pFET obeys the same
laws as in the nFET, we can derive the current-voltage characteristic in the
same way as described in Section 3.2. The corresponding I-V equation for
the pFET is

(3.3.1)

where Vw is the bulk voltage of the MOSFET. The pre-exponentials 10 and "'s
in the pFET and nFET equations are not equivalent. The", value for the pFET is
different than that for the nFET because of the different doping concentrations
underneath the gate in the two types of transistors.
If the pFET rests in a n-well, then the well voltage, Vw, is usually con­
nected to the highest potential, Vdd. But, as we have seen earlier, the bulk
of the pFET can also be used as an input as long as the pn junctions at the
source/drain regions and at the well are not forward-biased. If Vg , Vs, and Vd
are referenced to Vw, Eq. 3.3.1 can be written as

(3.3.2)

20 It can be shown Enz et aI. ( 1 995) that K, is given by

,,(/2
K, =I- .
JVg b - Vfb - "(2/4
21 Recall that in the pFET, the source is biased at a higher voltage than the drain.
MOSFET Characteristics 71

The equation for the pFET in saturation is then

1= Io e ( - KVg +vs ) /uT . (3.3.3)

The above-threshold equation for the pFET is

(3.3.4)

3.4 Small.Signal Model at Low Frequencies

G 0>----- r---D
--.----��-

gmgtNg

B o--------------�---
I"
Figure 3.14
Small-signal model of an nFET at low frequencies. The current due to %s flows in the opposite
direction to the currents that are due to the gate transconductance and the drain conductance. This
convention takes into account the fact that the current decreases when l{ increases.

We now look at the case of small variations in the terminal voltages so

that we can express the resulting currents using linear equations. The aim is to
replace the nFET by a linear circuit that is the small-signal equivalent model
of the transistor in a quasi-static condition (that is, at low frequencies) 22 . This
model is shown in Fig. 3.14. Since the signal variations are slow, we can ignore
the capacitive effects at the nodes.

Conductance Definitions

In the small-signal model, the change in the transistor current due to very small
changes in each of the terminal voltages, can be expressed as conductances.
Because three terminal voltages are are normally referenced to the voltage
22 We discuss the small-signal model of an nFET at moderate frequencies in Appendix 3.7.
72 Chapter 3

Table 3.2
Conductances of Subthreshold nFET at Low Frequencies.

Conductance Subthreshold
9mg '!d.
U'"
10e\"Vg sJ/uT
9mB U'"
10e\" 9 dJ/uT + L
9md u'" v.

at the fourth tenninal (either the source or the bulk), only three conductance
parameters are required. In most textbooks, the analysis assumes that the three
tenninals are referenced to the source voltage. However, we first look at the
case in which the voltages are referenced to the bulk voltage. The small-signal
gate transconductance, 9mg , is defined as
81
9mg= 8V; . (3.4.1)
9

This parameter describes how the current changes with a small change in the
gate voltage. It is a "transconductance" because the the gate tenninal only
indirectly determines the transistor current. The source conductance, 9 ms, is
the parameter that describes how the current changes with a small change in
the source voltage. The source and drain conductances are real "conductances"
because the current actually flows between the drain and source tenninals. The
source conductance is given by
81
9ms= 8Vs' - (3.4.2)

and the small-signal drain conductance, 9md, is given by

81
9md= 8Vd· (3.4.3)

In Fig. 3.14, the current due to the source conductance flows in the opposite
direction to those due to the drain conductance and the gate transconductance,
because the current decreases when the source voltage increases.
The total change in the current i due to small variations in Vg , Vd, and Vs,
(referenced to the bulk Vb) is
. 81 81 81
z= 8Vg �Vg + 8Vs � Vs + 8Vd � Vd
(3.4.4)
MOSFET Characteristics 73

This equation can be recast in tenns of conductances that are referenced to the
source:
(3.4.5)

The relationships between these conductances and the bulk-referenced con­

ductances are

aJ
gm= =gmg
avgs
--

aJ
gmb= s =gms - gmg - gmd
aVb
aJ
gds= Vds =gmd
-- (3.4.6)
a
where gm is the gate transconductance, gmb is the body or substrate transcon­
ductance, and g ds is the drain-to-source conductance. These bulk-referenced
conductances are summarised in Table 3.2 for the transistor operating in sub­
threshold and in Table 3.3 for the transistor operating in above threshold.

Bulk-Referenced Conductances

Gate Transconductance In subthreshold, the gate transconductance, 9 mg

(or gm ) can be derived by differentiating Eq. 3.2.10 with respect to Vg :

(3.4.7)

In above-threshold, the gate transconductance in the linear region can be

derived from Eq. 3.2.38:

(3.4.8)

and in the saturation region, from Eq. 3.2.39:

gmg=,B1\: (Vg - VT )
=,B (I\: (Vg - VTO) - Vs) . (3.4.9)
74 Chapter 3

Table 3.3
Conductances of Above Threshold nFET at Low Frequencies.

Conductance Above threshold

Triode Saturation
9mg f:J(Vd - Va) f:J(K(V, - VTO) - Va)
9mB e.(K(Vg - VTO) - Va) �(K(Vg - VTO) - VB)
9md ;(K(Vg - VTO) - Vd) �

Drain and Source Conductances In subthreshold, the drain and source

conductances can be derived from Eq. 3.2.10:
81 If
gmB = =
-
8VB UT
Io e ( K Vg -v. ) /uT
(3.4.10)
UT
81 Ir I
gmd = =
8Vd UT + Ve
L0 e ( K Vg - Vd ) /UT I
= +_ (3.4.11)
UT Ve
where the second tenn in Eq. 3.4.11 expresses the channel-length modulation
effect23 in the saturation region. The drain conductance reduces to 9 md = t..
In above-threshold linear regime, we use Eq. 3.2.35 to solve for these
conductances:

(3.4.12)

(3.4.13)

In the saturation regime, gmB is also given also by Eq. 3.4.12, while gmd is
given by the slope of the I-Vd curve:
I
gmd = (3.4.14)
Ve·
Source-Referenced Conductances

The gate transconductance gm and the drain-to-source conductance 9 dB are

equal to the gate transconductance gmg and the drain conductance gd respec­
tively. The latter two conductances were derived in Section 3.4. The bulk
23 This effect is described in detail in Section 3.5.
MOSFET Characteristics 75

transconductance, 9mb can be derived easily as shown in Eq. 3.4.6. In sub­

threshold, the bulk transconductance is

9mb = 9ms - 9mg - 9md

( 1 - ",)1
= (3.4.15)
UT
In the above-threshold linear regime,

1 1
9mb = ,B ("'(Vg - VTO) - Vs) ( l - ) -
'"
- -
v: ' (3.4.16)
e

and in the saturation regime, 9mb = O.

3.5 Second-Order Effects

Vs Vg Vd

I II,IlJIliI
'- - - - r
I '
I Jm,L'!i J
, - --_ /

L ell
, , , ,

, , , p
.
.--­•
" ,

, L
,.. ----. ,

Figure 3.15
The effective channel length LeU of a transistor operating in the above-threshold saturation region
decreases with increasing v,i because the pinchoff point moves into the channel, away from the
drain. The effective channel length can be described by the transistor length minus the length of
the pinchoff region in the channel.

In the derivation of the current-voltage characteristics of the MOSFET, we

assumed that certain properties (for example, the electron and hole mobilities)
are constant under all operating conditions. These ideal assumptions do not
apply in a real device. Some examples of the non-idealities in a MOSFET are
described below.
76 Chapter 3

Early Effect

When deriving the J-V characteristics of the nFET we assumed that the
current is constant in the saturation regime. This assumption is not sufficient,
particularly for short-length MOSFETs 24 . The drain voltage can modulate the
channel current even in saturation. In the above-threshold saturation regime,
the effective length of the transistor, L elf , decreases when Vd increases because
the pinchoff region extends further along the channel away from the drain
(see Fig 3.15). We describe L elf as the difference between the drawn channel
length, L, and the size of the pinchoff region. Hence, the transistor current
increases with Vd•

· Ve (Early voltage)

Figure 3.16
Plot of current versus drain-to-source voltage. showing the slope of the curve !lis in the saturation
regime. The intersection of the slope with the v.t. axis is called the Early voltage.

The same phenomenon occurs in subthreshold as is evident from the finite

slope of the J-Vds curve in Fig. 3.7. This slope is the output conductance of
the transistor:
oJ oJ oLe
9ds = Vds = oLe OVdslf (3.5.1)
O
-- -- -- .

lf
By taking the derivative of Eq. 3.2.7 with respect to L elf , we can make the
I
replacement, 8 8LI = - -L and rewrite Eq. 3.5.1 as
eff eff

J oLelf J
9ds = =
Lelf O Vds Ve
- -- --- -
(3.5.2)

24 As we will see later. when computing the gain of a two-transistor amplifier. it is important to
use a long transistor for high gain.
MOSFET Characteristics 77

where Ve = L elf - gy!;

is the absolute value of the voltage at which this slope
intersects the voltage axis of the curve in Fig. 3.16. This voltage is called the
Early voltage25• In the subthreshold domain, the saturation current, I sat , which
is given by Eq. 3.2.16 can be re-expressed as

(3.5.3)

to account for the Early effect.

Figure 3.17 shows the measured Early voltage plotted against transistor
length, for transistors operating in subthreshold. The Early voltage varies from
20 V to 750 V and shorter transistors have a smaller Early voltage. This de­
pendence of the channel current on V d is sometimes called the channel-length
modulation effect: . The modulation also occurs in transistors operating above
threshold. This effect is sometimes modeled as a reduction in the effective
threshold of the transistor, or as a reduction in the energy barrier at the source.
The latter is similar to the drain-induced barrier lowering (DIBL) effect26•

800
700
600
�500
g'
Q)

-'6 400
>

� 300
>.

200
1 00

5 10 15 20
Transistor length (j.!m)

Figure 3.17
Early voltage versus the transistor length of an nFET fabricated in a 0.8 JLm CMOS process.

25 It is named after Jim Early who first analyzed this effect in BJTs.
26 See Chapter 14 for a description of the DIBL effect.
78 Chapter 3

Mobility Dependence on the Vertical Field

In Section 3.2, we assumed that the electron mobility is constant. In reality,

the effective electron mobility, f..t elf , is a function of the electric field in the
channel. The horizontal component of the field drives the electrons from the
source to the drain while the vertical component (which depends on Vg ) attracts
the electrons to the oxide-silicon interface. When the vertical field increases,
the electrons collide more frequently with the interface, leading to a lower
effective mobility. The effective mobility can be derived from the current
versus Vgs relation of an nPET operating in the deep triode above-threshold
regime. The equation governing this relation is I = f..t elf Cox � (Vg - VT) Vds.

Velocity Saturation

Previously we assumed that the drift velocity v d of the electrons is linearly pro­
portional to the longtitudinal (or lateral) field component, E. This assumption
is only valid if the field is small. For fields above a critical value E c the velocity
saturates at a constant Vs, which has the same magnitude as the thermal veloc­
ity. This velocity saturation effect is also described in Section 2.5. The critical
value at which the velocity saturates depends on the doping concentration. As
a first-order approximation, the carrier-velocity saturation equation is

(3.5.4)

Because of velocity saturation, the current saturates at a smaller Vds for a given
Vg s than predicted by the current-voltage equations. This effect becomes more
pronounced for small-channel length transistors.

Narrow- and Short-Channel Effects

In our current-voltage derivations, we assumed that the channel of the transis­

tor is sufficiently long and wide that we could neglect any "edge" effects. As
the dimensions of transistors continue to shrink, this assumption is no longer
true. The field distribution in the channel becomes more complicated: The field
components along the channel, and along the width axis of the transistor be­
come significant. These changes result in short-channel and narrow-channel
MOSFET Characteristics 79

effects27 • In sub-micron processes, fabrication steps are engineered to mini­

mize these effects (See Chapter 13.).

Narrow-Channel Effects

The narrow-channel effect (due to narrow-width transistors) can be modeled by

an increase in the threshold voltage of the transistor. This effect is due to the
extension of the depletion region underneath the gate toward the sides, under
the bird's beak. The bird's beak is the part of the field oxide that encroaches
into the transistor's channel. A cross-section of the transistor showing this
encroachment can be seen in Fig. 13.1 in Chapter 13. The effective threshold
voltage appears higher because some of the field lines from the gate end in
the depletion region underneath the thick oxide (also called the fringing field)
rather than in the depletion region underneath the gate. As a result, it takes a
higher gate voltage to create the inversion layer beneath the gate.

Short-Channel Effects

We have previously discussed an example of a short-channel effect in Sec­

tion 3.5. Because the Early effect was studied before the recognition of short­
channel effects, it is often not recognised as such. Nonetheless, it is a short­
channel effect, because the drain voltage changes the current by modulating
the effective length of the transistor.
As the effective channel length decreases, the voltage drop across the pin­
choff region increases leading to an increasing field around the drain. Eventu­
ally, velocity saturation of the carriers occurs as described in Section 3.5. This
saturation leads to a decrease in the transistor current. As the field around the
drain increases further, hot-carrier effects occur: The carriers in the channel be­
come "hot" because they gain kinetic energy as they travel along the channel.
Hot-carrier effects include impact ionization, avalanche multiplication, hot­
electron injection, and punch-through. These effects are usually undesirable
but the hot-electron injection mechanism can be used for interesting pUlposes,
such as programming floating-gate memories.

Impact Ionization and Avalanche Multiplication In the presence of a high

longtitudinal or lateral field, the carriers collide with the lattice, breaking the Si

27 A channel is short if the sum of the depletion region widths around the source and drain
become comparable to the length. A channel is narrow when the width of the transistor becomes
comparable to the depth of the depletion region under the gate.
80 Chapter 3

bonds, and creating additional electron-hole pairs. This phenomenon is called

impact ionization in an nFET. The generated electrons are swept into the drain
and the generated holes are swept into the substrate creating a substrate current.
The drain current is now higher than the source current (I d "I- Is). This effect
prevents current sources from achieving a high output impedance. With even
higher electric fields, the impact ionization process turns into an avalanche
phenomenon: More and more electron-hole pairs are created even when the
field is kept constant.

Hot-Electron Injection The electrons in the channel under the high longti­
tudinal field can gain sufficient kinetic energy to surmount the barrier at the
interface of the silicon and the gate oxide and some carriers are injected into
the oxide. Most of these electrons will cross into the gate and the remainder
will be trapped in the oxide. The trapped charges alter the "threshold" of the
MOSFET. This phenomenon is usually undesirable because it leads to a effec­
tive threshold voltage change. But, as we will see in Chapter 4, this mechanism
can be used to our advantage: In the construction of a non-volatile memory.

Punch-through Punch-through is a form of transistor breakdown. This phe­

nomenon occurs when the depletion regions around the drain and source
merge. The gate becomes ineffective in controlling the current. The channel
length becomes effectively zero so a huge current flows even at low Vgs .

3.6 Noise and Transistor Matching

To fully characterise the operation of a circuit, the different noise sources that
contribute to the current in a transistor should be considered. This topic is
discussed in detail in Chapter 11. Transistor mismatch is also a major concern,
especially in analog circuit design. The mismatch can be reduced by using a
large layout area for the transistors and also by operating the transistors above
threshold. This means that good matching requires sacrifices in both chip area
and power consumption.
If small device geometries are used, the transistors are susceptible to spatial
variations from process-dependent parameters (see Section 12.2).
MOSFET Characteristics 81

3.7 Appendices

A. Small-Signal Model at Moderate Frequencies

,-------,--,,--0 G

Cgs Cgd

g mg6.vg
.-

g ms6.vs
S ---. >------>-- 0 Cgb

gmdt!. Vd
.-

Cbs

Figure 3.18
Small-signal equivalent model of a MOSFET operating at moderate frequencies.

The operation of the transistor is also detennined by the charges on its dif­
ferent parts. In addition to the currents from the conductance parameters, finite
capacitive currents arise due to voltage changes � V at any of its four ter­
minals. In computing these capacitive currents, we assume that the transistor
operates in the quasi-static condition: The change in the charge is proportional
to the change in the voltage. These capacitive currents (which become impor­
tant when the transistor operates at moderate frequencies) are incorporated into
the small-signal model in Fig. 3.14 by the addition of capacitors (as shown in
Fig. 3.18).
The MOSFET can divided into two parts; the intrinsic part, and the extrin­
sic part. The intrinsic part of a MOSFET is defined as the region between the
source and drain (the inversion layer and the depletion region), and the gate
oxide and the gate (see Fig. 3.19). The intrinsic part effects transistor action.
The undesirable parasitic elements constitute the extrinsic part of a MOSFET:
These elements include the drain and source resistances, the junction capac­
itances between the drain and source regions and the bulk, and the overlap
capacitances between the gate and the source/drain.
82 Chapter 3

The capacitive currents at the different nodes of a MOSFET become signif­

icant when the MOSFET is operated at moderate frequencies. We will consider
the contribution of only the intrinsic capacitances to these currents. The volt­
ages at the four terminals are referenced to a separate "ground". The voltages
at three terminals are kept constant while the voltage at the fourth terminal (for
example, Vs) is varied by � Vs. We consider how this voltage variation affects
the charge on the gate (Q 9)' the charge in the bulk (Q b), and the inversion
charge (Q i ). Here, the symbol, Q is the total charge rather than the charge per
unit area. Hence Q 9 is the total charge on the gate. The five trans-capacitors
that describe the change in the charges as a function of the terminal voltages are
shown in Fig. 3.182 8 • The capacitance symbol C defines the total capacitances
rather than the capacitance per unit area.
The gate-to-source capacitance, C9 S ' represents the change in the gate
charge, �Q 9' as a function of a change in Vs for a fixed V9 , Vb , and Vd :

(3.7.1)

The charge �Q 9 represents the change in the gate charge when Vs increases. A
positive increase in Vs means that the potential across the gate oxide decreases:
SO �Q 9 is negative, and C9 S is then positive (Eq. 3.7.1). Similar capacitances
can be associated with the gate for a given change in Vd and Vb respectively:

(3.7.2)

(3.7.3)

A positive � Vs also affects the depletion charge in the bulk. Because � Vs

is positive, the depletion region around the source increases, making the total
charge more negative (more acceptor ions are exposed in a p substrate). This
-

charge is balanced by an equivalent positive charge Q b leaving the substrate.

The capacitance Cbs represents the negative change in Q b resulting from the
increase in Vs :

(3.7.4)

28 These capacitances are not associated with a physical parallel-plate structure in the channel.
MOSFET Characteristics 83

Similarly the bulk-drain capacitance is:

(3.7.5)

A good discussion of these capacitances in both the subthreshold and above

threshold regimes of MOSFET operation can be found in Enz et al. (1995);
Tsividis (1998).

� T'-
,:
'- - -

- - - - - - r C9b
- - -'

,:
- - - - - -

B B B

(a) (b) (c)

Figure 3.19
Intrinsic capacitances of a nFET in weak inversion and strong inversion. (a) In weak inversion.
there is one dominant capacitance Cg b' When the FET operates above threshold. additional
capacitances appear both in the linear region (b) and in the saturation region (c).

Intrinsic Capacitances in Subthreshold

In subthreshold, the inversion charge Q i underneath the gate is negligible. Any

variation in the gate charge caused by a change in the terminal voltages affects
primarily the bulk charge. An explicit expression for four of the capacitances in
the subthreshold region can be found in Enz et al. (1995) and Tsividis (1998).
However, to a first-order approximation, we assume that these capacitances are
approximately zero. The gate-to-bulk capacitance, C g b , is the combination of
two Cox and Cd in series:

(3.7.6)
84 Chapter 3

Intrinsic Capacitances Above Threshold

Above threshold, the intrinsic capacitances depend on whether the MOSFET

is operating in the triode or the saturation regime.

Triode Regime Here, the inversion layer is approximately constant through­

out the channel. Any change in Vs or Vd affects the inversion charge, and hence
the gate charge. The total capacitance between the channel and the gate is given
by the oxide capacitance, C ox . By symmetry
Cox
Cgs = Cgd = 2 ' (3.7.7)

and similarly,

(3.7.8)

where Cb is the total capacitance of the reversed-bias depletion region formed

by the inversion layer and the bulk. C g b is equal to zero because the inversion
layer shields the bulk from the influence of the gate.

Saturation Regime In this case, the inversion charge disappears at the drain
end of the channel. Changing Vd has no effect on the intrinsic charges, so
Cbd = Cgd = O. However the source voltage affects the inversion layer, so
Cbs and Cgs are non-zero. The inversion layer decreases towards the drain end
of the channel, and the channel is depleted near the drain. Consequently, the
region of the channel near the drain is not affected by changes in Vs. From
detailed analysis in Tsividis (1998), Cgs = �Cox , and Cbs � �Cb. Because
of the pinchoff (depletion) region near the drain, C g b is equal to the series
combination of two capacitances:
1
Cgb = Ij3(Cox I I Cd) = 3 Cox ( 1 - ) K, . (3.7.9)

Unity-Gain Cutoff Frequency

The gate current of a MOSFET is close to zero at low frequencies, but at

higher frequencies, it becomes significant due to the current flow through the
capacitances (Cgs , Cg b, and Cgd) associated with the gate. The small-signal
equivalent model for the transistor in Fig. 3.20(a) is shown in Fig. 3.20(b). The
unity current gain cutoff frequency, fT, is the frequency at which the input
current (gate current) and output current (drain current) are equal. This cutoff
MOSFET Characteristics 85

G
B

E sin w t

(a)

(b)

Figure 3.20
Small·signal model of an nFET for unity-gain frequency calculation. (a) nFET operating in
saturation with a small-signal input voltage. (b) Small-signal equivalent circuit of (a).

frequency specifies the maximum operating frequency of the transistor.

iT is computed by supplying a small-signal input V i = € sin wt to the
gate29• The small-signal drain current, i d, is given by 9mVi and the gate current
is
(3.7.10)

We can ignore Cg b and Cgd because Cgb , Cgd < Cgs . Solving for IT with

29 The radian frequency. w is given in radians per second and is equal to 21T f where f is the
frequency in Hertz.
86 Chapter 3

=
Ih �' 1 (3.7.11)

For an above threshold transistor in saturation, the unity-gain frequency is

e V
fT '" 3JL ff (V9S 2- T) . (3.7.12)
4rr£
We see that fT is inversely proportional to the square of the channel length.
To obtain a high fT , we need a high carrier mobility and a short-channel
transistor3o•

B. The Enz-Krummenacher-Vittoz Transistor Model

There are a small number of models that describe the transistor's operation
continuously from subthreshold to above threshold. These models include
those of Maher and Mead (Mead, 1989; Maher, 1989) and others (Tsividis,
1998; Enz et al., 1995; Enz and Vittoz, 1997; Montoro et al., 1999). In Maher
and Mead's model, the current flow in a MOSFET is computed from the mo­
bile charge distribution in the channel as the terminal voltages are varied This
model is used in the SPICE (Simulation Program with Integrated Circuit Em­
phasis) simulation package from Tanner Tools. This circuit simulation program
can be used to perform many different types of analysis such as steady-state,
small-signal, time-domain, frequency, temperature, and noise analysis.
The Enz-Krummenacher-Vittoz (EKV) model (Enz et al., 1995) can also
be used with SPICE. This model provides a simple, closed-form expression
for the channel current of a MOS transistor in terms of the terminal voltages,
each of which is referenced to the transistor's bulk voltage. It is a continuous
model that is valid in all normal regimes of MOS transistor operation, that is,
the drain-bulk and the source-bulk junctions are reverse-biased. The model is
also symmetric with respect to the interchange of source and drain terminals.
The channel current is
(3.7.13)

where If and Ir are the forward and reverse components of the channel current
respectively. The forward component is a function only of the gate-to-bulk and
source-to-bulk voltages, whereas the reverse component is a function only of
30 We compute an approximate value for fT . When L = 0. 8J.!m ,Vg s=3V, VT =lV, J.!eff =

400cm2/ (V.s) , IT � 30 GHz.

MOSFET Characteristics 87

the gate-to-bulk and drain-to-bulk voltages. Under this fonnulation a MOS

transistor operates in saturation when If » Ir , which implies that I � If . In
this case, the channel current (neglecting the Early effect) is not a function of
the drain voltage. For a MOS transistor operating in the ohmic regime, I f � Ir.
In the EKV model with the nPET biased as shown in Fig. 3.4a, the forward

=L
component of the channel current is
W 2U. 2 f..t Cox lo 2 (1 )
=L
If g + e ( K ( V9 b - VTO ) - VS b ) / 2 UT
T 2�
W 2U. 2 f..t Cox lo 2
T g
2�
(1 + e ( K Vg + ( l - K) Vb - KVTo - Vs ) / 2 UT )
where VTO is the zero-bias threshold voltage (referenced to the bulk). The

= L T 2� ( 1
reverse component, Ir, is given by a similar expression
W 2U. 2 f..t Cox lo 2 )
= � �: (1
1r g + e ( K ( V9 b - VTO ) - Vdb ) / 2 UT

2 Uf
f..t x lo 2
g + e ( K Vg + ( l - K) Vb - KVTo - Vd ) / 2 UT ) •

The function log 2 ( 1 + e X / 2 ) 3 1 interpolates smoothly between an expo­

nential (subthreshold behavior) when x < 0 and a quadratic (above-threshold
behavior) when x > O. Thus, when Vg b < VTO + Vsb /�, the forward compo­
nent of the channel current becomes
W o e ( KVg + ( l - K) Vb - Vs ) /uT
If � I (3.7.14)
L

=
where 10 is the subthreshold pre-exponential current factor (Mead, 1989),
given by
10 2 Uf f..t Cox e - KVTO /UT /� .
+
On the other hand, when Vgb > VTO Vsb /�, the forward component of the
channel current becomes
W f..t Cox
L�
If � ( �Vg - ( 1 - � ) Vb - �VTO - Vs ) 2 . (3.7.15)

Likewise, the reverse component of the channel current becomes

W
1r "' L LO e (KVg + ( l - K) Vb - Vd ) /uT
.....,
(3.7.16)

31 This function was chosen for correct mathematical and qualitative properties and not for any
physical reason.
88 Chapter 3

when Vdb > '" (Vgb - VTO) , and it becomes

W f.L Cox
L � (",Vg - ( 1 - "' ) Vb - ",VTO - Vd)
Ir �
2 (3.7.17)

when Vdb < '" (Vgb - VTO) .

When Vg b < VTO + Vsb/'" and Vdb > ", (Vgb - VTO) , we can recover
the usual subthreshold MOS transistor model (Mead, 1989) by substituting
Eqs. 3.7.14 and 3.7.16 into Eq. 3.7.13 and factoring out the common part of
each tenn. This way, we obtain

I = � Io e (KVg + ( l - K) Vb ) /UT ( e - V. /UT - e - Vd / UT ) (3.7.18)

which covers both the the ohmic regime (when Vds < 4UT ) and the saturation
regime (when Vds > 4UT ) in subthreshold. To see that it does so, we write

I = � Io e ( KVg + ( l - K) Vb - V. ) /UT (1 _
e - Vd . /UT )

(3.7.19)

because if Vds > 4UT , then e - Vd . / UT < e - 4 « 1 and so 1 - e - Vd . /UT � 1,

thus removing the dependence of the channel current on the drain voltage.
When Vg b > VTO + Vsb / '" and Vdb < '" (Vgb - VTO) , the transis­
tor is operating in the above-threshold ohmic regime and by substituting
Eqs. 3.7.15 and 3.7.17 into Eq. 3.7.13 and factoring out the common part
of each tenn, we obtain an expression for the channel current:
w f.L oxC
L � ( ( "'Va + ( 1 - '" ) Vb - ",VTO - Vs ) -
I =
2

("'Va + ( 1 - "' ) Vb - ",VTO - Vd) ) .

2 (3.7.20)

If Vdb > '" (Vgb - VTO) , then the drain end of the channel will be only weakly
inverted and If » Ir • In this case, the transistor is operating in the above­
threshold saturation regime and the channel current is given by the first tenn
of Eq. 3.7.20:
w f.LCox
L � ( "'Vg + ( 1 - '" ) Vb - ",VTO - Vs ) .
I =
2 (3.7.21)

The (bulk-referenced) saturation voltage is thus

Vs at =
'" (Vg b - VTO) . (3.7.22)

Taken together, Eqs. 3.7.20 and 3.7.21 constitute an above-threshold model

MOSFET Characteristics 89

that differs from those found in most elementary circuit texts. This model has
the advantage that it exposes the source-drain symmetry of the MOS transistor.
It also makes clear how the above-threshold ohmic and saturation equations
are related to one another. Moreover, this model accounts directly for the body
effect (to first order) via the ", parameter without any auxiliary equations and
without adding too much complexity to the model. For a pFET, biased as

= WL 2 T ""2C",ox (1
shown in Fig. 3.4b, the forward component of the channel current is
U. 2 )
= WL 2 T ""2C",ox (1
l og 2 + e ( I« Vbg - I VTo l ) - VbS ) / 2 UT
If
U. 2 l og 2 + e ( Vs - I<Vg - ( 1 - I<) Vb - I< I VTo l ) / 2 UT )
where all parameters are defined as for the nFET, except that "" is the mobility

= WL 2 T ""2C",ox (1
of holes. Likewise, the reverse component is
1 U. 2 )
= � 2Uj. ""�:x (1
l og2 + e ( I« Vbg - I VTo l ) - Vbd ) / 2 UT
r

l og2 + e ( Vd - I<Vg + ( l - l<) Vb - I< I VTO 1 ) / 2 UT ) •

C. Bipolar Junction Transistor

Traditionally, microcircuits were implemented with bipolar junction transistors

(BITs); but in the past 30 years MOSFETs have replaced BITs in microcircuits
except for high-speed applications or in analog applications that require low
1/f noise or precise transistor matching. Bipolar CMOS (BiCMOS) processes
are available where both bipolar and MOS transistors are fabricated on the
same substrate. These technologies are attractive for the implementation of
mixed analog-digital designs. There are two types of BITs: The npn, and the
pnp transistors, whose symbols are shown in Fig. 3.21. We give a cursory
treatment of the basic operation of an npn bipolar transistor. As we will see in
Chapter 10, the bipolar transistor can also be used as a photo-sensing element
(phototransistor).

Basic Operation

The structure of an npn bipolar transistor is shown in Fig 3.22. It consists of

two back to back pn junctions. The emitter and collector terminals are both
90 Chapter 3

c E

B B

E c

(a) (b)
Figure 3.21
Bipolar transistor symbol. (a) An npn bipolar transistor. (b) An pnp bipolar transistor.

n-type regions but the emitter is doped higher than the collector. The base
is doped p type. In the normal operating mode, the emitter-base junction is
forward-biased and the collector-base junction is reverse-biased. If the base­
emitter junction is reverse-biased (Vbe <0.5 V) the transistor is in cutoff.
When the base-emitter voltage Vbe is greater than approximately 0.7 V, current
begins to flow from the base to the emitter. The current flow consists of
holes moving from the p base to the emitter and electrons moving from the
n+ emitter to the base. Because the emitter is doped higher than the base,
more electrons are injected from the emitter into the base. The collector-base
junction is reversed biased, and so the holes from the base are not attracted
to the collector. The base width is narrow, so that electrons injected from
the emitter will diffuse easily into the depletion region at the collector-base
junction. Once they are in this region, they are swept into the collector. Most
of the electrons from the emitter reach the collector, consequently, the collector
current Ie is approximately equal to the emitter current Ie . However some
of the injected electrons recombine with the holes in the base. This loss of
electrons is equivalent to the base current. For an almost ideal BJT, we require
the base current h to be much smaller than Ie . The common-emitter current
gain, (3, of the device is defined as

(3.7.23)
MOSFET Characteristics 91

The value o f (3 varies between 50 and 100. (3 depends o n the base width and
the lifetime of the carriers in the base. It also depends on the magnitude of the
collector current
(3.7.24)

where Ics is the pre-exponential constant.

An in-depth discussion of the operation of BJTs can be found in many
texts (Grove, 1 967; Gray et aI., 2001). The bipolar transistor has different
characteristics to the MOSFET. Because of its finite base current and its
higher output current drive, it has a lower input impedance than the MOSFET.
However, the BJT and the subthreshold MOSFET share similar exponential
current-voltage characteristics.

·
Ve o-�'!e� : : I :iil:f::f··::·:•:-·.:·.:.·· ::1 . . . · . . '
c
...
- -o
I-.... V

· · 22G]
:� c

:-:·-: :·-: :-::m

I U l :IiLmE .:
. . . .

I1ill1t :
.

Figure 3.22
Current components in an opo bipolar junction transistor.
4 Floating-Gate MOSFETs

Our goal in this chapter is to examine the physics of floating-gate devices, and
develop the technology of analog memory transistors.

4.1 Floating-Gate MOSFETs

The semiconductor industry has developed a variety of nonvolatile solid-state

memory devices, including the EPROM (electrically programmable read only
memory), EEPROM (electrically erasable, programmable, read-only mem­
ory), and flash EEPROM. A common feature of these memories is that they
use afloating-gate MOSFET (FGMOSFET) whose gate is isolated because it
is surrounded on all sides by an insulator (typically Si02). The basic organiza­
tion of such a floating gate is shown in Fig. 4.1(a).
Electrons on the floating gate are prevented from escaping by the surround­
ing insulator. Voltage inputs to a secondary control gate couple capacitively to
the floating gate, thereby modulating the transistor's channel current. From the
perspective of this control gate, the presence (or absence) of excess charge on
the floating gate causes a voltage offset and so an apparent shift in the MOS­
FET's transfer function (Fig. 4.1(b)).
The earliest experiments exploring charge storage on the insulated gate
of a MOSFET were performed in the 1960s (Kahng and Sze, 1967; Kahng,
1967; Frohman-Bentchkowsky, 1971). At that time, the goal was to find a
replacement for magnetic (core) memories and so work focused on the use of
floating gates for binary storage. Digital EEPROM chips achieved commercial
success roughly a decade later. In the 1980s, Alspector and Allen suggested
that the memory function in a neural network could be realized by using analog
charge on a floating gate (Alspector and Allen, 1987; Yang et al., 1992; Fujita
and Amemiya, 1993). Unfortunately, changing the charge on a floating gate
involves the difficult task of transporting electrons through the SiO 2 insulator.
It has taken yet another decade to achieve a measure of success in solving this
problem. Even now, analog floating-gate devices are confined primarily to the
research laboratory.
Neuromorphic engineers have adapted the floating-gate technology used
in digital EEPROMs both to allow nonvolatile analog storage, and to perform
a local learning function. In particular, a family of single-transistor floating­
gate devices called synapse transistors have been developed (Diorio, 1997;
94 Chapter 4

(Charge 019)
Source Gate Floating gate
oxide
(Si02)
Substrate

(a)

10-5

10--{)
�
C 10-7
�
� 10-8
Q)
u
� 10-9
CfJ
10-10

1 2 3 4 5
Control-gate-to-source voltage (V)

(b)

Figure 4.1
An n-channel floating-gate MOSFET (a) and its associated I-V transfer function (b). Voltage inputs

gate's perspective, changing the floating-gate charge Qf9 shifts the transistor's threshold voltage
to the poly2 control gate are coupled capacitively to the polyl floating gate. From the control

(bidirectionally).

Hasler, 1997; Diorio et aI., 1996, 1997c, 1998b,a, 1997a). These devices, like
neural synapses, implement long-term nonvolatile analog memory; allow bidi­
rectional memory updates; and learn from an input signal without interrupting
the ongoing computation.
Most electrically programmable solid-state memory technologies use two
basic mechanisms to modify the floating-gate charge 1 (Kerns et aI., 1991).

1 Other floating gate devices, such as programmable read-only memories (PROMs), uses UV light
Floating-Gate MOSFETs 95

Electron tunneling

Oxide barrier
(Si02)
Electron
tunneling

3.1
Meta\
Horizontal position

(a)

Electron injection

3.1

Horizontal position

(b)

Figure 4.2
Overcoming an SiO:! barrier. An electron can either (a) tunnel through the barrier, provided the
barrier is thin enough; or it can (b) acquire enough energy to inject over the barrier.

These two mechanisms are illustrated in the energy-band diagrams of Fig. 4.2.
Stated simply, moving electrons through Si02 requires overcoming the dif­
ference in electron affinities between a metal2 and the Si02. We can either

to read and/or write the floating-gate memory (Kerns et aI., 1991). In this case, UV photons excite
electrons to energy levels high enough to overcome the difference in electron affinities between
the metal and the SiO:!.
2 The "metal" can be an actual metal, a polysilicon gate, or a degenerately doped silicon implant.
96 Chapter 4

push the electrons through the potential barrier; or force them over the top of
the barrier. These two processes are called electron tunneling and hot-electron
injection, respectively (Takeda et al., 1995). Electron tunneling is a quantum­
mechanical process; we restrict our study to a particular form, called Fowler­
Nordheim (FN) tunneling (Lenzlinger and Snow, 1969), in which an electric
field is applied across the Si02 to facilitate tunneling. Hot-electron injection
comes in many flavors that differ in the mechanism by which the electrons are
made "hot". Synapse transistors use both tunneling and injection to alter their
floating-gate charge.

Electron Tunneling

The FN-tunneling process is illustrated in Fig. 4.3(a). In practice, the barrier

is Si02 gate oxide; the "metal" on one side of the barrier is the polysilicon
floating gate; and the "metal" on the other side is a heavily doped region of the
silicon surface called a tunneling implant. We tunnel electrons from the floating
gate to the tunneling implant. A potential difference between the floating gate
and the implant reduces the effective thickness of the SiO 2 barrier, facilitating
electron tunneling from the floating gate, through the barrier, into the oxide
conduction band. These electrons are then swept over to the implant by the
field across the barrier.
In general, we can tunnel electrons both onto and off the floating gate;
and we can tunnel them through any oxide (gate oxide, inter-poly oxide, etc.).
In practice we avoid tunneling electrons onto the floating gate, because to do
so we must either apply a large negative voltage to the tunneling implant, or
pull the floating gate to a high positive voltage. The former is unattractive
because most commercial processes are single-tub n-well. The latter is also
unattractive because of the large change in channel current during tunneling.
We also avoid tunneling electrons through the inter-poly oxide, because the
inferior oxide quality (compared with gate oxide) causes poor matching and
unreliable behavior.
Figure 4.3(b) shows the measured FN-tunneling current versus the recipro­
cal of the voltage across the oxide, for a 400 A gate oxide fabricated in a 2 J.Lm
CMOS process. We can approximate the data by

I 19 = It o e - ;;; I (4.1.1)

where 19 is the gate current, Vox is the oxide voltage, VI is a constant that
depends on the oxide thickness, and It o is a pre-exponential current. Equa-
Floating-Gate MOSFETs 97

FN-tunneling

3. t V
fi
'--'ll>Oe- / tunneling
Electron
"=_0 ,....=
____

.Q>
0:::- Oxide
::> '"
'O.�
o Q)
u15
c: c I
Metal
( 8 i0 2 )

I
��
w
�
2
70
L-______________________�.�
Horizontal position

(a)

Tunneling Data
3
E 10-1 o

:::l
10-14
::s
£
-15
g> 10
�
16
� 10-
"
10-17 Vf
��
5 10-18 Ig 8.18 x 10-2 � Vox
u
=

10-19
�
where Vf = 984V
'x
o 1O- 2 0':-:-: -'-__::-'::--;--'-__--::-C:::-:::---'-
:-- ---:c7;: -'---:�
-0.044 -0.04 -0.036 -0.032:::- -0.028
-1 lox ide voltage (1/v)

(b)

Figure 4.3
Fowler-Nordheim (FN) tunneling. (a) By applying a voltage across the oxide, the band diagram
is altered, thereby facilitating electron tunneling through the "thinned" barrier into the oxide
conduction band. (b) A plot of tunneling current versus reciprocal oxide voltage, measured from
the tunneling junction shown in Fig. 4.4 (a). Vox is the voltage across the gate oxide; Given an
oxide thickness of 4ooA, Vf=984 V is consistent with a survey (Mead, 1994) of Fowler-Nordheim
tunneling in Si�. We plot the data as oxide current divided by the gate-to-n+ edge length (in
lineal microns) of the tunneling implant because the floating gate induces a depletion region in
the lightly doped n- well, reducing the effective oxide voltage and with it the tunneling current.
The gate cannot appreciably deplete the n+ well contact, so the oxide field is higher where the
self-aligned floating gate overlaps the n +. Because tunneling increases exponentially with oxide
voltage, tunneling in analog memory transistors is primarily an edge phenomenon.
98 Chapter 4

tion 4.1.1 is a sufficient (for our pUlposes) approximation to a well-known

FN-tunneling equation (see, for example, (Sze, 1981)).

Electron Injection

We defer illustrating the hot-electron injection process (Sanchez and DeMassa,

1991) until we have described the nFET and pFET synapses, because the
injection process differs slightly between these two devices. We merely note at
this time that we accelerate electrons to high energies in the drain-to-channel
electric field of a transistor; that a fraction of these electrons scatter upward
into the gate oxide; and that a fraction of these scattered electrons inject into
the oxide's conduction band. The injected electrons can then be swept over to
the floating gate by an electric field across the gate oxide.

4.2 Synapse Transistors

The programming characteristics of floating-gate devices are quite variable,

even for nominally identical transistors on the same chip (Sin et al., 1992).
For digital memories, these variations can be compensated by applying ex­
cess charge during the write or erase processes. For analog memories, these
device variations require feedback to ensure accurate memory writes. Various
feedback mechanisms have been tried: Most employ either multi-step, iterative
writes (Sin et al., 1992); or single-step, open-loop writes with frequent oxide
calibration to compensate mismatch and degradation (Holler et aI., 1989). Nei­
ther approach permits the locally computed, parallel weight updates that are
needed for silicon learning arrays.
Synapse transistors are new types of floating-gate devices that permit si­
multaneous reading and writing of memory. The channel and oxide currents
are simultaneous, analog, and continuous. Consequently, we can use continu­
ous analog feedback to write the charge on the floating gate, and to implement
either a weight-update rule or a weight constraint in a silicon learning system.
Four-terminal nFET and pFET synapse transistors are shown in Fig. 4.4.
Both comprise a single MOSFET (with polyl floating gate and poly2 con­
trol gate) and an associated n-well tunneling implant. Both use hot-electron
injection to add electrons to their floating gates, and FN tunneling to remove
the electrons. The nFET synapse differs from a conventional n-type MOS­
FET by its use of a moderately-doped channel implant, which facilitates hot­
electron injection (Diorio et al., 1997a). The pFET synapse achieves a hot-
Floating-Gate MOSFETs 99

electron gate-current (Chung et al., 1997) using a conventional p-type MOS­

FET: No special channel implant is required. Various researchers have fab­
ricated nFET synapses in 2 JLm and 1.2 JLm n-well double-poly processes,
and pFET synapses in several double-poly processes ranging from 2 JLm to
0.35 JLm.
Both the nFET and pFET synapses store a weight as charge on their float­
ing gate. We can choose source current, drain current, or channel conductance
as the output. Typically we choose source current, although this choice is arbi­
trary because these quantities are related. Signal inputs are applied to the poly2
control gate, which in tum couples capacitively to the poly1 floating gate. Typ­
ically the synapses are operated in their subthreshold regime (Mead, 1989),
for three reasons. Firstly, subthreshold channel currents ensure low power con­
sumption; typically less than l OO nW per device. Secondly, because the source
current in a subthreshold MOSFET is an exponential function of the gate volt­
age, small changes to the floating-gate charge shift the transistor's operating
point measurably. Thirdly, the synapse output is the product of a stored weight
and the applied input:

(4.2.1)

(4.2.2)

where Isis the source current, lois the pre-exponential current, K, is the cou­
pling coefficient from the floating gate to the channel, Q /g is the floating-gate
charge, CT is the total capacitance seen by the floating gate, Ut is the thermal
voltage kT / q, Cin is the input (polyl to poly2) coupling capacitance, Vin is
the control-gate voltage; QT == CT Ut i K" K,' == K, Cin / CT, W==exp(Q/gIQT);
and, for simplicity, we assume the source potential to be ground (Vs=O). Equa­
tion 4.2.1 and Eq. 4.2.2 imply an n-channel MOSFET; if we change the sign
of all the variables, the equations describe a p-channel MOSFET.
The weight W is a learned quantity. Its value derives from the floating-gate
charge, which can change with synapse use. The floating-gate transistor output
is the product of W and the source current of an idealized MOSFET, which has
a control-gate input Vin , and a coupling coefficient K, ' from the control gate to
the channel.
Both the nFET and pFET synapses achieve bidirectional weight updates by
using FN-tunneling to remove electrons from the floating gate, and by using
hot-electron injection to add electrons to the floating gate. The magnitudes
100 Chapter 4

nFET synapse pFET synapse

(b) Side view (b) Side view

Inlerpoly Field-oxide Inlerpoly Field-oxide Interpoly
capacitor channel SlOP

��
OJ
.. Electron
Gate Electron Gate
p- substrate
oxide injection oxide tunneling

�I\� lK�)(
(c) Electron band diagram (c) Electron band diagram

�
Electron

+;
/� Jectlon

I
I0
s ource Channel e Drain

I, II
_ '
"
•
,

T' J
+v , Etectron/' lmpact lonlza!lOn
:>
_ ____ _ _ _ __

@
�
\
injection

,
"0 '-

31 v l
3 v
8 _3 3 1 $102
i 1 :)Y nnellng
+=c:,.
� !69
Drain
�
6
3 v Electron

1
3 V
g {ij
�E 0
�
lt
Floating gale e- Electron ��� �
Ec ___
I
_tunneling
80 Ey

1 implant
I�
W
SIO,
barner
1� T Tunneling
Implant
c

�
iIi
Source Channel
3.2V
l
Tun ng

330 L-________________�==��.. -------------------P"'

Si:C
tion7 (� m';J
. "----
330 = =
Position ().1m) OCC l�

Figure 4.4
nFET and pFET synapse transistors, showing the electron tunneling and injection locations.
The diagrams are aligned vertically; (a) and (c) are drawn to scale; the vertical scale in (b) is
exaggerated and all voltages in the conduction-band diagram are referenced to the source potential.
The transistors operate in subthreshold (Is < 100 nA). The oxide band diagrams, and the trajectory
of (scattered) injection electrons, both project vertically. To better illustrate the injection process,
we overlook the scattering and rotate the oxide band diagrams by 90', drawing them in the channel
direction (horizontally). The tunneling process in the pFET synapse is identical to that in the nFET.
The injection process is different in the two devices, as we describe in the text.

of the weight updates depend on the transistor's terminal voltages and source
current. Consequently, a synapse's future weight W varies with the terminal
voltages, which are imposed on the device; and with the source current, which
is the output. As a result, synapse transistors learn: Their future weight value
depends on both the applied input and the present weight value.
Synapse transistors retain all the attributes of conventional transistors, and
in addition, they have long-term nonvolatile analog memory; bidirectional
memory updates; they compute the product of their stored memory and the
Floating-Gate MOSFETs 10 1

applied input; and they learn from an input signal without interrupting the on­
going computation. Because synapse transistors permit both local computation
and local weight updates, they can be used to build autonomous learning ar­
rays in which both the system outputs, and the memory updates, are computed
locally and in parallel.

The nFET Synapse

The top and side views of the nFET synapse are shown in Fig. 4.4. Tunneling
increases the synapse's weight W. A high voltage applied to the n + well contact
causes electrons to tunnel off the floating gate. The n + is surrounded by a
lightly doped n- well to prevent reverse-bias pn-junction breakdown (from
n+ to substrate) during tunneling. The breakdown voltage of n + to substrate
is typically about one-quarter that of the breakdown of n - to substrate. For
example, in a typical 0.35 f..tm process, n+ breaks down at about 6V, whereas
n- breaks down at about 25V.
The weight W is decreased by injecting channel electrons onto the floating
gate. Hot-electron injection is well known in conventional MOSFETs (Sanchez
and DeMassa, 1991). It occurs in short-channel devices with continuous chan­
nel currents, when a high gate voltage is combined with a large potential drop
across the short channel. It also occurs in switching transistors, when both the
drain and gate voltages are transiently high. In neither case is the injection suit­
able for use in an analog learning system: The short-channel injection requires
large channel currents, consuming too much power; and the switching-induced
injection is poorly controlled, and transient. Instead, nFET synapse transistors
use the drain-to-channel electric field in a subthreshold MOSFET to accelerate
channel electrons to high energies, and a fraction of these electrons are injected
into the oxide conduction band. The process is shown in the energy-band dia­
gram of Fig. 4.4.
Channel electrons, accelerated in the nFET's drain-to-channel depletion
region, lose energy by colliding with the semiconductor lattice. A fraction
of these electrons scatter upwards toward the gate oxide; and a fraction of
these possess sufficient energy to overcome the 3.1 eV difference in electron
affinities between the Si and Si02 conduction bands and enter the Si02• These
electrons are then swept over to the floating gate by the oxide electric field. For
electrons to be collected at the floating gate, the following two conditions must
be satisfied: The electrons must possess the 3.1 eV required to overcome the
difference in electron affinities; and the oxide electric field must be oriented in
the direction required to transport the injected electrons to the floating gate.
102 Chapter 4

In a conventional n-type MOSFET we can easily satisfy the first require­

ment. We merely operate the transistor in its subthreshold regime, with a drain­
to-source voltage greater than about 3 V. Because the subthreshold channel­
conduction band is flat, the drain-to-channel transition is steep, meaning that
the drain-to-channel electric field is large. Channel electrons accelerate rapidly
in this field, and a fraction of them acquire 3.1 eV of energy. A fraction of
these 3.1 eV electrons scatter upward into the gate oxide. It is largely the oxide
field orientation requirement that prevents a gate current in a conventional sub­
threshold nFET. Subthreshold operation usually implies gate-to-source volt­
ages less than 0.8 V (assuming a 2 /Lm process). With the drain at 3 V, and the
gate at 0.8 V, the drain-to-gate electric field opposes transport of the injected
electrons to the floating gate and so the electrons return to the drain instead.
In the nFET synapse transistor, the transport of injected electrons to the
floating gate is promoted by adding a bulk: p-type implant ('" 10 17/cm3) to the
channel region. Here, an NPN bipolar-transistor base implant is used. This
implant serves two purposes: Firstly, it increases the drain-to-channel electric
field, thereby increasing the hot-electron population in the drain-to-channel
depletion region; and secondly, it raises the transistor's threshold voltage VT
from 0.8 V to 6 V, thereby allowing the MOSFET to operate with both high
floating-gate voltages and subthreshold source currents. Consequently, for typ­
ical floating-gate and drain voltages of about 5.5 V and 3 V respectively, the
channel current is still subthreshold, but the drain-to-gate electric field trans­
ports injected electrons over to the floating gate.
From the perspective of the control gate, raising the MOSFET's threshold
voltage is inconsequential, because the control and floating gates are isolated
capacitively. Therefore, we can use control-gate inputs in the conventional 0 V
to 5 V or 0 V to 3.3 V ranges, regardless of the requirement that the floating
gate be near 6 V.
Figure 4.5(a) shows injection efficiency (gate current I9 divided by source
current Is) for a 2 /Lm nFET synapse. The data are plotted as efficiency,
because the gate current increases linearly with source current over the entire
subthreshold range (see Fig. 4.6(a». The drain is referenced to the channel
potential because the hot-electron population derives from the drain-to-channel
electric field. When Vdc is less than 2 V, the hot-electron gate current is
exceedingly small, and the weight W remains nonvolatile. When Vdc exceeds
2.5 V, the hot-electron gate current causes measurable changes in the synapse
weight W. Neuromorphic systems typically use slow adaptation (small oxide
currents): Vdc for the nFET synapse is typically less than 3 V, and is always
Floating-Gate MOSFETs 103

nFET versus pFET injection

10-3
C 10-4
t':'
:; 10-5
u
OJ
� 10-6
:0
0
7
:;:, 10-
(f)

t':' 10-8
c:

:;
u
10-9
*
CJ 10-10

10-11
2 3
Drain-to-channel voltage (V)

(a)

pFET injection versus process

Channel-to-drain voltage (V)

(b)

Figure 4.5
(a) Hot-electron injection efficiency (gate current divided by source current) versus drain-to­
channel voltage, for both nFET (2 tLm process) and pFET (0.8 tLm process) synapses. The drain
voltage is referenced to the channel potential because the hot-electron injection probability varies
with the drain-to-channel electric field. The drain voltage can be re-referenced to the source voltage
using the relationship between source and channel potential in a subthreshold MOSFET (Enz et al.,
1995; Andreou and Boahen, 1994). For the purposes of deriving a synapse weight-update rule, we
fit the injection data empirically, using the simple exponential as shown. (b) Hot-electron injection
efficiency for pFET synapses fabricated in 2.0 tLm, 0.8 tLm, and 0.35 tLm processes. The injection
probability increases with decreasing process linewidth, due to higher drain-to-channel electric
fields (due to increased implant-impurity concentrations) and thinner gate oxides.
104 Chapter 4

less than 3.5 V. Consequently, we can approximate the data of Fig. 4.5 using a
simple exponential:

(4.2.3)

where Ig is the gate current, I8 is the source current, Vdc is the drain-to-channel
potential; and (3, Vinj are fit constants. Because of the nFET synapse's 6 V
threshold, the floating-gate voltage almost always exceeds 5 V. If Vdc <3.5 V,
then the drain-to-gate oxide electric field strongly favors the transport of in­
jected electrons to the floating gate. Consequently, we can safely omit gate­
voltage dependencies from Eq. 4.2.3.
The injection process that we have described is known in the literature
as channel hot-electron injection (CHEI). For more details of the injection
physics, see Hasler et a1. (1998).

The pFET Synapse

The top and side views of a pFET synapse are shown in Fig. 4.4. The tunnel­
ing implant is identical to that of the nFET synapse, and as in the nFET, the
electrons are removed from the floating gate when high voltages are applied to
the n+ well contact. However, because the pFET and nFET synapses are com­
plementary, tunneling has the opposite effect on a pFET synapse: It decreases,
rather than increases W.
The charge carriers in a pFET are holes, which cannot be injected onto a
floating gate. Furthermore electrons, not holes, must be added to the floating
gate to increase the pFET synapse's weight. So, to increase W, electrons are
first generated and then injected. Electrons are generated by a process called
impact ionization: Channel holes, accelerated in the transistor's channel-to­
drain depletion region, lose energy by colliding with the semiconductor lattice.
If the channel-to-drain electric field is large, as is the case for a subthreshold
pFET with large source-to-drain voltage, then a fraction of these holes collide
with sufficient energy to generate free electron-hole pairs. The silicon physics
then naturally provides electron injection: The liberated electrons, promoted
to their conduction band by the collision, are expelled rapidly from the drain
region by the same channel-to-drain electric field. These electrons can, if
scattered upward into the gate oxide, inject onto the floating gate. This impact­
ionized hot-electron injection (lHEI) process is illustrated in the energy band
diagram in Fig. 4.4.
Floating-Gate MOSFETs 105

nFET injection
10-13

? 1O-14

Ig = 7.0t x to-7 Is

10-10 10-9 10-8 10-7

Source current (A)

(a)

pFET injection

10-14

?1O-16
C
�
:;
� 10-15
iii Ig 2.32 x 10-7 Is
�
=

10-17

1O-18 ,---:-: -��uL -::-�

-,- �...w--::-
���...t...,:-,-��....J
,-'
10-11 10-10 10-9 10-8 10-7
Source current (A)

(b)

Figure 4.6
(a) Four-terminal nFET-synapse gate current 19 versus source current Is. (b) Four-terminal pFET­
synapse gate current 19 versus source current Is. For both devices. the gate current is linearly
proportional to the source current over the entire subthreshold range.

To collect electrons at the floating gate of a pFET synapse, the same two
conditions must be satisfied as for the nFET synapse: The electrons must
possess the 3.1 eV needed to overcome the difference in electron affinities
between the Si and Si02; and the oxide electric field must be oriented in
the direction required to transport the injected electrons to the floating gate.
106 Chapter 4

The first requirement is satisfied by using a sufficiently large source-to-drain

voltage in a subthreshold MOSFET. The silicon physics naturally satisfies
the second requirement: In a subthreshold pFET, the source-to-gate voltage
is typically less than 1 V, whereas the source-to-drain voltage is at least several
volts. Consequently, the floating gate is always a few volts higher than the
drain, and so the oxide electric field naturally transports the injected electrons
to the floating gate. Unlike conventional nFET transistors, conventional pFET
transistors do not need special channel implants to facilitate injection.
As shown in Fig. 4.5, a pFET synapse in a 2 JLm process achieves an IHEI
efficiency of about 10- 10 when operated with a drain-to-channel voltage of
about 6.5 V. An nFET synapse in a 2 JLm process achieves the same efficiency
at a drain-to-channel voltage of about 2.5 V. The pFET synapse's higher drain­
voltage requirement arises for three reasons: Firstly, for any given drain-to­
channel voltage, the nFET experiences a higher drain-to-channel electric field
than does the pFET. This higher field is a consequence of the bulk implant
that we add to the nFET synapse's channel region. Secondly, phonon loss
mechanisms in silicon are naturally more efficient for holes (the pFET charge
carriers) than they are for electrons (the nFET charge carriers). Consequently,
a pFET requires a larger voltage across its depletion region to achieve the
same hot-carrier population. And thirdly, a pFET synapse's two-step injection
process requires more energy, because of the loss associated with the impact
ionization.
A pFET synapse's gate current increases linearly with its source current
(see Fig. 4.6) because the gate current derives from the impact-ionized elec­
tron population, and this population increases linearly with the source current
(assuming subthreshold operation). Because the pFET synapse's floating-gate
voltage is high, the dependency of the gate current on the gate-to-channel po­
tential is small. Consequently, we ignore gate-to-channel dependencies in our
(approximate) fit. We use the same fit for the pFET synapse as we did for the
nFET:

(4.2.4)

The IHEI efficiency is plotted in Fig. 4.5(b) for p-channel synapse transistors
fabricated in 2 JLm, 0.8 JLm, and 0.35 JLm CMOS processes. These data show
clearly that the results for the 2 JLm process scale directly to more modem
processes3•

3 Three-teoninal silicon synapses have also been fabricated (Diorio, 2000). The nFET 3-terminal
Floating-Gate MOSFETs 107

The Gate-Current Equation

Because the tunneling and hot-electron gate currents flow in opposite direc­
tions, the final gate-current equation for both the nFET and pFET synapses is
obtained by subtracting Eq. 4.2.3 from Eq. 4.1.1:

(4.2.5)

This equation describes the gate current for both types of synapse, over the
entire drain-voltage and subthreshold channel-current ranges.
The gate current Ig changes a synapse's W (bidirectionally) by modifying
the floating-gate charge Q/g. We use Eq. 4.2.5 for both the nFET and pFET
synapses, although the sign of the weight updates is different in the two cases.
In the nFET, tunneling increases the weight W, whereas injection decreases it.
In the pFET, tunneling decreases the weight W, whereas injection increases it.

4.3 Silicon Learning Arrays

A synaptic array can form the basis for a silicon learning system, such as the
generic learning array shown in Fig. 4.7. Diorio et al. (1997b) have imple­
mented a 4 by 4 silicon array with this general structure, using a single nFET
synapse for each "synapse" in the array. Unlike traditional neural-network sim­
ulations that use continuous valued inputs, we use pulse inputs. The array com­
putes the inner product of the pulsed input vector and the stored analog weight
matrix. The synaptic weights are nonvolatile. Column input pulses that are co­
incident with row learn-enable pulses cause weight increases at synapses that
they address. Unbounded weight values are bounded by a constraint: The time­
averaged sum of the synaptic weights in each row of the array, is held constant.
This constraint forces row synapses to compete for floating-gate charge and so
stabilizes the learning.
Before considering the learning array, we derive a weight-update rule for
an nFET synapse. The weight-update rule for a pFET synapse is identical in
form to that for the nFET, except for a sign inversion due to the opposite effects
that tunneling and injection have on a pFET synapse.

device incorporates the tunneling function into the transistor's drain. The pFET 3-terminal device
incorporates the tunneling function into the well contact. Because 3-terminal synapses are not
fundamentally different from the 4-terminal synapses described here, and because they are more
difficult to use in practice, we will not consider them further.
108 Chapter 4

--+-----.----f---_---{ Error signal

--+----+-.---f-----hr--{ Learn-enable signal

)-----� Output = L W2!i

J
-+-----.----)---_---< Error signal
--+----+-.---f-----hr--{ Learn-enable signal

)-----� Output = L W,!i

J

X, Input vector X X2

Figure 4.7
A learning-array block diagram. Each synapse multiplies its column input with its nonvolatile
analog weight, and outputs a current to the row-output wire, which sums the synapse-output
currents along that row. Column inputs that are coincident with the row learn-enable signals
cause weight increases at selected synapses. The error signal constrains the time-averaged sum
of the row-synapse weights to be a constant, bounding the row weights by forcing the synapses to
compete for weight value.

Synaptic Weight Updates

Weight updates depend on the tunneling and injection oxide currents that
alter the floating-gate charge. Figure 4.8 shows the temporal derivative of
the source current versus the source current, for an nFET synapse with a
set of fixed tunneling voltages (Fig. 4.8(a)), and a set of fixed drain voltages
(Fig. 4.8(b)). In these experiments, the control-gate input Vin was held fixed.
Consequently, these data show the synaptic weight updates oW/at, as can be
seen by differentiating Eq. 4.2.2.
In Appendix 4.4 we show that the tunneling-induced weight increments
follow a power law:

oW � _1_ W (l-u) (4.3.1)

at Ttun

where (J and Ttun are defined in Eq. 4.4.5 and Eq. 4.4.6, respectively. In
Appendix 4.4, we show that the CHEI-induced weight decrements also follow
Floating-Gate MOSFETs 109

a power law:
aw = _ __1 w ( 2 e)
-
(4.3.2)
at Tinj

where e and Tinj are defined in Eq. 4.4.15 and Eq. 4.4.16, respectively.
Si02 trapping is a well-known issue in floating-gate transistor reliabil­
ity (Aritome et aI., 1993). In the synapse, oxide trapping decreases the weight­
update rates. Fortunately, our synapses require only small quantities of charge
for their weight updates and we can usually ignore oxide trapping.

The Learning Array

Figure 4.9 shows one row of the learning array which comprises a synapse
transistor at each node, and a normalization circuit at the row boundary. The
column inputs Xi and the row learn-enable signals Yj are digital pulses. Each
synapse multiplies its binary-valued input Xi with its stored weight Wij , and
outputs a source current Isij whose magnitude is given by Eq. 4.2.2. The total
row current lout is the sum of the source currents from all the synapses in
the row. Synapses are ordinarily on; low-true gate inputs Xi tum off selected
synapses, decreasing the current lout transiently. This decrease in lout in
response to an input vector X, is the row computation.
The row learn-enable inputs Yj are high-voltage inputs that can increase
the synapse weights when Vox == ytun - V, g (where ytun =
}j +25 and
V, g is the floating-gate voltage) is large enough to induce tunneling. Synapse­
weight increases occur only when both the row and column inputs, Yj and Xi ,
are true. To see why, first consider the case when the row learn-enable signal
}j is false (ytun is low). Because Vox =
ytun - V, g, Vox is small for every
synapse in the row when Vtun is low. In this case, the tunneling currents are
small, and there is no weight increase at any row synapse.
Now consider the case when Yj is true (ytun is high). Vox increases as V'g
decreases, and V,9 follows Xi . Therefore, if a low-true column input Xi is true,
then V, 9 is low; Vox is large; and electron tunneling causes a weight increase
at the selected synapse. If, on the other hand, the low-true column input X i is
false, then V,9 is high; Vox is too small to cause appreciable tunneling; and
there is little change in the synaptic weight.
Tunneling increases the weight value of a row-column selected synapse.
Because this weight update is single quadrant (tunneling only increases a
synapse's weight), tunneling allows unbounded weight values. To constrain
the array weights, we renormalize each row of the array. The array allows
1 10 Chapter 4

Electron tunneling
1 0-9
als oc aw � __
1 (1 - cr)

�
=
I
at Vjn=o at 'tun
W

10-11
C
�
::J
()

� 1O-13
::;
o
!J)
c

�
'
1O-15
c
Ol
L
()
10-17 ��������������������
1 0-10 10-9 1 0-8 10-7
Source current (A)

Hot-electron injection

Source current (A)

Figure 4.8
nFET synapse transistor (a) tunneling and (b) CHEI weight updates. We measured the synapse's
source current Is versus time, and plotted - oIs!8t - versus Is. We fixed the synapse's terminal
voltages; consequently, the change in Is is a result of changes in the synapse's weight W. In (a),
we applied \l;n=5 V, Vs=O V, Vds=2 V, and stepped Vtun from 29 V to 35 V in 1 V increments;
in (b), we applied \l;n=5 V, Vs=O V, Viun=20 V, and stepped Vcts from 2.9 V to 3.5 V in 0.1 V
increments. We turned off the tunneling and CHEI at regular intervals, to measure k. Because,
for a fixed \l;n, the synapse's weight updates oWlOt are proportional to oIs!8t (see Eq. 4.2.2),
these data show that the weight updates follow a power law. The mean values of (T and <: are 0.17
and 0.24, respectively.

unsupervised learning (Hertz et aI., 1991) under the constraint that the sum
of the row-synapse weights, averaged over time, is constant. CHEI feedback
along each row enforces the constraint.
Floating-Gate MOSFETs 111

r------------.,
Row weight- I Vdd
normalization :
circuit : M2
I
I
I
Synapse

:Q1
transistor
I
I�um
25V

Row output
current lout

Figure 4.9
One row of the learning array. The column input vector X comprises low-true, 5 V, lO J.Ls digital
pulses. The row input vector Y comprises high-true, 12 V, lO J.Ls digital pulses. Because the 2 J.Lm
CMOS process that we use has 400 A gate oxides, the tunneling voltages are high: To cause
measurable tunneling, we superimpose the row inputs onto a 25 V DC bias. The voltage coupling
between a synapse's control and floating gates is about 0.8. Consequently, a 5 V (low-true) input on
column wire Xl causes a 4 V decrease in synapseII's floating-gate voltage, which in turn causes a
4 V increase in synapse II's tunneling-oxide voltage. A column input Xi that is coincident with a
row learn-enable pulse 1'i causes a 16 V increase in the tunneling-oxide voltage at synapseII, but
only a 12 V increase at the other synapses. Because electron tunneling increases exponentially with
tunneling-oxide voltage (see Fig. 4.3), synapsel l 's floating gate receives about 100 times more
charge than do the other synapses' floating gates. Because W increases exponentially with floating­
gate charge (see Eq. 4.2.1), synapsel l 's weight increases much more than do the other synapses'
weights. The weight increase causes Isum to rise, which in turn causes the normalization circuit to
raise Vd. Because the CHEI efficiency increases with lI,is (see Fig. 4.5), a higher Vd causes CHEI
in all the synapses, decreasing all the weights. The array eventually settles back to equilibrium,
with Isum equal to h, but synapse II now takes a larger share of the total row current, and the
other synapses each take a smaller share. The inverting amplifier in the weight-normalization
circuit enhances loop stability, for reasons that we discuss in Section 4.3. Each row of the array
has its own normalization circuit.

Weight Normalization

The weight-normalization circuit (see Fig. 4.9) compares I su m, the sum of the
synapse drain currents in a row, with h, the bias current in transistor MI.
If Isu m > h, then the circuit uses CHEI to renormalize the weights. To
understand the renormalization, we begin by defining equilibrium: A row is
in equilibrium when Isu m h. In equilibrium, the drain voltage Vd causes
=

little or no CHEI in the row synapses.

The normalization circuit constrains I su m as follows: Assume that the row
is initially in equilibrium, and that tunneling then raises the weight values
1 12 Chapter 4

of selected synapses, increasing Isum. The excess drain current (Isum-h) is

mirrored by M2 and M3 into capacitor Cint , causing Vc to rise; Ql forces Vd
to follow Vc. When Vd rises, all the row synapses undergo CHEI, decreasing
all the weights and so causing Isum to fall. As Isum falls, Vd also falls, and
the row returns to equilibrium. The drain-current constraint requires that, over
time, lsum h. The nonnalization circuit creates a negative resistance at the
=

synapses' common drain node, causing Vd to rise when Isum increases.

We now show how the drain-current constraint renonnalizes the synapse
weights. We begin with the constraint

(4.3.3)

The renonnalization time constant Ta typically exceeds l Os; this value is 106
times longer than the 10 f..ts input pulses Xi (where Yin Xi ). Consequently,
=

for renonnalization, we replace Vin in Eq. 4.2.2 with its temporal average Vin ,
and we assume that Yin is time invariant, and has the same value for all the
row synapses. Substituting Eq. 4.2.2 into Eq. 4.3.3, we obtain

(4.3.4)

which implies that

h -,.'Vjn
= - e UT == Wsum = canstant. (4.3.5)
10

The drain-current and weight-value constraints are equivalent. Consequently,

row feedback renonnalizes the synapse weights.
Renonnalization forces the row synapses to compete for floating-gate
charge: When one synapse's weight value increases, the sum of the weight val­
ues of its row neighbors must decrease by the same amount. However, when a
selected synapse tunnels, thereby increasing its weight, renonnalization forces
all the row synapses to undergo CHEI, decreasing all the row-synapse weights.
The selected synapse undergoes both tunneling and CHEI. Because the expo­
nent in the CHEI weight-update rule is larger than that in the tunneling rule (see
Eq. 4.3.1 and Eq. 4.3.2), renonnalization constrains a synapse's weight-update
rate, in addition to its weight value.
Tunneling and CHEI effectively redistribute a fixed quantity of floating­
gate charge among the row synapse transistors. In Appendix 4.4 we derive the
Floating-Gate MOSFETs 1 13

array learning rule, for coincident (x,y) pulse inputs to synapsej:

(2
Wi , i#j (n + 1) = Wi, i#j (n) - Itear n Wi (n) -e) (4.3.6)

2
Wj (n + 1) = Wj (n) + ftearn L Wi (n)( -e) (4.3.7)
i, i
#j
where c and ilearn are defined in Eq. 4.4.15 and Eq. 4.4.25 respectively.
Figure 4. 10 shows unsupervised learning in one row of the 4 by 4 array. These
data highlight both the synapse weight and the update-rate constraints. The
data are fit by applying Eq. 4.3.6 and Eq. 4.3.7 recursively; the only inputs to
the fit equations are the synapse weights at n=O and the fit constants Ttun , tpw ,
(7, and c.

The parasitic coupling between a synapse's tunneling junction and its

floating gate is about 5 fF. With CT=1 pF, a 12 V row learn-enable pulse Yj
increases the floating-gate voltage of every row synapse by about 60 mV. This
coupling does not affect the row computation significantly, for two reasons:
Firstly, 5 V low-true column inputs Xi always tum off selected synapses,
regardless of Yj • Secondly, because row learn-enable pulses Yj increase the
floating-gate voltage of every de-selected synapse by a fixed 60 mV, we can
calculate the corresponding source-current increase using Eq. 4.2. 1 , and adjust
lout accordingly.

Normalization-Circuit Stability

The normalization circuit creates a negative resistance at the synapses' com­

mon drain node, Vd. The loop output is Vd and the loop feedback comprises
CHEI oxide currents. When I su m increases, Vd rises and CHEI decreases the
synapse weights, causing lsu m to fall. Because the CHEI oxide currents in­
crease exponentially with Vd, the loop dynamics are highly nonlinear. This
nonlinearity is inherent to synapse transistors, and is likely be present in any
circuit that employs oxide currents in the feedback path. A quantitative stability
analysis of this circuit is beyond the scope of this discussion. However, rather
than omit any discussion of stability, we describe qualitative loop-stability cri­
teria.
The normalization circuit employs positive feedback. To ensure stability,
we must make the loop gain less than unity for all frequencies. This require­
ment implies that the small-signal impedance Zd looking into the synapse drain
terminals, must be greater than the total impedance Z c at capacitor Cint. To see
why, we assume instead that Zc > Zd. A rising Vd induces a small-signal current
 1 14 Chapter 4

:?100
.s

"
0'
5 40
'"
"
'"
g- 20
c:
,.,
(/)
1000 2000 3000 4000 5000 6000 7000
C �nt input pulses (pulses x 103)

100 100
-
:?
Synapse 1: data & fit �
.s 1 �
80
�
c:
�
�
� -�
\ - �'
a 60 "
'-'
" "
0' 0'
5 40 5 10
'" '"
" "
'" �
g- 20 '"
c: c:
,., ,.,
(/) (/)
200 400 600 800 1000 1200 1400 1600
Coincident input pulses (pulses x 103) Coincident input pulses (pulses x 103)

Figure 4.10
Array learning behavior, with fits. We first initialized all synapses to the same source-current value.
We then applied a train of coincident (x,y) 10 J.ts pulses to synapseII, causing its weight value and
source current to increase. Renorrnalization caused the weight values and source currents of the
other synapses to decrease. Once synapseII had acquired 90% of the total row current, the pulse­
train stimulus was removed and applied instead to synapseI2, and then in turn to synapses 13 and
14. The synapse source currents were measured after every 10> input pulses. In the lower half
of the figure, the first 1600 data points are fit by applying Eq. 4.3.6 and Eq. 4.3.7 recursively.
The linear plot shows the fit accuracy over the entire sweep, whereas the logarithmic plot shows
that the weight values of deselected synapses do not saturate, but instead follow a power-law
decay as predicted by Eq. 4.3.2 and Eq. 4.3.6. The inputs to the fit equations are the initial
synapse source-current values (at n=O); the pulsewidth �w=1O J.ts; and the empirical constants
Ttun=1O ms, 0'=0.14, and <:=0.21. These data show that individual synapses can be addressed
with good selectivity, and that wide separation in the weight values of selected versus deselected
synapses can be achieved.

isu m Vd/Zd; isu m is mirrored by M2 and M3 into Cint , causing Vc to rise

=

by an amount vc=isu mzc. Because Vd follows Vc, if Zc > Zd, then Vc > Vd;
isu m will increase rapidly, causing Vc to rise toward Vdd.
The impedance Zd is limited by interconnect capacitances, and by synapse­
transistor channel-length modulation, ftoating-gate-to-drain overlap capaci­
tance, and drain-current impact ionization. We consider each of these limi­
tations in tum.
Floating-Gate MOSFETs 1 15

Interconnect Capacitance Interconnect capacitance at the synapses' com­

mon drain node causes Zd to decrease with frequency. We choose Cint to be
much larger than this parasitic capacitance, so the reactive impedance ratio,
Zc/zd, favors loop stability for all frequencies.

Channel-Length Modulation Channel-length modulation reduces a synapse's

drain impedance, limiting Zd. Fortunately, the synapse transistor's Early volt­
age exceeds 100 V, as a result of both the 10 f..tm channel length and the p-type
channel implant; consequently, the channel-length modulation is small.

Floating-Gate-to-Drain Overlap Capacitance Vd couples to a synapse

transistor's floating gate, by means of the floating-gate-to-drain overlap ca­
pacitance Cdg. The coupling coefficient is CdgjCT, where CT is the total
floating-gate capacitance. Because Is increases exponentially with VIg, Cdg
causes Isu m to increase exponentially with Vd, limiting Zd. To minimize the
effect, we use a large interpoly capacitor (CT=1 pF); we also apply inverting
feedback from Vd to the floating gate, increasing Zd (see Fig. 4.9). We used
an off-chip amplifier to generate this inverting feedback; an on-chip adaptive
floating-gate amplifier (Hasler, 1997) can be used instead.

Drain-Current Impact Ionization Channel electrons that possess sufficient

energy for CHEI also possess sufficient energy for impact ionization (Shock­
ley, 1961; Tam et al., 1984). In the synapse transistor, a drain-to-channel elec­
tric field that causes CHEI also creates additional electron-hole pairs, causing
Id to increase exponentially with Vdc. As a result, Isu m increases exponentially
with Vd, limiting Zd. If Vd becomes greater than about 4 V, the rate of drain­
current increase causes loop instability, and Vd rises rapidly. As Vd rises, CHEI
decreases all the synapse-transistor weights. As Vd saturates near Vdd, CHEI
causes Isu m to fall below Ib; Vd falls and the loop returns to a stable operating
regime. Loop instability causes Vd to undergo a single brief ( 1 0 f..ts) voltage
'"

spike, and reduces all the synapse weights substantially. Fortunately, because
the synapse CHEI efficiency is high, weight renormalization rarely causes V d
to exceed 3.5V; consequently, the loop is stable.

Local Learning in Silicon

The synaptic array that we have analyzed, although simplistic from a com­
putational point of view, demonstrates the potential for performing large-scale
1 16 Chapter 4

computation and unsupervised learning on a silicon chip. The computation and

synapse-weight modification occur locally and in parallel: The array perfonns
fast, single-transistor analog computation and slow, locally computed weight
adaptation. In addition, we can describe the computation and learning behavior
using rules derived directly from the silicon-MOS and silicon-oxide physics.
Although the present array affords unsupervised learning, it uses a feed­
back error signal to constrain the weight values. Feedback error signals typ­
ically are used in supervised neural networks, to adjust the array weights ac­
cording to a network learning rule. In future floating-gate arrays, rather than us­
ing unsupervised learning, we can use CHEI to adjust the synapse weights in a
supervised fashion, using either pulsed, or continuously valued analog (Hasler,
1997), inputs and row-error signals. Further examples of learning circuits can
be found in Cauwenberghs and Bayoumi ( 1999).

4.4 Appendices

We derive the tunneling weight-increment rule for an nFET synapse in Ap­

pendix 4.4; the CHEI weight-decrement rule for an nFET synapse in Ap­
pendix 4.4; and the nFET array learning rule in Appendix 4.4.

The Tunneling Weight-Increment Rule

We begin by taking the temporal derivative of the synapse weight W, where

W=exp(Q Ig/Q T) :
aw w a Q Ig
- - -
_ W
- - Ig
--
_

(4.4. 1 )
at QT at QT
and substitute Eq. 4. 1 . 1 for the tunneling gate current I g:

Ito
aw
= We-� . (4.4.2)
at QT
We substitute Vox yt un V, g (where yt un and V, g are the tunneling­
= -

implant and floating-gate voltages, respectively), assume that Vt un » V, g,

expand the exponent using ( 1 -x) - l � l +x, and solve:

Ito _ vtun _ V
aw
--

at
� - We v, � tun (4.4.3)
QT
Substituting V, 9 UT Q, gj",Q T, and solving for the tunneling weight-
Floating-Gate MOSFETs 1 17

increment rule yields

aw � 1
__
w (1 -u) (4.4.4)
at Ttun
where
_ V, U2 T
(j = --- (4.4.5)
1I:ytun
and
QT e �
Ttun ==
It o v'u" .
(4.4.6)

The parameters (j and Ttun vary with the tunneling voltage Vt un.

The CHEI Weight-Decrement Rule

We begin by defining a synapse transistor's drain-to-channel potential, Vdc,

in terms of Vd s and Is . In a subthreshold floating-gate MOSFET, the source
current is related to the floating-gate and source voltages (Mead, 1989) by

K V[ g - V.
Is =
Io e UT (4.4.7)

and the channel-surface potential iI! is related to the floating-gate voltage,

V'g (Andreou and Boahen, 1994; Enz et al., 1995) by

(4.4.8)

where 11: is the coupling coefficient from the floating gate to the channel, and
iI! 0 depends on the MOS process parameters.
Using Eq. 4.4.7 and Eq. 4.4.8, we solve for the surface potential iI! in terms
of Is and Vs :

(4.4.9)

We now solve for Vdc :

(4.4. 10)

The CREI gate current Ig is given by Eq. 4.2.3. We add a minus sign to Ig ,
because eREI decreases the floating-gate charge; and substitute for Vdc using
1 18 Chapter 4

Eq. 4.4.I O:
Vd. - >I' o - UT In ( I. II o ) � V� - >I' o( 1 - �)
Ig = - (3 Is e Vi nj = - (3 10 l n J e inj Is n
l J
.
(4.4.11)
Substitute for Is using Eq. 4.2.2, and solve:

(4.4.12)

Substitute Eq. 4.4.12 into aw/at (Eq. 4.4.1):

(4.4.13)

to obtain the final weight-decrement rule:

aw _ __1 w (2-e)
= (4.4.14)
at Tinj
where
UT
c :: -- (4.4.15)
Vinj
and
K/ Vi n _ Vd.s - ,, ' Vi n - · o
r.
QT e _ UT Vinj
inj -
=

(310
(4.4.16)

The paraIll eter c is a constant, whereas Tinj varies with Vin .

The Array Learning Rule

We consider the row-synapse weights at discrete time intervals t :: nT, where

n is the step number and T is the timestep and derive the row-learning rule
for a single coincident (x,y) input to a single row synapse. We begin with the
equilibrium condition for the row-weight normalization:

(4.4.17)

We assume that the normalization time constant Ta is fixed, for the follow­
ing reason: Coincident (x,y) input pulses cause a weight increase at a synapse;
the normalization circuit responds by establishing a drain voltage V d for which
the total weight decay, summed over all the row synapses, balances the weight
increase at the single synapse. If we assume that the mean density of the co­
incident input pulses is time-invariant, then Vd'S mean value, Vd, is constant,
Floating-Gate MOSFETs 1 19

and therefore the low-frequency loop time constant, T a , also is constant. We

assume that Ta « T.
The synapse weight values can violate Eq. 4.4. 17 for times t « T a « T,
but we require that they satisfy Eq. 4.4.17 at our measurement time intervals
t =nT. We permit array inputs at times (t+8t)==(n+8)T, immediately after the
synapse weight values at t =nT are measured. The array inputs comprise a
pulsed column vector X(n+8), where Xi E [0, 1] == [5V,OV ] , and a pulsed
row vector Y(n+8), where Yj E [0, 1] == [OV, 12V ] . Without loss of gener­
ality, we assume that at time t =nT, the circuit is in equilibrium, and that at
(t+8t+tpw )==(n+8pw )T, coincident row and column inputs of duration t pw have
caused synapsej's weight to increase:

(4.4. 18)

� Wj (n) + tpw
Wj (n/ 1-u) (4.4. 19)
Ttun

where in Eq. 4.4. 18 we have made the first-order approximation that ow/at
is constant over tpw , and in Eq. 4.4.19 we have substituted for ow/at using
Eq. 4.3.1. Because tpw « Ta , at time (n+8pw) the circuit no longer is in
equilibrium,
(4.4.20)

and the synapse weights inject down to reestablish equilibrium.

We wish to find the synapse weights at (n+l ), when the row again satis­
fies Eq. 4.4. 17. Using Eq. 4.4. 14 and Eq. 4.4. 19, we write weight-decrement
expressions for the row synapses

�w.t. , t..,...
. ...J.). (n + 1) = -�Tinj
w.t. , t..,...
. ...J. ). (n) ( 2- e) (4.4.21 )

T
( Wj (n) +
tpw
Wj (n) (l -u)
) (2- e) (4.4.22)
Tinj Ttun

where, because the row drain voltage Vd settles during renormalization, Tinj
may vary over T (recall that T » Ta » tpw ). For reasonable values of Vtun
and tpw , the weight increment from a single coincident (x,y) input is small;
120 Chapter 4

consequently, Eq. 4.4.22 can be simplified using ( l +x) n � 1 +nx,

�Wj (n + 1) � - ..I...-
Tinj
2
Wj (n)( -e) ( 1 + (2 - c-) tpw
Ttun
Wj (n)-u ) .
(4.4.23)
Because Tinj varies over T, T/Tinj can re-expressed in tenns of quantities
that we know at n. We equate the weight increment at synapsej (see Eq. 4.4.19)
to the sum of the weight decrements at synapses i,i-:f.j (Eq. 4.4.21) and j
(Eq. 4.4.23)

U
Wj (n)( l - ) (2
tpw
Ttun � L Wi (n) -e)
TmJ' i, i#
j

+ ..I...-
Tinj
2
Wj (n)( -e) ( 1 + (2 - c-) tpw
Ttun
Wj (n)-u )
(4.4.24)

and solve for T/Tinj :

T � W · (n)( l - u )
Tt u n 3
(4.4.25)
Tinj - (2 - c-) � W · (n)(2-e-u ) + L: Wi (n)(2-e)
nun 3 .
t

We define Ilearn =.T/Tinj , substitute Izearn into Eq. 4.4.21, and use
Eq. 4.4.17 to solve for the row-learning rule:
2
Wi , ih (n + 1) = Wi, ih (n) - !learn Wi (n)( -e) (4.4.26)

2
Wj (n + 1) = Wj (n) + izearn L Wi (n)( -e) . (4.4.27)
i, ih
Eq. 4.3.6 and Eq. 4.3.7 describe the row weight-update rule for a single
coincident (x,y) pulse input to synapsej.
II STATICS
5 Basic Static Circuits

In this chapter we present some basic analog VLSI circuits that are widely
used as building blocks of more complex circuits and that are more extensively
treated in various textbooks (Gregorian and Ternes, 1986; Horowitz and Hill,
1989; Mead, 1989; Johns and Martin, 1997; Maloberti, 2001; Gray et al.,
2001; Razavi, 2001; Allen and Holberg, 2(02). Other basic circuits that are
less commonly used, but particular to the type of structures presented in this
book, are introduced in the next chapter and some of the later chapters. Most of
the circuits in this chapter are described only in one configuration. An almost
equivalent circuit is obtained by exchanging the types of all MOSFETs and by
reversing the signs of all voltage differences. We restrict ourselves to a steady­
state analysis, which is valid if the largest signal frequency is much smaller
than the bandwidth of the circuit. Furthermore, unless otherwise noted, we only
consider the subthreshold domain and neglect second-order effects, such as
the Early effect. The equations governing the behavior above threshold can be
found in standard text books. We only point out qualitative differences between
the two domains. Note that the larger currents above threshold increase the
bandwidths of the circuits. In order to keep the equations simple, calculations
are for MOSFETs with unity width-to-Iength ratios, but extension to other
values is straightforward.
We saw in Chapter 3 that the natural reference potential for a MOSFET
is its bulk potential. However, for the analysis of circuits whose MOSFETs
do not all have the same bulk potential a common reference potential must be
chosen. In traditional CMOS circuits, all nFETs have a common bulk potential,
usually called Vss, and all pFETs have a common bulk potential called V dd. In
order to avoid large currents from the sources and drains into the bulks, all
source and drain diodes to the bulk are reverse-biased: Vss is the lowest and
Vdd is the highest potential in the circuit. The Vss and Vdd lines are therefore
called power rails. It is convenient to choose either Vss or Vdd as the reference
potential. In the following, we will reference all voltages to Vss such that all the
voltages in the circuit are positive. In our circuits, the bulks of the MOSFETs
do not necessarily have to be connected to the power rails, but we use the
convention that the bulk connections are omitted in the MOSFET symbols if
they are connected to the power rails. Note, however, that in standard CMOS
processes, all bulks of one MOSFET type are connected to the lightly-doped
silicon substrate and are thus at a common potential, which is one of the power
124 Chapter 5

rail potentials. MOSFETs of the other type rest in wells, whose potentials can
be individually chosen.
In the circuit schematics, nodes at larger potentials are drawn above nodes
at lower potentials for each branch of the circuit, such that the currents flow
downward. Connections to Vss are denoted by one of the common symbols
used for ground connections, and connections to V dd are denoted by a slanting
line. All other connections to external circuitry are denoted by circles. Some
of the input or output nodes of the circuits are not labeled with a voltage
parameter. If such a node is the drain of a MOSFET, the voltage value at
that node can vary freely. Unlabeled source nodes are assumed to be held at a
constant potential.
The steady-state subthreshold characteristics of MOSFETs (neglecting the
Early effect) are described by Eq. 3.2.10 and 3.3.1 for nFETs and pFETs
respectively. With the above conventions and under the assumption that the
bulks are connected to the respective power rails, we obtain for the drain
current of the nFET

(5.0.1)

where Ino denotes the nFET current-scaling parameter, "'n the nFET sub­
threshold slope factor, UT the thermal voltage, Vg the gate voltage, Vs the
source voltage, and Vd the drain voltage. The current is defined to be positive
if it flows from the drain to the source. The corresponding equation for the
pFET is

(5.0.2)

where the values of the pFET current-scaling parameter I pO and the subthresh­
old slope factor "'p are different from the corresponding nFET values.

5.1 Single- Transistor Circuits

The rich behavior of a MOSFET in its different regimes allow it to perform

several functions.

Current Source

One of the simplest functions of a MOSFET is obtained when its source and
gate potentials are held at constant values. As long as the difference between
the drain and source voltages is larger than approximately 4U T, the MOSFET
Basic Static Circuits 125

is in saturation and Eq. 5.0.1 reduces to

I 1 = Inoe(Kn vg-V.)/UT I (5.1.1)

(cf. Eq. 3.2.16). In this first-order approximation, the drain current is indepen­
dent of the drain voltage. A device that supplies an output current that is inde­
pendent of the voltage applied to the same terminal is called a current source.
The transistor deviates systematically from an ideal current source because of
the Early effect. This deviation can be reduced by making the length of the
transistor large. Above threshold, the saturation current is given by Eq. 3.2.39,
which we rewrite here as

(5.1.2)

Vg•
Above threshold, the drain voltage required for saturation is larger than below
threshold and depends on Furthermore, the Early voltage is smaller and
hence, the MOSFET is a less ideal current source than when it is operated
below threshold.

Linear Resistor

The drain current of the transistor in the triode regime strongly depends on the
relative values of the source and drain voltages. We can rewrite Eq. 5.0.1 for
subthreshold operation as

(
1 = InoeKn Vg/UT-(Vd+V.)/2UT e(Vd-V.)/2UT - e-(Vd-V.)/2UT ) (5.1.3)

= 2InO eKn Vg/UT-(Vd+V.)/2UT sinh (V�;:s) . (5.1.4)

If Vd - Vs is small enough to neglect third-order and higher order terms, a

Vd - Vs
Taylor series expansion yields

Kn Vg/UT-(Vd+V.)/2UT
'"., InO e
I,... (5.1.5)
UT
For a given common mode voltage (Vd Vs)/2
+ and a given gate voltage Vg, a
MOSFET acts as a linear resistor with resistance

R = UT e(Vd+V.)/2UT-Kn Vg/UT . (5.1.6)

I no

VB Vd,
Above threshold, the current-voltage relationship in the triode regime is de­
scribed by Eq. 3.2.38. Neglecting second-order terms in and the resistor
126 Chapter 5

is linear:
R = (,B(Vg - VT))-l. (5.1.7)

As in the subthreshold case, the resistance depends on Vg, and the common­
mode voltage through V T .
Hence, in the triode regime, the transistor is approximately linear both
above and below threshold The resistance is controlled by the gate voltage
of the transistor. Below threshold, the resistance is an exponential function
of the gate voltage. Above threshold, the resistance is a function of the in­
verse of the gate voltage. The triode regime of the transistor extends to larger
drain-to-source voltages above threshold than below threshold, but the linear
approximation holds only for first-order tenns, whereas below threshold the
approximation is valid up to second-order tenns.
Transistors operated in the linear regime are suitable for use as pseudo­
conductances in resistive networks (see Chapter 6).

Nonlinear Voltage-Current / Current-Voltage Converter

I
A MOSFET operating in saturation in the subthreshold region, as described
by Eq. 5.1.1 for an nFET, has a drain current that is an exponential function
of VB and of Vg. In this configuration, the MOSFET acts as an exponential
voltage-to-current converter if one of the two voltages is fixed and the other
node is the input. Above threshold, there is a quadratic voltage-to-current con­

I
version as described by Eq. 3.2.39. The inverse function (a current-to-voltage
conversion) is obtained by simply making the input signal and Vg or VB the
output signal. The MOSFET acts as a logarithmic current-to-voltage converter
below threshold, and a square-root current-to-voltage converter above thresh­
old. The current-to-voltage conversion function in subthreshold can be derived
by solving for the output voltage in Eq. 5.1.1:

VB = #l:nVg - UTlog
(I�o) (5.1.8)

if the source is the output tenninal and

(5.1.9)

if the gate is the output tenninal. In the latter case, we should remember that
Vg is detennined by the gate charge, which cannot be directly influenced by
the input current due to the infinite impedance between channel and gate. In
Basic Static Circuits 127

FigureS.1
Diode-connected nFET.

order to make the circuit work, the input node and the gate have to be in a
negative feedback loop that controls the gate charge and keeps the MOSFET
in saturation. When the input current is fed into the drain terminal, the feedback
from gate to drain is negative and thus the feedback from drain to gate must
be positive. As shown in Fig. 5.1, the gate can simply be shorted to the drain
to complete the negative feedback loop. The MOSFET is then reduced to a
two-terminal device with similar characteristics to a diode, and is said to be
diode-connected. Since the drain of the diode is always reverse-biased with
respect to the channel, a diode-connected MOSFET is always in saturation
as long as any appreciable current flows. If the source is chosen as the input
node, the feedback from source to gate must be made positive, which requires
additional transistors.

5.2 Two-Transistor Circuits

Now that we have seen how one transistor can perform multiple functions,
we show how two-transistor circuits can perform additional functions; for
example, replication, amplification, and reduction of either current inputs or
voltage inputs.

Current Mirror

Figure 5.2 shows a diode-connected MOSFET that has a common gate node
with another MOSFET of the same type. If both MOSFETs have fixed source
voltages and are in saturation, they act as current sources. Moreover, if both
MOSFETs are of the same size and have the same source voltage, they source
128 Chapter 5

Figure 5.2
Current mirror with adjustable amplification.

the same current, which is why the device is called a current mirror. The
input current lin through the diode-connected transistor M 1 sets the common
gate voltage Vg and hence the output current lout of the second transistor M2•
The output current can be scaled by choosing different transistor sizes, or by
choosing different source potentials Vsl and Vs2 for the two MOSFETs. The
dependence of the output on the difference in the source potentials is described
by

(5.2.1)

While this strategy scales the current by a factor that is exponential in the
source-voltage difference, as seen from Eq. 5.2.1, the transistor size strategy
scales the current by a fixed design factor that is linear in the ratio of the width­
to-length ratios.
Above threshold, the matching of input and output currents depends more
strongly on the drain voltage of M 2, due to the larger Early effect, and the
dependence of the gain on the difference of the source voltages is weaker.
For bidirectional input currents lin, a current mirror acts as a half-wave
rectifier, because the input transistor has a very high impedance in the opposite
direction. Half-wave rectification is useful in many applications, for example,
in the implementation of linear-threshold neurons (Chapter 6).

Source Follower

The source-follower circuit linearly transforms a voltage at a high-impedance

input terminal into a voltage at a lower-impedance output terminal, such that
the output signal is able to drive larger loads than the input signal. It is
Basic Static Circuits 129

o---i
Vin

Vout
Vb
0----9 Mb Vb
0----9 Mb

Vout
Vout

o---i
Vb

Ib
Vi n
0----9 M1 Vg
0----9 Vw

Vs

(a) (b) (c)

Figure 5.3
Source-follower circuits with Ca) range reduction, (b) unity gain, and (c) range reduction and
independent time-constant and offset control.

constructed by connecting a fixed current source to the source of a MOSFET

operated in saturation. As we saw in Section 5.1, a current source can easily be
implemented using a saturated MOSFET with fixed source and gate voltages.
This arrangement results in a circuit with two MOSFETs of the same type
connected in series (Fig. 5.3(a)), where Mb is the current source and MI is
the input transistor. The input voltage Vin is applied to the gate voltage of M I
and the output voltage Vout is the source voltage of MI. In the subthreshold
domain, the output voltage adjusts itself according to Eq. 5.1.1:

(5.2.2)

where "'b is the subthreshold slope factor of Mb and we assume that both
transistors have the same dimensions. Note that Vout is linearly related to
\lin with a positive slope of "'n < 1 and a fixed offset that depends on the
bias current h. The output voltage follows the input voltage with a gain of
"'n, hence the name of the circuit. The measured input-output characteristic
130 Chapter 5

of this circuit is shown in Fig. 5.4. The condition for saturation of the biasing
transistor, and thus proper operation of the circuit, is Vout > Vs + 4UT: That
is, Yin > ,,;;l ("bVb + 4UT). In this range, the residual non-linearity of the
characteristic is mainly due to the body effect of M 1: "n is not constant. The
dependence of "n on Yin, as obtained by differentiating the curve of Fig. 5.4, is
shown in Fig. 5.5. Given Eq. 5.2.2 we see that we can alternatively use V b as a
high-impedance input terminal to obtain a negative gain of -" b. However, this
arrangement has the disadvantage that the current flowing through the circuit
and thus the transient dynamics, depend on the input signal.

3.5

2.5

1.5

0.5

�n(V)
Figure 5.4
Transfer characteristic of source follower circuit.

It is possible to get around the" reduction factor in the transfer characteris­

tic, if the bulk potential of the input MOSFET can be controlled independently.
As mentioned previously, in a CMOS process this independence is possible for
only one type of MOSFET: The one that sits in a well with opposite doping
from the substrate. For a source-follower circuit constructed of well MOS­
FETs, the wells of the transistors can be connected to their respective sources,
as shown in Fig. 5.3(b) for pFETs in n-wells. Since the subthreshold slope
Basic Static Circuits 13 1

factors are equal for both MOSFETs,

Vout = Vin + �: log (:�) = Vin + Vs - Vb· (5.2.3)

The output voltage now follows the input voltage with unity gain and a fixed
offset Vs - Vb. The bias transistor stays in saturation for Vin < Vb - 4UT.

2 3 4 5

Vin(V)
Figure 5.5
Dependence of subthreshold slope factor of source follower circuit on input voltage.

The well of M1 ,vw, can be used as an additional input terminal as shown

in Fig. 5.3(c). This configuration affords an additional degree of freedom in
controlling the output voltage, which is now given by

Vout = (1 - Kp)Vw + KpVg + UTlog (:�) (5.2.4)

= (1 - Kp)Vw + KpVg + Kb (Vs - Vb ). (5.2.5)

The additional terminal allows the offset of Vout to be adjusted independently

of h and thus of the temporal dynamics of the circuit. We can either apply the
input voltage at the gate terminal and adjust the offset with V w, or apply the
input voltage at the well terminal and adjust the offset with Vg• In the former
case the gain is Kp and in the latter case it is 1 - Kp, which is usually smaller
132 Chapter 5

because subthreshold slope factors tend to be larger than 0.5. Since the well
has a smaller input impedance than the gate, due to its larger capacitance and
leakage currents, the well-input option should only be chosen if the voltage
source at the input is sufficiently strong. In either case, the circuit should be
operated such that V w is never significantly smaller than Vout, otherwise a large
leakage current flows from the output node into the well via the forward-biased
source node of M 1 .
The source-follower circuit can thus be used to introduce an adjustable
offset to a voltage, and to optionally reduce the voltage range with a gain
factor smaller than one. Its main function, however, is that of an impedance
converter: The circuit transforms a weakly-driven voltage signal into a more
strongly-driven voltage signal.

Inverting Amplifier

A completely different function is obtained if two MOSFETs of opposite types

are connected in series at their drains with their sources at fixed potentials, as
shown in Fig. 5.6(a). In this common-drain configuration, both MOSFETs act
as current sources, as long as they are in saturation. Since both transistors have
to source the same current, a small change aVg in the gate voltage of one
transistor results in a large change aVout in the common drain voltage in the
opposite direction. The circuit thus acts as an inverting voltage amplifier. The
gains An and Ap of the gate inputs to the two transistors can be computed from
the Early effect as

= aVout K-n VnE VpE

An
__

(5.2.6)
- aVgn - UT VnE + VpE
aVout = K-p VnE VpE
Ap =
_

(5.2.7)
- aVgp UT VnE + VpE
where K-n and K-p are the subthreshold slope factors and VnE and VpE are
the Early voltages of the nFET and the pFET respectively. The gains A n
and Ap are typically between -100 and -1000, so that both transistors will be
simultaneously in saturation for variations of only a few millivolts on one of
the gate terminals. In the above-threshold domain, the absolute values of the
gains are smaller due to the reduced Early voltages. The operating point and
the current through the circuit can be set by adjusting the source voltages of
the two MOSFETs accordingly.
Now consider the circuit in Fig. 5.6(b) where an input signal Vin is applied
to both gates. The resulting gain A = aVout /aVin is now the sum of the gains
Basic Static Circuits 133

(a) (b)

Figure 5.6
Inverting amplifier with (a) two input terminals and (b) one input terminal. The circuit in (b) is
also used as an inverter in digital CMOS logic.

for the individual gate inputs. This circuit, with the source voltages connected
to the respective power supply rails, is the basic building block of digital
CMOS logic. The circuit is called an inverter, because if Vin is close to one of
the power supply voltages representing one logic state, then Vout is close to the
other power supply voltage representing the other logic state. While the circuit
is in a given logic state the MOSFETs are turned off and nearly no power is
consumed, which is one reason why CMOS logic is popular for low-power
designs.

5.3 Differential Pair and Transconductance Amplifier

In the following, we will tum to some slightly more complex circuits in order
to introduce the principles of the transconductance amplifier, which is used in
a variety of circuit configurations in analog circuit design.

Differential Pair

The differential pair has the same basic structure as the source follower, ex­
cept that the bias current h is now shared by two MOSFETs Ml and M2
whose sources are connected to the drain of the bias MOSFET M b, as shown
134 Chapter 5

in Fig. 5.7. The sharing of the current between M I and M2 depends on their re­
spective gate voltages VI and V2• If all MOSFETs are operated below threshold
and in saturation and we assume that MI and M2 have the same subthreshold
slope factor K,n, we obtain

FigureS.7
Differential pair.

(5.3.1)

and

(5.3.2)

The dependence of these two currents on the difference of the input voltages is
shown in Fig. 5.8. The curves have a sigmoidal shape. They are almost linear
for small voltage differences and saturate at Ib for large voltage differences.
Such compressive nonlinearities are very useful for the implementation of
different functions, especially in the context of neural networks. What makes
the circuit even more useful is the fact that to a first approximation (neglecting
the Early effect), the output currents depend only on the difference of the input
voltages: The circuit has a small common-mode sensitivity. Given that voltages
are differential rather than absolute quantities such a property is very useful.
Basic Static Circuits 135

0.8

� 0.6

-'"

0.4

0.2

-8.3 -0.2 -0.1 0 0.1 0.2 0.3

V/-V2(V)

FigureS.S
Dependence of differential pair output currents on differential input voltage.

The condition for saturation of M b is

e-V./UT « 1.
(5.3.3)

With

(5.3.4)

the condition for saturation of Mb becomes

(5.3.5)

if IVI - V21 > 4UT. If Mb is not in saturation, the currents depend strongly on
the common mode of the input voltages.

Transconductance Amplifier

The two output currents in the differential pair circuit can be subtracted from
one another to form a single bidirectional output current. The subtraction is
performed by connecting a current mirror of the complementary transistor type
to the differential pair, as shown in Fig. 5.9(a). The resulting circuit is the
simplest version of a differential transconductance amplifier, whose symbol
is shown in Fig. 5.9(b). As long as all MOSFETs stay in saturation and the
136 Chapter 5

(a)

(b)

Figure 5.9
Simple differential transconductance amplifier. (a) Schematic diagram. (b) Circuit symbol with
inverting (-) and non-inverting (+) inputs.
Basic Static Circuits 137

differential pair is operated below threshold, the output current is given by

(5.3.6)
This relationship is confirmed by the the measured data shown in Fig. 5.10.
For small differential voltages it is approximately linear:

(5.3.7)

where

(5.3.8)

is the transconductance of the amplifier. The reason for this name is the
fact although 9m has the dimensions of a conductance, the output current is
measured at a different terminal to the pair across which the input voltage
gradient is applied.

-9
ax10

2
�
:; 0
_0

-2

-4

-6

=B.3 -0.2 -0.1 0 0.1 0.2 0.3

V1-V2(V)

Figure 5.10
I-V transfer characteristics of simple differential transconductance amplifier.
138 Chapter 5

If the differential pair is operated above threshold, it can be shown that

(5.3.9)

for IVI - V21 ::; y'2h /f3. The transconductance is thus given by

gm = ..fif4 . (5.3.10)

Because the input voltages are applied to insulated gates, the input conduc­
tances of the transconductance amplifier are negligible and the input currents
are close to zero under steady state conditions. The output conductance de­
pends on the Early voltages ofM2 andM4:

(5.3.11)

where V E2 and V E4 are the Early voltages of M2 and M4 respectively. For

VE2 = VE4 = VE this simplifies to

(5.3.12)

In addition to the saturation condition forM b in subthreshold (Eq. 5.3.5),

we have to consider the saturation conditions for the other transistors. For
practical purposes, MI and M3 will always be in saturation because M3 is
diode-connected and the drain of MI is thus at a high voltage. The saturation
conditions for M4 and M2 restrict the output voltage range for subthreshold
operation to

Figure 5.11 shows the relationship between output current and output voltage
for given equal input voltages. The output conductance is the absolute value
of the slope in the linear region. The limiting regions for Vout, according to
Eq. 5.3.13, are clearly visible.
In the mode described above, the transconductance amplifier is used as
a differential-voltage-to-current converter. However, it can also be used in
the open-circuit mode as a differential-voltage amplifier. In which case, for
Basic Static Circuits 139

2.5

-O.51'------ �2� --- �3--- -�4 --- ----"5

Vout(V)
Figure 5.11
I-V characteristics at output of simple differential transconductance amplifier for fixed equal input
voltages.

VI :::::; V2, we obtain

(5.3.14)

(5.3.15)

From this result we can compute the open-circuit voltage gain

(5.3.16)

which in the subthreshold regime can be expressed as

A = Kn VE2VE4 (5.3.17)
UT VE2 + VE4
Above threshold we obtain

(5.3.18)

Since Vout increases with increasing VI and decreases with increasing V2,
the gate of MI is called the non-inverting input terminal and the gate of
M2 the inverting input terminal of the amplifier. In the circuit symbol, the
140 Chapter 5

5
V2= V2= V2= V2= V2 V2= V2= V2= V2=
O.5V 1.0V 1.5V 2.0V 2.5V 3.5V 4.0V 4.5V 5.0V
4

2 3 4 5

Figure 5.12
Voltage amplification characteristics of simple differential transconductance amplifier.

two input terminals are denoted by a plus and a minus sign respectively,
as shown in Fig. 5.9(b). The open-circuit voltage gain increases with the
Early voltages, and therefore with the length of the output transistors. Typical
subthreshold values are between 100 and 1000. Given this large gain and
the unavoidable transistor mismatches introduced by the fabrication process,
the device is not normally used as an open-circuit voltage amplifier. In the
open-loop configuration, the transconductance amplifier is used mainly as a
comparator, which outputs a high voltage if VI > V2 and a low voltage if
VI < V2• However, these output voltages are not independent of the input
voltages and we therefore do not obtain an ideal compamtor with a binary
output.
Since in the open circuit Vout is normally at one of its limits, we will
now determine where those limits lie. If VI is larger than V2, M4 goes out
of saturation. For VI > V2 + 4UT, the current through M2 is much smaller
than the current through M I and M3. Hence, Vout goes almost all the way to
Vdd, shutting off M4. If VI is smaller than V2, M2 goes out of saturation; but
::::; h/2. If V2 is significantly larger
the current mirror acts, such that 11 ::::; 12
Basic Static Circuits 14 1

Figure 5.13
Wide-output-range differential transconductance amplifier.

than VI, the voltage drop across M2 is close to zero and Vout �
Vs. With

(5.3.19)

1
we see that Vout goes all the way to zero if VI < 1\,;; (I\,b Vb - (4 + log 2)UT )
and
(5.3.20)
1
if VI > 1\,;; (I\,b Vb + (4 -log 2)UT). The open-circuit voltage characteristics
of the simple transconductance amplifier are shown in Fig. 5.12 for different
fixed values of V2. As predicted by Eq. 5.3.20, the lower limit of the output
voltage range ramps up linearly with VI, with a slope of I\,n.
If an application requires an output range extending to both supply rails,
the circuit can be expanded as shown in Fig. 5.13. The current 1 2 is mirrored
142 Chapter 5

twice, such that the output stage is symmetric and decoupled from the input
stage. This decoupling also allows the design of output stages with large open­
loop gains (long MOSFETs) or large output currents (MOSFETs with large
width-to-Iength ratios or bipolars (see Chapter 12)). Disadvantages of this
extension are the need for almost twice as many transistors as for the basic
version, and the increased effect of mismatches due to fabrication tolerances.
Transconductance amplifiers are widely used as elements in circuit ap­
plications, where they are usually referred to as operational amplifiers. Most
commercially-available transconductance amplifiers are more sophisticated
and evolved than the versions presented here. They are multistage circuits,
which are optimized with respect to a set of performance criteria, some of
which relate to their dynamic behavior (stability, gain-bandwidth product,
etc.). The treatment of such circuits is beyond the scope of this chapter, but
can be found in the literature «Gregorian, 1999; Huijsing, 2(01)).

5.4 Unity-Gain Follower

Most circuit applications use the transconductance amplifier as part of a

negative-feedback loop. The negative feedback ensures that the amplifier stays
within its operating range. The simplest negative feedback loop is obtained by
short-circuiting the output terminal and the inverting input terminal, as shown
in Fig. 5.14. The transfer function of this circuit is given by

dVout = �1 �
1.
(5.4.1)
dVin A+
Due to the large open-loop voltage gain A, the transfer function is almost

Figure 5.14
Unity-gain follower.
Basic Static Circuits 143

0.05

0.04

0.03

:;-
-" 0.02
::,;-
1-
:::,5 0.01

-0.01

-0.02 L-----'----'---'
o 2 3 4

FigureS.IS
Deviation of output voltage from input voltage for simple unity-gain follower.

unity and Vout :=:::! Vin. The circuit configuration is therefore called unity­
gain follower. It is used as an impedance converter (also called a buffer)
and it converts a high input impedance into a lower output impedance. In
contrast to the source follower presented in Section 5.2, which is also used
as an impedance converter, the unity-gain follower does not introduce a large
voltage offset. The measured deviation of the output signal from the input
signal, Vout - Vin, as a function of the input voltage is shown in Fig. 5.15. The
deviation is about 5 m V, except for very low voltages, where the bias current of
the amplifier is practically shut off and the behavior of the circuit gets degraded
by leakage effects: At very high voltages, where the current mirror is shut off.
This deviation has a random component, which is due to circuit mismatches;
and a systematic component, which is due to the amplifier's finite open-loop
gain A, as shown by Eq. 5.4.1.
6 Current-Mode Circuits

During the last 40 years, the vast majority of analog circuits have used voltages
to represent and process relevant signals. However, recently, current-mode sig­
nal processing circuits, in which signals and state variables are represented
by currents rather than voltages (Tomazou et aI., 1990), have shown advan­
tages over their voltage-mode counterparts. Their advantages include higher
bandwidth, higher dynamic range, and they are more amenable to lower power
supplies.
In this chapter, we will describe some basic current-mode circuits com­
monly used in neuromorphic systems. Although the individual basic circuits
are relatively simple, they offer very interesting signal-processing properties
when connected to form networks. We start by first describing the current con­
veyor block which can be used to replace the traditional operational amplifier.

6.1 The Current Conveyor

In voltage-mode circuits, the main building block used to add, subtract, am­
plify, attenuate, and filter voltage signals is the operational amplifier. In
current-mode circuits, the analogous building block is the current conveyor
(Smith and Sedra, 1968; Wilson, 1990).

Figure 6.1
Current conveyor.

The original current conveyor (Fig. 6.1) was a three-terminal device (two
input terminals X and Y and one output terminal Z) with the following proper­
ties:

1. The potential at its input terminal (X) is equal to the voltage applied at the
other input terminal (Y).
146 Chapter 6

2. An input current that is forced into node X results in an equal amount of

current flowing into node Y.

3. The input current flowing into node X is conveyed to node Z, which has the
characteristics of a high output impedance current source.

The tenn conveyor refers to the third property above: Currents are conveyed
from the input tenninal to the output tenninal, while decoupling the circuits
connected to these tenninals.
The simplest CMOS implementation of a current conveyor is a single MOS
transistor (Fig. 6.2(a)). When used as a current buffer, it conveys current from
a low impedance input node X to a high impedance output node Z: And when
used as a source-follower, its source tenninal X can follow its gate Y. A more
elaborate current-controlled conveyor is shown in Fig. 6.2(b). This basic two­
transistor current conveyor is used in many neuromorphic circuits (Cohen and
Andreou, 1992; Boahen and Andreou, 1992; Delbriick and Mead, 1994) and is
a key component of the current-mode winner-take-all circuit that is analyzed
in Section 6.3. It has the desirable property of having the voltage at node X
controlled by the current being sourced into node Y. If the transistors are

operated in the subthreshold domain, the monotonic function that links the
voltage at the node X to the current being sourced into Y is a logarithm.
As voltages at the nodes Y and X are decoupled from each other, Va: can
be clamped to a desired constant value by chosing appropriate values of I y.
An example of a system level application is shown in Fig. 6.2(c): A current­
controlled conveyor is used to make multiple copies of the current I y. The input
current I y sets the value of Va: independent of Vy1. This property can be useful
for processing time-varying signals: If the current I y represents a time-varying
signal, and if the capacitance on node X is not negligible (for example, the
node X is connected to many transistors), then keeping the voltage Va: clamped
to a constant value ensures that the output currents IYl' Iy2, ••• , IY n faithfully
follow Iy •
Sedra and Smith (1970) refonnulated the definition of the current con­
veyor, describing a new circuit that combines both voltage and current-mode
signal processing characteristics. This new type of current conveyor (denoted
as conveyor of class II) is represented by the symbol shown in Fig. 6.1 and its

1 As opposed to simple current mirrors in which Va: == vy•

Current-Mode Circuits 147

y z

Iy
Z

Iz

Y 0-------1
Ix

X
Ix X

(a) (b)

y Z

I I IY
y y, 2

X
�---'�-----+r---�+-- - - -

(c)

Figure 6.2
Current conveyor implementations. (a) Single MOS transistor current conveyor. (b) Two MOS
transistor current-controlled conveyor. (c) System application of current conveyor for the genera­
tion of multiple copies of an input current.
148 Chapter 6

Table 6.1
Applications of the class II current conveyor to current·mode signal processing circuits.

Functional Element Function Realization Using Current Conveyor

"*"
R,
� x I,
Current Amplifier Iz = �h R, CCIl z f-----o
y
"*" .Li

Current Differentiator Iz = CRo/rr

�: ,� CCII

Current Integrator Iz =
CIR J Ildt
�: ,� CCII

:u I:
Current Summer Iz = - �j Ij
..f: ,� CCII

input-output characteristics are defined as

(6.1.1)

The input voltages Vx and Vy are linked by a unity gain relationship

(Vx =Vy) ; the tenninal Y has infinite impedance (/y 0) and the current =

forced into node X is conveyed to the high impedance output node Z with a ± 1
gain. Examples of CMOS circuits that comply with the definition of a class II
current conveyor are shown in Fig. 6.3. These circuits are typically used as
current precise rectifiers, but they (and similar circuits) can implement many
other signal processing functions. Table 6.1 shows some possible applications
of class II current conveyors.

6.2 The Current Normalizer

Signal normalization is another important signal processing function that can

be realized using current-mode circuits. This function is useful for rescaling
data from multiple sources into a standard reference frame. Nonnalization
Current-Mode Circuits 149

y
I,

(a)

'x
x

y o-------l+

(b)

Figure 6.3
CMOS implementations of class II current conveyors. (a) Uni-directional current conveyor: This
circuit operates correctly only for positive values of k (that is currents that are sourced into
node Z). (b) Bi-directional current conveyor: This circuit conveys both positive and negative input
currents.
150 Chapter 6

is often used in artificial signal processing systems and is also ubiquitous in

biological nervous systems.
A current-mode circuit that normalizes its input signals is shown in
Fig. 6.4. It receives analog continuous time input currents and provides nor­
malized output currents. The circuit is based on the Gilbert normalizer cir­
cuit (Tomazou et al., 1990) that was originally designed using bipolar transis­
tors (see Chapter 7). It is a modular circuit that can be extended to an arbitrary
number of cells by simply connecting additional current mirrors to the com­
mon node Vc. If the input currents are subthreshold, the circuit is characterized
by the equations

(6.2.1)

where i is the index of the i th cell of the circuit, UT is the thermal voltage and K,

is the subthreshold slope coefficient (see Section 3.2). By applying Kirchhoff's

current law to the common node Vc we obtain
N
L lout; = h (6.2.2)
i=l

where h is a constant current set by the control voltage V b. We use this

constraint to solve Eq. 6.2.1 for Vc, and to derive the dependence of the output
current on the input currents:

(6.2.3)

The output current of each cell l out; is directly proportional to its input
current (with a proportionality constant I b), but scaled by the sum of all the
input currents Lj lin; .

6.3 Winner-Take-All Circuits

A winner-take-all (WTA) circuit is a network: of competing cells (neural,

software, or hardware) that reports only the response of the cell that has the
strongest activation while suppressing the responses of all other cells. The
circuit essentially implements a maxO function. These circuits are typically
used to implement and model competitive mechanisms among populations
Current-Mode Circuits 15 1

'in, linz

lout, loutz

Vd Vd
, z

M1 M3 M2 M4

Figure 6.4
Two-input current normalizer circuit.

of neurons. For example, they are used to select specific regions of an input
space (Yuille and Geiger, 1995). Many WTA networks have been implemented
both in software (von der Malsburg, 1973; Nass and Cooper, 1975; Grossberg,
1978; Kaski and Kohonen, 1994) and in hardware (Lazzaro et aI., 1989; Choi
and Sheu, 1993; Starzyk and Fang, 1993; Serrano and Linares-Barranco, 1995;
Lau and Lee, 1998; Demosthenous et aI., 1998; Liu, 2000; Hahnloser et aI.,
2000; Indiveri, 2001a).
In this section we analyze a class of WTA networks that emulate biolog­
ical networks, consisting of a cluster of excitatory neurons that innervate a
global feedback inhibitory neuron. These networks have been implemented in
aVLSI and applied to a wide variety of tasks, including selective attention (Bra-
152 Chapter 6

jovic and Kanade, 1998; Wilson et aI., 1999; Indiveri, 2000, 2001b), auditory
localization (Lazzaro and Mead, 1989), visual stereopsis (Mahowald, 1994),
smooth pursuit/tracking (Etienne-Cummings et aI., 1996; Horiuchi and Koch,
1999), and detection of heading direction (Indiveri et aI., 1996; Mudra et aI.,
1999).

WTA Neural Models

We shall focus on a particularly simple yet powerful model that describes a

population of N homogeneous excitatory units that excites a single global
inhibitory unit which feedbacks to inhibit all the excitatory units (Fig. 6.5).
For sake of simplicity, we neglect the dynamics of the system and examine
only the steady-state solutions. Dynamic properties of these networks and of
other physiological models of competitive mechanisms are described in detail
in Kaski and Kohonen (1994); Yuille and Geiger (1995); Grossberg (1978);
Ben-Yishai et ai. (1995).

Figure 6.5
Network of N excitatory neurons (empty circles) projecting to one common inhibitory neuron
(filled circle), which provides feedback inhibition. Small filled circles indicate inhibitory synapses
and small empty circles indicate excitatory synapses. XI . . . XN are external inputs; Yet . . . YeN
are the outputs of the excitatory neurons; Yi is the output of the inhibitory neuron; 'Ukt . . . weN
are the excitatory synaptic weights of the external inputs; 'U./t • • . WIN are the excitatory weights
onto the global inhibitory neuron; and Wit . . . WiN are the inhibitory weights from the inhibitory
neuron onto the excitatory neurons.
Current-Mode Circuits 153

Consider a network (Fig. 6.5), in which the external input to the j th

excitatory neuron is x j, the response of the j th excitatory neuron is Yej ,
the response of the inhibitory neuron is Y i; and in which the weights of the
synapses from the external inputs to the excitatory neurons, from the inhibitory
neuron to excitatory ones and from the excitatory neuron to the inhibitory one
are wej' Wij and Wlj respectively. We can write this network as

(6.3.1)

Yi = f (t )
3=1
WljYej (6.3.2)

where f( .) denotes the transfer function of both excitatory and inhibitory

neuronS. This system of coupled equations describes the recurrent interactions
between excitatory neurons and the inhibitory neuron. We explore the behavior
of the system by considering three special cases:

1. The case in which all neurons have a linear transfer function (j(x) x). =

2. The case in which the neurons are linear-threshold (j(x) max(O, x)),
=

and all external inputs are identical.

3. The case in which the neurons are linear-threshold, and one external input
is much larger than all others.

More general cases using non-linear transfer functions are difficult to solve
analytically; however, they can be studied using numerical simulations.

Linear Units lithe neurons are fully linear (j(x) = x) we can solve the
system analytically:

Yej = WejXj - WijYi

Yi = L Wlj ( wejXj - WijYi) (6.3.3)
j
which implies that

(6.3.4)
154 Chapter 6

In the simplified case, we assume that all the weights of each kind are the same:

We; = We Vj
Wi; = Wi Vj
Wl; = Wo Vj

and so

(6.3.5)

The output of each neuron is proportional to its input, but has a nonnalizing
tenn subtracted. Equation (6.3.5) shows that the response Y e; of a linear
excitatory neuron can have both positive and negative values, depending on
the inputs Xk, on its connection weights W e, Wi, Wo and on the total number of
excitatory neurons N.

Linear Threshold Units with Uniform External Inputs The half-wave

rectification function (f(x) = max(O, x» is a more biologically realistic
function than the linear one of the previous case. Neurons with this transfer
function have a response of only positive values.
In this case, the system of linear equations (Eqs. 6.3.2) becomes a system
of non-linear coupled equations, and it is not longer possible to obtain a
general closed fonn solution. However if all external inputs are identical
(X j = Xo Vj), we can reduce the system to

Ye; = max(O, we;Xo - Wi;Yi)

y, = max (0, � ) >l; y,; (6.3.6)

and if we make the working hypothesis that

we obtain the linear system:
(W e;Xo - Wi;Yi) > 0 V j, then

Ye; = we;Xo - Wi;Yi

Yi = L Wl; (We;XO - Wi;Yi) (6.3.7)
j
Current-Mode Circuits 155

which yields

(6.3.8)

If the synapses from external inputs and those from the inhibitory neuron
have equal strength (w ej = Wij = Wo 'V i), then

(6.3.9)

The hypothesis used to obtain Eq. 6.3.7 is satisfied for all values of x 0 > 0,
Wo > 0, and Wlj > 0 'V i. In summary, if all inputs are equal, then all
excitatory linear threshold units have identical outputs which are equal to the
input nonnalized by a tenn that is directly proportional to the weights W Ij and
inversely proportional to Wo.

Linear Threshold Units with One Input Much Greater than All Others
Now consider the case in which one input (say the external input to unit i o ,
Xjo) is much greater than all other external inputs (x jO »Xj 'Vi ¥- io ) and
the synaptic weights are as described above. Again, we assume a priori that
the weighted external excitatory input to unit i 0 exceeds the inhibitory input to
the same unit (wejoXjO - WijoYi > 0) and that the weighted external inputs to
all other excitatory inputs don't(w ejXj - WijYi < 0 'V i ¥- iO). Under these
assumptions, Eq. 6.3.6 can be rewritten

YejO = (wejoXjO - WijoYi)

Yej = 0 'Vi ¥- iO
Yi = Wljo (WejoXo - WijoYi) (6.3.10)
156 Chapter 6

which can be simplified to yield

WejoXjO
YejO ...,--::":""
1 + Wljo-=- W--
ijo
=

Yej = 0 'Vj # jO
wejo WljoXjO
Yi (6.3.11)
1 + Wljo Wijo
=

This solution satisfies the assumption that W ejoXjO > WijoYi for all values of
wejo' XjO, and Wijo. It also satisfies the a priori assumption that WejXj < WijYi
as long as the external input x jO is sufficiently large with respect to all other
Xj inputs.
Summarizing: if one external input is much greater than the other inputs,
then all excitatory linear threshold units, except the one receiving the strongest
input, are suppressed. The output of the winning unit is a normalized version of
the input, and the normalizing factor is directly proportional to the connection
weights Wijo' Wljo' and inversely proportional to w ejo•

0.8

0.6
Z.
:;
.

U
ell

·c 0.4
-

::J

0.2

0
0 10 0

Figure 6.6
Simulations of a WTA network comprising 100 linear-threshold units ordered along one spatial
dimension. The input (solid line) is composed of 3 Gaussians. The outputs are shown for two
cases: Wej = 1, Wij = 1 and Wlj = 0.0250 Vj (dashed line); Wej = 1, Wij = 1 and
Wlj = 0.0325 Vj (dotted line).
Current-Mode Circuits 157

Numerical Simulations It is not possible to obtain a closed form solution

for networks with linear threshold units and any arbitrary input distribution, or
networks with arbitrary transfer functions, however numerical simulations are
useful for providing insight into the general computational properties of the
network.

4 2

}\
�1.5
'>

J:
n
�:> 1
c.
:;
00.5

0 .
0 10 0 %���--��-----1�00
Ca) Cb)

Figure 6.7
Numerical simulation the same WTA network shown in Fig. 6.6 now with weight values tIk;
1, Wi; = 1 and WI; = 0.0275. Ca) is the input distribution of increasing amplitude. Cb) Network
responses to the three inputs shown in Cal.

For example, the simulations shown in Figs. 6.6 and 6.7 explore the re­
sponse of a network with f(x) = max(O, x) and N=IOOto a more complicated
input distribution, consisting of three Gaussians centered at unit positions 20,
50, and 80, and having maximum values of 0.75, 0.5, and 0.35 respectively
(see solid line of Fig. 6.6). The simulations of Fig. 6.6 show the effect of mod­
ifying the excitatory to inhibitory weights WI; (with all other weights set to
one). When WI; = 0.0250 '<ij, the output is a thresholded version of the input,
consisting of 3 peaks of activity. However, when WI; is increased to 0.0325
'<ij, only the strongest input peak is reflected in the output.
In the simulations of Fig. 6.7(a), the excitatory to inhibitory weights WI;
are set to an intermediate value of to 0.0275 '<ij, and the network responds to
the two strongest peaks in the input. The form of the response is invariant to
the input strength (or alternatively, the strength of the we weights) as shown
;

in Fig. 6.7.
158 Chapter 6

Non-linear Programming Formulation

The competitive mechanism that emerges from the neural architecture of

Fig. 6.5 can also be described mathematically. The following set of non-linear
equations select the largest number among N real numbers by multiplying the
neuron output signals Yej by binary-valued constants (a j, either 0 or 1):

min (- t )
3=1
ajYej with constraint:

N
L aj= 1 (aj E {O, I}). (6.3.12)
j=l

Systems of non-linear equations with discrete constraints are difficult to solve.

We can simplify the system if we extend the domain of a j to the continuous
interval [0, 1] and include an additional constraint that forces the continuous
values of aj to tend either toward zero or one. :

min (- � ) a;y" with constraints:

L aj=I
j
L ajln aj= O. (6.3.13)
j

These types of systems can be solved using the Lagrange multipliers

method (Bertsekas, 1982). Solving Eq. 6.3.13 is equivalent to finding min (L)
aj
and max (L), where A1 and A2 are parameters called Lagrange multipliers, and
Al,A2
L is the cost function

(6.3.14)
Current-Mode Circuits 159

If we set A2 to a constant, then we can find min (L) and max(L) by solving:
O'.j )..1

(6.3.15)

which implies
Yej ->'1 -'>"2
e );2OJ =

LeYej 1)..2
�
e� = (6.3.16)
j
and
eYej 1)..2
(Xj =

L:k eYekl)..2 • (6.3.17)

This equation approaches the solution of Eq. 6.3.13 when A2 is sufficiently

small.
Systems of constrained non-linear equations can be implemented using
MOS transistors in the subthreshold domain, The response of the circuit is
analogous to the solution of Eq. 6.3.16. Figure 6. 8 shows an example of such
a circuit, which solves the system of equations (Eqs. 6.3.13). The solution can
be confirmed by applying Kirchhoff's current law at node Vc of the circuit
(h L:j loutj)' and observing that the output current of each branch of the
=

WTA network can be expressed as a fraction (X j of the total bias current:

N
loutj (Xj h
= = (Xj L loutk .
(6.3.18)
k=l

Because loutj can also be written as

(6.3.19)

the circuit's response and the system of equations (Eqs. 6.3.16) are equivalent,
if
160 Chapter 6

and

(6.3.20)

Current Mode WTA Circuit

The circuit of Fig. 6. 8 is a continuous time, analog circuit that implements

a WTA network:. It was originally designed by Lazzaro et al. (1989) and is
extensively used in a wide variety of applications. The circuit is extremely
compact and elegant: It processes all the (continuous-time) input signals in
parallel, using only two transistors per input cell, and one global transistor
that is common to all cells. Collective computation and global connectivity is
obtained using one single node common to all cells.
An example of a WTA circuit containing only 2 cells is shown in Fig. 6.8.
Each cell comprises a current-controlled conveyor (Section 6.1). The WTA
network: is modular and can be extended to N cells, by connecting additional
cells to the node Vc. Input currents are applied to the network through current
sources which are implemented for example using subthreshold pFETs. The
output signals are encoded both by the l outl and lout2 currents, and the Vdl and
Vd2 voltages. The voltage Vb sets the bias current h . Transistors Ml and M2
discharge nodes Vd and so implement inhibitory feedback. Transistors Mg and
� implement an excitatory feedforward path by charging node V The overall c.

circuit selects the largest input current lin; because cell j provides lout; Ib, �

and so suppresses all other output voltages and currents (Vd;#i 0, loutdi
� �

0). Cell j wins the competition because its voltage Vd; determines Vc by the
exponential characteristics of the transistor that sinks the output current l out;
(for example, Mg or M4).
We will analyze the behavior of the circuit in the steady-state case using the
methods that we applied for the network model: By providing constant input
signals and measuring the outputs after the circuit has settled. We consider
three cases: Both inputs are equal; one input much larger than the other; and
two inputs that differ by a very small amount (small-signal regime).

Both Inputs Equal If the two input currents are equal (lin I = lin2 = 1m)
then the currents flowing into transistors Ml and M2 of Fig. 6. 8 are also
Current-Mode Circuits 16 1

lin, lin2

lout, lout2

Vd Vd
, 2
M3 M4

Vc Vc
M, M2

Figure 6.8
Two cells of a current mode WTA circuit.

equal. In this case, because the gates of M 1 and M2 are tied to the same
common node Vc, the drain voltages of Ml and M2 must take the same value
(Vdl = Vd2 Vm). As a result, the output transistors M3 and M4 will
=

have the same gate-to-source voltage difference (Vgs3 VgS4 Vm - Vc).

= =

If both output transistors are in saturation then the output currents must be
identical. Moreover, Kirchhoff's current law requires that, at the common node
Vc, Ioutl = Iout2 = h/2 (Eq. 6.2.2).

One Input Much Greater than The Other Recall that the subthreshold cur­
rent flowing through a transistor can be divided into a forward component, If,
162 Chapter 6

and a reverse component, Ir, (Eq. 3.2.11). When the transistor's source voltage
Vs is approximately equal to its drain voltage Vd, Ir becomes comparable to
If·
With this property in mind, we can consider the case in which I inl » Iin2 .
In this case, the drain voltage of M1
(Vdl) will be greater than the drain voltage
of M2 (Vd2). If the transistor Ml is in saturation (Vdl > 4UT), the dominant
component of its drain current will be in the forward direction and its gate
\C
voltage Vc will increase such that I dl = 1/1 = Ioel<i$ = Iin1. Although the
two input currents Iin1 and Iin2 are different, the forward component of the
drain currents of Ml and M2are equal (l/1 = I h) because the two transistors
have a common gate voltage Vc, and both their sources are tied to ground. The
drain current Id2 of transistor M2 can only be equal to the input current Iin2
under the following conditions:

which implies that

which implies that

The reverse component of I d2 becomes significant only if Vd2 decreases

enough for M2 to operate in its ohmic region (Vd2 � 4UT). In this case, the
output transistor M4 is effectively switched off, and Iout2 O. Consequently,
=

M3 sources all the bias current (loutl Ib), with Vd1 satisfying the equation
=

Ioel<Vdl -Vc h.
=

The experimental data of Fig. 6.9 shows the output voltages (V d,l and
Vd,2) and output currents (lout,l and Iout,2) of the circuit, in response to the
differential input voltage d V which encodes the ratio of the input currents. In
this experiment, the input currents were provided by pFETs operating in the
subthreshold regime: The gate voltage Vinl of the pFET sourcing current into
the first cell was set to 4.3V, while the gate voltage Vin2 of the pFET sourcing
current into the second cell was set to Vin2 = Vinl + d V . The two traces in
each plot show the responses of the two cells as d V was swept from -8mV to
+8m V. When d V is zero (the input currents are identical), the output signals of
both cells are also identical. When d V is large (one input current dominates),
a single cell is selected.
Current-Mode Circuits 163

18 .

1.6
1.4
�1.2
Q)
Ol
.!9 1
"0
>
:; 0 8 .
c.
80.6
0.4
0.2
-� -6 -4 -2 0 2 4 6 8
Differential input voltage (mV)

(al

250

200
�
c:
�150
�
OJ
()
:; 100
c.
:;
0
50

0
-8 -6 -4 -2 0 2 4 6 8
Differential input voltage (mV)

(b)

Figure 6.9
Responses of the two-cell WTA circuit shown in Fig. 6.8. (a) Voltage output Oal and Vd2 ) versus
the differential input voltage. (b) Current output (kutl and Iout2 ). The bias voltage \i = O.7V.
The small difference in the maximum output currents is due to device mismatch effects in the
read-out transistors of the two cells.
164 Chapter 6

If � V is small, the above description is not adequate. Instead, we can

compute the outputs signals of the cells using small-signal analysis (Lazzaro,
1990).

Two Inputs DitTer by a Small Amount To analyze the circuit in this regime,
we must consider the Early effect of the transistor operating in the saturation
region (Eq 3.5.3):

Ids Isat(l + Vd
=
s
Ve ) (6.3.21)

where Ve is the Early voltage.

Assume that the two input currents Iin1 and Iin2 are initially equal. In this
case, the transistors Ml and M2 operate in the saturation region: The output
voltages Vdl and Vd2 will settle to a common value, and the output currents
Iouh and Iout2 will both be equal to h/2. If we now increase the input current
Iin1 by a small amount 8[ and apply Eq. 3.5.3 to transistor Ml of Fig. 6.8, then
its drain voltage Vdl will increase by

(6.3.22)

As Vdl is also the gate voltage of transistor M3 , the Iout,l will be amplified
by an amount proportional to eOv. The constraint of Eq. 6.2.2 requires that
Iout2 decrease by the same amount in steady state. This reduction means the
gate voltage Vd2
of M4 must decrease by 8v.
The gain of the competition mechanism ( � ) in the small signal regime is
directly proportional to the Early voltage Ve and inversely proportional to I sat.
The Early voltage depends on the geometry of the transistors and is fixed at
design time. On the other hand I sat depends on Vc, which changes with the
amplitude of the input currents.

6.4 Resistive Networks

Conventional methods of implementing resistors in VLSI technology include

using single MOSFETs as described in Section 5.1; or using more complex
circuits such as the transconductance amplifier of Section 5.3. These methods
have the disadvantage of emulating linear resistors for only a very limited
range of voltages, and of resistance values. If we consider currents, and not
Current-Mode Circuits 165

voltages, to representi input and output signals of MOSFETs, then we can

implement resistive networks using single transistors instead of resistors. In
this configuration, the transistor is linear for a wide range of current values.
Furthermore, if the transistor is operated in the subthreshold regime, then the
resistance (or conductance) can be varied by changing its gate voltage.
A conventional conductance, G, is defined by the relationship

lab = G ( V a - Vb)
where lab is the current flowing from terminal a to terminal b, and V a , Vb the
voltages at the corresponding terminals. If the two terminals a and b are the
source and the drain of a subthreshold nFET, the current I ab can be expressed
by the usual transistor relationship:

� II:� �
lab = 10e"'UT-rr; - 10e"'UT- UT (6.4.1)

where Vg is the transistor's gate voltage.

v
If we define the pseudo-voltage (Vittoz and Arreguit, 1993) V * Voe- TJT
=

(where Vo is an arbitrary scaling voltage), and the pseudo-conductance G * =

II:
-(?;e"'il} , then we can write
(6.4.2)

where the value of pseudo-conductance G*depends exponentially on the tran­

sistor's gate voltage Vg' Using Eq. 6.4.2 we can map any resistive network into
an equivalent transistor network: Each resistor Ri of the resistive network can
be replaced by a single transistor Mi, provided that all the transistors share the
same substrate (that is, they are all either nFETs or pFETs). If the gate volt­
ages V�1i of all the transistors are equal, then the transistor network is linear
with respect to current (Vittoz, 1994). This linear behavior holds for the entire
range of weak inversion, which may be as much as 6 orders of magnitude in
transistor current. Because all Vg; must be the same, the values of the indi­
vidual conductances can only be adjusted by changing the W/L ratio (which
modulates 10) of each transistor.
An alternative intrepretation of the mapping between resistive and transis­
tor networks uses the concept of a current diffusor (Boahen, 1997) illustrated
in Fig. 6.10.
The currents linland lin2are inputs to the circuit. Assuming that the three
nFETs are identical (that is, their 10 and", parameters are equal), and solving
166 Chapter 6

Figure 6.10
Current diffusor circuit. The current h. proportional to (h - It). diffuses from the source to the
drain of M3.

the circuit equations, we obtain:

(6.4.3)

If VI = V2 = Vref, this relationship can be simplified to yield

(6.4.4)

The diffusion current h through M3 is proportional to (h II)' The pro­ -

portionality factor can be modulated by either V ref or V3 • The current-mode

diffusor network (Fig. 6.11(a)) is composed of multiple instances of the circuit
of Fig. 6.10. In this network, current injected at a node j diffuses laterally and
decays with distance (Mead, 1989). Consequently the network acts as spatial
low-pass filter (see Section 8.1); and because the network is linear, the effects
of currents injected at different nodes superimpose.
The diffusor network (Fig. 6.11(a)) has the same network response as the
resistive network in Fig. 6.11(b). This equivalence can be demonstrated by
comparing the transfer functions of the two circuits. Applying Kirchhoff's
current law at node V j of Fig. 6. 11(a):

(6.4.5)
Current-Mode Circuits 167

(a)

(b)

Figure 6.11
Similarities between (a) current-mode diffusor network, and (b) resistive network.

Using Eq. 6.4.4, we can express Ij and Ij-1 in terms of the output currents:

Ij-l ...!L(VO-VR)(I
e UT Iout;_l ) (6.4.6)
= out; -

Ij ...!L(VO-VR)(I
e UT Iout; ) . (6.4.7)
= out;+l -
168 Chapter 6

Substituting these two relationships in Eq. 6.4.5 yields

Ioutj - Iinj = e �(VG-VR) (Iouti+l - 2Ioutj + Ioutj_l) (6.4.8)

Similarly, we can apply Kirchhoff's current law at node V j of Fig. 6.11(b):

(6.4.9)

Because Ij = G( Vj - Vj+d and Ioutj = Vj / R, Ij can be expressed as a

function of Iout and of Iouti+l:
j
1
Ij = (loutj - Iouti+l). (6.4.10)
RG
Combining this equation with Eq. 6.4.9 yields

1
Ioutj - Iinj = (louti+l - 2Ioutj + Ioutj_l)· (6.4.11)
RG
The term (lOUti+l - 2Ioutj + Ioutj_l) in Eq. 6.4.11 is the discrete ap­
proximation of the � operator. Both circuits of Fig. 6.11 approximate the
diffusion equation that characterizes the properties of a continuous resistive
sheet (Mead, 1989):

(6.4.12)

where A is the diffusion length. In the discrete resistive network of Fig. 6.11(b)
the diffusion length A = 1/ JRG, while in the diffusor network of Fig. 6.11(a)
the diffusion length is A = e
2UT (VG-VR).

6.5 Current Correlator and Bump Circuit

The current correlator measures the correlation between unidirectional input

currents, whereas the bump circuit measures the similarity or dissimilarity of
input voltages. Both of these circuits have been used in many aVLSI designs.
For example, these circuits have been used in a stereoscopic vision system
(Mahowald, 1994) to disambiguate between real and false targets, and in an
auto-focusing system to measure the input signal power (Delbriick, 1989).

Simple Current Correlator

Mead recognized that in subthreshold operation, the current-correlator circuit

in Fig. 6.12 computes a measure of the correlation between its two input
Current-Mode Circuits 169

Figure 6.12
The simple current-correlator. S is the strength ratio (see text) between the transistors in the middle
leg. and the transistors in the outer legs. I is large only when both II and h are large.

currents h and 12• Intuitively, the series-connected transistors perfonn an

analog AND-like computation. If either of the gate voltages on these series­
connected transistors is low, then the output current is shut off. Conversely,
if both of the input voltages are high, then the output current is large. In the
intennediate regions, the circuit computes an approximation to the product of
the input currents.
We will analyze the simple current-correlator for subthreshold transistor
operation. We refer to the effective WI L ratio for a transistor as the strength of
the transistor. For circuit configurations like the simple current-correlator, the
strength ratio is the ratio of the strength of the transistors in the middle leg, to
the strength of the transistors in the outer legs. This parameter is given by

s (WI L)middle (6.5.1)

(WI L)outer
=

and is an important circuit parameter for both the simple current-correlator and
the bump circuit.
To compute the output current 1 , we assume that the top transistor M 2 in
Fig. 6.12 is saturated, and that the currents through M1 and M2 are identical.
170 Chapter 6

Using the subthreshold equation, we obtain

eV1eV2
I = Se- V. eV1 +eV2
--;-;--"""';7'"

1112
= S 11 +12 . (6.5.2)

While the transistors are operating in subthreshold, the circuit operation

is symmetric in the two input currents, despite the apparent asymmetry in the
stacking order of the transistors M1 and M2. Above threshold. the function is
more complicated and is no longer symmetric in the input currents.
This simple current-correlator circuit computes a self-normalized correla­
tion. The output current is proportional to the product of the two input currents,
divided by the sum of the inputs.
We can extend the simple current-correlator to more than a pair of inputs.
The output current for n input currents (a stack of n series-connected transis­
tors) is
n
1 1
-
LI -· (6.5.3)
=
=

lout k l k
The n-input current correlator computes the parallel combination of the n
input currents. The maximum number of inputs can be large, because the
only requirement for correct circuit operation is that the top transistor in the
correlator be saturated. However, the output current scales as 1/ n.

Bump-Antibump Circuit

Figure 6.13 (a) shows the bump-antibump circuit. It has three outputs (Fig. 6.13
(b»: 11, 12, and Imid. Output Imid is the bump output. Outputs h and 12
behave like rectifier outputs, becoming large only when the corresponding
input is sufficiently larger than the other input. If 11 and 12 are combined,
they form the antibump output, which is the complement of the bump output.
Intuitively, we can understand the operation of this circuit as follows: The
three currents must sum to the bias current I b; hence, the voltage Vc follows
the higher of V1 or V2• The series-connected transistors M1 and M2 form the
core of the same analog current correlator that is used in the current-correlator.
When Ll V =0, current flows through all three legs of the circuit. When III VI
increases, the common-node voltage Vc begins to follow the higher of V1 or
V2• This action shuts off Imid, because the transistor whose gate is connected
to the lower of V1 or V2 shuts off. Both V1 and V2 can rise together and I mid
Current-Mode Circuits 17 1

v1-++----I

v2- v" v, held constant (V)

(a)
(b)

Figure 6.13
(a) The bump-antibump circuit. (b) Output characteristics of bump-anti bump circuit. The plots
show data points together with theoretical fits of the form given in the text. The curve pointed to
by the arrow shows the fit that would result from using S=5.33 derived from the drawn layout
geometry, before any process correction. The two theoretical curves shown for /mid are the
result of computing the best numerical fit to the entire curve (S=22.4), and using the ratio of
the measured maximum to minimum current II + h (S=28). The two numerically fit curves
are nearl y indistinguishable, and both are very different from the theoretical curve derived from
the layout geometry. The width and length reduction parameters from this process run were of
the order of 0.5 jlm. Using these parameters, we compute S= 6.6, which is still far short of the
observed behavior. The curve labeled Sum is h + 12 + Imid. The slope on the Sum curve is
due to the drain conductance of the bias transistor.

does not increase, because the common-node voltage Vc increases along with
VIand V2•
Using the subthreshold transistor equation; the input-output relation for
the simple current-correlator (Eq. 6.5.2); and Kirchhoff's current law applied
to the common node (h = II + 12+ Imid) we can compute the current Imii
h
Imid =
2 .a.v (6.5.4)
1 + 1.
S cosh -"-2

( ) _ e� ±e - � .s chx
2coshx ( )_ 1
- 2 e -COSh(x)'
172 Chapter 6

Of course, the antibump outputs sum together to make the current

h
Ii +12 = Ib - Imid = 8 2T
av
. (6.5.5)
'4sech +1
When � V = 0, the cosh 2 term is 1 and the bump current becomes
h
Imid = -- 4 (6.5.6)
1+S
while the antibump outputs sum to

h
Ii +12 = -8--' (6.5.7)
'4 +1
Now we can observe the effect of the transistor strength ratio, S, on the
circuit behavior. The width of the bump, measured in input voltage units,
depends on this ratio. S controls the fraction of the bias current I b that is
supplied by I mid when � V O. By examining the denominator of Eq. 6.5.4,
=

we see that the width of the bump scales approximately as10g(S), when S � 1.
If we define the width of the response as the � V that makes I mid drop to 1/2
of the value it takes when � V 0, we obtain
=

2
� Vi/2 � -log S
K,
(6.5.8)

in the limit of very large S. The units for � V are kT/ q.

The dynamic range of the antibump output, I i +12 also depends on S. The
antibump output ranges from the minimum value in Eq. 6.5.7, to I b. The ratio
of maximum to minimum output (the dynamic range) is � + 1. It appears that
making S very large gives a very wide dynamic range. This is true, but if S
is very large, the bottom of the antibump output characteristic becomes very
flat. To make an output characteristic that is maximally parabolic on a linear
scale requires a suitable compromise on S. A suitable value for S, and thus
the actual transistor dimensions, should be carefully chosen using simulation
of circuit performance.
Figure 6.13(b) shows measured operating curves for the bump circuit. The
theoretical form for I mid fits the data quite well; there is a discrepancy between
the measured and expected value for S. The bump-antibump circuits behaves
as though S is much larger than it should be given the layout of the circuit.
This effect is fortuitous, because it means the circuit layout need not be as
bulky as would be required for very large S. In practice, if oneuses the bump
circuit output (Imid), the effective value of S is not particularly relevant unless
Current-Mode Circuits 173

x10-8 Large geometry (/.. = 6 J.!m)

1

0.8

� 0.6
-'"
+ Sratio
-"
0.4 / effective
/ Drawn
0.2 I
131164
0
-0.6 -0.4 -0.2 o 0.2 0.4 0.6
V, - V2 (V)

x10--8 Small geometry (/.. = 3 J.!m)

1

0.8

0.6
�
-'"
+
Sratio
-" 0.4
/ effective
0.2 Ib = 9.4 x 1Q-9A
1/ Drawn
K = 0.64

574/64
0
-0.6 -0.4 -0.2 o 0.2 0.4 0.6
V, - V2 (V)

Figure 6.14
Antibump outputs from 8 bump-anti bump circuits with different geometries. The solid curves
show the theoretical fits derived from Eq. 6.5.5. The numbers beside each curve are the actual
and expected S values for the circuit. The different bump circuits in each set of graphs all had
transistors of the same width in the outer legs, and transistors of the same length in the middle
leg. For the top set of curves: Minimum transistor dimension was 6 f.tm; correlating transistors
had widths 6, 12, 24, and 48 f.tm; and outer transistors had lengths 6, 12, 24, and 48 f.tm. For the
bottom set of curves, all transistor dimensions were halved. The discrepancy between measured
and expected S values are larger for the circuits with smaller dimensions. The curves were fit by
minimizing the total squared error, using a single common � and K,.

S is really large, because the width of the bump is only weakly dependent on
S. On the other hand, if the antibump outputs are used, the output current at
zero input difference depends critically S. Since the dynamic range of output
174 Chapter 6

10-9

-,
, I \ /' "" /--
I / I / I /
I / \ / I /
I / \ / I /
I / I / I I
I I I / I I
/
I I I / I I
/ I
I / I I /
I / I I I /
I / I I I I
I / I I I I I
1/ I I / I I
1/ I I I I
I I / /
I I I I
I I / /
I I /
I I /
I I
I /
I / /
I / /
I / I
I / /
I / /
/
\. ./

10-12 L-_---'-__-'-_----'L-_--'-__-L-_---'-__...L__L-_-'
0.5 1.5 2 2.5 3.5 4 4.5 5
Inpu1 voltage (V)

Figure 6.15
The back-gate effect on antibump circuit operation. Each curve shows the antibump output current
for a different setting of the Vi input. The output current is minimum when Vi = V2 . The bottom
curve shows the output when Vi = V2 .

depend on S, it is important that S should be correct. Obtaining the correct

S value is not easy since S depends on two transistor effects: The short
and narrow channel effects, which arise from the electrostatics of the fields
in the channel3. These effects make the channel dimensions behave smaller
than they are drawn. For logic designs, the channel effects show up primarily
as shifts of the threshold voltage: Minimum-length transistors have a lower
threshold voltage than longer ones, and minimum-width transistors have a
higher threshold voltage than wider ones. The threshold may vary by as much
as several hundred millivolts. In subthreshold operation, a threshold voltage
decrease of 100 mV is equivalent to a 10-fold increase in 10, The best method
we have found to estimate the actual threshold shift for a given geometry and
operating point is to use the most accurate SPICE transistor model available
(presently BSIM3v3) to carefully simulate the design. Even then, the models
are not very good at predicting the actual value of S. Figure 6.14 shows

3 Sec. 13. 1 shows process formation of the bird's beaks and lateral diffusion of the field implant
that lead to the narrow channel effect.
Current-Mode Circuits 175

measured and calculated S values for a number of different bump circuits built
in a MOSIS 2 f..tm process.
The short and narrow channel effects are affected by back-gate bias. The
back-gate voltage acts differentially on the short and narrow channel effects;
consequently S changes with operating point. Figure 6.15 shows this effect:
this effective S increases with common-mode voltage. This effect can be used
to advantage for tuning the response if one has control over the common mode
input voltage.
7 Analysis and Synthesis of Static Translinear Circuits

A Short History of Translinear Circuits In 1975, Barrie Gilbert coined

the word translinear to describe a class of circuits whose large-signal behavior
hinges on the extraordinarily precise exponential current-voltage characteristic
of the bipolar transistor; and the intimate thermal contact and close matching of
monolithically integrated devices (Gilbert, 1975). The unusual functions per­
formed by these circuits-including multiplication (Paterson, 1963; Gilbert,
196 8a,c; Briiggemann, 1970; Schlotzhauer and Viswanathan, 197 2; Gilbert,
197 4), wideband signal amplification (Gilbert, 196 8a,b), and various power­
law relationships (Barker and Hart, 197 4)-were utterly incomprehensible in
terms of linear current amplifiers composed of bipolar transistors. At the same
time, Gilbert enunciated a general circuit principle, the translinear principle
(TLP), whereby the (steady-state) large-signal characteristics of these circuits
can be quickly analyzed, usually with only a few lines of algebra, by consider­
ing only the currents flowing in the circuits.
The word translinear refers to the exponential current-voltage character­
istic of the bipolar junction transistor that is central to the functioning of these
circuits: The bipolar junction transistor's transconductance is linear in its col­
lector current. Gilbert also used the word to refer to analysis and design tech­
niques (for example, the translinear principle) that bridge the gap between the
well-established domain of linear-circuit design and the largely uncharted do­
main of nonlinear-circuit design, for which precious little can be said in gen­
eral (Gilbert, 1990, 1993, 1996). As we shall see in Sect. 7.3, the translin­
ear principle is a translation of a linear constraint on the voltages in a circuit
(Kirchhoff 's voltage law), into a product-of-power-law constraint on collector
currents flowing in the circuit, using the exponential current-voltage relation­
ship.
Since Gilbert coined the word translinear, the translinear principle has
been the basis of a plethora of useful nonlinear circuits, including wideband
analog multipliers (Gilbert and Holloway, 19 80; Huijsing et al., 19 8 2; Gilbert,
19 83), translinear current conveyors (Fabre, 19 8 4, 19 85; Normand, 19 85),
translinear frequency multipliers (Ashok, 1976a; Genin and Konn, 1979; Konn
and Genin, 1979; Surakampontorn et al., 19 88), operational current ampli­
fiers (Fabre, 19 86; Fabre and Rochegude, 19 87; Fabre, 19 8 8; Toumazou et al.,
19 89), RMS-DC converters (Gilbert and Counts, 1976; Seevinck et al., 19 8 4;
Wassenaar et al., 19 88; Frey, 1996a; Mulder et al., 1996, 1997c), and vector-
178 Chapter 7

magnitude circuits (Ashok, 1976b; Gilbert, 1976; Doorenbosch and Goinga,

1976; Wong et al., 19 83; Seevinck et al., 19 8 4). In the 19 80s, Seevinck ( 19 8 1,
19 88) made significant contributions to translinear-circuit design by develop­
ing systematic techniques for the analysis and synthesis of these circuits.
Since the mid- 1990s, there has been a growing interest in translinear cir­
cuits, primarily because of the development of the class of dynamic translinear
circuits, which had its origins in 1979 in the work of Adams ( 1979). Although
he does not appear to have made a connection between his own ideas and the
growing body of work on translinear circuits, Adams proposed a method of
implementing large-signal-linear, continuous-time filters using linear capaci­
tors, constant current sources, and translinear devices. He called this method
log-domainfiltering, because all the filtering occurred on log-compressed volt­
age state variables using translinear devices. The concept of log-domain filter­
ing remained in obscurity for over a decade, only to be independently redis­
covered by Seevinck. In 1990, Seevinck presented a first-order filter, which
he dubbed a companding current-mode integrator (Seevinck, 1990). Unfortu­
nately, it appears that neither Seevinck nor Adams had a clear idea of how to
generalize their ideas to implement filters of higher order. In 1993, encouraged
by Adams to pursue the idea of log-domain filtering, Doug Frey introduced a
general method for synthesizing log-domain filters of arbitrary order using a
state-space approach; and he presented a highly modular technique for imple­
menting such filters (Frey, 1993).
Jan Mulder et al. coined the phrase dynamic translinear circuits and
made the clearest connection between translinear circuits and log-domain fil­
ters (Serdijn et al., 1997b; Mulder et al., 1997b; Serdijn et al., 1999; Mul­
der et al., 1999). They extended Seevinck's translinear analysis and syn­
thesis methodology to encompass dynamic constraints based on what they
called the dynamic translinear principle, with which capacitive currents em­
bedded within translinear loops can be expressed directly in terms of prod­
ucts of the currents flowing through translinear devices, and their time deriva­
tives. Dynamic translinear circuit techniques have been successfully applied to
the structured design of both linear dynamical systems (e.g., log-domain fil­
ters (Frey, 1993, 1996b; Perry and Roberts, 1996; Liu, 1996; Drakakis et al.,
1997; Mulder et al., 1997a; Drakakis et al., 199 8, 1999; Enz and Punzen­
berger, 1999)) and nonlinear dynamical systems (for example, RMS-DC con­
verters (Frey, 1996a; Mulder et al., 1996, 1997c), oscillators (Pookaiyaudom
and Mahattanakul, 1995; Thanachayanont et aI., 1995; Serdijn et al., 1997a,
199 8), phase detectors (Payne and Thanachayanont, 1997), and phase-locked
Analysis and Synthesis of Static Translinear Circuits 179

loops (Thanachayanont et al., 1997; Payne et al., 199 8)). Although dynamic
translinear circuits are beyond the scope of this chapter, the principles that we
shall develop in this chapter directly extend to, and fonn a foundation for, the
analysis and synthesis of this emerging class of circuits.

7.1 The Ideal Translinear Element

Figure 7.1(a) shows a circuit symbol for an ideal translinear element (TE).
This symbol, has a gate, an emitter, and a collector. It is commonly used in
power electronics to represent an insulated-gate bipolar transistor (IGBT),
which is a hybrid bipolar/MOS device that combines the high input impedance
of an MOSFET and the larger current-handling capabilities of a power bipolar
transistor. Although it may be possible to build translinear circuits with IGBTs,
that possibility is not what we are presently considering. Instead, we use the
circuit symbol shown in Fig. 7.1(a) for two reaSOnS: Firstly, the ideal TE
should have the nearly inviolate exponential current-voltage relationship of the
bipolar transistor and the infinite input impedance of the MOSFET; the hybrid
symbol of Fig. 7.1(a) is highly suggestive of precisely this mixture of bipolar
and MOS qualities. Secondly, even though translinear circuits were originally
implemented with bipolar transistors, they can also be implemented using
subthreshold MOSFETs. By using a symbol for the ideal TE that resembles
both types of transistors, we remind ourselves continually of these properties.
We shall assume that the ideal TE produces a collector current I that is
exponential in its gate-to-emitter voltage V and is given by

(7.1.1)

where Is is a pre-exponential scaling current, >. is a dimensionless COnstant

that scales Is proportionally, 'fJ is a dimensionless COnstant that scales the gate­
to-emitter voltage V, and UT is the thennal voltage kT/ q. To demonstrate that
the ideal TE is translinear in the sense that its transconductance is linear in
its collector current, we can calculate its transconductance by differentiating
Eq. 7.1.1 with respect to V:

9 =
8I = >'1s e",V/UT x ..!l..- =
'fJI
m 8V UT UT·

Figures 7.1(b) through 7.1(e) show five practical circuit implementations

of the ideal TE. The first of these TEs is the pn junction diode, shown
180 Chapter 7

{y (b)
.. •

��,
�-
(a)
.. •

I / \ (e)

i (e)
i� (d)

Figure 7 .•
Translinear elements (TEs). (a) Circuit symbol for an ideal TE. Such a device produces a current
I that is exponential in its controlling voltage V. Parts (b) through (f) show five practical TE im­
plementations comprising (b) a diode, (c) an npn bipolar transistor, (d) a subthreshold nFET with
its source and bulk connected together, and (e) a compound TE comprising an npn and a pnp with
their emitters connected together. Of course, for the TEs shown in parts (c) and (d), the appropriate
complementary transistors are also TEs.

in Fig. 7.1(b). Although the forward-biased diode does have an exponential

current-voltage characteristic, it is a two-terminal device and does not, strictly
speaking, have a transconductance. Moreover, diodes seldom actually appear
in translinear circuits. Instead, for the sake of device matching, we almost in­
variably use diode-connected transistors in place of diodes. Nonetheless, for
simplicity, many presentations of the translinear principle begin by consider­
ing a loop of diodes. For the diode, A corresponds to the relative area of the pn
junction and"., is typically close to unity.
Most people consider the bipolar transistor, biased into its forward-active
region, to be the quintessential TE (Fig. 7.1(c)). This transistor usually has
a precise exponential relationship between its collector current and its base­
to-emitter voltage, over more than eight decades of current. For the bipolar
transistor, A corresponds to the relative area of the emitter-base junction and".,
is typically close to one. The main limitation of the bipolar transistor as a TE
is its finite base current, which is often what limits the range of usable current
levels in bipolar translinear circuits.
Analysis and Synthesis of Static Translinear Circuits 18 1

The subthreshold MOSFET with its source and bulk connected together,
as shown in Fig. 7.1(d), and biased into saturation, also has an exponential
current-voltage characteristic. In this case, >. corresponds to the W/ L ratio of
the MOSFET and 11 is equal to K,. In the majority of CMOS technologies, we
fabricate one type of MOSFET (either nFET or pFET ) in a global substrate
that is maintained at a single common potential. The other type is fabricated
inside isolated local wells that may be biased at different potentials. In such
technologies, we can only connect the source and bulk together for the type of
transistor that is fabricated inside the well. For instance, in an n-well CMOS
technology, the pFETs are fabricated in n-wells. By fabricating pFETs in sep­
arate wells, their individual sources and bulks can be connected together, and
different TEs can operate simultaneously with different source potentials. All
of the nFETs are fabricated inside a global p-type substrate. Consequently, if
all of the sources and bulks must be shorted together, different TEs cannot have
their sources at different potentials. Fortunately, for certain translinear-loop
topologies, the source and bulk of all of the MOSFETs within the translinear
loops need not be connected together (see Sect. 7.3).
A variety of compound TEs can be constructed by combining two or more
transistors in various ways. Figure 7.1(e) shows one such compound TE, com­
prising an npn transistor and a pnp bipolar transistor with their emitters con­
nected together. For this TE, the controlling voltage is the voltage difference
between the base of the npn and that of the pnp; the output current is available
at the collector of either transistor. In this device >. corresponds to the geomet­
ric mean between the relative emitter area of the npn and the relative emitter
area of the pnp and 11 = �.

7.2 Translinear Signal Representations

Translinear circuits are current-mode circuits: Their input and output signals
are represented as currents. More precisely, we represent a dimensionless
quantity z as the ratio of a signal current, Iz, to a unit current, Iu. In other
words, we define a number:
Iz
z .
== Iu

We call Iu the unit current precisely because it is the current level that repre­
sents unity in our number system: z = 1 if and only if I z = Iu. The value
of the unit current ultimately determines the power dissipation, computational
182 Chapter 7

throughput, and the precision, of a translinear analog infonnation-processing

system. By allowing the unit current level to change over time, we can build
computing systems that can adaptively trade computational throughput and
precision, for power dissipation (Andreou and Boahen, 199 4).
The representation of signals by currents in translinear circuits is anal­
ogous to a floating-point number representation in the digital domain. For
translinear circuits, current signals span a large dynamic range (typically sev­
eral decades), but noise restricts the precision available at any given signal­
current level to a certain finite fraction of that current level. This fraction is
characterized by the noise-to-signal ratio (NSR). The manner in which the
noise current scales with the signal current depends on the type of noise that
dominates the system (Sarpeshkar, 1998). If white noise dominates, then NSR
scales inversely with the square root of the signal current; so we can trade
power consumption for precision. On the other hand, if 1/ f noise dominates,
then the NSR is roughly independent of the signal-current level; so we cannot
obtain additional precision by increasing power consumption. In the digital do­
main, floating-point numbers also span a large dynamic range, but because the
mantissa is represented by a finite number of bits, the precision for any given
exponent is limited to a constant fraction of two raised to that exponent.
Usually voltages in translinear circuits are considered to be irrelevant.
However, these voltage signals are implicitly logarithmic representations of
the various input and output signals; and so they are directly analogous to
logarithmic number systems in the digital domain (Mitchell, 196 2; Swartzlan­
der and Alexopoulos, 1975; Edgar and Lee, 1979; Arnold et al., 1990), where
numbers are represented internally by their binary logarithms to some finite
precision. Many researchers (Hall et al., 1970; Kingsbury and Rayner, 197 1;
Swartzlander et al., 19 83; Lang et al., 19 85; LaMaire and Lang, 19 86; Vainio
and Neuvo, 19 86; Taylor et al., 19 8 8; Kurokawa and Mizukoshi, 199 1; Lai,
199 1; Lai and Wu, 199 1; Yu and Lewis, 199 1; Lewis, 1995) have proposed
using logarithmic number systems in special-purpose digital signal processors
for applications such as digital filtering, fast Fourier transfonn computation,
and computer graphics; which typically require a large dynamic range, and in
which operations such as multiplication, division, squaring, and square-rooting
occur more frequently than do addition and subtraction.
The appeal of such logarithmic number systems is that operations of mul­
tiplication and division can be implemented using only fixed-point adders; and
the operations of squaring and square-rooting using only shifters. However,
in such digital arithmetic units, conversion between conventional number for-
Analysis and Synthesis of Static Translinear Circuits 183

mats and the logarithmic number format is quite expensive, typically involving
large lookup tables. Also, whereas multiplication and division are relatively
inexpensive in such systems, addition and subtraction can only be approxi­
mated and involve additional lookup tables, making them quite cumbersome.
By contrast, in translinear analog information-processing systems, conversion
between the logarithmic (that is, voltage) signal representation and the linear
(that is, current) signal representation is extremely inexpensive, requiring only
a single translinear device. Consequently, with translinear circuits, operations
like multiplication, division, squaring, and square-rooting are inexpensive and
the operations of addition and subtraction, which we can implement simply
using Kirchhoff 's current law on a single wire, are also inexpensive.
Because of the logarithmic relationship that exists between the controlling
voltage and the output current of a translinear device, we must ensure that the
currents flowing through all translinear devices remain strictly positive at all
times. In order to represent both positive and negative quantities by currents,
we can follow one of two basic approaches. Firstly, we can add an offset
current Iy to a signal current Iz so that their sum Iz + Iy remain positive at all
times, as shown in Fig. 7.2(a). In this case, the signal current is a bidirectional
current. Note that the condition that Iz + Iy > 0 implies that Iz > -
Iy
.

Thus, while there is no restriction on the magnitude of positive values of I z ,

negative values of Iz cannot exceed the magnitude of the offset current I y .
Secondly, we can represent a quantity that can be both positive and negative as
a differential current: The difference between two strictly positive components,
each of which is represented independently as the ratio of a signal current to
the unit current, as shown in Fig. 7.2(b). In other words, we represent a number
z == z+ - z-, where

7.3 The Translinear Principle

In this section, we shall derive the translinear principle for a loop of ideal TEs
and illustrate its use in analyzing translinear circuits. We shall then consider a
loop of subthreshold MOSFETs with their bulks all connected to the common
substrate potential to determine how the translinear principle is modified by
the body effect.
184 Chapter 7

(a)

Input Output

(b)

V;o--1� )�v;
z
,+ ,z-
,z+ ,z-

Input Output

Figure 7.2
Translinear representations for quantities that can take on both positive and negative values. (a)
The quantity z is represented by a bidirectional current lz offset by another current so thatIy.
Iz + Iy > O. (b) The quantity z is represented by a differential current lz == I't - I;. where
I't > 0 and I; > O.

Translinear Loops of Ideal Translinear Elements

Consider the closed loop of N ideal TEs, shown in Fig. 7.3. The large arrow
shows the clockwise direction around the loop. If the emitter arrow of a TE
points in the clockwise direction, we classify that TE as a clockwise element. If
the emitter arrow of a TE points in the counterclockwise direction, we classify
that TE as a counterclockwise element. We denote by CW the set of c1ockwise­
element indices, and by CCW the set of counterclockwise-element indices.
As we proceed around the loop in the clockwise direction, the gate-to­
emitter voltage of a counterclockwise element corresponds to a voltage in-
Analysis and Synthesis of Static Translinear Circuits 185

Figure 7.3
A conceptual translinear loop comprising N ideal TEs. The large arrow indicates the clockwise
direction around the loop. If a TE symbol's emitter arrow points in the direction opposite to that
of the arrow, then we consider the element a counterclockwise element. If a TE symbol's emitter
arrow points in the same direction as the large arrow, then the element is a clockwise element. The
translinear principle states that the product of the currents flowing through the clockwise elements
is equal to the product of the currents flowing through the counterclockwise elements.

crease, whereas the gate-to-emitter voltage of a clockwise element corresponds

to a voltage drop. One way of stating Kirchhoff 's voltage law is that the sum
of the voltage increases around a closed loop is equal to the sum of the volt­
age drops around the loop. Consequently, by applying Kirchhoff 's voltage law
around the loop of TEs shown in Fig. 7.3, we have that

(7.3.1)
nECCW nECW
By solving Eq. 7.1.1 for V in terms of I and substituting the resulting expres­
sion for each Vn in Eq. 7.3.1, we obtain

(7.3.2)
186 Chapter 7

Assuming that all TEs are operating at the same temperature, we can cancel
the common factor of UT fry in all the terms of Eq. 7.3.2 to obtain

(7.3.3)

Because log x + log y = log xy, Eq. 7.3.3 can be rewritten as

log II In
= l og II
In
(7.3.4)
nECCW A n 1S
nECW n s
>:T.
Exponentiating both sides of Eq. 7.3.4 and rearranging yields

IIIn -
-
JSNccw-Ncw II In
(7.3.5)
A
nECCW n A
nECW n
where Nccw and Ncw denote the number of counterclockwise elements,
and the number of clockwise elements respectively. It is easy to see that, if
Ncw = Nccw, then Eq. 7.3.5 reduces to

(7.3.6)

which has no remaining dependence on temperature or device parameters.

Equation 7.3.6 is the translinear principle, which can be stated as follows:

In a closed loop of ideal TEs comprising an equal number of clockwise and counter­
clockwise elements, the product of the (relative) current densities flowing through the
counterclockwise elements is equal to the product of the (relative) current densities
flowing through the clockwise elements.

Further, if each TE in the loop has the same value of A, then Eq. 7.3.6 reduces
to

(7.3.7)
nECCW nECW
Equation 7.3.7 is an important special case of the translinear principle that can
be stated as follows:

In a closed loop of identical ideal TEs comprising an equal number of clockwise and
counterclockwise elements, the product of the currents flowing through the counter­
clockwise elements is equal to the product of the currents flowing through the clockwise
elements.
Analysis and Synthesis of Static Translinear Circuits 187

This derivation of the translinear principle can be characterized as a

translation of a linear set of algebraic constraints on the voltages in the cir­
cuit (that is, Kirchhoff 's voltage law applied around the loop of Fig. 7.3) into
a product-of-power-Iaw constraint on the currents flowing in the circuit. This
characterization of the translinear principle is one way to state the second sense
of the word translinear as originally used by Gilbert ( 1990, 1993, 1996, 1975).

(a) (b)

Figure 7.4
Two translinear-loop circuits. (a) A simple circuit with one translinear loop comprising two
clockwise elements and two counterclockwise elements arranged in a stacked topology. (b) A
circuit with two overlapping translinear loops each of which comprises two clockwise elements
and two counterclockwise elements arranged in a stacked topology.

To illustrate the use of the translinear principle, we shall analyze the

two translinear-loop circuits shown in Fig. 7.4. First, consider the circuit of
Fig. 7.4(a). This circuit has a single translinear loop comprising four identi­
cal TEs, two of which face in the clockwise direction and two of which face
in the counterclockwise direction. Input current II passes through both COun­
terclockwise elements. Input current 12 passes through One of the clockwise
elements. The output current 13 passes through the other clockwise element.
Consequently, to analyze this circuit, we apply the translinear principle, as
stated in Eq. 7.3.7, and write that
188 Chapter 7

which can be rearranged to obtain

13 = I>
12

Thus, the circuit of Fig. 70 4(a) is a squaring-reciprocal circuit.

Next, consider the circuit shown in Fig. 7 o 4(b). This circuit has two over­
lapping translinear loops, each of which comprises four identical TEs with
two in the clockwise direction and two in the counterclockwise direction. By
Kirchhoff 's current law, we have that the output current 1 3 is given by

(7.3.8)

In the first loop, input current IIpasses through both counterclockwise ele­
ments. Intermediate current Ix passes through the first clockwise element and
the output current 13 passes through the second clockwise element. So, by the
translinear principle,

Ix =....!.
which implies that
1 2
. (7.3.9)
13
By a similar argument involving the second translinear loop,

Iy = I13i. (7.3.10)

Substituting Eqs. 7.3.9 and 7.3.10 into Eq. 7.3.8 and solving for 1 3:

13 = Jrf + Ii·
Thus, the circuit of Fig. 7 o 4(b) computes the length of a two-dimensional
vector. This circuit can be extended to handle an arbitrary number of inputs
and could be useful to compute vector magnitudes.

Translinear Loops of Subthreshold MOSFETs

Consider the closed loop of N saturated subthreshold MOSFETs whose bulks

are all connected to a common substrate potential, shown in Fig. 7.5(a). Here,
Vn represents the gate-to-source voltage of the nth MOSFET, and Un is the
voltage on the nth node relative to the substrate potential. Again, the large
Analysis and Synthesis of Static Translinear Circuits 189

(a) 14 (b)
In
s Us� s"Un
14 15

�
-� l V
15
u"_,�
�};-

U6 In
V' Clockwise
'
element

l��nU"
( c)

Un-1 --;';vy
11

Counterclockwise
element

Figure 7.5
A translinear loop of subthreshold MOSFETs with their bulks tied to a common substrate potential.
Here Vn refers to the gate-to·source voltage of the nth MOSFET and u,. refers to the voltage
on the nth node referenced to the common substrate potential. (a) A conceptual translinear
loop comprising N subthreshold MOSFETs with their bulks tied to a common substrate. (b) A
clockwise element is one whose gate·to·source voltage is a voltage drop in the clockwise direction
around the loop. (c) A counterclockwise element is one whose gate-to-source voltage is a voltage
increase in the clockwise direction around the loop.

arrow in Fig. 7.5(a) indicates the clockwise direction around the loop. As
shown in Fig. 7.5(b), we shall consider a clockwise element to be one whose
gate-to-source voltage is a voltage drop in the clockwise direction around the
loop, and we shall consider a counterclockwise element to be one whose gate­
to-source voltage is a voltage increase in the clockwise direction around the
loop( Fig. 7.5(c)).
Recall that the channel current, of a saturated nMOSFET, operating in
subthreshold, is given by
190 Chapter 7

From this equation, if the nth MOSFET is a clockwise element, we have that

In = )..nIOe(KUn-l-Un)IUT

(eUn-t/UT ) K ( )..;:0 )
which can be rearranged to yield

eUnlUT = . (7.3.1 1)

Equation 7.3.1 1 expresses a recurrence relationship between the nth node

voltage to the (n - l)st node voltage for clockwise elements. On the other
hand, if the nth MOSFET is a counterclockwise element, we have that

which can be rearranged:

eUnlUT = (eUn-t/UT )11K ( )..In�o ) 11K (7.3.1 2)

Equation 7.3.1 2 likewise expresses a recurrence relationship between the nth

node voltage and the (n - 1)st node voltage for counterclockwise elements.

Uo

(a) (b)

Figure 7.6
Two subthreshold MOS translinear loops comprising two clockwise transistors and two counter­
clockwise transistors. In each case, the bulks of all four transistors are tied to a common potential.
(a) A stacked loop topology. (b) An alternating loop topology.

We can use the recurrence relationships, expressed in Eq. 7.3.1 1 and

Eq. 7.3.1 2, to build up the translinear-loop constraint equation for the sub­
threshold MOS translinear loop (Fig. 7.5(a», as follows. We begin at one of
the nodes in the loop, say Uo, and proceed sequentially around the loop in the
Analysis and Synthesis of Static Translinear Circuits 19 1

clockwise direction, recursively applying Eq. 7.3.1 1 or Eq. 7.3.1 2 to get to the
next node, depending on whether the current element is clockwise or coun­
terclockwise. When we encounter a clockwise element, we raise the partially
formed translinear-loop equation to the", power and multiply it by A n IO / In , as
expressed in Eq. 7.3.1 1. W hen we encounter a counterclockwise element, we
raise the partially formed translinear-loop equation to the 1/", power and mul­
l
tiply it by (In / Anlo) /K , as expressed in Eq. 7.3.1 2. Finally, when we return to
the node from which we started, we stop and simplify the resulting expression.
To illustrate this process, we shall consider two simple subthreshold MOS
translinear loops, as shown in Fig. 7.6. Each of these loops comprises four
transistors, two of which face in the clockwise direction and two of which face
in the counterclockwise direction. The first loop, shown in Fig. 7.6(a), has a
stacked topology: All of the gate-to-source voltage drops are stacked up. The
second loop, shown in Fig. 7.6(b) has an alternating topology: We alternate
between clockwise and counterclockwise elements, as we go around the loop.
First, consider the stacked MOS translinear loop, shown in Fig. 7.6(a).
Starting with node U0 and proceeding around the loop in the clockwise direc­
tion, we encounter two counterclockwise elements followed by two clockwise
elements before we finish back at node Uo. Following the procedure just de­
scribed, we have that

which can be simplified to obtain

By canceling common factors and grouping the counterclockwise currents on

the left-hand side of the equation and grouping the clockwise currents on the
192 Chapter 7

right-hand side of the equation, we obtain the translinear-loop equation:

(7.3.13)

which has no remaining temperature dependence and no dependence on 1 0, but

does depend on the subthreshold slope factor, K-.
Next, consider the alternating MOS translinear loop, shown in Fig. 7.6(b).
Again, starting with node U0 and proceeding around the loop in the clockwise
direction, we first encounter a counterclockwise element followed by a clock­
wise element, followed by another counterclockwise element, followed by an­
other clockwise element. Following the procedure just described, we have that

((((eU"IUTt'v�orrc;:o) r(A��Orr
'"

eUl/UT ..
,

(A;:O ) eUO/UT
=

which can be simplified to obtain

By canceling common factors and grouping the counterclockwise currents on

the left-hand side of the equation and grouping the clockwise currents on the
right-hand side of the equation, we obtain the translinear-loop equation:

(��) (�:) (�:) (�:)

'-v-"
=

'-v-"
(7.3.1 4)

CCW CW
which has no remaining temperature dependence, no dependence on 10, and no
dependence on K-. From these examples, we can make a number of observations
about subthreshold MOS translinear loops. Firstly, if the number of clockwise
Analysis and Synthesis of Static Translinear Circuits 193

elements is equal to the number of counterclockwise elements in a closed loop,

then the eUO/UT factor will cancel on both sides of the equation. In general,
each time a clockwise element is traversed, this factor is raised to the", power;
and each time a counterclockwise element is traversed, this factor is raised by
1 /"'. Thus, if there are New clockwise elements and Nccw counterclockwise
elements, the factor eUO/UT will appear on the left-hand side of the translinear­
loop equation raised to the ",Ncw-Nccw power, which is equal to unity if
New = Nccw. Therefore, this factor, which appears on both sides of the
final translinear-Ioop equation will cancel.
Secondly, if the number of clockwise elements is equal to the number of
counterclockwise elements, then the translinear-Ioop equation will be inde­
pendent of 10. To see why, consider first the power to which 10 is raised af­
ter we traverse a clockwise element followed by a counterclockwise element.
Suppose that 10 appears in the initial translinear-Ioop equation raised to the
0: power. After traversing a clockwise element, lois raised to the "'0: + 1.
Then, after traversing a counterclockwise element, 10 will be raised to the
("'0: + 1) / '" - 1/", = 0: power. Thus, the power to which lois raised in the
translinear-loop equation remains unchanged when we pass from a sequence
of clockwise elements to a sequence of counterclockwise elements. Next, COn­
sider what happens to the power to which lois raised after we traverse a
counterclockwise element followed by a clockwise element. Again, suppose
10 appears in the initial translinear-loop equation raised to the 0: power. After
traversing a counterclockwise element, we will find 10 raised to the 0:/'" - 1 / '"
power. Then, after traversing a clockwise element, we will have 10 raised to
the '" (0:/ '" - 1 / "') + 1 = 0: power. Thus, the power to which lois raised in
the translinear-Ioop equation also remains unchanged when we pass through
a boundary from a sequence of counterclockwise elements to a sequence of
clockwise elements.
Now, if there are an equal number of clockwise and counterclockwise
elements, then there will be at least two such boundaries; this lower bound
is achieved when the loop has a purely stacked topology. If there are 2N
elements in the loop (that is, New = Nccw = N), then there can be at
most N such boundaries; this upper bound is achieved when the loop has a
purely alternating topology. Because the elements at the boundaries between
runs of clockwise and counterclockwise elements do not alter the power to
which 10 is raised in the translinear-Ioop expression, we can omit the elements
at each of these boundaries in determining the overall power to which lois
raised. However, the elements adjacent to those that we have omitted then
194 Chapter 7

fonn another boundary. Consequently, if there is a clockwise element in the

loop for each counterclockwise element, then the overall power to which 10
is raised as we go around the loop remains unaltered. Moreover, when we
begin accumulating the translinear-Ioop expression, 10 does not appear in the
expression (that is, it is raised to the zeroth power). Therefore, because the
power to which lois raised begins at zero and it remains unaltered for a loop
comprising an equal number of clockwise and counterclockwise elements, the
overall translinear loop expression remains independent of 10• An identical
argument holds if all the MOSFETs have the same W/ L ratio (that is, Al =
. .= AN = A): If all of the transistors have the same W/ L ratio, and if
.

the number of clockwise elements is equal to the number of counterclockwise

elements, then the translinear-Ioop equation is independent of the transistors'
common W/ L ratio.
Finally, consider a purely alternating subthreshold MOS translinear loop
comprising an equal number of clockwise and counterclockwise elements with
their bulks connected to a common potential. For such a loop, the translinear­
loop equation will be independent of K,. To see why, consider the loop-equation
construction procedure. If we begin by traversing a counterclockwise element,
the first current factor in the equation will be raised to the 1/ K, power. After
traversing the second element, which by hypothesis is a clockwise element,
the first current factor will be raised to the K, power, canceling the initial 1/ K,.
The second current factor will be added to the product raised to the first power.
After traversing the third element, the first two current factors will again be
raised to the 1/ K, power as will the third current factor. After going through
the fourth element, we raise the first three current factors to the K, power,
again canceling the 1/ K, power on each factor from the previous step. A fourth
current factor is then added to the partially fonned product, raised to the first
power. Thus, in an alternating loop, we alternate between raising each factor
to the 1/ K, power and raising each one to the K, power, effectively removing
the K, dependence from the partially fonned product after every other step.
Moreover, if there are an equal number of clockwise and counterclockwise
elements in such a loop, then it follows that the final translinear-Ioop equation
will be independent of K" because there must be an even number of steps. A
similar argument would hold if we begin with a clockwise element. This result
has been demonstrated previously using other arguments by Vittoz ( 1996) and
by Andreou and Boahen ( 1996).
With these considerations in mind, we can simplify our translinear-Ioop
equation construction procedure for loops that comprise an equal number of
Analysis and Synthesis of Static Translinear Circuits 195

clockwise and counterclockwise subthreshold MOSFETs whose bulks are all

tied to a common potential. First, we do not have to write down the initial
and final eUO/UT factors. Instead, we can replace them by unity, effectively
dividing out the factor that will be common to both sides of the equation ahead
of time. Next, we can dispense with keeping track of all of the 10 factors,
because we have shown that the final expression will be independent of 10• To
summarize this reduced procedure: We pick a starting point in the loop and
begin the translinear loop expression with unity. Then, we proceed around the
loop from node to node in the clockwise direction. If we traverse a clockwise
element, we raise the partially formed expression to the K, power and then we
multiply it by An /In . If we traverse a counterclockwise element, we raise the
partially formed expression to the 1/K, power and multiply by (I n /An)1/"'.
When we arrive back at the starting node, we equate the product that we have
accumulated to unity and simplify it as necessary.

7.4 ABC's of Translinear-Loop-Circuit Synthesis

In this section, we give an overview of the basics of translinear-Ioop circuit

synthesis. As with all synthesis problems, the task of constructing a translinear
circuit that implements some desired functionality is underconstrained. For any
given function, there will be a variety of circuit solutions with design decisions
to be made and many trade-offs to consider. Thus, our brief discussion cannot,
by any means, be exhaustive. Rather, we shall attempt to make the basic
procedure clear and we shall illustrate it with a simple example. In the process,
we shall discuss some design decisions and trade-offs.

Synthesizing Static Translinear-Loop Circuits

The basic procedure for synthesizing translinear-loop circuits can be described

as follows: Firstly, we acquire a set of translinear-Ioop equations from the
relationship(s) that we want to implement. Secondly, we build a translinear
loop for each of the translinear-Ioop equations. Then, we bias each of the
translinear loops. Finally, if possible, we consolidate the resulting circuits by
merging some of the loops. We shall discuss each of these steps in tum, and
then we shall use them to synthesize a simple two-quadrant translinear-loop
multiplier circuit.
196 Chapter 7

Acquiring a Set of Translinear-Loop Equations The starting point for syn­

thesizing a translinear-loop circuit is a static linear or nonlinear mapping be­
tween dimensionless variables. The class of translinear circuits is capable of
embodying a wide range of useful linear and nonlinear relationships. However,
not all functions are directly implementable by translinear circuits; we can di­
rectly realize products, quotients, power-law relationships, polynomials, ratio­
nal functions, and various combinations of such relationships. In many cases,
we may need to find an acceptable approximation for one or more nonlinear
functions in terms of polynomials, rational functions, continued fractions, or
some other suitable mathematical form before we can realize the required rela­
tionship with translinear circuits. In his book on translinear circuits, Seevinck
( 19 8 8) provides a good discussion of suitable approximation techniques and
approximations of various transcendental functions.
Once we have obtained a set of relationships that can be realized with
translinear circuits, we then represent each of the dimensionless quantities as
the ratio of a signal current to the unit current or as the ratio of a differen­
tial current to the unit current, as described in Sect. 7.2. Then, we decompose
the resulting equations into a set of translinear-Ioop equations of the form of
Eq. 7.3.6. In the decomposition process, it is sometimes convenient to intro­
duce intermediate currents, which serve to parameterize some of the relation­
ships, allowing us to further decompose the system of equations.

Building the Translinear Loops Once we have a set of translinear-Ioop

equations, we construct a closed loop of TEs for each one. In general, there will
be more than one translinear loop that implements any given translinear-Ioop
equation. Our choices of loop topology and current ordering must be guided by
experience and other system-level design considerations. For instance, loops
with an alternating topology are better than stacked loops for systems that
require a low power-supply voltage. On the other hand, stacked loops are easier
to bias in the case that the same current must pass through multiple TEs that
face in the same direction. In a stacked loop, we only have to supply a single
copy of the input current to the circuit and we can pass the same current from
one element to the next in a stack of TEs. In an alternating loop, there are
no runs of clockwise or counterclockwise TEs, so if we must pass the same
current through N TEs facing in the same direction, then we must supply N
matched copies of the input current to the circuit.
Analysis and Synthesis of Static Translinear Circuits 197

Biasing the Translinear Loops The process of biasing a translinear loop

involves forcing input currents into the emitter or collector of each input TE
in the loop and arranging some type of local negative feedback around the TE,
thereby adjusting its gate-to-emitter voltage so that the TE passes the input
current. Once again, there are many possible feedback arran gements that will
properly bias each TE in any given loop. In some cases, operational amplifiers
may be used to bias a translinear loop. Here, we shall consider only the three
simplest possibilities, which are shown in Fig. 7.7.
Figure 7.7(a) shows the ubiquitous diode connection. Here, we force a
current into the collector of a counterclockwise TE, which corresponds to
a voltage increase, and we feed the collector voltage back to the gate. Any
mismatch between the input current and the TE's collector current causes the
gate voltage to charge up or down in such a way to reduce the mismatch. If
the collector current is bigger than the input current, the gate will discharge,
reducing the gate-to-emitter voltage and thereby reducing the collector current.
If the collector current is smaller than the input current, the gate will charge
up, increasing the gate-to-emitter voltage, and thereby increasing the collector
current. Of course, we can introduce additional circuitry between the collector
and the gate, such as one or more buffer stages, as long as changes in the
collector voltage induce like changes in the gate voltage. In some cases, we
actually use other TEs in the loop with their input-current sources to form
emitter-follower buffer stages in making a diode connection. In some cases,
we may introduce buffer stages to eliminate base-current errors in bipolar
translinear-loop circuits.
Figure 7.7(b) shows an emitterjollower connection. Here, we force a
current out of the emitter of a clockwise TE, which corresponds to a voltage
drop. In such an arrangement, for any given gate voltage, the emitter voltage
will adjust itself up or down so that the emitter current just balances the input
current. If the emitter current is larger than the input current, the emitter voltage
will charge up, reducing the gate-to-emitter voltage, thereby reducing the
emitter current, until the currents match. If the input current is larger than the
emitter current, the emitter will be discharged, increasing the gate-to-emitter
voltage, and thereby increasing the emitter current until the currents balance.
Figure 7.7(c) shows a simple alternative to the emitter-follower connection
for biasing clockwise TEs. Here, we force a current into the collector of a
clockwise TE and we feed the collector voltage back through another transistor
to adjust the emitter current. An nFET transistor is shown in Fig. 7.7(c), but a
bipolar transistor could be used instead. A similar feedback arrangement (that
198 Chapter 7

In

In

In
Un-1- Vn

Un-1 + Vn
In

Un-1

(a) (b)

(c)

Figure 7.7
Three simple biasing arrangements for TEs. (a) Collector current forcing with the diode connec·
tion. (b) Emitter current forcing with the emitter-follower connection. (c) Collector current forcing
with the Enz-Punzenberger (EP) connection.

is, forcing the collector current by feedback to the emitter terminal) involving
operational amplifiers has been used in biasing translinear circuits for many
years (Gilbert, 1990). The use of this particularly elegant implementation of
such a feedback arrangement in the context of trans linear-circuit biasing was
first proposed by Punzenberger and Enz ( 1996) to bias low-voltage log-domain
Analysis and Synthesis of Static Translinear Circuits 199

filters. Consequently, this connection is referred to as the Enz-Punzenberger

(EP) connection. For a given gate voltage, any imbalance between the collector
current and the input current will cause the feedback transistor to adjust the
gate-to-emitter voltage in such a way as to reduce the imbalance. If the input
current is too large, the collector voltage will increase, causing the feedback
transistor to pull a larger current out of the emitter. Conversely, if the collector
current is too large, the collector voltage will decrease, reducing the amount
of current flowing through the feedback transistor. This decreased current, in
tum, will cause the emitter to charge up, reducing the gate-to-emitter voltage,
and thereby reducing the collector current. Note that if some additional current
I is injected into the emitter node, the collector voltage will adjust itself so
that the feedback transistor sinks both the current flowing through the TE and
the additional current. This feature of the EP connection greatly facilitates the
biasing of translinear loops with an alternating topology.
In general, any translinear loop can be biased using only two of the three
simple arrangements of Fig. 7.7. We can use the diode connection for biasing
counterclockwise TEs, and we can use either the emitter-follower connection
or the EP connection to bias the clockwise TEs. In general, the emitter-follower
connection involves only one node, whereas the EP connection involves two
nodes, which can cause more complicated settling behavior with the EP con­
nection. On the other hand, in an alternating translinear loop, by using the
diode connection to bias the counterclockwise elements and the EP connection
to bias the clockwise elements, we are always forcing collector currents and
taking outputs from collector currents. This uniform use of collector currents
allows us to freely use translinear devices whose emitter current and collector
current are not nearly the same, such as the compatible lateral bipolar transis­
tor (CLBT), which is always available in any CMOS process (Vittoz, 19 83).
Such devices have at least two collectors, a lateral collector and a parasitic
vertical collector, each of which collects only a finite fraction of the emitter
current.

Consolidating the Circuit In some cases, after we have biased each of the
translinear loops in the circuit, we will recognize some redundancy between
the loops in the circuit. For example, if two TEs in different loops pass the same
current and are at the same voltage level, then these devices are redundant and
may be shared between the loops. Such consolidation is a good idea, because it
usually results in smaller circuits, and fewer opportunities for errors resulting
from device mismatch.
200 Chapter 7

Synthesis of a Two-Quadrant Translinear Multiplier

Suppose that we want to implement a circuit that multiplies two quantities, x

and y. Further, suppose that x can be either positive or negative and that y is
strictly positive. Their product
z = xy (7.4.1)

can be either positive or negative. We shall represent y by I y /Iu and we use a

differential representation for x and z, as described in Sect. 7.2: We represent
x by

where x+ = It /Iu and x- = I;; /Iu . Likewise

z = z+ - z-
where z+ = Ii /Iu and z- = I; /Iu.
Next, we substitute these definitions for x, y, and z into Eq. 7.4.1 to get
( Ii _ I; ) ( It
=
_ I;; ) ( Iy )
Iu Iu Iu Iu Iu
which we can rearrange to obtain

Iuli - luI; = Iylt - IyI;; .

One straightforward way to decompose this equation into a pair of
translinear-Ioop equations is to equate individually the positive and negative
terms on each side of the equation. Using this decomposition, we obtain the
pair of translinear-loop equations

Iuli = Iylt and (7.4.2)

'-v-' '-v-'
CW CCW
Figure 7.8(a) shows a pair of translinear loops with alternating topologies
that implements Eq. 7.4.2. Figure 7.8(b) shows one possible biasing arrange­
ment for these loops; we use diode connections for each of the counterclock­
wise elements, and EP connections for each of the clockwise elements. Again,
the exact value of the voltage Vref is not critical: It should be high enough
allow an adequate swing on the common-emitter node while keeping the feed­
back transistor acting properly. Note that in each loop, there is a copy of I y
forced into a diode-connected TE whose emitter is fixed at Vref. Thus, we
can eliminate one of these stages and connect the diode-voltage to both loops.
Analysis and Synthesis of Static Translinear Circuits 20 1

Additionally, we supply a copy of Iu to both loops, and as we just observed,

the gate of both TEs will be at the same potential. Thus, one of the I u states
can also be eliminated. The consolidated two-quadrant multiplier is shown in
Fig. 7.8(c).

(a)

(b)

(c)

Figure 7.8
Synthesis of a two-quadrant translinear multiplier based on two alternating translinear loops. (a)
A pair of alternating translinear loops that implement luIt = IyIt and luI; = IyI;. (b)
A biasing scheme based on collector-current forcing with diode connections and EP connections.
(c) The final consolidated two-quadrant translinear multiplier circuit. Here the.t and Iu circuitry
could be shared between the two translinear loops.
202 Chapter 7

7.5 The Multiple-Input Translinear Element

Inspired originally by Shibata and Ohmi's neuron MOSFET concept (Shibata

and Ohmi, 199 2), we recently introduced a new translinear circuit primitive,
called the multiple-input translinear element (MITE) (Minch, 1997; Minch
et al., 199 8). Such an element produces an output current I that is exponential
in a weighted sum of its K input voltages, Vi through VK :

(7.5.1)

where Vk is the kth input voltage, W k is a dimensionless positive weight

that scales Vk , and U is the emitter voltage of the MITE. Here Is, >., and
UT are the same as for the ideal TE. So defined, the MITE has K different
transconductances, each of which is linear in the MITE's output current.
Figure 7.9(a) shows a circuit symbol for an ideal K-input MITE. This
symbol looks like an ideal TE whose gate voltage is set by a K -input capacitive
voltage divider, where the kth divider ratio is denoted by W k . The inputs do
not need to be capacitive, but we shall assume that the input terminals draw
a negligible amount of DC current and that we can control the values of the
weights proportionally. In many cases, we shall be interested primarily in the
number of identical unit weights, each with value w, coupling an input voltage
into a MITE rather than the actual weight values involved. In such cases, each
weight is an integer multiple of W and we shall omit the W associated with each
of the inputs.
Figures 7.9(b) through 7.9(g) show six different practical circuit imple­
mentations of the MITE. For the first of these MITEs, shown in Fig. 7.9(b), we
use a resistive voltage divider to implement the weighted voltage summation,
and a bipolar transistor to implement the exponential voltage-to-current trans­
formation. In this case, the weight associated with each input is proportional
to the conductance through which that input couples into the base of the bipo­
lar. Here, we must buffer the input voltages into the resistive network: so the
network: neither supplies current to nor sinks current from the input nodes.
In subthreshold, the drain current of the K -input floating-gate MOS (FG­
MOS) transistor is proportional to the exponential of a weighted sum of its
K control-gate voltages (Andreou and Boahen, 199 4). Consequently, a MITE
can be implemented using a single subthreshold FGMOSFET (FGMOS tran­
sistor), as shown in Fig. 7.9(c). In this case, the weight of each input is propor­
tional to the capacitance through which that input couples into the floating gate.
Analysis and Synthesis of Static Translinear Circuits 203

�
(9)
I

�
(I)

Figure 7.9
Multiple-input translinear elements (MITEs). (a) Circuit symbol for an ideal K -input MITE. Such
an element produces an output current that is exponential in a weighted sum of its input voltages.
Parts (b) through (g) show six different MITE implementations comprising (b) a resistive voltage
divider and a bipolar transistor, (c) a single subthreshold floating-gate MOS (FGMOS) transistor,
(d) a cascoded subthreshold FGMOSFET, (e) a subthreshold FGMOSFET and a bipolar transistor,
(f) a floating-gate source follower and a subthreshold MOSFET, and (g) a floating-gate source
follower and a bipolar transistor. For each of the five FGMOS MITE implementations, shown
in parts (c) through (g), we can use the amount of floating-gate charge to store electronically
adjustable, non-volatile multiplicative scale factors that we can use to build adaptive information­
processing systems, or to compensate for device mismatch.

The subthreshold FGMOSFET is a good MITE implementation over the entire

range of subthreshold currents. The main limitation of the subthreshold FG­
MOSFET as a MITE is the existence of a parasitic gate-to-drain capacitance,
which results from a small region of overlap between the polysilicon gate and
the drain diffusion region that arises during processing. Because the gate of a
FGMOSFET is floating, an increase in the drain voltage couples into the float­
ing gate through this overlap capacitance, thereby increasing the subthresh­
old drain current exponentially. In principle, this coupling can be decreased
by making the FGMOSFET narrower (thereby decreasing the overlap capac­
itance); or by making the control-gate capacitances larger (thereby increas­
ing the total floating-gate capacitance and so decreasing the drain capacitive­
divider ratio); or by using both techniques. However, in practice, neither of
204 Chapter 7

these solutions are attractive. A better solution to this problem is to cascode

the subthreshold FGMOSFET, as shown in Fig. 7.9(d). We can think of the
cascode transistor as a source follower with a constant input voltage Vcas: It
reduces the swing of the drain voltage of the FGMOSFET (that is, the source
follower's output voltage), and so decreases the change in current through both
transistors. The cascoded subthreshold FGMOSFET is an excellent MITE im­
plementation over the subthreshold range of currents.
Figure 7.9(e) depicts a two-transistor MITE comprising a K-input sub­
threshold FGMOSFET and a bipolar transistor. Intuitively, this bipolar­
FGMOS MITE works as follows: The subthreshold FGMOSFET produces
a current that is exponential in the weighted sum of the input voltages; again,
the weight of each input is proportional to the capacitance through which
that input couples into the floating gate. The bipolar transistor then acts as
a current-gain stage by multiplying the subthreshold FGMOSFET current by
the bipolar's forward current gain. Consequently, the upper end of the current
range over which this two-transistor circuit is a good MITE implementation
is extended beyond that of the subthreshold MOSFET by the bipolar's current
gain, f3. Because the drain of the FGMOSFET is held at a fixed potential, this
MITE is insensitive to the parasitic drain-overlap capacitance.
The final two MITEs, shown in Figs. 7.9(t) and 7.9(g), are similar: Each
comprises a two-transistor FGMOS source follower and a third transistor that
has an exponential current-voltage characteristic. Intuitively, the floating-gate
voltage develops as a weighted sum of the K input voltages via a capacitive
voltage divider. In the source-follower configuration, the FGMOSFET's source
voltage is approximately a linear function of the floating-gate voltage. Con­
sequently, the source voltage is also a weighted sum of the input voltages.
The third transistor then generates a current that is exponential in this source
voltage. In the MITE of Fig. 7.9(t), the exponential element is a subthreshold
MOSFET: Whereas, in that of Fig. 7.9(g), the exponential element is a bipolar
transistor. Because the drains of the FGMOSFETs are held at a fixed potential,
these MITEs also do not suffer from the drain-overlap capacitance problem.
The source-follower circuit configuration does not depend on the fonn of
the current-voltage relationship of the MOSFET. Consequently, these three­
transistor circuits are good MITE implementations even when the FGMOS
source follower is biased with an above-threshold current. For the circuit of
Fig. 7.9(f), biasing the FGMOS source follower with an above-threshold cur­
rent allows us to make the output MOSFET as wide as necessary to obtain a
larger range of exponential currents, without having to make the FGMOSFET,
Analysis and Synthesis of Static Translinear Circuits 205

and hence the floating-gate capacitance, large. The above-threshold bias gives
the FGMOS source follower enough bandwidth to drive the large gate capac­
itance of a wide output MOSFET. The circuit of Fig. 7.9(g) is a good MITE
implementation only when the base current is negligible compared with the
source-follower bias current. Thus, biasing the FGMOS source follower with
above-threshold currents allows this MITE to operate at high current levels and
so with high bandwidth.

7.6 Multiple-Input Translinear Element Networks

In this section, we introduce three basic circuit stages, each constructed from
a single MITE. These three circuit stages are the bricks from which we build
a class of low-voltage translinear circuits, which we call MITE networks, that
are equivalent to the class of translinear-Ioop circuits. Then, we shall examine
how we can compose these stages to make translinear circuits.

Basic MITE Circuit Stages

Consider the three basic MITE circuit stages that are depicted in Fig. 7.10.
The first of these circuits is a voltage-in, current-out (VICO) stage, shown in
Fig. 7.IO(a). Here, we apply input voltages Vi and Vk to two different input
terminals of Q n , which generates an output current I n. To see how In depends
on Vi and Vk , we use Eq. 7.5.1 to write

(7.6.1)

The second of the three basic MITE stages, shown in Fig. 7.1 O(b), is
a current-in, voltage-out (CIVO) stage. Here, we source an input current I i
into the output of Q i , and we feed the output voltage Vi back through the
self-coupling weight Wii. This feedback configuration adjusts Vi, so that the
current sunk by Q i just balances the input current Ii. A MITE in this feedback
configuration is analogous to a diode-connected transistor; so we say that it is
diode connected through Wii. To determine how the output voltage Vi depends
On the input current Ii , we begin with Eq. 7.5.1 and solve for Vi in terms of Ii :

which can be rearranged to obtain

UT
Vi = - log Ii - •
(7.6.2)
Wii
. . .
206 Chapter 7

C a use Ca use
/ /
Effect '; ';

C a use Effect

I�
V; W --i-",-<_-<> V
I-W
;
�
V;

VK OWnk I
!
\Cause: Effect

(a) (b) ( c)

Figure 7.10
Three basic circuit stages. each comprising a single MITE. (a) A voltage-in. current-out (VICO)
stage. (b) A current-in. voltage-out (CIVO) stage. (c) A voltage-in. voltage-out (VIVO) stage.

The third basic MITE stage is a voltage-in, voltage-out (VIVO) stage,

shown in Fig. 7.1O(c). This configuration is identical to the CIVO stage of
Fig. 7.1O(b), except that we now hold the current I i fixed, and we are concerned
instead with how the output voltage Vi depends on an input voltage Vj , which
we apply to another of the input terminals of Q i. Beginning with Eq. 7.5.1, we
write

which can be rearranged to solve for Vi in terms of Vj :

Vi =
Wij Vj . . . . (7.6.3)
Wii
- - -

We can use the circuit stage of Fig. 7.1O(c) both as a CIVO stage and as a
VIVO stage simultaneously. In this case, it is easy to see that Vi depends on Vj
and Ii through a linear combination of of Eqs. 7.6.2 and 7.6.3 as follows:
UT l I Wi
Vi = og i -j Vj . . . . (7.6.4)
Wii Wii
- - -

Elementary MITE Networks

In this section, we shall examine two simple current-mode MITE circuits,

each comprising two CIVO stages and a single VICO stage. These two basic
current-mode circuits illustrate all of the basic intuition behind the operation
Analysis and Synthesis of Static Translinear Circuits 207

Ik Ii

Vk Vi In

'�
Ik

I (a)
I;
Ij Ii

Vi In
�
Ij

I; (b)
I,
Figure 7.11
Two basic current-mode circuits comprising two CIVO stages and one VICO stage. These two
circuits illustrate all of the intuition underlying the class of MITE networks. (a) A product-of­
power-law circuit. (b) A quotient-of-power-Iaw circuit.

of MITE networks.
In the first current-mode circuit, shown in Fig. 7.1 1(a), the outputs of two
different CIVO stages are connected directly to a single VICO stage through
separate inputs. To analyze this circuit, we apply Eq. 7.6.1 to the output stage:

(7.6.5)
208 Chapter 7

Substituting Eq. 7.6.2 into Eq. 7.6.5 for each of Vi and Vk , we obtain

[ wni (UT- )] [ wnk ( UT log Ik - . . . )]

In ex: exp -

UT Wii log Ii - UT Wkk

. . . exp -- -- .

Using the first tenn in each of the two summations and regrouping, we obtain

[UT Wni ] [UT Wnk ]

Inex: exp
UT Wii log Ii UT Wkk log Ik .
TT - exp TT - (7.6.6)

Note that, if MITEs Q i , Q k , and Q n are operating at the same temperature,

then the primary temperature dependence of the relationship among I i , Ik ,
and In disappears from Eq. 7.6.6. In this intuitive analysis, we have not kept
track of the scaling currents Is, which can be strongly temperature dependent.
As we have demonstrated previously (Minch, 1997), when properly designed,
the relationship between the output current and the input currents for one of
these circuits is generally insensitive to isothennal variations. Consequently
Eq. 7.6.6 can be rewritten as

(7.6.7)

Thus, the output current is proportional to the product of the two input currents,
each of which is raised to a power that is set by a ratio of weights.
For the second basic current-mode MITE circuit, instead of connecting
the output of the second clva stage directly to a second input of the output
vlca stage (as in the circuit of Fig. 7.1 1(a)), we now connect the output of
the second clva stage to the output vlca stage through the first clva stage,
as shown in Fig. 7.1 1(b). This first clva stage both generates a voltage that is
logarithmic in the input current Ii, and serves as a vIva stage for the second
clva stage. This connection allows us to obtain negative powers. To illustrate
this property, we apply Eq. 7.5.1 to the output vIva stage:

(7.6.8)

Substituting Eq. 7.6.4 into Eq. 7.6.8, we obtain

[ wni (UT- log Ii - -

Wij Vj - . . . )]
In ex: exp -

UT Wii Wii .

Substituting Eq.7.6.2 into the preceding equation for Vj and rearranging yields
Analysis and Synthesis of Static Translinear Circuits 209

which becomes

(7.6.9)

Thus, the output current is proportional to the quotient of the two input cur­
rents, each of which is raised to a power that is set by ratios of weights. Here,
the powers are not completely independent of each other. However, for any
value of Wni / Wii , we can adjust the value of Wij / Wjj to set the power of
Ij to any value. This quotient-of-power-Iaw relationship is also insensitive to
isothermal variations.

I Wnk Vk
�
0
Vk
� In Vk 51 In

Vi
�
Vi Vi Un
:
wni �
wni
:
Wni :

(a) (b) (c)

Figure 7.12
Three possible ways of traversing a MITE embedded in a MITE network. We can go (a) from an
emitter to a control gate, (b) from a control gate to an emitter, or (c) from a control gate to another
control gate.

These two basic current-mode circuits capture all of the intuition under­
lying MITE-network operation: Voltages that are logarithmic in the input cur­
rents are generated using diode-connected MITEs. Power laws are set through
ratios of weights, and negative powers are obtained through voltage-inversion
stages. Products are obtained by summing two or more logarithmic voltages
on MITEs.
We have formalized this intuitive analysis and have obtained systematic
analysis and synthesis procedures for this class of nonlinear circuits (Minch,
1997; Minch et aI., 1999).
2 10 Chapter 7

7.7 Analysis of MIT E Networks

In this section, we shall develop a by-inspection analysis procedure for MITE

networks by extending the analysis procedure for translinear loops of sub­
threshold MOSFETs discussed in Sect. 7.3. We have previously published
analysis procedures for MITE networks (Minch et al., 1996; Minch, 1997).
The analysis method that we develop here is somewhat more general and re­
quires fewer initial definitions. Additionally, we can use this procedure directly
to analyze subthreshold MOS translinear circuits that make use of the back gate
(that is, the substrate) in addition to the front gate (van der Gevel and Kuenen,
199 4; Mulder et al., 1995; Andreou and Boahen, 1996; Fried and Enz, 1996;
Serrano-Gotarredona et al., 1999). To do so, we simply view the four-terminal
subthreshold MOSFET as a two-input MITE with a weight of K, for the front
gate and of 1 - K, for the back gate and apply the procedure that we shall

develop.
As we go around loops in a MITE network, a MITE can be traversed in
the three possible ways depicted in Fig. 7.1 2. Q n can be traversed by going
from its emitter Un to one of its control gates Vk (Fig. 7.1 2(a)); this traversal
corresponds to going through a counterclockwise element in a translinear­
loop circuit. In this case, Eq. 7.5.1 can be rearranged to obtain the following
recursion relation:

Conversely, Q n can be traversed by going from one of its control gates Vk

to its emitter, Un (Fig. 7.1 2(b)); this traversal corresponds to going through
a clockwise element in a conventional translinear-Ioop circuit. In this case,
Eq. 7.5.1 can be rearranged to obtain the following recursion relation:

(7.7.2)

Finally, Q n can be traversed by going from one of its control gates Vk to

another of its control gates Vi (Fig. 7.1 2(c)); this transition has no analog in
conventional translinear-Ioop circuits. In this case, Eq. 7.5.1 can be rearranged
Analysis and Synthesis of Static Translinear Circuits 211

to obtain another recursion relationship:

e Vi' / VT = ( / VT ) / ( AnInIs ) 1/ n (e U" / VT )1/w,,;

w ;
e Vik -W" k W,,;

. II (eVj/UT ) -Wn j /w,,;

__

•
(7.7.3)
#i, k

The final product in each of these three recursion relationships accounts

for the contributions of peripheral control gates, which do not lie directly on
the particular path through the MITE network that we have chosen. Multiple
translinear loops can flow together through these extra control gates, like
tributaries joining to fonn a river. Translinear loops can split into multiple
paths and merge back together again. We shall call such translinear loops
confluent. The existence of confluent translinear loops makes the analysis
of MITE networks slightly more involved than conventional translinear-loop
circuits, because we may have to traverse several confluent translinear loops
to analyze a given circuit completely. To analyze a network, we first identify a
loop through the circuit that traverses most of the MITEs. We proceed around
the loop from node to node, building up a translinear-Ioop expression as we go,
by applying the recursion relationship appropriate to each transition. If there
is a confluence of translinear loops, we trace through each One until we have
built a complete translinear-loop expression.
We shall nOw illustrate this analysis procedure by applying it to several
simple networks. Consider the network shown in Fig. 7.13( a), which comprises

eU,, / UT
three two-input MITEs. Note that all of the emitters are grounded in this
circuit. In this case, all of the factors in the recursion relationships
evaluate to unity. Consequently, we can ignore them in applying the recursion
relationships. Moreover, we shall show by construction in Sect. 7.8 that any
translinear-loop equation can be realized by a network with grounded emitters.
However, in some cases, it may prove beneficial to have some emitters at a
potential other than ground. In such cases, we would have to keep track of the
emitter factors. To analyze the circuit of Fig. 7.13(a), we first identify a loop
through the circuit that traverses as many of the MITEs as possible. We begin
at the emitter of Q 1 , which is grounded, and proceed to node V1 through Q 1 .
Then, we move to node V2 through Q3 . Finally, we return to ground by moving
to the emitter of Q2 . This single loop traverses each MITE in the circuit; there
212 Chapter 7

(a) (b)

(e)

Figure 7.13
Three networks comprising two-input MITEs that can be analyzed completely by tracing a single
loop. (a) A two-input geometric-mean circuit. (b) A squaring-reciprocal circuit. (c) A multiply­
reciprocal circuit.

are no confluent loops. By following the procedure just described, we have that

which can be simplified to obtain

By rearranging the preceding equation, the following translinear-Ioop expres­

sion is obtained:
Analysis and Synthesis of Static Translinear Circuits 2 13

which can be solved for the output current

Thus, the circuit of Fig. 7.13(a) is a two-input geometric-mean circuit. If each

MITE has the same value of A (that is, A i = A2 = A3 = A) , then the output
current is simply given by

13 = J11 h
Next, consider the network shown in Fig. 7.13(b), which also comprises
three two-input MITEs. To analyze this circuit, we first identify a loop through
the circuit that traverses as many of the MITEs as possible. We begin at the
emitter of Ql , which is grounded, and proceed to node Vi through Ql . Then,
we move to node V2 through Q2. Finally, we return to ground by moving to the
emitter of Q3. This single loop traverses each MITE in the circuit; once again,
there are no confluent loops for us to trace. If we go around this loop, applying
the recursion relationship appropriate to each move, we find that

which can be simplified to obtain

By rearranging the preceding expression, we obtain the translinear-Ioop equa­

tion:

(7.7.4)

which, apart from a simple renumbering of the currents, is identical to the

equation we derived for the circuit of Fig. 7.13(a). This result should not be
too surprising, because the two networks shown in Fig. 7.13 have the same
basic topology; they are merely biased differently. We can rearrange Eq. 7.7.4
2 14 Chapter 7

to obtain the output current

I3 _A1 A3 I�
- A� 11 •
Thus, the circuit of Fig. 7.13(b) is a squaring-reciprocal circuit. Again, if each
MITE has the same value of A (that is, A 1 = A2 = A3 = A), then the output
current is simply given by
12
13 =
I� .
Next, consider the network shown in Fig. 7.13(c), which comprises four two­
input MITEs.To analyze this circuit, we first identify a loop through the circuit
that traverses as many of the MITEs as possible. We begin at one of the control
gates of Q1 , which is connected to Vref, and proceed to node V1 through Q 1 .
Then, we move to node V2 through Q2 . Then, we move to V3 through Q3 .
Finally, we return to Vref through Q 4 . This single loop traverses each MITE in
the circuit; once again, there are no confluent loops to trace. If we go around
this loop, applying the recursion relationship appropriate to each move, we find
that

• (�
A4 Is
) 1

W
= e Vrer/UT

which can be simplified to obtain

Rearranging the preceding expression, we obtain the translinear-Ioop equation:

( �: ) (�:) (�:) (�:) .

= (7.7.5)
Analysis and Synthesis of Static Translinear Circuits 2 15

We can rearrange Eq. 7.7.5 to obtain the output current

Thus, the circuit of Fig. 7.13( c) is a multiply-reciprocal circuit. Again, if each

2
MITE has the same value of A (that is, >'1 = A = A3 = A4 = A), then the
output current is simply

For each of the networks that we have analyzed so far, we only had to trace
a single loop around the network to fully characterize the circuit. We shall
now consider a simple example where we must trace at least two confluent
translinear loops to fully analyze the circuit. Figure 7.1 4 shows a network
comprising four two-input MITEs. We cannot identify a single loop through
this circuit that traverses each MITE: We must consider at least two loops that
are confluent with one another. To analyze this circuit, we first identify a loop

to node 1. 1,
through the circuit that traverses as many of the MITEs as possible. As shown

V1
in Fig. 7.1 4(a), we begin at the emitter of Q which is grounded, and proceed
through Q Then, we move to node V3 through Q 3. Then, we return
to ground through Q 4 . We have not traversed Q 2 at all in this loop. If we go
around this loop, applying the recursion relationship appropriate to each move,
we find that

/ 2 /
( ((1) W (iT.)"'}W W C!�. t r (�:.) (eV,/UT r 1
1 �

..

e V2 / UT,
(7.7.6)

e V2 / UT
which has a factor of that we would like to express in terms of the
collector currents. We can derive a suitable expression for by traversing
the confluent loop shown in Fig. 7.1 4(b). If we go around this second loop,

e V2 / UT
applying the recursion relationship appropriate to each move and substitute
the resulting expression for directly into Eq. 7.7.6, we find that
216 Chapter 7

..

eV2/UT
The preceding equation can be simplified:

and can be rearranged to obtain the translinear-Ioop equation:

( �:) (�:) ( �� ) (�:) .

= (7.7.7)

We can rearrange Eq. 7.7.7 to obtain the output current

A4 A1 12 13
14 = .
A2 A3 11
Thus, the circuit of Fig. 7.1 4 is also a multiply-reciprocal circuit. Again, if each
MITE has the same value of A (that is, A 1 = A2 = A3 = A4 = A), then the
output current is simply

7.8 ABC's of MITE-Network Synthesis

In this section, we shall consider the basics of MITE-network synthesis. As

was the case with translinear-Ioop circuit synthesis, the problem of synthe­
sizing MITE networks is underconstrained and there are design trade-offs in­
volved in the process. Once again, in our brief discussion, we cannot be ex­
haustive and we shall endeavour to make the basic procedure clear by illus-
Analysis and Synthesis of Static Translinear Circuits 2 17

(a)

(b)

Figure 7.14
A network comprising four two·input MITEs. To analyze this circuit, at least two loops that are
confluent with one another must be considered: (a) The primary loop used to analyze the circuit,
and (b) the confluent loop that we trace to complete the analysis.

trating it with some simple examples. The starting point for MITE-network
synthesis is identical to that of translinear-Ioop circuit synthesis: A set of
translinear-loop equations is derived from some functional or behavioral de­
scription of the system to be implemented. The final steps are also similar: The
networks can often be consolidated in the same way that translinear-Ioop cir­
cuits can, by merging redundant parts of the circuits that we have synthesized.
Because the initial and final steps are similar, we shall focus on the middle
steps in the synthesis procedure: The construction of MITE networks from
translinear-loop equations.
2 18 Chapter 7

Synthesizing Static MIT E Networks

As with the synthesis of translinear-Ioop circuits, we can summarize the syn­

thesis of MITE networks as follows: First, we acquire a set of translinear-Ioop
equations from a behavioral or functional description of the system that we
want to implement. Next, we build a network for each of the translinear-Ioop
equations. This construct involves a building phase, a balancing phase, a bi­
asing phase, and a completion phase. Finally, if possible, we consolidate the
resulting networks by merging parts of them where possible. We shall discuss
each of these steps in tum, then we shall use them to synthesize two simple
multiplier circuits.

Acquiring a Set of Translinear-Loop Equations We start the construction

process with a set of translinear-loop equations, each of the fonn

II II (7.8.1)
nE " CW " nE " CCW "

where "cw" denotes a set of clockwise currents and "ccw" denotes a set of
counterclockwise currents 1 , and the kn are positive integer powers to which
the currents are raised, such that

(7.8.2)
nE " CW " nE " CCW "

With translinear-Ioop circuits, the reason for restricting the powers to be inte­
gers is obvious: A current is raised to a given power because it passes through
an integer number of TEs facing in the same direction around a loop. A cur­
rent cannot pass through a fractional number of TEs. However, with MITE
networks, it is certainly possible to allow these powers to be positive real num­
bers subject to the constraint expressed in Eq. 7.8.2, but we shall restrict our
attention here to the case of integer powers for two reasons: Firstly, integer
powers suffice for many practical purposes. Secondly, with MITE networks
these powers are set by ratios of weights and we obtain the most accurate ra­
tios by connecting an integer number of identical unit cells in parallel with one
another. The procedure by which we obtain such translinear-loop equations is
identical to the one described in Sect. 7.4.

1 In the context of MITE networks, such designations are not as meaningful as they are in
translinear-loop circuits.
Analysis and Synthesis of Static Translinear Circuits 2 19

(a) (b)

(d)

(e)

Figure 7.15
Steps in the construction of MITE networks. Here k; , kj ' and kk are integers that each denote
some number of identical inputs and K denotes the total number of control gates for each MITE.
(a) Beginning the network. (b) Building the network. (c) Balancing the network. (d) Biasing the
network. (e) Completing the network.

Building MIT E Networks For each translinear-loop equation, we build a

MITE network. We begin a network by picking a current from each set (for
example, current Ii from the "clockwise" set and current Ij from the "counter­
clockwise" set). We create a new MITE for each one and make a new node in
the circuit, coupling it into Q i through kj unit inputs and into Q j through ki
unit inputs, as shown in Fig. 7. 15(a). If k i and kj have a factor in common, we
can divide both by that factor in determining the number of unit inputs for each
connection. For each additional current in the translinear-loop equation (for ex­
ample, current h from the "clockwise" set), we create a new MITE, and we
220 Chapter 7

make a new node in the circuit, connecting it to an existing MITE whose cur­
rent is from the opposite set (for example, Q j ) through k k unit inputs and to
Qk through kj unit inputs, as shown in Fig. 7.15(b). Once again, if kj and kk
have a factor in common, we can divide both by that factor in determining the
number of unit inputs for each connection. We continue adding MITEs in this
way until we have exhausted all of the currents in the translinear-Ioop equa­
tion. The order in which we add MITEs and the existing MITEs to which we
connect them affect the structure of the final network and the number of inputs
that fan-in to each MITE.
Once we have built the basic network for a translinear-Ioop equation, as
just described, we then balance the fan-in of all MITEs in the network. Suppose
that the largest MITE fan-in is K. We then add a sufficient number of unit
inputs to each MITE, connected to an appropriate voltage Vref, so they each
have a fan-in of K, as shown in Fig. 7.15(c). As long as the translinear-Ioop
equation from which we started conforms to Eq. 7.8.2, the exact value of Vref is
not critical: The quiescent collector voltages in the MITE network will depend
on the value of Vref, but as long as all of the collector voltages stay sufficiently
far away from the power supply rails, the network's behavior is independent of
the value of Vref.
We need to balance the number of inputs to each MITE in the network
because of the way in which we implement the weighted voltage summation.
If we implement the weighted voltage summation using a capacitive voltage
divider, as discussed in Sect. 7.5, then each weight is equal to a coupling ca­
pacitance divided by a total floating-gate capacitance. The power-law relation­
ships implemented by a network are given by ratios of weights. As design­
ers, we would like these powers to be independent of the total floating-gate
capacitances, because they include (nonlinear) parasitic capacitances. By re­
quiring the total floating-gate capacitance of each MITE to be the same, the
total floating-gate capacitances will cancel in the weight ratios, making them
depend only on ratios of coupling capacitors. The best way to ensure that the
total floating-gate capacitances are the same is to require that each MITE have
an identical complement of inputs. In the context of integer numbers of unit
inputs, we would give each MITE the same number of unit inputs.

Biasing MITE Networks The process of biasing a MITE network is consid­

erably simpler than that of biasing a translinear-Ioop circuit. We simply force
input currents into the collectors of some of the MITEs and diode-connect
them by connecting some of their control gates to their collectors, as shown
Analysis and Synthesis of Static Translinear Circuits 221

in Fig. 7.15(d). Those MITEs that are diode connected become inputs, while
those that are not diode connected are outputs. Other biasing schemes are cer­
tainly possible, but are never needed.

Completing MITE Networks It may seem that by adding unused MITEs

in the process of balancing the fan-in in a network, we are wasting resources.
Indeed, such unused inputs can account for a significant fraction of the total
transconductance of a MITE. Leaving them unused leads to larger collector­
voltage swings and a higher required power-supply voltage. It so happens that,
as long as the translinear-loop equation from which we started conforms to
Eq. 7.8.2, we can utilize all of the extra control gates that we add during the
balancing phase (Minch, 1997). Intuitively, because the behavior of a MITE
network is unaffected by the value of Vref, we can short Vref to one of the
collector voltages in the network without affecting the behavior of the circuit.
This MITE-network transformation is called completion (Minch, 1997). We
connect all of the unused inputs to one of the collector voltages, as shown in
Fig. 7.15(e). In doing so, we should generally avoid the creation of feedback
loops in the network that could affect its stability. We can always do this by
choosing a MITE that only has self connections.

Consolidating MIT E Networks In some cases, as with translinear-loop cir­

cuits, after we have biased each of the MITE networks in the circuit, we will
recognize some redundancy between them. For example, if two MITEs in dif­
ferent networks pass the same current and their control gates are connected in
the same manner, then these MITEs are redundant and may be shared between
the networks. Such consolidation is usually a good idea, because it usually
results in smaller circuits and fewer opportunities for errors resulting from de­
vice mismatch. Other, more subtle forms of MITE-network consolidation are
possible (Minch, 2000b), but are beyond the scope of this chapter.

Synthesis of One-Quadrant MITE-Network Multipliers

Suppose that we want to implement a multiplication operation with strictly

positive inputs using a MITE network. That is, we want to find a network that
implements the relationship
z = xy (7.8.3)

where x > 0, y > 0, and z > ° are dimensionless quantities. Here x and y are
the independent variables (that is, the inputs) and z is the dependent variable
222 Chapter 7

(that is, the output). First, we represent x by Ix /Iu , y by Iy /lu and z by Iz /Iu ,
where Iu is the unit current. Then, we substitute these definitions of x , y , and
z into Eq. 7.8.3, obtaining

Iz _ Ix Iy
Iu Iu Iu
which can easily be rearranged to obtain the translinear-loop equation

IzIu = Ixly . (7.8.4)

'-v-' '-v-'
"CW " "CCW "

Starting from Eq. 7.8.4, we shall synthesize two different MITE networks
to illustrate how the building order can influence the structure of the final
network. We begin the first network by selecting I z from the "CW " set and
Ix from the "CCW " set and make a MITE for each one. Then, we make a
new node in the circuit and couple it into Q z through one unit input, and into
MITE Qx through one unit input, as shown in Fig. 7.16(a). Next, we select
Iy from the "CW " set and make another MITE for it. We make a new node
and couple it into Q z through one unit input, and into Q y through one unit
input, as shown in Fig. 7.16(b). Next, we select Iu from the "CCW " set and
make another MITE for it. We make a new node and couple it into MITE
Q y through one unit input, and into Q u through one unit input, as shown in
Fig. 7.16(c). Next, we balance the fan-in of all MITEs. In this case, MITEs
Qz and Qy have two inputs, whereas MITEs Q x and Qu have only one. To
balance the fan-in in the network, we add another unit input to MITEs Q x and
Qu, each connected to Vref, as shown in Fig. 7.16(d). Next, we bias the network
by forcing Iu into Qu, Iy into Qy, and Ix into MITE Qx; and diode connect
each one through one control gate, as shown in Fig. 7.16(e). This network
implements Eq. 7.8.3, but it has two unused inputs. We can utilize these two
inputs and complete the network by connecting Vref to the collector of MITE
Qu, as shown in Fig. 7.16(t). This network also implements Eq. 7.8.3, but has
no unused inputs. Finally, no consolidation is possible.
We begin the second network by selecting I z from the "CW " set and Ix
from the "CCW " set and make a MITE for each one. Then, we make a new
node in the circuit and couple it into Q z through one unit input, and into Q x
through one unit input, as shown in Fig. 7.17(a). Next, we select I u from the
"CCW " set and make another MITE for it. We make a new node and couple it
into Q x through one unit input, and into Q u through one unit input, as shown
in Fig. 7.17(b). Next, we select Iy from the "CW " set and make another MITE
Analysis and Synthesis of Static Translinear Circuits 223

(d)

(e)

(f)

Figure 7.16
Synthesis of a one·quadrant MITE-network multiplier. (a) Beginning the network. (b, c) Building
the network. (d) Balancing the network. (e) Biasing the network. (f) Completing the network.
224 Chapter 7

for it. We make a new node and couple it into Q y through one unit input, and
into Qu through one unit input, as shown in Fig. 7.17(c). Next, we balance
the fan-in of all MITEs. In this case, MITEs Q x and Qu have two inputs,
whereas MITEs Q y and Q z have only one. To balance the fan-in in the MITE
network, we add another unit input to MITEs Q y and Q z , each connected to
Vref ' as shown in Fig. 7.17(d). Next, we bias the network by forcing I u into
Qu, Iy into Qy, and Ix into Qx; and diode connect each one through one
control gate, as shown in Fig. 7.17(e). This network implements Eq. 7.8.3,
but it has two unused inputs. We can utilize these two unused inputs and
complete the network by connecting Vref to the collector of Q y, as shown in
Fig. 7.17(f). Note that, if we had connected Vref to the collector of Q u , then we
would have introduced a positive feedback loop into the network with a loop
gain of unity, making the desired network equilibrium an unstable one. If we
had connected Vref to the collector of Qx, then we would have introduced a
negative feedback loop, creating the potential for instability. This network also
implements Eq. 7.8.3, but has no unused inputs. Finally, no consolidation is
possible.
We have synthesized four different MITE networks (Figs. 7.16(e), 7. 16(f),
7.17(e), and 7.17(f)) that each implement Eq. 7.8.3. The circuits of Fig. 7.16
are more symmetric with respect to how many stages separate the x and y in­
puts from the z output than those of Fig. 7.17. Intuitively, we should expect
that a network with fewer stages on average between the inputs and an out­
puts would be less sensitive to mismatch in MITE weight values, than would
be a network with more stages. We have demonstrated this fact for these four
one-quadrant multiplier circuits elsewhere (Minch, 1997). The circuits shown
in Fig. 7.16 differ from those shown in Fig. 7.17 in the order in which we se­
lected the currents in the building process and where we chose to connect their
MITEs. Because the number of ways in which currents can be chosen from a
translinear-Ioop equation grows rapidly in the number of currents, it is difficult
to say general things about how the chosen order affects the performance of the
final network. However, we shall make some observations: The more MITEs
that we connect to any given MITE, the larger the required fan-in per MITE
in the network as a whole, but the fewer the average number of intermediate
stages between any two MITEs. We have shown previously (Minch, 1997) that
any translinear-Ioop equation can be implemented as a MITE network with
a maximum of one MITE between any pair of MITEs. Networks with fewer
intermediate stages should be less sensitive to offset and noise accumulation
than networks with more intermediate stages. Additionally, because of para-
Analysis and Synthesis of Static Translinear Circuits 225

(a) (b)

(d)

(e)

(I)

Figure 7.17
Synthesis of a one·quadrant MITE-network multiplier. (a) Beginning the network. (b, c) Building
the network. (d) Balancing the network. (e) Biasing the network. (0 Completing the network.
226 Chapter 7

sitic node capacitances, the response time of a network with fewer intermediate
stages will be faster than that of a network with more intermediate stages.

(a)

(b)

Figure 7.18
Synthesis of a two-quadrant MITE-network multiplier. (a) Two independent copies of the one­
quadrant multiplier of Fig. 7 . 1 6(0. (b) The final consolidated two-quadrant MITE-network multi­
p
ier circuit. Here the Iy and Iu circuitry are shared between the two MITE networks.

Synthesis of a Two-Quadrant MITE-Network Multiplier

Suppose that we want to implement a circuit that multiplies two quantities, x

and y, where x can be either positive or negative and that y is strictly positive.
Analysis and Synthesis of Static Translinear Circuits 227

Thus, their product

z = xy, (7.8.5)

can be either positive or negative. We shall represent y by I y flu and we shall

use a differential representation for x and z, as described in Sect. 7.2. That is,
we represent x by

where x + == Ii" flu and x - == I; flu. Likewise, we represent z by

where z + == Ii flu and z - == I; flu .

Next, we substitute these definitions for x , y , and z into Eq. 7.8.5:
( Ii _ I; ) ( Ii"
=
_ I; ) ( Iy )
Iu Iu Iu Iu Iu
which can be rearranged to obtain

Iuli - luI; = Iyli" - IyI; .

One straightforward way to decompose this equation into a pair of
translinear-loop equations is to equate individually the positive and negative
terms on each side of the equation. Using this decomposition, we obtain the
pair of translinear-Ioop equations:

Iuli = Iyli" and

'-v-' '-v-'
CW CCW
By following the procedure shown in Fig. 7.16 for each of these translinear­
loop equations, we obtain the two independent MITE networks shown in
Fig. 7.18(a). Note that in each MITE network, we supply a copy of I u to a
MITE that is diode connected through two unit inputs. The collector voltages
of these two MITEs should be identical, so a single Q u can be shared between
the two networks. Also, in each network, we supply a copy of I y to a MITE that
is diode connected through a single control gate and its other control gate is
connected to an identical voltage. Thus, their collector voltages are also iden­
tical, and we should be able to share a single Q y between the two circuits. The
consolidated two-quadrant MITE-network multiplier is shown in Fig. 7.1 8(b).
III DYNAMICS
8 Linear Systems Theory

In this chapter we will review some properties of linear time-invariant systems.

We consider their input/output relationship in the time domain, the impulse
response, and the convolution theorem. We also review basic concepts of
complex number theory; the Laplace transform, system's transfer function, and
frequency domain analysis. More extended descriptions of this material can be
found in many standard textbooks (Carlson, 1986; Poularikas and Seely, 1994;
Oppenheim et aI., 1996).

8.1 Linear Shift-Invariant Systems

In linear systems theory, a system is treated as a black box that does not reveal
its internal states, and is characterized only by the relationship between its
input and output (see Fig. 8.1). If a system has no internal stored energy, then
its output response y(t) is forced entirely by the input x(t):

y(t) = F[x(t)] (8.1.1)

where F[·] is the transfer function.

Linearity A system is linear if it obeys the two fundamental principles:

Homogeneity, and additivity.
The principle of homogeneity states that output scales linearly with the
input:
F[o:x(t)] = o:F[x(t)]. (8.1.2)

Usually, linear systems theory is applied to time-varying signals. However the

same methods can be applied to input and output signals that are distributed

x ( t) SYSTEM y(t)
• •

FigureS.1
Typical black-box representation of a linear system. Its input is the signal x(t) and its output is the
signal y( t).
232 Chapter 8

Input Output

Figure8.2
Graphical example of the homogeneity principle of a linear system. The signals in the left
quadrants represent the system's input, and the signals in the right ones represent its output. An
increase in the input signal causes a proportional increase in the output signal.

over space rather than time. For example, in Fig. 8.2, the input signal is a spatial
unit impulse and the output is a spatial Gabor function (a Gaussian modulated
by a cosine function))
The principle of additivity states that if the input signal is composed
of elementary signals, then the system's response is the composition of its
responses to each of the elementary signals:

Figure 8.3 shows a graphical example: If the response of the system to a

spatial impulse is a Gabor function, and if the system's input signal is a linear
combination of spatial unit impulses, then the system's response will be a linear
combination of Gabor functions.
The principles of homogeneity and additivity taken together are commonly
referred to as the principle of superposition, which state that a system is linear

1 Gabor functions are commonly used to model the (linear) response properties of a particular
class of neurons in the visual cortex.
Linear Systems Theory 233

Input Output

I I

Figure8.3
Graphical example of the additivity principle of a linear system. The signals in the left quadrants
represent the system's input, and the signals in the right ones represent its output.

if
y(t) = L akF[Xk(t)] (8.1.4)
k
for input x(t) = I:k akxk(t), and ak constant for all k.
In other words, a system is linear if its response function F is a linear
operator:

(8.1.5)

Shift Invariance A system is said to be shift-invariant if its responses to

identical stimuli shifted in time are also identical, except for the corresponding
time shift (Fig. 8.4).
If a system is shift-invariant, then its response function is also shift­
invariant: Given input signal x(t), its time-shifted variant x(t - r) will produce

F[x(t - r)] = y(t - r). (8.1.6)

The system's output signal is unchanged, except for a time shift.

234 Chapter 8

Input Output

Figure8.4
Graphical example of a time-invariant system's response.

Time invariance and linearity are two independent characteristics. Not all
linear systems are time-invariant and, similarly, not all time-invariant systems
are linear.

8.2 Convolution

Convolution is an important mathematical operator used in linear systems

analysis. The convolution of two time-varying signals, v(t) and w(t), is

v(t) * w(t) =
1+-0000 V(A)W(t - A)dA (8.2.1)

where A is the integration variable, and t is the independent variable.

Figure 8.5 shows a graphical representation of the convolution process
between signals v(t) (Fig. 8.5(a» and w(t) (Fig. 8.5(b» at three different time
steps. The result of the convolution is shown in Fig. 8.6. Note how the overlap
between the two curves is null for t < 0, increases for 0 < t < T, peaks at
t = T and decreases again for t > T.
If the independent variable for both input signals is the same, then it can
be omitted and we express the convolution between to the two signals v(t) and
Linear Systems Theory 235

viti wit)

(a) (b)

v(!..)
(kO)

V(!..) I

(O<ld)

(t>T)

(c)

FigureS.S
Graphical representation of the convolution between v(t) and w(t) for three different values of t.
Note that the integration variable>. in (c) is on the abscissae of the plots. Modified from Carlson.
A. B. (1986).
236 Chapter 8

v*w

I I T
I I I
(kO) (O<kT) (/>T)

Figure8.6
Result of the convolution between the two signals v(t) and w(t) of Fig. 8.5. The three dashed
lines are at the three values of t used in Fig. 8.5.

w(t) simply as v * w. The convolution operator is linear and has the following
properties:
commutative: v*w w*v
associative: v * (w * z) (v * w) * z
distributive: v * (w+z) (v * w)+(v * z)

8.3 Impulses

The unit impulse or Dirac delta function 8(t) is not a function in the strict
mathematical sense. It is defined by a set of assignment rules.

• If v(t) is a continuous function at t = 0 then

{ t2 v(t)8(t)dt
tl < 0 < t2
lt l
= {V(O)0 otherwise.
(8.3.1)

• If € is an arbitrary small number

(8.3.2)
Linear Systems Theory 237

From these rules we can infer that 8(t) has unit area at t 0 and that
=

8(t) = 0, for all t ¥- O. We can also note that the Dirac delta function has
no mathematical or physical meaning, unless it appears under the integral
operator.

Impulse Integration Properties

When used in conjunction with the integral operator, the Dirac delta function
has the following properties:

• Replication:

v(t) * 8(t - r) = v(t - r) (8.3.3)

• Sampling:

r+oo v(t)8(t - r)dt

1-00 = v(r) (8.3.4)

where v(t) is a continuous time-varying signal.

Impulses in the Limit

There are many (proper mathematical) functions 8 E(t) that approach the Dirac
delta function 8(t), in the limit:

Elim
--+O
8E(t) = 8(t) (8.3.5)

An example of such a function that is commonly used is

sin(!.E )
8E = __ _. (8.3.6)
t
Figure 8.7 shows how 8E of Eq. 8.3.6 approaches the Dirac delta function
as to decreases.

8.4 Impulse Response of a System

We can now use the notions of convolution and unit impulse to define the
impulse response of a linear time-invariant system. If y(t) is the system's
response to its input x(t) we can write

y(t) = F[x(t)]. (8.4.1)

238 Chapter 8

FigureS.7
Plot of the function sin(t/ E)/t for three decreasing values of E.

If the input signal is the Dirac delta function (x(t) = o(t)), then the system's
response to the unit impulse is defined as

h(t) == F[o(t)]. (8.4.2)

If x(t) is continuous in time, the replication property of the unit impulse

allows us to rewrite x(t) as x(t) * o(t). With this reformulation of the system's
input signal, Eq. 8.4.1 becomes

y(t) = F [[:00 X(A)O(t - A)dA . ] (8.4.3)

If the system is linear, Eq. 8.4.3 is equivalent to:

y(t) =
1+-0000 x(A)F [o(t - A)] dA (8.4.4)

If we substitute Eq. 8.4.2 into Eq. 8.4.4, and if the system is time-invariant,
Linear Systems Theory 239

then

y(t) r+oo x()")h(t - )")d)"

1-00 r+oo h()..)x(t - )")d)". (8.4.5)
L
= =

oo
This property of linear time-invariant systems is extremely powerful. It states
that if a system's impulse response h(t) is known, the response of the system
to any arbitrary signal x(t) can be computed simply by performing the convo­
lution of its impulse response with the signal itself:

1 y(t) = h(t) * x(t) I· (8.4.6)

Step Response

We can define a system's step response in the same way we defined its impulse
response. If the input signal x(t) is the step function

u(t) =
{I 0
ift� O
(8.4.7)
otherwise

then we define its step response to be

g(t) == F[u(t)]. (8.4.8)

A first interesting property can be obtained by exploiting the system's impulse

response (see Eq. 8.4.6):

g(t) = h(t) * u(t). (8.4.9)

By applying the derivative operator to this equation, and noting that the deriva­
tive of the step function is the unit impulse, we obtain

d d
g(t) = h(t) * u(t) = h(t) * 8(t). (8.4.10)
dt dt
If we apply the unit impulse replication property (Eq. 8.3.3), then we obtain

d
h(t) = g(t). (8.4.11)
dt
Thus, a system's impulse response can be obtained by computing the derivative
of its step response. This property is extremely useful in practical situations be­
cause unit impulses are impossible to generate with physical instruments but
it is easy to generate waveforms that approximate ideal step functions. Conse­
quently, a physical linear time-invariant system is characterized experimentally
240 Chapter 8

R
(�VW-I"'-T c--�) r+11
\
c :t R +:�}
1
x(t) y(t) x(t) _ _ y(t)
a 0 o�------L-----�,

(a) (b)

FigureS.S
Resistor capacitor (RC) circuits. The signals x(t) represent input voltages. and the signals y(t)
represent output voltages. (a) Integrator circuit; (b) Differentiator circuit.

by measuring its step response and then deriving its impulse response from
Eq. 8.4.11.

8.5 Resistor.Capacitor Circuits

The resistor-capacitor (RC) circuits of Fig. 8.8 represent first order, linear,
time-invariant systems. In both circuits, the input signal is x(t) and the out­
put signal is y(t). The circuits of Fig. 8.8(a) and (b) are referred to as RC
integrator and RC differentiator respectively. In this section we focus only on
the properties of the RC integrator. The properties of the RC differentiator will
be described in Chapter 9.
The integrator circuit of Fig. 8.8(a) is governed by the differential equation:

d
RC y(t)+y(t) = x(t). (8.5.1)
dt
By solving Eq. 8.5.1 for a unit impulse input signal (x(t) = 8(t)), we obtain
the circuit's impulse response:
1
h(t) = e -t/RC . u(t) (8.5.2)
RC
-

where u(t) is the step function. Similarly, solving Eq. 8.5.1 for a step input
signal (x(t) = u(t)), we obtain the circuit's step response

g(t) = (1 - e -t/RC) . u(t). (8.5.3)

Figure 8.9 shows the impulse response and the step response. The value
RC is defined as the system's time-constant and is often labeled T. As pointed
out in Section 8.4, the response of the circuit to an arbitrary input signal can be
Linear Systems Theory 241

h(t) get)
1/RC - - - - - - --.:-:;..;;;-...--

't=RC 't=RC

(a) (b)

Figure8.9
Impulse response (a) and step response (b) of an RC circuit.

obtained by the convolution between the input signal and the circuit's impulse
response:

y(t) = x(t) * h(t) = 10(JO R1C e -A/RC . x(t - 'x)d'x. (8.5.4)

8.6 Higher Order Equations

Time-domain analysis becomes increasingly difficult for higher order systems.

Fortunately there is a unified representation in which any solution to a linear
system can be expressed: Exponentials with complex arguments. All solu­
tions to linear homogeneous (undriven) equations are of the form e st where s

is a complex number (see Fig. 8.10):

S = (J +jw = M cos(¢) +jM sin(¢) (8.6.1)

where j = A, (J is the real part of the complex number, w is the imaginary

part, M represents its magnitude, and ¢ its phase. Magnitude and phase of a
242 Chapter 8

I
I�
I�
en

5'

()'

Figure8.10
Complex number representation. The complex number s has magnitude M and phase </>. Its real
part is (T and imaginary part is w.

complex number obey the following relationships:

M = V0'2 +w 2 (8.6.2)

¢> = arctan (;) . (8.6.3)

The magnitude of a complex number s is often denoted as lsi. Furthermore,

applying the properties of complex exponentials, one can observe that

e j,p = cos(¢» +j sin(¢» (8.6.4)

e -j,p = cos(¢» - j sin(¢». (8.6.5)

It follows that s can be also written as

(8.6.6)

These notations can be used to solve higher order differential equations.

As an example, we consider the second order linear homogeneous equation

(8.6.7)
Linear Systems Theory 243

Assume that est is an eigenjunction2 and substitute for V:

(8.6.8)

Solving for s we obtain

- 0: ± J
0:
2 - 4/3
s = -----'----
(8.6.9)
2

Consequently, if 0:
2 4/3 � 0, s is real, otherwise s is a complex number. In
-

practice, if Eq. 8.6.7 is a linear system, we could measure its response V = e st

with a real instrument: but if est was a complex exponential, we would measure
only its real component:

Re { est } = {
Re e(u+iW)t } = eut Re { eiwt } . (8.6.10)

So the measured response of the system would be

Vmeas = eut cos wt. (8.6.11)

Figure 8.11 shows the possible kinds of response of Vmeas for different values
of w and a. If a < 0 all the solutions are stable and decay to zero with
time. If a > 0 all solutions are unstable and diverge with time. If a = 0
the solutions are naturally stable (they neither decay, nor diverge). All physical
passive linear systems will have stable solutions (a < 0). The w axis scales the
oscillation frequency f of a solution (w = 211" f).

8.7 The Heaviside-Laplace Transform

By analyzing the example of the previous section (see Eq. 8.6.7) we can make
the following observation: Any time we substitute the eigenfunction e st into a
linear differential equation of order n, the following property obtains:

dn st
_ e = s n est . (8.7.1)
dtn
In other words:

We can consider s as an operator meaning derivative with respect to time. Similarly,

we can view � as the operator for integration with respect to time (Heaviside).

2 An eigenfunction is a nonzero solution of a second order linear homogenous differential

equation
244 Chapter 8

«)

v � V � n R II
�,
t
v �
" lJ II
V t t V t

I� t 0
J t cr

Figure8.11
The possible kinds of measured responses for a first order linear system.

This observation was made by Heaviside, when trying to analyze analog

circuits: but was also formalized by Laplace when he introduced the Laplace
Transform. The Laplace transform is a useful operator that links functions that
operate in the time domain with functions of complex variables:

£[y(t)] = Y(s) = i: y(t) e-stdt. (8.7.2)

8.8 Linear System's Transfer Function

Now that we have introduced the concepts of convolution (Section 8.2), im­
pulse response (Section 8.4), and the Laplace transform (Section 8.7), we can
define a linear system's transfer function. It is a function defined in the com­
plex domain:
Y(s)
H(s)
_

= (8.8.1)
X(s)
Linear Systems Theory 245

Input System Output

x(t) H(s) y(t)

X(s) Y(s)

Figure8.12
Typical representation of a linear system with input and output signals both in the time domain
(x(t), y(t») and in the Laplace domain (X(s), Y(s»).

where Y(s) is the Laplace transform of the system's output y(t) and X(s) is
the Laplace transform of the system's input x(t) (see Fig. 8.12). Conversely,
we can say that the output of any linear time-invariant system is determined by
multiplying the system's transfer function with its input:

1 Y(s) = H(s)X(s) I· (8.8.2)

Transfer Function and Impulse Response

Consider the special case in which the system's input signal x(t) is the unit
impulse x(t) = 8(t).lts Laplace transform X(s) is

(8.8.3)

In this case, following the definition of Eq. 8.8.1, the system's response in the
complex plane is

Y(s) = H(s) . (8.8.4)

On the other hand, the system's response in the time domain is (by definition)
its impulse response:

y(t) = h(t). (8.8.5)

Because Y(s) is the Laplace transform of y(t), we can substitute Eq. 8.8.5
into Eq. 8.7.2:

(8.8.6)
246 Chapter 8

and so

H(s) = i: h(>.) e-ASd>' = £[h(t) ] . (8.8.7)

The transfer function H(s) is the Laplace transform of the impulse response
h(t).

Summary Given a linear time-invariant system with input x(t) , output y(t) ,
and impulse response h(t) :

y(t) =x(t) * h(t)

Y(s) = X(s)H(s)

where

X(s) =£x
[ (t) ]
Y(s) =£y
[ (t) ]
H(s) =£[h(t) ].

8.9 The Resistor-Capacitor Circuit (A Second Look)

Consider again the RC circuit of Fig. 8.8. As mentioned in Section 8.5, this
circuit is governed by
d
T y(t) +y(t) = x(t) (8.9.1)
dt
whereT = Re.
In the complex domain, we have

Y(S) (TS + 1) = X(s) . (8.9.2)

Therefore the circuit's transfer function is

1
H(s) = - . (8.9.3)
1 +TS
-

Consider now how this circuit responds to sinusoidal signals of different

frequencies. Sinusoids have a very special relationship to shift-invariant linear
systems, such as the one we are analyzing. When a sinusoidal signal is applied
Linear Systems Theory 247

10° r----__

10° 101
Angular frequency (Hz)

Ca)

0
-10
-20
-30
Oi
�-40
Q)
�-50
.r:
a..
-60
-70
-80
-90
10° 101
Angular frequency (Hz)

(b)

Figure8.13
Bode plot of a first order linear system, such as the RC circuit of Fig. 8.8. Ca) Magnitude, Cb) Phase.
248 Chapter 8

as input to a shift-invariant linear system, then its response will be another

sinusoidal signal, with possibly a different amplitude and a different phase,
but certainly with exactly the same frequency! That is, if the input is x(t) =

sin(wt) , the output will be y(t) = Asin(wt + cjJ), where A and cjJ determine
the scaling and shift.
When we analyze a system using sinusoidal signals of different frequen­
cies, we are working in the frequency domain. In this domain s = jw and the
circuit's transfer function is
1
H(jw) = . (8.9.4)
1 +JWT
From this transfer function, we make two useful observations:

1. If the frequencies of the sinusoidal signals are small with respect to the
circuit's time-constant (WT« 1), then the circuit's output will resemble its
input ( Y(jw) � X(jw) .
2. On the other hand, if the frequencies are large with respect to the circuit's
time-constant (WT» 1), then

Y(jw) 1
jWT'
� (8.9.5)
X(jw)
These observations are also reflected in the plots of the transfer function's
magnitude and phase (Fig. 8.13). These plots are referred to as Bode plots
and they are used to analyze the response of a dynamic system in terms of its
transfer function. The magnitude of the transfer function is

(8.9.6)

and its phase is

cjJ = arctan( -WT) . (8.9.7)

The frequency w = � is defined as the cutofffrequency . In Fig. 8.13 it is set

to one.
The RC circuit of Fig. 8.8 is a low-pass filter, because it allows sinusoidal
signals with frequencies lower than the cutoff frequency to pass virtually
unchanged. On the other hand, the frequency components of the input signals
that are above the cutoff frequency are attenuated. The phase lag between the
input and the output of the system increases with w (see Fig. 8.13(b» and
saturates at - 90°. Figure 8.14 shows experimental data measured from an RC
Linear Systems Theory 249

1.2
- Input
-- 50 Hz
-+- 100 Hz
CD __ 200 Hz
(J
c ....... 400 Hz
�0.8
�

00 0.2 0.4 0.6

Normalized time units

Figure8.14
Response of an RC low-pass filter (R = lOM!1. C =InF) to input sinusoids of different fre­
quencies. The input signals have been normalized to unity. and the outputs have been normalized
with respect to the input. The time axis has also been normalized so that the responses to all the
frequencies could be presented on the same graph.

lowpass filter with R = lOMf! and C = InF. Sinuosoids of increasing

frequency were applied to the circuit and the corresponding responses were
measured. To show the effect of a range of input frequencies on the circuit's
response, all the data are plotted on a normalized scale. The responses have
been normalized with respect to the input and time has been normalized to
unity. As expected, the output signal is attenuated as the input frequency
increases; and the phase lag between the input and output signals increases
with increasing frequency.

8.10 Low-Pass, High-Pass, and Band-Pass Filters

The RC circuit analyzed in the previous sections is the simplest example of

a passive filter. Filters are typically used to alter the frequency spectrum of
their input signals. Specifically, filters allow one or more frequency bands
to pass unchanged (except for a multiplicative gain factor), whereas others
are attenuated. Passive filters do not amplify the input signal, whereas active
filters can also amplify the frequency components of the input signal. If a filter
transmits low frequency components (from DC to a lower cutoff value WI), it is
said to be a low-pass filter. If it transmits high frequency components (higher
250 Chapter 8

than a cutoff value w u ), it is said to be a high-pass filter. Filters that transmit

only frequency components between a lower cutoff and an upper cutoff are
said to be band-pass filters.
The analysis of linear systems often assumes ideal filters. These filters have
distortionless signal transmission over one or more frequency bands, and have
zeros responses at all other frequencies. For example, the transfer function of
an ideal bandpass filter is:

H(w) =
{Ke-jwtd WI ::; Iw l ::; Wu
(8.10.1)
o otherwise.

Although ideal filters cannot be implemented in practice, their use in theoreti­

cal analysis simplifies the study of linear systems.
9 Integrator-Differentiator Circuits

Integrators are a very useful class of low-pass circuits suitable for filtering out
the high-frequency components of the signal (often present due to noise). On
the other hand, differentiators filter out the low-frequency components of the
input signal and respond best to its changes. Integrators and differentiators can
also be used to implement adaptation in neuromorphic systems. Adaptation
is ubiquitous in neural systems and allows a system to optimize its dynamic
range against the characteristics of the prevailing input signal.

_o+�
) Vout
o
Figure 9.1
Resistor-capacitor (R-C) integrator circuit.

The simplest type of integrator is the RC circuit in Fig. 9.1. This circuit's
transfer function (Section 8.9) is
1
H(s) = -
1 -
+TS
(9.0.1)

where T = RC is the circuit's time-constant. This time-constant can be

controlled by setting the capacitance or the resistance values of the elements
of the RC circuit.
In VLSI technology, the time-constant of RC circuits implemented with
passive elements cannot be changed once the chip has been fabricated. Both
resistance and capacitance are fixed at design time by the geometries of the
layout mask layers (see Section 12.1). By contrast, the transconductance am­
plifier (Section 5.3) has a transconductance that depends on its bias voltage
which can be set on the operating chip. Therefore, this device can be used to
design an integrator circuit that has an adjustable time-constant: Thisfollower ­
integrator (Mead, 1989) is shown in Fig. 9.2.
252 Chapter 9

Figure 9.2
Follower-integrator circuit. The bias voltage li, which sets the transconductance amplifier's bias
current lb can be used to modify the integrator's time-constant.
'

9.1 The Follower-Integrator

This circuit comprises a unity-gain follower (see Section 5.3), and a capacitor
connected to the follower's output node. The input voltage is applied to the '+'
terminal of the follower. If the circuit operates in subthreshold, we can apply
Kirchhoff's current law at the circuit's output node, and write

C�v:
dt out -
Ib tanh (Ii:(Vin - Vout) )
2UT
(9.1.1)

as derived from Eq. 5.3.6 of Section 5.3.

Small Signal Behavior

In the small signal regime where the transconductance amplifier operates in its
linear range I , Eq. 9.1.1 can be simplified to

G(Vin - Voud
d
C dt Vout = (9.1.2)

where G = ;t� is the amplifier's transconductance (see also Eq. 5.3.8 of

Section 5.3). Under these conditions, the follower-integrator has a transfer
function that is identical to that of the R-C integrator (Eq. 9.0.1), with a time­
constant 7 = f:
1
(9.1.3)
1 + 78

too rapidly. Specifically, Eq. 9.1.2 is valid as long as ftVin

The amplifier will operate in its linear range provided that Vin does not change
< 4�T (Mead,

1 The DC component of Yin is in a range in which the amplifier is well behaved, and the AC
component of Yin is sufficiently small.
Integrator-Differentiator Circuits 253

1989). Under these conditions the transfer functions of the integrators of

Fig. 9.1 and Fig. 9.2 are identical. However, the two circuits differ in a key
property: the composition of transfer functions.

Composition Property

We can compose (connect) multiple instances of the R-C integrator circuit (see
Fig. 9.3), or multiple instances of the follower-integrator circuit (see Fig. 9.4),
in sequence to form a delay line.

Figure 9.3
Delay line formed by connecting a large number of R-C integrator circuits.

Each section in the delay line of Fig. 9.4 is modular (from a functional
point of view) and independent of the other sections. The current out of
the transconductance amplifier of one section can charge only the capacitor
connected to the output node: It is not affected by the other sections connected
to the output node. On the other hand, the sections of the R-C delay line of
Fig. 9.3 are tightly coupled to one another. Current of one section flows into
both the capacitor of that section and the resistor of the next. Because the
characteristics of the single R-C circuit change when connected to other R-C
circuits (as in Fig. 9.3), the transfer function of the composition of R-C circuits
is not equal to the composition of individual transfer functions.

>--+_-<>Vout

Figure 9.4
Delay line formed by connecting a large number of follower-integrator circuits.
254 Chapter 9

Due to the modularity of the follower-integrator elements, the transfer

function of the composition of follower-integrator circuits corresponds to the
composition of their individual transfer functions. For example, if the number
of elements in the delay line of Fig. 9.4 is n, we can write

(9.1.4)

---Jn
2.5
I - l
out

2.4

�2.3
.l!l
(5
>
2.2

2.1

0.1 0.2
Time (ms)

Figure 9.5
Large signal behavior of a follower-integrator. Response of the circuit to a large negative step input
(dashed line). The output voltage Yout decreases linearly for large difference Yout - \l;n values
and asymptotes exponentially for small differences.

This is a very useful property, because it allows the frequency response

properties of the delay line to be evaluated analytically. Consider the case
where sinusoidal inputs are applied to the circuit of Fig. 9.4
delay line's transfer function is
(8 = jw). The

1
(1 jWTt·
+
(9.1.5)

Exploiting the fact that

(1 jWT) (1 �(WT)2) ejwT ( wT 1)

+ � + for « (9.1.6)
Integrator-Differentiator Circuits 255

we can write the transfer function explicitly in terms of magnitude and phase:

Vout 1 e-jnwT
-- � (9.1.7)
Vin 1 + ¥UWT)2
where the pre-exponential ratio is the magnitude and the exponential's argu­
ment is the phase. From this equation, we can conclude that the magnitude of
the output signal is attenuated by the factor ! (WT)2 as it crosses each section of
the follower-integrator delay line, and the phase delay introduced by each sec­
tion corresponds to WT radians (equivalent to a time delay of T). This analysis
is valid provided that each follower-integrator operates in its linear region.

Large Signal Behavior

When the AC component of V in is large, the transconductance amplifier is

no longer linear and Eq. 9.0.1 is invalid. For very large variations in V in,
the output current of the transconductance amplifier saturates at ±h (the
asymptotes of Eq. 9.1.1). In this condition, the transconductance amplifier
acts as a constant current source rather than as a linear conductance. While
the difference l Vout - Vin l is greater than 4UT , Vout changes linearly with
time. As the the difference enters the small signal regime, the amplifier begins
to behave as a linear conductance and Vout begins increasing (or decreasing)
exponentially (see Fig. 9.5).

Figure 9.6
Current-mirror integrator. The current lin is the input signal while the output signal is the current
lout .
256 Chapter 9

Figure 9.5 shows the typical response of the follower-integrator circuit to

a large step voltage input. The rate of change of output voltage d�1"t in the
region of constant slope, is defined as slew rate and is usually specified in units
of V / f.lsec. Slew rate is one measure of the performance limit of an operational
amplifier, and is proportional to the maximum output current of the amplifier.

9.2 The Current-Mirror Integrator

This circuit is a non-linear integrator. It does not have the composition property
of the follower-integrator circuit but it is the advantage that it is compact, and
so various forms have been used to implement dense arrays of synaptic circuits
for spiking neural networks (Boahen, 1998; Hiifliger and Mahowald, 1998;
Indiveri, 2(00).

<1Jout
dt

'out

Figure 9.7
Plot of -it lout as a function of lout. according to Eq. 9.2.3. Note the slopes of the parabola at
. � � .
ordmate values lout = O. lout = 2a and lout =
a

The circuit comprises only two transistors and one capacitor (see Fig. 9.6).
Input current lin is applied to Ml and its low-pass filtered version passes
through M2• The circuit can be thought of as a diode-capacitor filter, rather
than a resistor-capacitor one. As its response is not linear we cannot apply
the methodology introduced in Chapter 8 to obtain analytical solutions of its
response to any arbitrary input. However, we can obtain analytical solutions
for responses to some typical input signals by solving the following system of
Integrator-Differentiator Circuits 257

equations:

(9.2.1)

For simplicity, we assume that Ml and M2 are identical (lo and K are the same).
Of particular interest is the circuit's response to a pulse input (that could
represent incoming action potentials from a silicon neuron). We shall subdivide
the pulse function into two step functions: an step lin(t) =linoU(tO); and a
step lin(t) = lino (1 - u(h)) (see Eq. 8.4.7 for the definition of u(t)).

£lout £lout £lout

dt dt dt

(a) (b) (c)

Figure 9.8
Profiles of -itlout as a function of lout. and locally estimated profiles of lout as a function of
time. (a) Local estimate of Iout{t) for values of lout close to zero. (b) Local estimate of Iout{t)
for values of lout close to IinoevT/uT. (c) Estimate of the profile of Iout{t) for all values of
lout.

This case is solved by analyzing how lout changes with time. By using the
chain-rule for differentiation:

(9.2.2)
258 Chapter 9

Using the relationships of Eq. 9.2.1; expressing ft Ve in terms of Ie; and

substituting Iewith (lin - It} we obtain
d K,
I
d out CUTlout (lino - I).
/ = (9.2.3)

When IIis expressed in terms of lout, Eq. 9.2.3 becomes

d K,

d out CUTlout (lino - aloud

/ = (9.2.4)

where
(9.2.5)

is the circuit's gain. Equation 9.2.4 can be re-written in the simplified form

d
-d tlout =
1 lout
-lout 1 - a-.I
T
( ) (9.2.6)
�no
G/UT .
where T =
Ii., ,no Plotting -ddtlout as a function of lout (Fig. 9.7), we obtain a
parabola that intersects the abscissa at lout = 0 and lout = iIno / a.

,
1 "

,
" _ V,=OV
" V,=0.05V
" '

----
___

"
0.8 \ " " V =0.1 OV
\ " "
'- ' -'
--
--
--
---
----
0.6 - --
----
\\ __
__
__
__
_

...... ..
__

::::::.... "
-
_0 ,
:J
,

'--

" ..........
"
"

" "
" "
" "
"
" " " " " - " " '- ,
0.2 0.4 0.6 0.8
Time (s)

Figure 9.9
Response profiles of the current-mirror integrator to a downward current step (from to OA). lino
The three curves show the responses for three values of v;. (see Figure legend). The curves have
been normalized to show the effect of v;. on the time-course of the response.
Integrator-Differentiator Circuits 259

The slope of the parabola at these two intersections is � and - � respec­

tively. The maximum of the parabola (point of zero slope) is at l out = �.
These points can be used to derive the equation that describes how l out changes
with time. We know that for l out close to zero, lout(t) must resemble an expo­
nential et/r (see Fig. 9.8(a)); for l out close to lino /0: it must resemble an expo­
nential e-t/r (see Fig. 9.8(b)); and forI close to the center of the parabola,
out
lout(t) must be linear. Furthermore because the parabola is continuous, l out(t)
must change smoothly from one profile to the other (see Fig. 9.8(c)). The form
of lout(t) in Fig. 9.8(c) resembles a hyperbolic tangent function. We can test
whether the solution contains a tanh(· ) function, by substituting the generic
function lout [tanh(bt + c) + d] into Eq. 9.2.6 and testing for parameters
= a

a, b, c, and d for which the equality holds.

Correct parameters do exist and they form the solution

lout(t) =
lino
2
0:
[
1 + tanh ( 1 (t - to) )]
27 (9.2.7)

which describes the response of the circuit to a step input current from Iin = 0
to lin = lino at time to.

�
- out
F-}l
O.B

1:
�0.6
:;J
()
'0
Q)
N
�0.4
E
o
z

0.2

-8.5 0.5 1 .5 2
Time (5)

Figure 9.10
The output (solid line) of the current-mirror integrator's in response to a pulse input current (dashed
line).
260 Chapter 9

In this case there is a step change that brings the input current from I in = Iino
to lin = 0 at time t1. For t > t1 we have lin = O. In this case the system of
equations 9.2.1 can be simplified to obtain

dV';, .l:£. -UT

.lL.
C -IoeI<UT . (9.2.8)
dt
- =

Integrating both sides of the equation yields

lVc(t) e-I<U;:- dV';,

II: L
- ...£e-#
II: it dt (9.2.9)
C
=

Vc(td tt
and by expressing e -I<Vc in terms of lout (through the last relationship of
Eq. 9.2.1), we obtain

(9.2.10)

where Itl is the output current at t = tt, ItlIoevc(td/uT, and

= Iout(tt} =

V';, (t1) is the voltage generated by the charge stored on the capacitor at t t 1. =

If we define

(9.2.11)

then Eq. 9.2.10 can be rewritten as

Itl
Iout(t) . (9.2.12)
1+7t
= --

When no input current is applied, for t » T the output current decreases

with a t profile. The shape of the profile can be modulated (exponentially) by
the control parameter VT• Figure 9.9 shows the profile of Eq. 9.2.10 for three
different values of VT•
Figure 9.10 summarizes graphically the current-mirror integrator's re­
sponse properties when a current pulse input is applied. This circuit is com­
monly used to implement models of synapses that integrate pulses (spikes).
Figure 9.11 shows experimental data obtained from one of these synapses
when stimulated with uniformly distributed spikes at three different frequency
settings.
Integrator-Differentiator Circuits 261

160

140

�120
.s
c 100
�
:::J
U 80
"5
a.
"5 60
o
40

20

1.5 2 2.5 3 3.5 4 4.5

Time (5)

Figure 9.11
Response of a synaptic circuit based on a current-mirror integrator to input spike trains of different
frequencies. The bottom trace shows the circuit's response to uniformly distributed spikes of 25Hz.
The middle trace and the top trace show the circuit's response to an input spike train of 50Hz and
of 100Hz respectively.

9.3 The Capacitor

The simplest differentiator is a capacitor (see Fig. 9.12). The transfer function
of a capacitor is

(9.3.1)

This element converts an input voltage signal into an output current, computing
the exact derivative of its input signal. Ideally, the capacitor should provide
an infinite output current in response to an input voltage step. This ideal
performance cannot be realized in physical capacitors, because they have a
finite output impedance that limits l out. Furthermore, if we need the same
type of signal (voltage, for example) at both the input and the output of the
differentiator circuit, then the output current must be converted back into a
voltage. A resistor can be used for this purpose, leading to the capacitor-resistor
(C-R) circuit shown in Fig. 9.13, described by

d
Vout{t) Rlout RC (Vin{t) - Vout{t)). (9.3.2)
dt
= =
262 Chapter 9

Figure 9.12
The perfect differentiator.

When a step function Vin = Vo u(t) is applied (see Eq. 8.4.7), the circuit's
step response is
(9.3.3)

where 7 is the circuit's time-constant (7 = RC). The differentiator's transfer

function is

H(8)
Vout � .
(9.3.4)
Vin 1+78
= =

Recalling that the operator 8 stands for the derivative operator ft, we
observe that at low frequencies (78 «1) H(8):::::; 78 which approximates
the transfer function of an ideal differentiator. On the other hand, at high
frequencies (78 » 1), H(8):::::; 1 and the circuit behaves like a unity-gain
follower. The circuit is commonly referred to as a "high-pass filter" because
it allows high-frequency signals to pass unaltered, while suppressing low­
frequency signals. Typically, the resistors fabricated in VLSI technology are
restricted to less than a few kn. This range provides fairly good differentiator
circuits, but small Vout• As in the case of integrator circuits with passive

Figure 9.13
The capacitor-resistor (C-R) circuit.

resistors, differentiators of the type of Fig. 9.13 have limited flexibility when
implemented using linear resistors. An adjustable time-constant requires that
the differentiator circuits use active elements, such as the transconductance
Integrator-Differentiator Circuits 263

amplifier, to implement the resistor component. The simplest of these circuits

is thefollower-differentiator (see Fig. 9.14).

�----�----�
�) Vout

Figure 9.14
The follower-differentiator circuit.

9.4 The Follower-Differentiator Circuit

Like its cousin the follower-integrator (see Section 9.1), the follower-differentiator
comprises a unity-gain follower and a capacitor. However in this case, the ca­
pacitor is connected to the input node rather than the output. Furthermore, the
input signal is applied to the negative input terminal of the amplifier, while the
positive terminal is connected to the reference voltage.
Intuitively we can see that the unity-gain follower tries to clamp Vout to the
reference potential, while changes in the input signal are capacitively coupled
to Vout and act against the clamp. The output of this circuit in response to a
step input is shown in Fig. 9.15.
In the small signal regime, the behavior of the circuit is described by

lout GVout (9.4.1)

d
=

lout = C dt (Vin Voud - (9.4.2)

where G is the amplifier's transconductance. The Heaviside-Laplace transform

(Section 8.7) is used to write the circuit's transfer function:

H(s)
Vout � (9.4.3)
Vin 1+TS
= =

whereT is C / G. Like the C - R differentiator circuit, the follower-differentiator

acts as a high-pass filter (H(s) � TS for TS «1 and H(s) � 1 for TS » 1).
264 Chapter 9

0.55

1--- �in 1
1-- � --- ..... - ... -- ............. - _ .. - ....... - -- ..... - ..... ...

- out

�
$0.375
"0
>

2 4 6 8 10
Time (ms)

Figure 9.15
Step response measured from a follower-differentiator circuit.

This circuit pennits 'T to be adjusted over several orders of magnitude, but its
output saturates to ±h for large input variations leading to distortion in the
circuit's response.

r-------�+

V ;n A1
0---+--4-----1

Figure 9.16
The diffl circuit.

9.5 The diffl and diff2 Circuits

These two types of circuits, previously described in Mead (1990), exploit

the fact that an approximation of a temporal derivative can be obtained by
Integrator-Differentiator Circuits 265

comparing a signal to its time-averaged version (for example, by subtracting

a low-passed version of the signal from its original version). The most direct
way to compute a time-averaged of a signal, using VLSI circuits, is to use
the follower-integrator circuit of Section 9.1. The simplest way to subtract
two voltage signals is to apply them to a transconductance amplifier: Vout =

A(Vin+ - Vin_) (where A is the amplifier's voltage gain). The diffl circuit,
shown in Fig. 9.16, does exactly this: It subtracts the low-passed version of the
input signal Vc from the input signal Vin. The problem with this arrangement
is that amplifier A2 is in an open-loop configuration. Consequently, if the open
circuit voltage gain is high (as is typically the case - see Section 5.3), then any
small input offset will be greatly amplified and the output voltage of the diffl
circuit will be clamped at one of the power-supply rails, even with steady-state
inputs.

>------_-0{) V out

Figure 9.17
The diff2 circuit.

On the other hand, the diff2 circuit has both of its amplifiers arranged
in a negative-feedback configuration, and subtracts the time-averaged version
of the output voltage from the input (see Fig. 9.17). In this case, amplifier
A2 is configured as a follower-integrator and its output, which is the low­
passed version of Vout, is fed back to amplifier AI. For constant and slowly
varying signals, this circuit is simply a unity-gain follower (see Section 5.3):
The negative feedback on Al drives Vout to values very close to Vin. In this
case, offset voltages are multiplied by unity gain (A/(A+1))2 rather than by
the amplifier's open loop gain A.

2 This is close to unity if \1, is biased properly (that is. voltage offsets are not amplified).
266 Chapter 9

100 �--- °
---���--�-
4
10-2 10 10
2 10
Angular frequency (Hz)

(a)

100

80

g> 60
E
Q)
If)
I1l
if 40

20

°
10 102
Angular frequency (Hz)

(b)

Figure 9.18
Bode plot of the diff2 circuit's transfer function. (a) Magnitude (b) Phase.
Integrator-Differentiator Circuits 267

The diff2 circuit's characteristic equations are

Vout = A{Vin - Vc) (9.5.1)

1
Vc = ---
1+rs
Vout (9.5.2)

where A is the open circuit voltage gain of AI, r CjG, and G is the
transconductance of A2. Substituting Eq. 9.5.2 into Eq. 9.5.1 we obtain

A 1+rs
(9.5.3)
A+1 1+{rj{A+1))s·

{
This transfer function can be simplified for the following three domains:

!
A l forrs « 1
Vout,...., A
A+ l rs for ATS «1 «rs (9.5.4)
Vin '" +l
A for ;�l » 1.

The frequency response of this circuit can be found by setting s = jw

and evaluating the magnitude and phase of the circuit's transfer function. The

10

O ����������--�����
1� 1� 1� 1if 1� 1�
Angular frequency (Hz)

Figure 9.19
Actual frequency response measured from a diff2 circuit.

Bode plot of Fig. 9.18 illustrates the approximations made in Eq. 9.5.4: At
low frequencies the diff2 circuit acts as a unity-gain follower; at intermediate
frequencies as a differentiator; and at high frequencies as an amplifier.
268 Chapter 9

The transfer function of Eq. 9.5.3 (and the corresponding frequency re­
sponse of Fig. 9.18) does not take into account the effect of parasitic capac­
itances that are present in physical implementations of the circuit. The mea­
sured frequency response curve from a real diff2 circuit is shown in Fig. 9.19.
The main difference between this plot and the one obtained from a first or­
der analysis of the circuit, is the distinct resonant peak at high frequencies. As
the input frequency increases, rather than smoothly saturating to the gain A of
the amplifier, the response first peaks and then decreases well below A. This
behavior is largely due to the parasitic capacitance at Vout•

0.
1

� IA.
'5 o
o
>
<J

0.
- 10L- -
-�
0. - - - 0-.�
002 00- 4- - 0-.0 - � .0�08- - -0�
� 0-6- 0 .0-1 - - 0
� .01 2
Time (5)

Figure 9.20
Response of a diff2 circuit (solid line) to a small (.6.l{n=30 mY) step input signal (dashed line).

Small-Signal Step Response

Figure 9.20 shows the response of the diff2 circuit to a small step input
(�Vin 30mV). The circuit's response has high transient gain, linear decay
=

and ringing. The transient gain is high because there are frequency components
in the input signal high enought that l�l » 1 (see Eq. 9.5.4). On the other
hand, the linear decay on the other hand is a consequence of the slew-rate
limited of the follower-integrator in the diff2 circuit (see Section 9.1). Finally,
the ringing is present due to the same parasitic capacitance on the Vout node
that caused the resonant peak in Fig. 9.19.
Integrator-Differentiator Circuits 269

1.5

v.,.-�- - - - - - - - - -

-1

-1.5

o 0.002 0.004 0.006 0.008 0.01 0.0 12

Time(s)

Figure 9.21
Response of a diff2 circuit (solid line) to a large (t. \{n=600 mY) step input signal (dashed line).

Large Signal Step Response

If the gain A of the amplifier is of several hundred, a change of Vin by a value

greater than a few millivolts brings Vout to the power supply rails. The response
peaks are clipped by these limits (see Fig. 9.21). The voltage swing would be
much higher, if the transient gain was not limited by the power supply rails.
Except for the regions where the voltage is clipped, the profile of the circuit's
response to large input steps is qualitatively equivalent to the one obtained for
small input steps: there is a high transient gain, a linear decay and ringing.

Figure 9.22
Hysteretic differentiator circuit.
270 Chapter 9

9.6 Hysteretic Differentiators

The unity-gain follower in the diff2 circuit can be regarded as a resistive el­
ement with a linear region for small signals, and a compressive non-linearity
for larger signals. An entirely different characteristic is obtained if this element
is replaced by an element with an expansive non-linearity, in which the resis­
tance is large for small signals and gets smaller for larger ones. Elements of
this type, that have exponential current-voltage characteristics in both direc­
tions, can be constructed with simple MOSFET circuits. An obvious solution
is a configuration of two antiparallel diode-connected MOSFETs of the same
type (Fig. 9.22) (Mead, 1989). Another implementation of a bidirectional ex­
ponential element, involving a single MOSFET and parasitic bipolar junction
transistors, will be discussed in Section 10.4.

2.4

2:
Q)
2. 1
N
(5
>
-------- ••
2
V.
In

1.9

2 3 4 5
Time (s) X 10-3

Figure 9.23
Small-signal sine-wave response of the hysteretic differentiator. showing open-loop amplification.

A hysteretic differentiator with an exponential resistive element can exhibit

a variety of different characteristics depending on the frequency and amplitude
of the input signal. Due to the strong non-linearity of the resistive element,
the time-constant of the feedback loop varies over several orders of magnitude
as a function of the voltage difference between the output terminal and the
Integrator-Differentiator Circuits 271

feedback terminal. Small voltage differences result in large time-constants and

large voltage differences in small time-constants.

,. , ,- , ,
, , , ,
\ I \
,
2.5 " \ \
I I , \
?:: I
\
I
I \
OJ 'V. ,
CJ) In I I \
.!!l 2 I I .
0 I I
I I
I
> I
I I
I I
1.5

2 3 4
Time (5)

Figure 9.24
Large-signal sine-wave response of the hysteretic differentiator, showing hysteretic following
behavior.

We will first examine the circuit characteristics for a sine-wave input.

For small signals and high frequencies the time-constant of the feedback
loop is much larger than that of the input signal, and the transconductance
amplifier acts as an open-loop circuit in which the signal is amplified linearly
by the open-loop voltage gain of the amplifier, as plotted in Fig. 9.23. If
the frequency is lowered, or the amplitude increased, then the voltage on the
feedback node will begin to follow the input voltage. Since the time constant
is strongly dependent on the voltage difference between the output and the
feedback node, Vc will lag behind \lin for small differences and then suddenly
catches up when Vout has a large enough excursion. These dynamics result
in a following-behavior with possible overshoots during the catch-up phase.
The following-behavior occurs whenever Vout and Vc are sufficiently different
that the time-constant of the feedback circuit is small. During following, Vout
thus has an offset with respect to Vin, whose sign depends on the direction
of the current flow, which in tum depends on whether Vin is increasing or
decreasing. This jump of the DC offset in the following-mode as a function
272 Chapter 9

of the direction of the current flow can be regarded as hysteretic behavior

and gives the circuit its name. For low frequencies and high amplitudes the
overshoots disappear and the circuit remains in the hysteretic following mode
with the signal discontinuities appearing whenever the sign of the derivative of
the input signal changes. This situation is shown in Fig. 9.24.

3.5

3 V
out
1\ \ 1\
2.5
2: ,-__..J- ,r-----J -------,
Q)
______

OJ vIn
� 2
(5
-------
,

> ,

1.5

1
0 2 3 4 5
Time (5) X 10-3

Figure 9.25
Large-signal square-wave response of the hysteretic differentiator, showing differentiating and fol­
lowing behavior. The asymmetric behavior and the DC offset are due to the use of an asymmetric
resistive element in the circuit.

For square-wave inputs, small signals are also amplified by the open­
loop gain, while larger signals are followed with overshoots at the transitions
(Fig. 9.25). These overshoots can be regarded as a differentiating character­
istic. The data plotted in Figs. 9.23-9.25 were obtained from a hysteretic
differentiator with an asymmetric resistive element (see Fig. 1O.11(a) in Sec­
tion 10.4). This asymmetry explains the DC offset between input and output
signals and the asymmetric response to square-wave input signals.
An interesting variant of the circuit is obtained by replacing the resistive
element with a bidirectional source-follower circuit (Fig. 9.26) (Mead, 1989).
This element has a very high impedance at the Vout node: The currents to
charge and discharge the capacitance at the feedback node are obtained from
the power rails via the source followers, and not from the transconductance
Integrator-Differentiator Circuits 273

Figure 9.26
Hysteretic differentiator circuit with rectifying current outputs.

amplifier. Thus, the circuit has a faster large-signal response than the circuit of
Fig. 9.22. Furthermore, if the currents I;tut and I�t are used as output signals
the circuit acts a rectifying temporal differentiator (Kramer et aI., 1997) with

for ltVin > 0

(9.6.1)
for 1, Vin < 0 .

The output signals for positive and negative temporal transients appear at two
different terminals, both of which has very small DC responses. This property
reduces DC offsets due to variations in fabrication parameters, which often
limit the performance of analog integrated circuits. In addition, the responses
are thresholded, because a substantial excursion of Vout is required to draw a
significant current at either terminal. This property is useful for suppression
of low-amplitude temporal noise introduced by thermal effects and statistical
fluctuations in the input signal.
 10 Photosensors

Photosensors convert electromagnetic radiation into a different physical form,

usually electrical charge. The growth in the market for optical communica­
tion and electronic imaging has brought a hugh importance to the photosen­
sor as optoelectronic interfaces. Consequently, a wide range of photosensors
designed for different applications are available. Here, we will consider only
photosensors that can be fabricated with semiconductor processes used for the
implementation of transistor-based electronic devices.

10.1 Photodiode

A photodiode is a diode that is used as a photosensor. Commercial photodiodes

are specifically tailored to different application domains. However, since the
diode forms part of every transistor, photodiodes are also available in standard
semiconductor processes, and so they can be monolithically integrated with
electronic circuitry.
In a semiconductor, an incident photon and therefore its energy can be ab­
sorbed by an electron; a process known as the inner photo-electric effect. A
photon with an energy larger than or approximately equal to the bandgap en­
ergy can excite an electron from the valence band into the conduction band.
In the energy-band diagram, this process corresponds to the generation of an
electron-hole pair. Illumination of a semiconductor therefore increases the con­
centration of mobile charge carriers above the thermal equilibrium value in
the exposed area. If the motion of the carriers is driven by diffusion, then the
generation is balanced by recombination. However, in the presence of an elec­
tric field, electrons and holes can separate and some of the separated carriers
contribute to an electrical output signal. This phenomenon is called photocon­
duction. The simplest device that implements the phenomenon is the photo­
conductor, which is a slab of semiconductor in an externally applied electric
field. Photoconductors exhibit a large dark current, which is a background cur­
rent which is still present in the absence of optical stimulation. This current is
due to the relatively large doping concentration used in most modem semicon­
ductor processes. This doping results in high conductivity values and a poor
signal-to-noise ratio of photoconductors.
A diode is a much more suitable photosensor, because it has a depletion
region with a low conductivity and a built-in electric field. The presence of
276 Chapter 1 0

a depletion region substantially reduces the dark current, while the built-in
electric field in the depletion region performs charge separation even in the
absence of an externally applied voltage. Electron-hole pairs generated in the
depletion region and within a diffusion length of it, are likely to be separated
(Fig. 10.1). The resulting reverse current component is called photocurrent.
As in the photoconductor, an incident photon cannot contribute more than one
electron to the photocurrent.

Photon

0_0-0
- -
o 0- 0
-
-0- ® 0
-
0- � 0 -
p reg ion Depletion reg ion n region

Hol e Electron
Drift
diffusion d iffusion

Figure 10.1
Principle of operation of a photodiode. Electron-hole pairs generated by incident photons in or
within a diffusion length outside the depletion region become separated and contribute to a reverse
generation current.

If the photodiode is open-circuited, that is, no external current is allowed

to flow, then generated charge accumulates at the boundaries of the depletion
region until a steady state is attained. In steady state, a forward bias appears
Photosensors 277

across the junction, such that the photocurrent is balanced by a forward diffu­
sion current. If the photodiode terminals are short-circuited the photocurrent
can be measured as a reverse diode current. In the presence of an applied ex­
ternal bias the photocurrent is superimposed onto the diode current discussed
in Chapter 2. The current-voltage characteristic of a photodiode has the same

I
shape as that of a normal diode, but the curve is displaced along the current
axis by the value of the photocurrent ph (Fig. 10.2).

II

III IV

Figure 10.2
Steady-state current-voltage characteristics of a photodiode. The upper curve is the normal diode
characteristic (dark characteristic). The lower curve shows the characteristic under illumination.
Photodiodes are usually operated either in quadrant III as photosensors, or in the quadrant IV as
solar cells.

A photodiode has two principal modes of operation, depending on its

application. If the photodiode is used to convert optical power into electrical
power it is called a solar cell and operated in the photovoltaic mode. In this
mode, a load is connected between the two terminals, such that a reverse
current flows through the diode in the presence of a forward voltage. A solar
cell is thus operated in the fourth quadrant of the current-voltage characteristic.
278 Chapter 1 0

The generated power is given by the product of the reverse current and the
forward voltage. Commercial solar cells are complicated devices, which are
optimized with respect to optical-to-electrical power-conversion efficiency that
is typically between 10% and 20% 1 •

In the other mode of operation, the photosensing mode, the photodiode

is used to estimate the photon flux. In steady state, the diode is typically
either open-circuited and the forward voltage is read out or a reverse (or zero)
bias is applied to the diode and the reverse current is read out. In the latter
case, the photodiode is operated in the third quadrant of the current-voltage
characteristic. Here, the photodiode is quite a good current source, because the
photocurrent is almost independent of the applied reverse bias.
In electronic imaging applications, the outputs of an array of photodiodes
are read out in sequence onto a shared wire. In these applications, the photo­
generated charge is usually integrated while the photodiode is not addressed.
During readout the charge is either transferred as a current pulse onto the out­
put wire or the voltage change across the diode due to the accumulated charge
is sensed and the charge is drained after readout before a new integration cycle
is started.

Photodiode Characteristics

In order to compute the photodiode current we have to consider how light

is absorbed in the material. At any given position in the semiconductor the
number of absorbed photons is proportional to the total number of photons for
a given photon energy. This results in an exponential decrease of the photon
flux � per unit area of the material with penetration depth x, according to

(10.1.1)

where �o denotes the flux of photons per unit area penetrating the semicon­
ductor surface and a is called optical absorption coefficient. The optical ab­
sorption coefficient is the inverse of the distance over which the photon flux
is reduced to a fraction of lie from its initial value. Figure 10.3 shows the
optical absorption coefficients and light penetration depths for different pho­
tosensing materials as a function of optical wavelength. Photons with larger
wavelengths penetrate deeper, because they have less energy and are thus less
likely to generate an electron-hole pair at any given location in the material.

1 High-efficiency solar cells are built with a layering of different semiconductors of different
bandgaps.
Photosensors 279

105 10-1

,-
E E
�
1:$
.3
C �
Q)
'0
104 �
ii
�0 Q)
"0
U C
C 0
0
.E- �
a;
(; c
(J) Q)
.0 10 3 101 0-
ro
Cii :c
u Ol
::i
·li
0

10 2 10 2

Wavelength (�m)

Figure 10.3
Optical absorption coefficients for different semiconductor crystals at room temperature as a func·
tion of wavelength of the incident light. Figure adapted from H. Melchior (1972), Demodulation
and photodetection techniques, in Laser Handbook, Vol. I, 725-825. ©1972, with permission
from Elsevier Science.

The wavelength at which the absorption coefficient drops to zero is called

the cutoff wavelength Ac and corresponds to a photon energy that equals the
bandgap energy. For silicon, Ac 1 . 1 f.lm, which is in the near infrared. The
=

term «l>o can be computed from the incident optical power P opt as

1 -R A
«l>o =
--y he Popt (10.1.2)

where R is the reflection at the semiconductor surface, A is the cross-section of

the semiconductor perpendicular to the direction of the photon flux, and he/ A
is the photon energy as given by the photon wavelength A, the Planck constant
280 Chapter 1 0

h, and the speed of light c . The generation rate G for electron-hole pairs can
be computed from the attenuation of the photon flux as

G(X) -- - dcJ(x) -a�o

dx
-- J,._
e-ax . (10.1.3)

The photocurrent density J ph consists of a drift component J ph ,drilt due to

carriers generated in the depletion region and a diffusion component J ph ,dil I
due to carriers generated outside the depletion region. We can compute these
components under the assumptions that the thermal generation current is much
smaller than the photocurrent and that the neutral semiconductor layer the light
has to cross to reach the depletion region is much thinner than 1/a. For the drift
component we obtain

Jph,drilt = -q loW G(x)dx = q (cJ(W) - cJ(O)) = - qcJo ( 1 - e-aW)

(10.1.4)
where W is the depth of the depletion region in the direction of the photon flux.
The diffusion current density can be computed from the continuity equations
(Eqs. 2.5.20 and 2.5.21) under the Shockley approximation stated in Chapter 2.
It consists of the reverse diffusion current density under dark conditions, which
for a sufficiently large reverse bias is computed from the Shockley equation
(Eq. 2.6.22) to be -Js, and the photogenerated contribution from the bulk
region, which is obtained from the one-dimensional diffusion equation as
aL -aW
1 +aL e
Jph ,dill = -qcJ0 (10.1.5)

where L is the minority carrier diffusion length in the bulk. The total photodi­
ode current density is then given by

J = Jph - Js = Jph ,drilt + Jph ,dil I - Js = (

- qcJo 1 - l
e a
: :L ) - Js .
(10.1.6)
Note that J depends only on the applied reverse bias via the width of the
depletion region W. For an ideal photodiode Js, as given by Eq. 2.6.23, is the
only contribution to the dark current density. In a real photodiode, the dark
current density is significantly larger than J s and depends somewhat on the
applied reverse bias. However, the operating conditions of photodiodes are
usually chosen such that the photocurrent is much larger than the dark current.
The reverse current density is then proportional to the incident photon flux
cJo per unit area This linear relationship is expressed by the definition of the
Photosensors 28 1

100
05 X:\" Responsivity (A/W)
\ 1
\
\
\ 2
\
\
\
\ Ge \
80 \ \ \
\ \ \
\ \ \
\

()
� \ \
� \ \

�
>. \ \
0.2 '.
c 60 \
Q) \
� \ \
Q; \ IOG'AsP \
\ \
E \ , \
:::J \ ,
C ,
cu 40 , , ,
:::J , ...... ......
a ...... ...... ......
...... ...... ...... .
" "
" "
" "
20
.......... ' ...... ,
...... , "
' - --
--.

0
0.4 4
Wavelength (�m)

Figure 10.4
Quantum efficiencies and responsivities of photosensors fabricated from different semiconductors.
Silicon exhibits a very good quantum efficiency peaking in the near infrared. Also shown are lines
of equal responsivity. For a given responsivity the relationship between quantum efficiency and
wavelength is inverse, as can be seen from Eq. 10.1.7. Figure adapted from S. M. Sze (1981),
Physics of Semiconductor Devices, 2nd Edition. ©1981 by John Wiley & Sons, Inc. Reprinted by
permission of John Wiley & Sons, Inc.

quantum efficiency"., as the number of electron-hole pairs contributing to the

photocurrent generated per incident photon. The quantum efficiency can be
computed from Eqs. 10.1.2 and 10.1.6 as the ratio of the generated charge
carrier flow density J ph / q to the incident photon flux density >'P opt / Ahe as

".,
=
Ahe Jph
q>. Popt
=
he [ph
q>. Popt
= )
(1 _ R 1 _
( e-aW
1 + o:L
) (10.1.7)

where [ph =
AJph is the photocurrent. The ratio of photocurrent to incident
optical power [ph / Popt is called responsivity. Typical quantum efficiencies and
responsivities of photosensors fabricated with different semiconductor mate­
rials are shown in Fig. 10.4. Silicon has a very good quantum efficiency in
the visible and near infrared, which may approach 100% in a certain spectral
282 Chapter 1 0

range. The corresponding responsivity is on the order of 0.5 A/W. The above
analysis does not take into account that charge carriers generated near the semi­
conductor surface are likely to recombine due to surface effects. Photocurrent,
quantum efficiency, and responsivity for blue light are therefore much lower in
practice than expected from the above formulas.

Types of Photodiodes

Photodiodes designed to be operated in the continuous-current photosensing

mode are usually optimized with respect to their quantum efficiency and their
response time, which are two partly conflicting requirements. The quantum
efficiency can be optimized by applying an anti-reflection coating to the semi­
conductor surface to reduce R and by generating a thick depletion region close
to the surface. The response time can be kept small by minimizing the junc­
tion capacitance, the carrier transit time through the depletion region, and the
carrier diffusion time to the depletion region. A small junction capacitance re­
quires a thick depletion region, while a short transit time favors a thin depletion
region and a large drift velocity. Diffusion times to the depletion region can be
minimized if the depletion region extends close to the surface. It is thus ad­
vantageous to operate a photodiode at a large reverse bias in order to increase
the depletion region width and the drift velocity. The depletion region width
can also be increased by having a low impurity-doping concentration in the
junction region, as is done in most commercially-available photodiodes. Such
photodiodes are known as p-i-n photodiodes, because they have a (nearly) in­
trinsic region between the n-type and p-type region.
In the photodiode operation range described so far, the quantum efficiency
is smaller than unity: Each photon cannot produce more than one electron-hole
pair. However, if a diode is operated in the avalanche multiplication regime in
the vicinity of reverse junction breakdown, the photogenerated carriers multi­
ply in the depletion region due to impact ionization and the quantum efficiency
can be significantly larger than unity. Avalanche photodiodes are photodiodes
designed to be operated in this domain. They have small response times and
better signal-to-noise ratios than normal photodiodes. A normal semiconductor
process provides very poor avalanche photodiodes with instability and match­
ing problems. A photosensor with an internal gain mechanism that may be
implemented more efficiently with standard processes is the phototransistor,
described in the following section.
Photosensors 283

10.2 Phototransistor

A phototransistor is usually fabricated as a bipolar junction transistor (BfT)

structure. One type of phototransistor that is available in standard CMOS
processes is the parasitic BIT formed by a source/drain implant as the emitter,
a well as the base, and the substrate (or another source/drain implant) as the
collector.

100

0.1

10-10 10-9 10-8 10-7 10-6 10-5 10-4 10-3 10-2 10-1 10-0
Ic(A)

Figure 10.5
Static common-emitter current gain hFE and small-signal common-emitter current gain hIe
versus collector current Ie of a phototransistor. Figure adapted from A. S. Grove (1967), Physics
and Technology of Semiconductor Devices. ©1967 by John Wiley & Sons, Inc. Reprinted by
permission of John Wiley & Sons. Inc.

In such a bipolar phototransistor, the photon-generated electron-hole pairs

are separated in the depletion region of the reverse-biased base-collector junc­
tion, resulting in a collector current that is the photodiode current pd of the
junction. In a phototransistor, the base is floating and its potential is deter­
I
mined by the concentration of photogenerated majority carriers in the base,
and therefore by the photon flux. This irradiance dependence of the base volt­
age causes a light-induced modulation of the collector current. An additional
contribution to the collector current is thus obtained by carrier injection from
the emitter through the base. This component is an amplification of the photo-
284 Chapter 1 0

diode current by the common-emitter current gain h FE2, which is typically on

the order of 100. The resulting total photocurrent is

(10.2.1)

Phototransistors are useful in applications where significant leakage cur­

rents generated by the connected circuitry have to be overcome or large capac­
itive loads have to be driven directly by the photocurrent, because they amplify
a typical photodiode current in the pA range into the nA range. However, the
phototransistor has several disadvantages with respect to the photodiode. It
uses more area and its response is not linear with the photon flux, because
hFE depends on the current level, as shown in Fig. 10.5. In addition, it has an
inherently slow response because the base potential changes with illumination.
This necessitates the charging and discharging of the large base-collector ca­
pacitance, while in a photodiode the voltage at both terminals can be clamped,
making it typically about two orders of magnitude faster than a phototransistor,
as we will see in the next section. Furthermore, the amplification mechanism
degrades the signal-to-noise ratio of the phototransistor with respect to that of
the photodiode, and it worsens with increasing gain.

10.3 Photogate

The Metal-Insulator-Semiconductor (MIS) structure discussed in Chapter 2

can be used as a photosensor, where one type of photogenerated charge carriers
is collected in a depletion region underneath the conductor plate, while the
other type leaves the semiconductor via the substrate. In this application, the
MIS structure is referred to as a photogate.
Since in most cases the light impinges onto the structure through the
conductor plate, the quantum efficiency, particularly for blue light, is reduced
by absorption in and reflection at the conductor plate. This effect is mimimized
by using thin polysilicon as conductor plates instead of metal.
As opposed to the photodiode, the photogate is by nature an integrating
photosensor, which accumulates the charge until it is read out by a change
in the applied voltages. It is therefore not surprising that it is mainly used in
image sensors. As a matter of fact, it is the basis of the most common image
sensor, the Charge-Coupled Device (CCD), where the MIS structure is not only
used for charge integration but also for charge transportation by appropriate

2 The common-emitter gain is often called /3.

Photosensors 285

clocking of the bias voltages applied to neighboring MIS gates, as we will see
in Section 10.5.

-�1¥-- -
- - - inversion
n d iffu s ion

-
��-�- -� -
- _ depletion _ - - _ - -_ depletion
-
- -

- _0 0:0� 0_0
0:0� 0
-_ 0-_0 -00 p·substrate - 0- 0
_ -0� p. substrate
0_ - 0� 0 -
0-_ - 0� - 0
l.U :::c "*
(a) (b)

Figure 10.6
Cross-sections of photogates with (a) surface charge storage and (b) bulk charge storage. The
energy E of the conduction band edge in the semiconductor as a function of depth is shown on the
left of each cross- section.

The simplest photogate collects the charge carriers in a uniformly doped

semiconductor substrate. The gate voltage is chosen such that the semiconduc­
tor surface is depleted of majority carriers. In the depletion region, and within
a diffusion length from it further in the the bulk the photogenerated electron­
hole pairs are likely to get separated such that the minority carriers drift to
the surface and the majority carriers to the bulk, as depicted in Fig. 1O.6(a)
for a p-type substrate. The minority carriers stay at the surface as an inversion
layer, but due to structural imperfections at the semiconductor-insulator inter­
face caused by lattice mismatches and impurities, some of them get trapped
in interface states with different time constants, which causes smearing effects
when the collected charge is removed from underneath the gate electrode. In
applications where high bandwidth is required and where the charge needs to
be transported over large distances, as in the case for CCDs, it is therefore de­
sirable to collect the charge in a region removed from the interface, because
the density of trapping centers is much smaller in the bulk than at the surface.
Such a charge distribution is accomplished by collecting the charge carriers
in a depleted diffusion layer between the interface and the bulk with a doping
286 Chapter 1 0

opposite from the bulk doping. The collected charge carriers are majority
carriers in this diffusion layer and replenish the depleted states, as shown in
Fig. 1O.6(b) for an n-type diffusion layer. Biasing of such a bulk-storage device
is a little bit trickier than for a surface storage device. In the case of an n­
type diffusion layer, for example, the diffusion layer has to be biased with a
much more positive potential than the gate and the bulk to achieve depletion
at the semiconductor surface and at the junction between diffusion layer and
bulk. The charge carriers then collect within the diffusion layer between the
two depletion regions. Furthermore, the collected charge has to be laterally
confined to prevent it from leaking directly to the diffusion contacts, as in
a photodiode. This can be done with additional gate electrodes at nearby
locations that are held at a lower potential than the charge-collecting electrode,
or with lateral p-type diffusions. While surface storage has a large charge­
retaining capacity, the maximum charge-storing density of a bulk storage
device cannot exceed the doping density of the diffusion layer. Additional
photo-generated charge spills over to neighboring locations and the diffusion
contacts. Despite this disadvantage, bulk storage is much more common in
commercial CCDs than surface storage due to the reduction of the above­
mentioned trapping effects.
Surface-storage photogates can be fabricated with the same silicon­
processing technology as standard MOSFETs, but for commercial devices spe­
cial processes with a cleaner silicon-insulator interface are used. Bulk-storage
requires a moderately doped diffusion layer underneath the gates, which is not
available in standard MOSFET technology, but is provided as an additional
implant in some processes.
As we will see in Section 10.5, CCDs using surface-storage photogates
are called surface-channel CCDs and those using bulk-storage photogates are
referred to as buried-channel CCDs.

10.4 Logarithmic Photosensors

In this section we will make use of the fact that photodiodes can be readily
integrated with transistors in standard MOS technologies to show different
ways to convert the photocurrent logarithmically into a voltage signal.
The logarithmic characteristic turns out to be quite for environments in
natural ambient lighting. Such lighting conditions are characterized by a time­
varying intensity, due the continuously shifting angle of incidence of the
Photosensors 287

Table 10.1
llIurninance and irradiance values under typical lighting conditions

Light source Illuminance [Ix] Irradiance [W/nr]

Full moon 1 0.01
Street lighting 10 1
Office lighting 100-1000 10-100
Direct sun 100000 1000

sunlight and changing optical transmission of the atmosphere (obscuration by

clouds, etc.). Such temporal variations simultaneously affect large parts of the
environment, but do not carry any information about it. In order to robustly
characterize an environment in this situation, the characterization has to rely
on parameters that are insensitive to global illumination changes.
Object surfaces are optically characterized by their spectral reflectivity dis­
tribution. Since the ambient light level is usually unknown, but can be assumed
to be locally approximately constant (except at edges of shadows), the relative
reflectivity of nearby areas in object space becomes an important parameter for
the description of the objects. A measure for this relative reflectivity as mapped
by an optical system into an image space is called contrast. There are differ­
ent ways of defining contrast, but they all have in common that they are based
on ratios of irradiances or illuminances 3 in the image and do thus not depend
on the ambient scene illumination. A logarithmic mapping has the property
that a ratio is mapped onto a difference. Consequently, in a circuit where the
photocurrent is logarithmically converted into a voltage, contrasts are encoded
as voltage differences, while the ambient light level sets the absolute voltage
value of the operating point. Illuminances and irradiances for typical lighting
conditions are shown in Table 10.1.
To simplify the equations in this section, we will assume that the width-to­
length ratios W/ L of the MOSFETs are taken into account by their respective
current-scaling constants 10 and we will use a first-order approximation of
the subthreshold MOSFET characteristics, neglecting the Early effect, unless
stated otherwise.

Basic Implementations

As we saw in Section 10.1, the current through an ideal photodiode is approx­

imately linear with the incident power for a given spectral distribution of the

3 Irradiance denotes the incident radiant flux (power) per unit area, while illuminance denotes the
incident luminous flux (obtained by weighting the radiant flux withthe human photopic sensitivity
curve) per unit area.
288 Chapter 10

incoming radiation. This approximation turns out to be quite good over several
orders of magnitude in the photocurrent, corresponding to a substantial part
of the illuminance range encountered in natural environments. It is thus possi­
ble to build electronic image sensors that operate over a large dynamic range
without the need for any additional range-reducing devices, such as aperture
stops, typically used in video cameras or light amplifiers. Depending on the
particular application the photocurrent may have to be converted into a volt­
age. For certain applications, a linear conversion may be desirable, but given
a power supply voltage of a few volts and typical noise levels of integrated
circuits of the order of a few millivolts, the dynamic range in the voltage do­
main spans only about three orders of magnitude. In order to retain the large
dynamic range of the photodiodes it is thus appropriate to use a compressive
mapping, which is easily available via the voltage-current characteristics of
transistors (Chamberlain and Lee, 1983). The most straightforward implemen­
tation of such a mapping is by connecting one or more transistors in series
with the photodiode, such that the photodiode acts as a current source and the
transistors as current sensors. Several possible MOSFET-based configurations
of such a circuit are shown in Fig. 10.7. A diode-connected pFET is used as
a current sensor in Fig. 1O.7(a). A stack of two diode-connected pFETs, as
shown in Fig. 1O.7(b), increases the voltage signal's amplitude and moves its
operating point further away from the power supply rail. The source-follower
configuration of Fig. 10.7(c) sets the operating point with a bias voltage Vb, but
further compresses the signal with respect to the other implementations due to
the subthreshold slope factor of the MOSFET (see Chapter 3), This compres­
sion can be avoided if a unity-gain source follower (see Section 5.2) with the
source connected to the bulk is used. Of course, all these configurations can
also be implemented by simultaneously exchanging the type of MOSFET, the
direction of the photodiode, and the power supply rails, or by using other types
of transistors.
Typical irradiances under ambient lighting conditions (see Table 10.1)
elicit photocurrents in the pA or nA range in photodiodes with areas of the
order of (10 JLm)2 as employed in imagers. At these current levels typical
MOSFETs operate in their subthreshold domain, such that the current-to­
voltage conversions performed by the circuits of Fig. 10.7 show a logarithmic
characteristic. We then obtain

Vout = Vdd -
UT
-log
K,
( Ip10h)
- (1004.1)
Photosensors 289

(a) (b) (c)

Figure 10.7
Photosensors with logarithmic irradiance-to-voltage conversion for subthreshold photocurrents,
consisting of a photodiode and a current-to-voltage conversion stage implemented as (a) a diode­
connected MOSFET, (b) two diode-connected MOSFETs in series, (c) a MOSFET in source­
follower configuration.

1ph10
for the circuit of Fig. 1O.7(a), where U T denotes the thermal voltage, K the sub­
threshold slope factor of M 1 and the current-scaling constant of the current­
sensing MOSFET. We define to be positive if it is a reverse photodiode
current, as indicated in Fig. 10.7. We obtain

Vout = Vdd -
UT
K ;1 log
(1;Oh) (10.4.2)

for the circuit of Fig. 1O.7(b) under the assumptions that both MOSFETs are
identical and that the voltage dependence of K can be neglected; and we obtain

Vout = KVg -
UT log
(To1Ph) (10.4.3)

for the circuit of Fig. 1O.7(c). The contrast-encoding property can be readily
290 Chapter 1 0

seen when the current-to-voltage conversion characteristics are written in dif­

ferential form. For example, differentiating Eq. 10.4.3 gives
dlph
dVout = -UT-1- (10.4.4)
ph .

The photodiodes in the circuits of Fig. 10.7 can also be replaced by photo­
transistors. However, the logarithmic characteristic is lost in this case because
of the nonlinear current-irradiance dependence of the phototransistor. In addi­
tion, if MOSFETs are used as current sensors they may be forced into their
above threshold operating domain by the amplified photocurrents provided by
the phototransistor, which further distorts the logarithmic characteristic.

Figure 10.8
Logarithmic photosensor with feedback loop increasing the bandwidth by clamping the voltage
Vs across the photodiode.

Feedback Implementation

A limitation of the photosensors depicted in Fig. 10.7 is their small bandwidth

for time-varying signals. This is due to the photodiode junction capacitance
and the parasitic capacitances of the output node onto the different MOSFET
terminals. These capacitances have to be charged or discharged by the typi-
Photosensors 29 1

cally very small photocurrent. If a phototransistor is employed, the amplified

collector current can easily handle the MOSFET capacitances, but the base­
emitter junction capacitance still has to be driven by the non-amplified pho­
tocurrent. Since the base voltage of a phototransistor has to be modulated to
cause transistor action the problem is inherent. In the case of the photodiode,
however, the response speed can be enhanced by adding a high-gain negative
feedback loop from the source to the gate of the current-sensing MOSFET in
the source-follower configuration of Fig. 1O.7(c). The voltage output signal
then appears at the gate of the MOSFET, while the source and therefore the
voltage across the photodiode is practically clamped. Such an arrangement is
depicted in Fig. 10.8, where the source node drives a two-transistor amplifier
with a bias voltage Vb and a voltage gain A whose output is connected to the
gate node. The output voltage is then given by

Vout = ( Iph
K,-1 Vs + UT log 10 ( )) (10.4.5)

with the source Vs being held nearly clamped to the constant value

Vs = (
K,�1 K,P(Vdd - Vb) + UT log (�::)) (10.4.6)

where K,n and K,p are the subthreshold slope factors of M 2 and M3 respectively
and Ion and lop are the corresponding current-scaling constants.
A differential photocurrent change dI ph results in a differential output
voltage change

dV,out - UT
A dlph ,..., UT dlph
K,A - 1 Iph '" K, Iph
. (10.4.7)

From Eqs. 10.4.4 and 10.4.7 we see that the feedback circuit amplifies the
response of the simple source-follower configuration by a factor 1/ K, and
inverts it, such that an increase in I ph results in an increase in Vout • The
measured steady-state irradiance-voltage characteristic of such a photosensor
is shown in Fig. 10.9. The differential source voltage change is
UT dlph � UT dlph
dVs . (10.4.8)
K,A - 1 Iph K,A Iph
=

Hence, the feedback circuit reduces the voltage variations on the source node
by a factor of K,A. If the bias current of the amplifier is chosen such that
it is much larger than the photocurrent then the time constant of the circuit
is determined by the dynamics of the source node of the current-sensing
292 Chapter 1 0

1.85

1.8

1.75

1.7
?:
:;

�o 1.65

1.6

1.55

0.01 0.1

Figure 10.9
Steady-state voltage response Vout of an implementation of the logarithmic photosensor of
Fig. 10.8 for different irradiance levels. The response function is to a good approximation log­
arithmic over the measured irradiance range spanning 4.5 orders of magnitude.

MOSFET as in the case of the simple source-follower version. If the additional

parasitic capacitance introduced by the feedback circuit can be neglected, the
feedback circuit reduces the response time by a factor of Ii:A. Note that this
increase in bandwidth decreases the time over which the signal is integrated
and correspondingly increases the noise at the output.

Adaptive Photosensor

In the following, we will illustrate how adaptation, an ubiquitous phenomenon

in biological nervous systems, can enhance the performance of a photosensor.
This example shows some benefits of adaptation that can be applied to other
data processing systems, as well.
A major concern in the design of photosensor arrays for imaging applica­
tions is mismatches between the characteristics of identically-designed pho­
tosensing elements. These mismatches are due to inaccuracies of the semi­
conductor processing. The effect of these mismatches on the output signal
is known as fixed-pattern noise (FPN). In the logarithmic photosensors de­
scribed in this section the current-sensing MOSFETs that are operated in their
subthreshold domain are the main source of FPN. An e-fold increase in the
photocurrent corresponding to the effect of a medium optical contrast elicits
Photosensors 293

a voltage change of the order of U T, which can be of the same order of mag­
nitude as the gate-to-source voltage mismatch between the MOSFETs for a
given current. The photocurrents are proportional to the areas of the corre­
sponding photodiode junctions, whose spatial variations are the main source
of photocurrent mismatches and are typically limited to a few percent.

....---DVout

Vs _-------i

E
J"'J""'+ D1

Figure 10.10
Adaptive logarithmic photosensor with amplified transient response. The amplification stage
consists of a capacitive divider and a resistive element, which typically shows a non-linear current­
voltage characteristic.

The FPN can be reduced by individual tuning of each photosensor. How­

ever, in addition to a calibration procedure, such a solution requires either in­
dividual biasing of each sensor or local memory at each element to store the
calibration parameter values. A more economic solution is the enhancement
of transient signals with respect to the mismatched steady-state signals using a
well-matched amplifier stage. This strategy increases the signal-to-noise ratio
and makes better use of the voltage range provided by the power supply. Such
a well-matched transient gain stage can be fabricated with a capacitive divider.
Adaptation to the DC value can be provided by a resistive element. The gain
294 Chapter 1 0

stage can be integrated into the feedback loop of the photosensor of Fig. 10.8,
as shown in Fig. 10.10 (Delbriick, 1993; Delbriick and Mead, 1994). The tran­
sient output voltage change dVout is amplified with respect to the DC output
voltage change dVfb (Eq. 10.4.7) by the capacitive divider ratio

A - C1 +C2 (10.4.9)
c =
C2
as long as the adaptation effect can be neglected. The transient output voltage
change is thus
A dlph
dv'out -
- A e UT 1 1
'"
'" A C UT dlph
1 . (10.4.10)
'"A - ph '" ph
The photosensor adapts to variations in the photocurrent on a long time scale,
which usually reflect slow changes in the background illumination that are
typical of natural lighting conditions. The adaptation state is represented by
the charge Qfb stored on the capacitor plates of the feedback node. The
output voltage Vout depends on this adaptation state and on the input signal
represented by Vfb. It can be expressed as

v,out AC V b Qfb +C1 Vdd

f -
-
-
(10.4.11)
C2
where Vfb is given by

(10.4.12)

The adaptation dynamics are determined by the characteristics of the resistive

element. A large resistance results in slow adaptation: a small resistance in fast
adaptation. A linear resistor with a resistance R makes Vout adapt exponen­
tially after an irradiance step. This can be seen from the differential equation
that is obtained when Vfb remains constant following an initial step change
.6. Vfb: Differentiating Eq. 10.4.11 and using Ohm's law yields

dVout _C-1 dQfb l

2 dt - (RC2 )- (v,out - Vfb·
) (10.4.13)
dt = =

If the sensor was fully adapted before the irradiance step at time t =
0 we
obtain the initial condition that immediately after the step

(10.4.14)
Photosensors 295

VIbO
lib
...
�M,
leu,
... OVout

(a)

Vib

(b)

(c)

Figure 10.11
Resistive elements suitable for implementation in the adaptive photosensor of Fig. 10.10. (a) Ex­
pansive element with low common-mode sensitivity. (b) Expansive element with more symmetric
characteristics. (c) Compressive element.

Using this condition, Eq. 10.4.13 is easily solved:

Vout(t) = Vfb + Ac �Vfbe-t/RC2. (1004.15)

With the specific capacitance and sheet resistance values provided by typ­
ical semiconductor technology the decay time constants achieved with linear
resistors would be much too small for practical applications. The resistive el­
ements are therefore more conveniently built from transistors, which can be
operated at low currents that are matched to typical capacitance values. The
296 Chapter 1 0

Figure 10.12
Current· voltage characteristics of the resistive element of Fig. IO.lI(a). The currents at both
terminals are shown for both directions of current flow. For Vaut > V fb the element operates
as a diode-connected MOSFET. For Vfb > Vout the MOSFET is shut down, but current flows
through a lateral BJT between the two terminals and through a vertical BIT from the feedback
node to the substrate.

current-voltage characteristics of these elements are non-linear, with the resis­

tance in most cases decreasing or increasing monotonically with the absolute
value of the voltage difference across the element. If it decreases we speak of
an expansive resistive element; if it increases we call it a compressive resis­
tive element. In the case of an expansive element the adaptation is faster than
exponential, and for a compressive adaptive element it is slower than exponen­
tial. Expansive elements provide rapid adaptation to new lighting conditions,
but do not adapt out low-contrast signals, which usually correspond to signifi­
cant image details rather than the effects of changes in the scene illumination.
Compressive elements may take a long time to adapt to large changes in illumi­
nation and can therefore support a larger voltage swing of the amplified signal,
which may be due to either a larger amplifier gain or to a larger input signal.
The price to be paid for this enhanced response is slow adaptation to new light­
ing conditions. During this adaptation, the response of the output node may be
saturated until a new equilibrium is reached.
Three different implementations of non-linear resistive elements are shown
in Fig. 10.11. The element of Fig. 1O.11(a) (Delbrtick, 1993; Delbrtick and
Photosensors 297

2.2

2.1

1.9
?::
:;
;:,.0
1.8

1.7

1.6

1.5
0 50 100 150 200
Time (5)

Figure 10.13
Voltage response Vout to an irradiance pulse of an implementation of the adaptive photosensor of
Fig. 10.10 with the resistive element of Fig. 10. I I(a). The response shows large transients followed
by gradual adaptation to the DC values.

Mead, 1994) is expansive. For current flow from the output node to the feed­
back node it has the exponential characteristics of a diode-connected pFET. In
the reverse direction the pFET is turned off, but the diffusion at the feedback
node acts as the emitter and the well as the base of a BJT with two collectors,
one being the diffusion at the output node, the other being the semiconduc­
tor substrate. The current-voltage characteristic, shown in Fig. 10.12, is thus
also exponential in this direction, but without the subthreshold slope factor K
in the exponent and with a much larger linear current-scaling parameter. The
response of an adaptive photoreceptor with such a resistive element to an ir­
radiance pulse is shown in Fig. 10.13. The asymmetry between upward and
downward adaptation can clearly be seen. A more symmetric expansive ar­
rangement, where the resistance can be modulated by a bias voltage Vb, is
shown in Fig. 1O.11(b) (Liu, 1999). The behavior in both directions is gov­
erned by the pFET characteristics with the remaining asymmetry that the gate
voltage is set by Vout , via a source follower. A further difference between
the two elements is the fact that the element of Fig. 1O.11(a) is a good ap­
proximation of a two-terminal device and its resistance thus only depends on
Vout - Vl b, the only common-mode effect being the well-to-substrate voltage,
298 Chapter 1 0

Vs ._------l

E
J'\..r'+

Figure 10.14
Adaptive logarithmic photosensor with cascode transistor increasing the bandwidth for small
photocurrents.

which is the emitter-to-collector voltage of one of the BJTs. The other ele­
ment, however, is subject to the body effects of the source-follower nFET and
the resistive pFET and therefore changes its characteristics with irradiance:
an effect that may be beneficial if faster adaptation at higher light levels is
desired. Furthermore, it is more sensitive to leakage currents to the power sup­
ply rails induced by photon-generated minority carriers. The resistive element
of Fig. 10.11(c) (Delbriick, 1993) is a bipolar transistor with a floating base,
which can be regarded as two diodes with reversed polarity connected in se­
ries. In either direction, the current is limited by the dark current of the reverse­
biased diode. The current-voltage characteristic has a compressive, sigmoidal
shape, which for ideal diodes saturates at the corresponding reverse diffusion
currents Js, as described by the Shockley equation (Eq. 2.6.22). However, as
Photosensors 299

we saw in Chapter 2, the reverse currents of real diodes are substantially larger
than predicted by the Shockley equation and do not saturate completely.
The transient voltage variations at the output node of the adaptive pho­
tosensor can be quite large, depending on the chosen capacitive divider ratio
and resistive element. The expansive resistive elements of Fig. 1O.II(a) and
(b) support variations of the order of 1 V, while with the compressive resistive
element of Fig. 10.11(c) and a large enough capacitive divider ratio almost the
entire range between the power supply voltages can be used. The parasitic ca­
pacitance Cp from the source of M1 onto the output node via the gate-to-drain
capacitance of nFET M2 gives rise to the Miller effect (Gregorian and Ternes,
1986). Hereby, the apparent capacitance, as seen from the source of M 1, is in­
creased by Cp multiplied by the voltage gain A from gate to source. For small
photocurrents this effect leads to a slow response. The introduction of a cas­
code nPET M4 with a fixed gate voltage Vc (Fig. 10.14) clamps the voltage
on the drain of M2 , because the current through the amplifier is approximately
constant, and largely nullifies the Miller effect. However, for certain biasing
conditions the presence of the cascode can make the circuit unstable.

10.5 Imaging Arrays

There is a large market for electronic imaging devices, which consequently

have reached a high degree of sophistication. In the following, we will present
a brief overview of the basic principles of the most commonly used types of
imagers, most of which are typically designed to yield a linear measure of
the irradiance distribution of the image projected onto them. An individual
photosensing element, called pixel (for picture element), is typically small, so
that a high image resolution is achieved. Because of its small size, a pixel can
usually only contain transistors of one type.
Large photosensor arrays integrated on a single semiconductor substrate,
as used in electronic imaging applications, cannot usually be read out in par­
allel, due to the large number of pixels and the limited number of output ter­
minals. The number of pixels scales with the square of the linear imager size,
while the number of output terminals scales only linearly. The outputs of the
individual pixels have to be multiplexed onto few signal lines: in most cases a
single one, and in some cases one per row. The readout duty cycle for each pho­
tosensor is typically quite small. Given this data sampling requirement, there
are two basic strategies for pixel operation: In the continuous-time mode the
300 Chapter 1 0

photogenerated charge is converted into a steady state current, which is sensed

whenever the pixel is addressed. In the integration mode the photogenerated
charge is collected on a capacitor until it is read out, whereafter the capaci­
tor charge is reset to a given value and a new integration cycle is started. In
continuous-time operation the photogenerated charge is not sensed when the
pixel is not addressed. By contrast, in the integration mode all the collected
charge contributes to the output signal. This mode offers better sensitivity and
signal-to-noise ratio provided that the integration time is larger than the re­
sponse time of the sensor.
Integration of low-Ievel parallel image-processing circuitry with photosen­
sor arrays (focal-plane processing) is a method of efficiently implementing a
host of functions that are also found in biological retinal structures, such as
adaptation to background illumination or color, edge enhancement and mo­
tion extraction. Such functions often rely on local spatial coupling among the
pixels, where the instantaneous value of one pixel influences the value of its
neighbors to implement a spatial filtering function. A simple integrate-and­
reset scheme for the photosensors distorts the computation of such functions 4 ,
unless all pixels are reset simultaneously or the readout period is made much
smaller than the integration period. Modem imaging devices often have an
electronic shutter function, in which the image is synchronously transferred to
an array of shielded storage nodes at the end of the integration cycle. However,
in the interest of simplicity we will only present the basic version of each type
of pixel.

Passive Pixel Sensors

A passive pixel typically consists of a photodiode and a binary transfer

gate (Weckler, 1967), as shown in Fig. 1O.15(a). The transfer gate M 1 leads to
a shared readout line, which is held at a fixed potential Vrej . When the transfer
gate Ml is open, the photodiode is reverse-biased to Vre j . Closing the transfer
gate open-circuits the photodiode and the photogenerated charge accumulates
across the capacitance of the photodiode's depletion region, decreasing the
voltage Vph across the diode. After a certain integration time the transfer gate
is opened again and the charge is sensed by an amplifier connected to the read­
out line, while the reverse bias is restored to Vrej . The readout is destructive,
which means that the signal charge is lost during the readout.

4 It also induces a temporal distortion to the raw image, which becomes apparent for long
integration periods.
Photosensors 301

If the integration time is fixed, the sensed charge is proportional to the ir­
radiance, provided that the collected charge does not saturate the photodiode.
Saturation sets in when Vph approaches zero and charge carriers start to re­
combine due to forward diffusion. The maximum charge that can be collected
without significant saturation effects is given by

Qmax = C Vre / (10.5.1)

where C is the depletion capacitance across the diode at zero external bias.

V,ef

V, o-1 v' s 0-1

vph

E
J\..r'+ D1
� V,o

(a) (b)

Figure 10.15
Integrating photodiode pixels for imaging applications. (a) Passive pixel for charge readout and
reset by a shared line biased at v,.e f and connected to the pixel via a transfer gate MI . (b) Active
pixel with a reset gate MI for voltage readout via a source-follower transistor M! and a transfer
gate M3.

In a two-dimensional array of the passive pixels described above, each

column has its own readout line and an entire row is read onto the column lines
simultaneously. The column lines are then multiplexed onto a common line
with an additional transfer gate per column (Fig. 10.16). If random access to
each pixel is required the pixel has to include a second transfer gate addressed
by additional column lines to decouple the readout along the rows.
If spatially coupled pixels are not read out simultaneously, which is usually
the case if they share a readout line, their response is distorted with respect to
302 Chapter 1 0

V� O--'----��--'---JP--

�, o--,-----fll,---,---JI'--

--------- . ----------<�_o V,ef

....--

Figure 10.16
Architecture of a two-dimensional passive pixel array. The signals of each row are transferred in
parallel to column readout lines via the transfer gates controlled by voltage signals VI and Vr 2 ,
respectively. The signals from the different columns are read out serially via the transfer gates
controlled by voltage signals Vcl and Vc 2 , respectively. The readout lines are biased at a fixed
voltage v" e ! '

the response for true parallel coupling. Sequential addressing of adjacent pix­
els, which is normally used for reading entire images, minimizes the distortion.
Passive pixels have quite a large fill factor (the ratio of the photosensitive
area and the total area) because they use only one or two small transistors
and two or three global wires per pixel. Furthermore, their FPN is rather small,
because the photodiodes can be reasonably well matched and the transistors are
only used as switches. Disadvantages of passive pixels are the large parasitic
Photosensors 303

capacitance of the readout lines resulting in a relatively large readout noise 5

and the low sensitivity and slow response of the charge amplifier. The parasitic
capacitance is proportional to the length of the readout line and thus increases
with array size.

Active Pixel Sensor

Active Pixel Sensor (APS) (Fossum, 1993, 1997) is the name given to a
class of image sensors, where each pixel contains at least a buffer or an
amplifier (Noble, 1968). An implementation of an integrating APS pixel with
voltage readout is depicted in Fig 1O.15(b). The voltage signal Vph generated
by the collected charge on the diode capacitance is buffered by a source
follower consisting of transistor M 2 and a current source (not shown) common
to all pixels connected to the same readout line. A pixel is read while pass
transistor M3 is open and reset when pass transistor M 1 is opened. In this
configuration, the signal voltage is roughly linear with irradiance for fixed
integration intervals. Non-linearities are introduced by the voltage-dependence
of the photodiode capacitance and by the body effect of M 2 .
Standard readout strategies for active pixel arrays are similar to those for
passive pixel arrays. However, readout and reset signals are now separated.
This separation complicates the addressing of the binary gates but it allows
for additional modes of operation. In particular, data readout is not destructive.
That is, the pixel is not automatically reset upon readout.
The photodiode APS pixel has a smaller fill factor than the passive pixel,
but much better signal-to-noise ratio and larger bandwidth due to the buffering.
The mismatch of the transistors M 2 that are operated in their analog domain
causes a significantly increased FPN with respect to the passive pixel sensor.
Hence, active pixel sensors are only practical if they include offset correction
circuits, such that the signal of each pixel is measured with respect to a value
that represents the pixel's response to a given reference input signal (Nixon
et aI., 1995).
Integrating active or passive pixels can also be implemented with a photo­
gate instead of a photodiode (Mendis et al., 1997). In this case the photogener­
ated charge is collected underneath the photogate. At the end of the integration
period the charge transferred to a reverse-biased diode by changing the voltage
on the photogate. From the diode, it is read out in the same way as in a pho­
todiode pixel. Photogate pixels may have lower noise (due to smaller readout

5 The charge noise on a capacitance C is proportional to vkTC.

304 Chapter 1 0

capacitances) but they also have lower quantum efficiencies than photodiode
pixels.
The logarithmic photosensors (Section l OA) can also be used in active
pixels. The simple versions of Fig 10.7 should be buffered, for example with
the source-follower circuit of Fig 1O.15(b), while the feedback versions may
not need a buffer, because they do not use the photocurrent to drive the output
node. Instead they use an independent current source. The feedback versions
include both types of transistors, at least one of which requires a well (see
Chapter 13), and therefore cannot be implemented compactly.

Charge-Coupled Device

A charge-coupled device (CCD) (Boyle and Smith, 1970; Theuwissen, 1995)

consists of an array of photogates that integrate the photogenerated charge.
At the end of the image acquisition period, which is the common charge­
integration period of all pixels in the array, the signal charge packets of the
different pixels are simultaneously moved underneath the semiconductor sur­
face towards the readout circuitry. Each charge packet stays together in a lo­
cally confined volume until it is read out and does not spread out on a long
signal line. This method of readout achieves high bandwidth, high sensitivity,
and low noise. Most CCDs are buried-channel CCDs (Walden et al., 1972) that
consist of bulk-storage photogates and thus integrate and transport the charge
in the bulk of the semiconductor. They are usually preferred to surface-channel
CCDs, where storage and transport occurs at the semiconductor-insulator in­
terface, because the interface has a much larger trapping density than the bulk,
due to lattice mismatch and surface effects during fabrication. The main dis­
advantage of buried-channel CCDs is their small charge-storing capacity. If a
charge packet gets overfilled the excess charge may spill into adjacent packets,
an effect known as blooming. Buried-channel CCDs cannot be fabricated with
standard CMOS technology, because no moderately-doped buried channel dif­
fusion is available. Surface-channel CCDs are in principle CMOS-compatible,
but in practice, the silicon-oxide interface of a standard CMOS process does
not have a good enough quality to guarantee an efficient charge transport across
hundreds of stages, required for CCD imagers. Charge-coupled devices for
imaging applications are therefore fabricated with specialized processes, which
are more expensive than standard CMOS processes.
The charge packets are either transported underneath the photogates them­
selves, as in the case of the Jrame-transJer CCD (Fig. 1O.17(a)), or first trans­
ferred to parallel running lines of MIS gates via a transfer gate and transported
Photosensors 305

imaging
array

vertical
vertical readout
readout registe r
reg i ster

""h+llV'AOHI�""'� output �1----t�*--��� output

horizontal readout reg ister horizontal readout register

(a) (b)

Figure 10.17
Architecture of (a) frame-transfer CCO and (b) interline-transfer CCO. The hatched areas indicate
MIS gates that are shielded from light. In the frame-transfer CCO. the photogenerated charge
distribution is rapidly shifted into a shielded readout register adjacent to the imaging array at the
end of the image acquisition period. The readout register is then scanned out at the speed required
by the video standard. In the interline-transfer CCO. the shielded readout register is interlaced
with the imaging array. The image is latched from the imaging array into the readout register by
a single transfer gate. Interline-transfer CCOs have a smaller fill factor of the imaging area than
frame-transfer CCOs. but require less complicated timing and suffer less from image smearing
effects.

underneath those (Fig. 10.17(b)) in which case the device is called an interline­
transfer CCD. In either scheme, the transport is carried out by shifting the local
energy minima for the collected charge type in the semiconductor from under­
neath one gate to an adjacent one in a given direction, such that the charge
packets stay separated, each traveling with a different local minimum. The
shifting is accomplished by appropriately clocking the voltages of the differ­
ent gates between two or more voltages. Two different clocking schemes are
illustrated in Fig. 10.18 (Amelio et aI., 1971). After the integration period, the
frame-transfer technique shifts the charge packets rapidly along the columns
from the imaging array into a readout register that is shielded from light. The
interline-transfer technique uses a single transfer gate to shift the charge into
a readout register, whose columns are interlaced with those of the imaging ar­
ray. The frame-transfer CCD offers a better fill factor than the interline-transfer
306 Chapter 1 0

(a) (b)

Figure lo.t8
Charge-packet transport strategies in CCDs with (a) three-phase clock and (b) two-phase clock
with stepped oxide. Cross-sections of the CCDs with indicated connections to the clock phases
(� 1 . �2 . � 3 ) are shown together with snapshots at different times (tl . t2 . t3 . t4 . t5 . t6) of the
potential distributions underneath the photogates. The charge packets. symbolized by the hatched
areas. are shifted from left to right as time progresses. By using a stepped oxide with different
thicknesses for adjacent photogates connected to the same clock phase. the directionality of charge
transfer can be built into the device and a two-phase clock is sufficient. Two-phase operation can
even be accomplished with a single clock by keeping one clock phase at a fixed reference potential.
and clocking the other one to potentials lower and higher than the reference.

CCO. However, the interline-transfer CCO has simpler clocking and does not
suffer as much from light-induced smearing effects.
In a two-dimensional CCO image sensor the charge is transported in par­
allel, synchronously clocked column CCO shift registers. In some applications
requiring a large readout bandwidth, each column has its own charge-sensing
amplifier, but more typically a horizontal CCO along the edge of the array
transports the charge towards a common charge-sensing amplifier. The single­
amplifier scheme limits the FPN to the matching of the photogates and smear­
ing effects due to incomplete charge transfer.
Charge-coupled devices are suitable for the implementation of certain local
image processing operations (Fossum, 1989), which can be programmed with
appropriate clocking schemes. Since signal processing occurs in the charge
domain, operations such as addition, subtraction and averaging are easily
implemented. For example, acquisition of image pyramids, that is, the same
image at different resolutions, in successive frames can readily be achieved by
switching adjacent pixels together to decrease the resolution by a step (Seitz
et aI., 1993).
Photosensors 307

The CCD has dominated the solid-state image sensor market since the
1970's (Boyle and Smith, 1970), because of its low FPN and readout noise
and its high sensitivity. However, the CCD's large power consumption and the
inconvenience of integrating peripheral circuitry on the same substrate with
the imaging array, due to the large capacitive load of the gates and the high
processing costs, make it more expensive and less suitable for miniaturization
than the APS. It is generally predicted that the consumer market will eventually
be taken over by the APS (Chute, 2(00).

10.6 Limitations Imposed by Dark Current on Photosensing

Junction leakage current in the dark (dark current) is the main limitation to
photosensitivity at low light levels. In typical CMOS processes, which are not
2
optimized for low junction leakage, large area junctions leak about 1 nA/cm .
This number seems to be fairly constant over processes with which we have
experience, varying over perhaps a factor of 5. Some fabrication houses have
specialized processes developed for DRAM or imagers that have much lower
2
claimed leakage currents, down to perhaps 25-50 pA/cm . However these
processes are presently not available to multiproject wafer customers. Dark
current is generally not a very strong function of junction reverse bias voltage;
We have seen at most a doubling of current as the reverse voltage is swept from
o to the power supply.
In addition, dark currents from different pixels can vary by factors of 10
or 100: It is common to see a subpopulation of pixels that fonn outliers in
a histogram. These hot spots are thought to arise from as little as a single
generation/recombination state at a level that happens to be right in the middle
of the band gap. In an imager, these hot pixels cause isolated white spots that
are very noticeable in an image and must be corrected for by interpolation of
surrounding pixel values 6 •

The dark current in a junction is dominated by leakage around the edges

of the junction, where the junction meets the interface between Si and SiO 2 .
It is thought the increased leakage here in the sidewall region arises from
the additional stress induced in the Si crystal and from interface states at the
interface itself (Hawkins, 1985). Measurements have shown wide variation in
the relative size of the sidewall leakage relative to the area leakage, but a factor

6 Around 1990, some CCD video camcorders were already equipped with nonvolatile storage of
the location of these hot pixels and interpolation from surrounding pixels to correct for them.
308 Chapter 10

anywhere between 5 and 100 is possible. A factor of 10 would mean that 1 f..tm
along the edge leaks as much as 10 f..tm2 of area. For a small photodiode, the
sidewall leakage completely dominates the total, so claims of low junction
leakage sometimes given by vendors can be quite misleading, since because
claims are based on measurements from large area junctions. The replacement
of LOCOS isolation by STI 7 , if the STI is done correctly, is now generally
believed to lead to lower sidewall leakage.
Since dark current is a thermal process, it is a strong function of temper­
ature, doubling about every 8 ° C (Theuwissen, 1995) 8 . For an imager or pho­
tosensor that must operate at ambient temperatures of 60 °C, the dark current
will be about 20 times larger than at room temperature. CCD manufacturers
usually quote their dark current figures at a temperature of 50--60 °C to keep
their customers from being dissatisfied when the camera is operated after lying
in a hot car all afternoon.
This junction leakage acts like a "glow in the dark" that degrades image
contrast and contaminates outputs from a storage pixel array during readout.
We can estimate the equivalent illuminance of a scene that corresponds to
a given dark current. First we will compute the chip illumination, then we
will calculate the scene illumination. Take a photodiode with area 10 f..tm 2 . If
the junction leakage is I nA/cm2 , corresponding to about 50 electrons/f..tm2 /S,
and the sidewall leakage is 10 times larger, 500 electrons/f..tm/s, then this
photodiode will leak about 8000 electrons/so Only 500 of these electrons come
from the area leakage, the other 7500 come from the sidewall leakage. The
illumination of the chip corresponding to this leakage is computed as follows.
We assume the photodiode has a quantum efficiency (QE) of 50%. That QE
means that the flux is doubling the resulting current density. So, the photon
flux is 8000*2 photons in an area of 100 f..tm 2 (160 photons/f..tm2 /s). "White"
light of illuminance 1 lux is about 10 4 photons/f..tm2 /s (Rose, 1973). This
conversion is approximate, because it depends the definition of " white", but it
is good enough for our estimate. Using this conversion, we compute that the
equivalent illuminance of the chip is about 20 mlux. Now we must estimate
the illuminance of a scene that would result in this illumination of the chip. We
can do this by using a very useful formula that can by derived from spherical
geometry and some understanding of the units of photometry. A perfectly
transparent lens with a given f number imaging a perfectly white diffusively

7 See Section 13.1.

8 Generation processes depend on the dominant (most likely) activation energy, usually between
Eg and Eg /2. Lower leakage generally corresponds to higher activation energy.
Photosensors 309

reflecting surface produces an image illuminance I image related to the scene

illumination Iscene by
Iscene
Iimage (10.6.1)
4/ 2
=

If we are using a fairly fast 1/2 lens, the image illumination is only 1/16
that of the scene. We can now compute that the dark-current-equivalent scene
illumination is about 0.25 lux. Under full-moon conditions, the illumination
is about 0.1 lux (Rose, 1973). We can see that the dark current will limit our
photosensor performance! Even with full moon, white objects out in the world
will only generate about one third the photodiode's leakage current. It is clear
that to make commercially viable devices that can operate over a truly wide
range requires access to low-leakage processes.
Since low-leakage is a strong Darwinian survival trait for DRAM and com­
mercial imagers, there has been a huge effort to reduce the junction leakage.
Fastidious wafer cleaning before fabrication, gettering of contaminants, and
other proprietary tricks have been developed to lower the junction leakage.
Plummer et al. (2000) have written a very clear description of some of these
tricks. Research users have little access to these specialized process flows.
One trick that is particularly interesting was discovered almost accidently
by Teranishi et al. (1984) in pursuit of an interline transfer CCD with lower
image lag. The trick is to make a buried photodiode: A photodiode which
is covered with a thin implant of the substrate doping type. This covering
implant, which is shorted to the photodiode substrate, acts to nearly eliminate
the sidewall leakage in the photodiode and also reduces the leakage due to
interface states at the SilSi02 interface. Unfortunately for ordinary users, this
trick requires a thin surface implant which can overlap the edge of the buried
junction and connect to the bulk, and it does no particular good unless it is also
combined with all the other contamination and defect reduction techniques.
This technology is called a Hole accumulation diode, HAD for short, because
the buried implant is n-type and it is covered with a p-type accumulated layer
that fills the interface states, so that they cannot contribute to dark current. This
technology is also called pinned photodiode, because the surface potential is
pinned to the bulk potential. HAD has its drawbacks: It requires additional
processing steps, and the blue response is poor and badly controlled, because
the blue photons make minority carriers near the surface where they are very
likely to recombine with the abundant holes. Since blue photons are always in
short supply, this is a big problem for good color imaging.
310 Chapter 1 0

For imagers designed for still pictures, dark current can be measured in one
frame with the mechanical shutter closed, and then later subtracted from the
pixel output. This still leaves the shot noise from the dark current that cannot
be calibrated away, but substantial improvements in image quality can still be
realized by removal of transistor mismatch effects. The problem remaining is
the noise in the dark current: If the dark leakage averages N electrons per
frame, the variation will be -IN. This noise due to dark current cannot be
calibrated away. In any case, subtracting uncorrelated noise sources, whatever
the source of the noise (thermal or photon shot) only increases the noise.
Finally, for APS imagers with storage pixels and electronic shutter, dark
current can considerably degrade the image during the storage and readout
phase, particularly for a large array. Here it is the leakage on the drains of the
transistor storage switch that matters, and not the photodiode leakage. A HAD
photodiode does not help the readout problem.
IV SPECIAL TOPICS
11 Noise in MOS Transistors and Resistors

Many people find the subject of noise mysterious. Although the fundamental
physical concepts behind noise are simple, much of this simplicity is often
obscured by the mathematics invoked to compute expressions for the noise.
In this chapter, we cover the basics of noise in electronic devices and circuits
and we also discuss theoretical and experimental results for white noise in
the low-power subthreshold region of operation of a MOS transistor (see
Chapter 3) (Godfrey, 1992; Sarpeshkar et aI., 1993; Sarpeshkar, 1997)
We solve the mystery of how a shot-noise answer derives from a thermal­
noise viewpoint by taking a fresh look at noise in subthreshold MOS transis­
tors. We then rederive the expression for thermal noise in a resistor from our
viewpoint. We believe that our derivation is simpler and more transparent than
the one originally offered in 1928 by Nyquist who counted modes on a trans­
mission line to evaluate the noise in a resistor (Nyquist, 1928). The derivation
here leads to a unifying view of the processes of shot noise and thermal noise
in electronic devices. Specifically we show that white noise, whether it is la­
beled as shot or thermal, is completely accounted for as shot noise due to inter­
nal diffusion currents in electronic devices. These internal diffusion currents
are thermally generated and scale linearly with temperature. Internal diffusion
currents are present in all devices independent of whether the dominant cur­
rent flow mechanism is by drift (which causes almost no noise) or by diffusion
(which causes noise)1.

11.1 Noise Definition

Noise is considered to be any unwanted excitation of a circuit. It comes from

both external sources and internal sources. Noise from external sources appears
because of unintended coupling of the circuit with other parts of the physical
world; noise from internal sources appears because of unpredictable micro­
scopic events that occur in the devices that constitute the circuit. In principle,
noise from the former can be eliminated through careful design whereas noise
from the latter can be reduced but never eliminated. It is important to consider
noise in the design of low-power systems because the signal levels (voltages

1 Some parts of this chapter were taken from Sarpesbkar et a1. (1993), White noise in MOS
transistors and resistors, IEEE Circuits and Devices Magazine. ©1993 IEEE. Reprinted with
permission from IEEE.
314 Chapter 11

or currents) are small. The noise level sets the size of the smallest signal that
can be processed meaningfully by a physical system.
The amount of noise in a signal x(t) is characterized by its root-mean­
square (RMS) value. It is computed by calculating the mean-square variation
of the signal about its mean value:

�X 2 =
l
lim -T
iT(x(t) X)2 d t (11.1.1)
T->oo 0
-

where x is the mean of the signal. The RMS value of the signal is then
J�X 2. The total noise in a system caused by independent noise generators
can be computed by adding up the mean-square deviations from these noise
generators:
(11.1.2)

In the equation above, we assumed two independent noise sources in the

system.

Constant power for

Sf') = c a fixed D. ,

Figure 11.1
White noise spectrum. The spectrum is fiat over all frequencies.

Power spectral density One way of characterizing the amount of noise

in a system is by computing the noise power over the frequency spectrum.
This function is also called the power spectral density (PSD) of the noise.
It describes how much power the noise waveform carries in each frequency
interval. The PSD ( Sx (I» of the noise is defined as the average power carried
by x(t) in a one-Hertz bandwidth around f, or the power per unit frequency.
Noise in MOS Transistors and Resistors 315

This spectrum is obtained by applying the noise waveform to a bandpass filter

with center frequency It and a I-Hz bandwidth; squaring the output; and
calculating the average power over a long time to obtain S x(It). By repeating
the measurements for filters with differing center frequencies, we obtain the
PSD Sx(f). Because each value of Sx(f) is measured for a I-Hz bandwidth,
we express Sx(f) in units of V 2 / H z. It also common to show the noise
spectrum of a system in terms of the root of S x(f), hence units of V/ �.
The total noise power Px which has units of V2 is obtained by integrating
Sx(f) over the whole frequency range2:

(11.1.3)

Constant power for

a fixed 'I /'2

log'

Figure 11.2
Ricker noise spectrum. The noise power is constant for a fixed ratio of frequencies.

There are two distinct classifications of noise spectra: White noise and
pink noise. We call noise "white" if its power is spread uniformly across the
spectrum (a flat spectrum), (see Fig. 11.1) and "pink" if the noise power is
concentrated at lower frequencies. The most studied pink noise is flicker noise
which is also known as 1/ f noise3. The spectrum for flicker noise is shown in
Fig. 11.2.

2 We call Px the noise power even though the units are in v2 because we have disregarded the
resistor. Of course, the signal can also be, for example, current, light intensity, or sound pressure.
3 Other forms of pink noise include avalanche noise and burst noise.
316 Chapter 11

White Noise

Although the spectrum of white noise is defined to be flat over the entire fre­
quency range, in reality, such a spectrum does not exist. We consider any noise
spectrum that is flat in the region of interest as white noise. In conventional
literature, white noise in electronic devices is assumed to be composed of two
possible types: Thermal Noise and Shot Noise.

Thermal Noise Thermal noise is believed to be due to random thermal mo­

tions of carriers. In a resistor, the thermal noise current fl.I R 2 can be modeled
by a noise current source in parallel with the noiseless resistor (Fig. 11.3) and
described by

fl.IR 2 = 4k T G fl.! (11.1.4)

where k is the Boltzmann constant, G is the conductance, and fl.! is the

bandwidth. The units of the PSD S(f) = fl.I R 2 I fl.! are A21Hz. The noise
spectrum of the resistor is shaped by the circuit of which the resistor is a part.
An example of this calculation is shown in Section 11.5.

Figure 11.3
Model of thennal noise in a resistor.

Shot Noise Shot noise is believed to be due to discrete random arrivals of

electrons traversing an energy barrier. Shot noise is believed to require D.C.
current flow whereas thermal noise is believed to require no D.C. current flow.
The shot noise current (derived later in the text) in a device is given by

fl.I2 =
qIfl.! (11.1.5)

where q is the charge on the electron, I is the mean current flowing through
the device, and fl.! is the system bandwidth.
Noise in MOS Transistors and Resistors 317

1 /( noise dominant

/ White noise dominant

..,
"
,
- I, -
j
1

fe log f

Figure 11.4
Noise spectrum of a device is a mixture of flicker noise and thermal noise. The 1/ f noise corner
fe is process and bias dependent.

Flicker Noise

Flicker noise is also called 1/ f noise because its power spectrum varies
inversely with the frequency. It is found in all active devices and some discrete
passive devices. Flicker noise has different origins. In MOSFETs, it is widely
believed that it arises from electrons in the channel moving into and out of
surface states, and into and out of impurities or defect traps in the gate oxide.
It is always associated with a DC current and has the form
1m
tl.12 = K tl.f (11.1.6)
r
where K is a constant for a particular device, I is the current in the device,
m is between 0.5 to 2 and n is approximately 1. Further details can be found
in Gray et al. (2001). This noise is dominant at low frequencies. However, it
can also dominate at higher frequencies if the current in the device is large.
Usually, the noise in a device is a mixture of white noise and 1/ f noise.
The spectrum of the noise is as shown in Fig. 11.4. The 1/ f noise comer or f e
is process and bias dependent.

11.2 Noise in Subthreshold MOSFETs

Noise in MOS transistors is composed of both white noise and flicker noise.
We first discuss the formulation for flicker noise in a subthreshold MOSFET,
and then the formulation for shot noise.
318 Chapter 11

Flicker Noise

We compute the flicker noise in a MOSFET by looking at the probability of

a charge carrier in the channel tunneling in and out of defect traps in the gate
oxide. If we assume that there is a uniform density p of traps, then a charge
carrier can enter or leave a trap at a time scale set by the tunneling probability

(11.2.1)

where a is a constant, x is the distance from the bulk Si-Si02 interface, and f
is the fluctuation frequency. By taking the natural logarithm of Eq. 11.2.1, we
get
df
logf oc - ax � j oc -adx. (11.2.2)

For a fixed frequency interval df, the total amount of trapped charge Q will
fluctuate as .JAjidX where A is the area of the gate oxide, so the mean-square
deviation in the charge is

-2 df
D..Q oc Apdx oc A j. (11.2.3)

We can think of the trapped charge as modulating the surface potential and
hence, the threshold voltage of the transistor V T:4

2
--2 D..Q
D..VT oc (j2 (11.2.4)

where C is the capacitance of the gate oxide. A larger area of the transistor
leads to a larger oxide capacitance C and a smaller effect of any one fluctuating
charge on the transistor's threshold voltage. However, the larger transistor area
also leads to more flu ctuating charges because of the constant trap and defect
densities. So the increased-capacitance effect reduces the noise power like
1/A2, and the increased total-charge effect increases the noise power like A,
thus
--2 A df 1 df
D..VT oc j oc j. (11.2.5)
A2 A

4 We can view 1/ f noise as noise due to a dynamically changing threshold voltage.

Noise in MOS Transistors and Resistors 319

The mean-square deviation in the current due to the deviation in the threshold
voltage can then be computed as

<X <X g� df
6.12 g2m 6. VT2 (11.2.6)
A f
where gm is the transconductance of the transistor. Refer to Chapter 3 for
the definition of transconductance. Note that this equation also applies to the
above threshold MOSFET. In a subthreshold MOSFET where 9 m = '" 1 JUT,

(11.2.7)

where the constant K includes UT; "'; the gate oxide capacitance per unit
area; and also the effects of typical offsets in MOS technology which lead
to offsets between transistors. These mismatches scale inversely with the area
of the transistor. The equivalent voltage noise is

-- K' df
6. Vn2 = - (11.2.8)
C 2ox A f

where K' is a constant and C ox is the oxide capacitance per unit area 5.

Shot Noise

A formula for subthreshold noise in MOS transistors has been derived by Enz
(1989) and Vittoz (1990) from considerations that model the channel of a
transistor as a series of resistors. The integrated thermal noise of all these
resistors yields the net thermal noise in the transistor, after some fairly detailed
mathematical manipulations. The expression obtained for the noise, however,
strongly suggests that the noise is really "shot noise", conventionally believed
to be a different kind of white noise from thermal noise.
In this section, we show how one generates a shot-noise answer from a
thermal-noise derivation by taking a fresh look at noise in subthreshold MOS
transistors. We then rederive the expression for thermal noise in a resistor from
our viewpoint. The derivation here leads to a unifying view of the processes
of shot noise (noise in vacuum tubes, photo diodes and bipolar transistors) and
thermal noise (noise in resistors and MOS devices).
We also show noise measurements in a subthreshold transistor. The mea-

5 Depending on the theory for the origin of flicker noise, the c;,,,, variable in the equation for
flicker noise is often equal to 1 or 2 (Tsividis, 1998).
320 Chapter 11

surements were taken at current levels in the 100 fA-loo pA range. W hite
noise was the only noise observable even at frequencies as low as 1 Hz. (Re­
imbold, 1984) and (Schutte and Rademeyer, 1992) have measured noise for
higher subthreshold currents (> 4 nA), but have reported results from flicker­
noise measurements only.
We will show that measurements of white noise in subthreshold transistor
operation are consistent with theoretical predictions. We also show measure­
ments of noise in photoreceptors (a circuit containing a photodiode and an
MOS transistor) that are consistent with theory. The photoreceptor noise mea­
surements illustrate the intimate connection of the equipartition theorem of
statistical mechanics with noise calculations.
The measurements of noise corresponding to miniscule subthreshold tran­
sistor currents were obtained by conveniently perfonning them on a transistor
with W/ L � 104• The photoreceptor noise measurements were obtained by
amplifying small voltage changes with a low-noise high-gain on-chip ampli­
fier.
Imagine that you are an electron in the source of an nFET. You shoot
out of the source, and if you have enough energy to climb the energy barrier
between the source and the channel, you enter it. If you are unlucky, you might
collide with a lattice vibration, surface state, or impurity and fall right back
into the source. If you do make it into the channel you will suffer a number of
randomizing collisions. Eventually, you will actually diffuse your way into the
drain. Each arrival of such an electron at the drain contributes an impulse of
charge.
Similarly, electrons that originate in the drain may find their way into the
source. Thus, there are two independent random processes occuring simulta­
neously that yield a forward current If from drain to source, and a reverse
current Ir from source to drain. A detailed discussion of these currents can be
found in Chapter 3. Since the barrier height at the source is smaller than the
barrier height at the drain, more electrons flow from the source to drain than
vice-versa and I f > Ir. The channel current I is given by

By using the subthreshold current equation Ir = If e - � ( Vds is the drain-

Noise in MOS Transistors and Resistors 321

to-source voltage), we obtain

(11.2.9)

where 1/ = I sat is the saturation current of the transistor, and U T = kT/q is

the thennal voltage.
Because the forward and reverse processes are independent, we can com­
pute the noise contributed by each component of the current separately and
then add the results. Thus, we first assume that Iris zero, or equivalently that
Nd, the concentration of electrons per unit width at the drain end of the chan­
nel, is zero. The arrival of electrons at the drain can be modelled by a Poisson
process with an arrival rate A. A small channel length L, a large channel width
W, a large diffusion constant for electrons Dn, and a large concentration of
electrons at the source Ns, all lead to a large arrival rate. Because the current
in a subthreshold MOS transistor flows by diffusion, the electron concentration
is a linear function of distance along the channel. The forward current I / and
arrival rate A are given by

Ns
I/=qDnWL (11.2.10)

A= It!q. (11.2.11)

Powerful theorems due to Carson and Campbell, as described in standard

noise textbooks such as van der Ziel (1970), allow us to compute the power
spectrum of the noise. Suppose that each arrival event in this Poisson process
causes a response F(t) in a detector sensitive to the event. Lets(t) be the
macroscopic variable of the detector that corresponds to the sum of all the
F(t)'s generated by these events. Then, the mean value ofs(t) and the power
322 Chapter 11

spectrum P(f) of the fluctuations ins(t) are given by

s(t)= A i: F(t)d t (11.2.12)

-(s-(t-) ---s-(t-) )"2= A i: F2(t)d t (11.2.13)

00
= 2A 1 1'IjJ(f)12df (11.2.14)

00
= 1 P(f)df (11.2.15)

where 'IjJ(f)= J�: F(t)e-j21Tftd t is the Fourier transfonn of F(t). We have

used Parseval's theorem and the symmetry of the Fourier transfonn for positive
and negative frequencies in obtaining the last equation. Each electron arrival
event at the drain generates an impulse of charge q that corresponds to F(t):
that is, F(t)=q 8(t). Thus, we obtain
l=qA (11.2.16)
f
(I _1)2= 2q2 A ra
df (11.2.17)
10
= 2qlD.f (11.2.18)

where D.f is the bandwidth of the system. Equation 11.2.18 is the well­
known result for the shot-noise power spectrum. Thus, the noise power6 that
corresponds to our forward current is simply given by 2qI fD.f. Similarly, the
noise power that corresponds to the reverse current is given by 2qI rD.f. The
total noise in a given bandwidth D.f is given by

(11.2.19)
KVg-VS
where Isat = If = I oe UT corresponds to the saturation current at the
given gate voltage. Note that as we transition from the linear region of the
transistor ( Vds < 4UT) to the saturation region, the noise is gradually reduced
from 4qIsatD.f to 2qIsatD.f. This factor of two reduction occurs because
6 Again, we have disregarded the influence of the resistor for convenience's sake and we call this
measure "noise power".
Noise in MOS Transistors and Resistors 323

the shot-noise component from the drain disappears in saturation. A similar

phenomenon occurs in junction diodes where both the forward and reverse
components contribute when there is no voltage across the diode; as the diode
gets deeply forward or reverse biased, the noise is determined primarily by
either the forward or reverse component, respectively (Robinson, 1974).

2.25

0
2.00

1.75 Normalized current noise power

1.50

1.25

1.00

0.75

0.50
Normalized current
0.25

0.0

-0.25 +---+--+--1----1
-0.25 0.0 25 50 75 100 125 150 175 200
Drain-to-source voltage (mV)

Figure 11.5
Measured current and noise characteristics of a subthreshold MOS transistor. The lower curve is
the current, normalized by its saturation value /sat, so that it is 1 .0 in saturation and zero when \.ds
is O. The upper curve is the noise power t::.P normalized by dividing it by the quantity 2q1satt::.J,
where t::.J is the bandwidth and q is the charge on the electron. We see that as the transistor moves
from the linear region to saturation, the noise power decreases by a factor of two. The lines are
fits to theory using the measured value of the saturation current and the value for the charge on the
electron q = 1.6 X 10-19 C.

The flatness of the noise spectrum arises from the impulsive nature of the
microscopic events. We might expect that the flat Fourier transform of the mi­
croscopic events that make up the net macroscopic current would be reflected
in its noise spectrum. Carson's and Campbell's theorems express formally that
this is indeed the case. The variance of a Poisson process is proportional to the
324 Chapter 11

rate, so it is not surprising that the variance in the current is just proportional
to the current. Further, the derivation illustrates that the diffusion constant and
channel length simply alter the arrival rate by Eq. 11.2.11. Even if some of
the electrons recombined in the channel (corresponding to the case of a bipo­
lar transistor or junction diode) the expression for noise in Eq. 11.2.19 is un­
changed. The arrival rate is reduced because of recombination. A reduction in
arrival rate reduces the current and the noise in the same proportion. The same
process that determines the current also determines the noise.

10-23

10-27 +-........
.. +<+<!
... -....+-++
.. H4--+-........
.. _ .. ........
.. � ..
10-9 10-8 10-7 10-6 10-5

Current (A)

Figure 11.6
The noise power per unit bandwidth t:.I 2 / t:.j plotted against the saturation current Isat.The
MOS transistor is operated in saturation. Theory predicts a straight line with a slope of 2q =
3.2 X 10-19 C, which is the line drawn through the data points. The small but easily discernible
deviations from the line increase with higher levels of !.sat due to the increasing levels of 1/ j
noise at these current values.

Experimental measurements were conducted on a transistor with W/ L �

104 at a saturation current of 40 nA (Fig. 11.5). A gigantic transistor was used
to scale up the tiny subthreshold currents to 10 nA-l /-LA levels and make
them easily measurable by a low-noise off-chip sense amplifier with commer­
cially available resistor values. The shot noise scales with the current level,
so long as the transistor remains in subthreshold. The noise measurements
were conducted with a HP3582A spectrum analyzer. The data were taken
Noise in MOS Transistors and Resistors 325

over a frequency range of 0-500 Hz. The nonnalized current noise power
6.12/( 2q1sat6.f) and the nonnalized current 1/1 sat are plotted. The lines
show the theoretical predictions of Eqs. 11.2.9 and 11.2.19. Using the mea­
sured value of the saturation current, the value for the charge on the electron,
and the value for the thennal voltage we were able to fit our data with no free
parameters whatsoever. Notice that as the nonnalized current goes from 0 in
the linear region to 1 in the saturation region, the nonnalized noise power goes
from 2 to 1 as expected. Figure 11.6 shows measurements of the noise power
per unit bandwidth 6.1 2/6.f in the saturation region for various saturation cur­
rents 1sat. Since we expect this noise power to be 2q1 sat. we expect a straight
line with slope 2q which is the theoretical line drawn through the data points.
As the currents start to exceed 1 p,A-I0 p,A for our huge transistor, the pres­
ence of 1/f noise at the frequencies over which the data were taken begins to
be felt. The noise is thus higher than what we would expect purely from white
noise considerations.

11.3 Shot Noise versus Thermal Noise

We have taken the trouble to derive the noise from first principles even though
we could have simply asserted that the noise was just the sum of shot-noise
components from the forward and reverse currents. We have done so to clarify
answers to certain questions that naturally arise:

• Is the noise just due to fluctuations in electrons moving across the barrier or
does scattering in the channel contribute as well?
• Do electrons in the channel exhibit thennal noise?
• Do we have to add another tenn for thennal noise?

Our derivation illustrates that the computed noise is the total noise and
that we need not add extra tenns for thennal noise. Our experiments confinn
that this is indeed the case. The scattering events in the channel and the
fluctuations in barrier crossings at the source and drain ends of the channel
all result in a Poisson process with some electron arrival rate. Both processes
occur simultaneously, and are caused by thennal fluctuations, resulting in
white noise. Conventionally, the fonner process is labelled "thennal noise"
and the latter process is labelled "shot noise". In some of the literature, the two
kinds of noise are often distinguished by the fact that shot noise requires the
presence of a DC current while thennal noise occurs even when there is no
326 Chapter 11

DC current (Gray and Meyer, 1993). However, we notice in our subthreshold

MOS transistor that when If = Ir, there is no net current but the noise is at
its maximum value of 4qI f6.f. Thus a two-sided shot noise process exhibits
noise that is reminiscent of thermal noise. We will now show that thermal noise
is shot noise due to internal diffusion currents in electronic devices.

I = If - 'r = 0
n ............ If- ...............l,- ............;;r ...
./ If
, ./ 'r
, ./ - -
, ./
, -
./
,./
./,
./ ,
- ./ ,
./ ,
./ ,
./
' ,

L x

Figure 11.7
Model for thermal noise. The figure on the left shows the concentration per unit volume of
electrons diffusing from both ends of the resistor. We can approximate this scenario by a resistor
whose total current is zero in the presence of diffusion of carriers.

New Derivation of Thermal Noise in a Resistor

Let us compute the noise current in a resistor shorted across its ends as
shown in Fig. 11.7. Since there is no electric field, the fluctuations in current
must be due to the random diffusive motions of the electrons. The average
concentration of electrons is constant all along the length of the resistor. This
situation corresponds to the case of a subthreshold transistor with V ds = 0,
where the average concentrations of electrons at the source edge of the channel,
drain edge of the channel, and all along the channel, are at the same value and
therefore the current in the resistor is I = O.
In a transistor, the barrier height and the gate voltage are responsible for
setting the concentrations at the source and drain edges of the channel. In
a resistor, the concentration is set by the concentration of electrons in its
conduction band. The arrival rate of the Poisson process is, however, still
determined by the concentration level, diffusion constant, and length of travel.
This is so because, in the absence of an electric field, the physical process
of diffusion is responsible for the motions of the electrons. Thus, the power
Noise in MOS Transistors and Resistors 327

spectrum of the noise is again given by 2q( I f Ir). The currents If and Ir
+
are both equal to qDnAj L (see Fig. 11.7) where D is the diffusion constant of
electrons in the resistor, n is the concentration per unit volume, A is the area of
cross section and L is the length. Einstein relation yields Djf..t = kTjq, where
f..t is the mobility.
Thus, the relation between the transistor noise power and its conductance
is given by

-2 qDnA
� I = 4q If �f= 4q- - �f
L
A A
= 4q f..tkTn �f = 4 k T (qf..tn) �f
L L
A
= 4 k T ( 0" ) �f
L
where 0" is the conductivity of the material. This equation further reduces to

(11.3.1)

where G = 0" 1: is the conductance of the resistor. Thus, we have re-derived

Johnson and Nyquist's well-known result for the short circuit noise current in
a resistor! The key step in the derivation is the use of the Einstein relation
Djf..t =
kTjq. This relation expresses the connection between the diffusion
constant D, which determines the forward and reverse currents, the mobility
constant f..t, which determines the conductance of the resistor, and the thermal
voltage kTjq.
The Einstein relation shows that at zero temperature, D = 0, there are no
internal diffusion currents, and there is no noise. This prediction is in accord
with that obtained from the conventional view of thermal noise. It is because
of the internal consistency between thermal noise and shot noise that formulas
derived from purely shot noise considerations agree with those derived from
purely thermal noise considerations (Enz, 1989).
Our derivation suggests that in a resistor with DC current flowing through
it, the noise is still described by internal diffusion currents in the resistor, which
are barely changed by the small systematic drift velocities superposed upon
the huge thermal velocities of the electrons. Thus, even if there is a DC electric
field in the resistor, so long as this field is not large enough to elicit hot-electron
behavior, the thermal noise of the resistor is unchanged. Our derivation also
explains why it is unwise to add a shot noise term 2q VDCj R to the thermal
noise term 4kTG to get the resistor's PSD at a particular DC voltage VDC.
328 Chapter 11

Thus the current noise PSD in a resistor with DC voltage across it remains at
4kTG as is experimentally observed. Nyquist's derivation of thermal noise is
valid only in thermal equilibrium and would be unable to predict the noise in
a resistor with a DC voltage across it unlike our derivation. Furthermore, our
derivation implies that, contrary to what is believed, a DC current flow is not
necessary to observe shot noise. Devices that have internal diffusion currents
such as resistors exhibit shot noise with no DC current flow, but this noise gets
relabeled as "thermal noise".
In summary, white noise in electronic devices is caused by shot noise
due to thermally generated diffusion currents in these devices, independent
of whether the noise is labeled as "thermal noise" or "shot noise". There is no
need to double count various forms of white noise once the shot noise due to
diffusion currents has been accounted for.
Our discussion has focused on white noise in semiclassical electronic
devices. It is possible to have shot noise statistics without any thermal process:
For example, perfectly coherent light generates photons with Poisson statistics
due to the fundamental randomness inherent in discrete quantum phenomena.

11.4 The Equipartition Theorem and Noise Calculations

No discussion of thermal noise would be complete without a discussion of

the equipartition theorem of statistical mechanics, which lies at the heart of
all calculations of thermal noise: Every state variable in a system that is not
constrained to have a fixed value is free to fluctuate. The thermal fluctuations
in the current through an inductor or the voltage on a capacitor represent the
state-variable fluctuations in an electrical system. If the energy stored in the
system corresponding to state variable x is proportional to x
2, then x is said
to be a degree of freedom of the system. Thus, the voltage on a capacitor
constitutes a degree of freedom, since the energy stored on it is C V 2/2.
Statistical mechanics requires that if a system is in thermal equilibrium with
a reservoir at temperature T, then each degree of freedom of the system will
have a fluctuation energy of kT/2. Thus, the mean square fluctuation � V 2 , in
the voltage of a system with a single capacitor must be such that

C � V2 kT
=
2 2
kT
� � V2
_

(11.4.1)
- C'
Noise in MOS Transistors and Resistors 329

This simple and elegant result shows that if all noise is of thermal origin,
and the system is in thermal equilibrium, then the total noise over the entire
bandwidth of the system is determined just by the temperature and capaci­
tance (Rose, 1973). If we have a large resistance coupling noise to the ca­
pacitor, the noise per unit bandwidth is large but the entire bandwidth of the
system is small. If we have a small resistance coupling noise to the capacitor,
the noise per unit bandwidth is small but the entire bandwidth of the system is
large. Thus, the total noise which is the product of the noise per unit bandwidth
4kT R and the bandwidth of the circuit 1c is constant and independent of R.
We will illustrate for the particular circuit configuration of Fig. 11.8 how the
noise from various devices interact to yield a total voltage noise of kT/ C.
Figure 11.8 shows a network of four transistors all connected to a capacitor,
C at the node Vs' We use the sign convention that the forward currents in each
transistor flow away from the common source node, and the reverse currents
flow toward the common source node 7• The gate and drain voltages, Vg i and
Vdi, respectively are all held at constant values. Thus, Vs is the only degree of
freedom in the system. Kirchhoff's current law at the source node requires that
in steady state
n n

L I}= L I:. (11.4.2)

i =l i=l
The conductance of the source node is given by
n
Ii
gs= L ,J . (11.4.3)
i =l T
The bandwidth tl.f of the system is then

tl.f= � gs
x � X = .!!.!... (11.4.4)
21T 2 C 4C
where the factor 1/ 21T converts from angular frequency to frequency, and factor
1T/ 2 corrects for the rolloff of the first-order filter not being abrupt 8. Thus, the
7 Our sign convention is for carrier current, not conventional current. Thus. in an nFET or a pFET.
the forward current is the one that flows away from the source irrespective of whether the carriers
are electrons or holes. This convention results in a symmetric treatment of nFETs and pFETs.
8 Use
330 Chapter 11

I
Figure 11.8
A circuit with four transistors connected to a common node with some capacitance C. By
convention, the common node is denoted as the source of all transistors, and the forward currents
of all transistors are indicated as flowing away from the node while the reverse currents of all
transistors are indicated as flowing towards the node. Only the voltage l;; is free to fluctuate, and
all other voltages are held at fixed values, so that the system has only one degree of freedom. The
equipartition theorem of statistical mechanics predicts that if we add the noise from all transistors
over all frequencies to compute the fluctuation in voltage Lll;; 2 the answer will equal kT/ C no
matter how many transistors are connected to the node, or what the other parameters are, so long
as all the noise is of thermal origin. We show in the text and in the data reported in Fig. 11.9 that
our expressions for noise yield results that are consistent with this prediction.

total noise is

6.12 =
L 2q (I} + I;) :�
i =l
n

=
L 4q1}
i =l
:� (11.4.5)

where we have used Eq. 11.4.2 to eliminate I r. The voltage noise is just
 Noise in MOS Transistors and Resistors 331

-60

-80
N
I

�
co
�
OJ

N
� -100

OJ
<f)
'0
c

:;
c.
:;
o -120

1/t instrumentation

-140 ����WL �-U���-L�llll�-L��llll �LLLU�_______

� �
10-1 10 0 101 10 2 10 3 104
Frequency (Hz)

Figure 11.9
Measured noise spectral density in units of dBV/rtHz (0 dBV = IV, ·20dB = O.IV ) for the voltage
Vs in the circuit above . The current source is Iight·dependent and the curves marked 0, -I and -2
correspond to bright light (high current), moderate light, and dim light (low current) respectively.
The intensity levels were changed by interposing neutral density filters between the source of the
light and the chip to yield intensities corresponding to 1.7 W/nr, 0.17 W/m 2 , and 0.017 W/m2
respectively. The 1/ f instrumentation noise is also shown: Its effects were negligible over most
of the range of experimental data . We observe that the noise levels and bandwidth of the circuit
change so as to keep the total noise constant: That is, at low current levels, the voltage noise is high
and bandwidth low, and the converse is true for high current levels. Thus, the area under the curves
marked 0, -I and -2 are the same. The theoretical fits to the lowpass filter transfer functions are
for a temperature of 300" K and a capacitance of 3 IO iF, estimated from the layout. These results
show that the kT/ C concept, derived from the equipartition theorem is correct.

(11.4.6)

The fact that the total noise is kT/ C implies that this circuit configuration is
332 Chapter 11

a system in thennal equilibrium. Typically, most circuit configurations yield

answers for total voltage noise that are proportional to kT/ C.

Noise Measurements i n a Photoreceptor Circuit

We obtained direct experimental confinnation of the kT/ C result from our

measurements of noise in photoreceptors. A detailed theoretical and exper­
imental analysis of noise in different photoreceptor circuits can be found
in Delbriick and Mead (1995). Figure 11.9 shows a source-follower configura­
tion which is analogous to the case discussed previously with two transistors
connected to a common node. The lower current source is a photodiode which
has current levels that are proportional to light intensity. (This photoreceptor
circuit was previously described in Section 10.4.) The voltage noise is mea­
sured at the output node, Vs. The voltage Vg is such that the MOS transistor
shown in the figure is in saturation and its reverse current is zero. The pho­
todiode contributes a shot-noise component of 2qI �f. Thus, we obtain equal
shot-noise components of 2qI �f from the transistor and light-dependent cur­
rent source, respectively. The theory described in Section 11.4 predicts that in­
dependently of the current, the total integrated noise over the entire spectrum
must remain constant. We observe that as the current levels are scaled down
(by decreasing the light intensity), the noise levels rise but the bandwidth falls
by exactly the right amount to keep the total noise constant. The system is a
lowpass filter with a time constant set by the light level. Thus, the noise spectra
show a lowpass filter characteristic. The voltage noise level per unit bandwidth
� Vs2 proportional to � I 2/9s2 or to 1/1, and the bandwidth is proportional
to 9s and therefore to the photocurrent I. The product of the noise per unit
bandwidth and the bandwidth is proportional to the total noise over the entire
spectrum and is independent of I. Consequently, the area under all three curves
in Fig. 11.9 is the same. The smooth lines in Fig. 11.9 are theoretical fits using
a temperature of 3000K and a value of capacitance estimated from the layout.
It is possible to extend our way of thinking about noise to the above­
threshold region of MOS operation as well. However, the mathematics is
more difficult because the presence of a longitudinal electric field causes non­
independence between the noise resulting from forward and reverse currents.
Further, the modulation of the surface potential by the charge carriers results
in a feedback process that attenuates fluctuations in the mobile charge con­
centration; an effect that is referred to as space-charge smoothing. A simpler
approach to the problem is described in Tsividis (1998) and yields what we
Noise in MOS Transistors and Resistors 333

believe to be a mathematically correct answer.

1M2 =
4kTJ.1 !f Qt::.f l (11.4.7)

where Q is the average charge per unit area.

In a subthreshold MOSFET, Eq. 11.4.7 becomes

I� � 4kTP !f (Q, ;Q') at f (11.4.8)

In the above threshold MOSFET, Eq. 11.4.7 becomes

(11.4.9)

unit area of Q, and G J.1 �

= Q
Eq. 11.4.7 behaves like the noise in a sheet resistor with an average charge per

I t::.I2 =
4kTG t::.f. 1 (11.4.10)

11.5 Noise Examples

Here we show some examples of noise calculation in simple circuits. We look

at four cases: An RC circuit, a MOSFET noise model, a MOSFET inverter,
and a transconductance amplifier.

R R

(a) (b)

Figure 11.10
Noise in an RC circuit.
334 Chapter 11

o o

Ga--O---1
Flicker noise Input-referred noise
Shot noise

S S

(a) (b)

Figure 11.11
Two possible noise models for a MOSFET at low frequency. (a) A noise current source is included
in parallel with the MOSFET together with a noise source in series with the gate. (b) The output
noise current is converted to a input-referred noise source in series with the gate.

Noise in an R C circuit

We can compute the noise spectrum in an R C circuit as shown in Fig. 11.10.

The transfer function of the R C circuit in the Laplace domain is

Vo 1
( )= + (11.5.1)
VR S l R Cs
The noise spectrum of the resistor, SR(f) is shaped by the transfer function of
the circuit. The noise power at the output, So(f) is

So(f)= SR(f) I �: 1 (jW)

2

1
= 4 k TR 7r2 2 2 2 + (11.5.2)
4 R C J 1

and the total noise power at the output is 9

(11.5.3)

The total RMS noise voltage at the output is given by ffo.

= (00) = 7r /2 to solve this equation .

9 Use fooo xg�l arctan - arctan (0)
Noise in MOS Transistors and Resistors 335

(a) (b)

Figure 11.12
Noise model for an inverter. (a) Circuit for an inverter with the input going to the nFET, M..
The bias voltage to the pFET, M2 is constant. (b) Noise current sources added to the transistors.
The output impedance at the output is the parallel combin ation of the output conductances of the
transistors.

Noise Model of MOSFET

The noise calculations so far refer to the output-referred noise, Vn,out. Unfor­
tunately, this figure depends on the gain of the circuit, Av. A better figure of
merit would be to compute the "input-referred" noise, Vn,in' Hence, one can
take the total output-referred noise which is caused all the noise sources in the
circuit and by dividing this output noise by the gain, we get only one noise
voltage source at the input, that is, V;,in = V;,outlA�. Notice that this input
noise source does not exist in the physical circuit.
The noise in a MOSFET at low frequencies comes from two different
sources: white noise and flicker noise. The noise in a MOSFET can be modeled
by a white noise current source in parallel with the transistor (Fig. 11.11(a))
and a flicker noise source in series with the gate. The current noise can be
converted to an input-referred voltage noise Vsn by dividing the current noise
In by the transconductance of the transistor g m: That is, Vs2n = 1;)g ;'.
Assuming that the transistor is in saturation and using Eqs. 11.2.7 and 11.2.19,
we can replace these two noise sources by an input-referred noise source as
336 Chapter 11

I�

(a) (b)

Figure 11.13
Noise model for a transconductance amplifier. (a) Circuit for the amplifier. (b) Noise current
sources added to the transistors.

shown in Fig. 11.11 (b) where

(11.5.4)

So by increasing the transistor's transconductance and the area of the transis­

tor's gate, we can decrease the input-referred noise voltage.

Noise in an Inverter

We now show the noise calculation in an inverter comprising of an nFET M 1

and a pFET M2 (Fig. 11.12). The input Vin goes to the gate of M1 and the tran­
sistor M2 acts as a current source. The voltage, Vb, determines the magnitude
of this current source. In Fig. 11.12(b), the noise current sources associated
Noise in MOS Transistors and Resistors 337

1121b - n3
'_"------1:) Va
Ina -ln3
=

(a) (b)

Figure 11.14
Noise current analysis. (a) Noise current analysis for M3. (b) Noise current analysis for Mt.

with the MOSFETs are added to the circuit. For each noise source, we com­
pute the ac transfer function between its current source and the differential
output current. The total output current noise is the sum of the squares of the
current noise for each source. The input-referred voltage noise per unit band­
width is given by the total output current noise divided by 9 m' The total voltage
noise is then computed by integrating the input-referred voltage noise over the
bandwidth of the circuit.
Assuming that these noise sources are uncorrelated and that the subthresh­
old transistors are operating in saturation, we get

(11.5.5)

where Isatl and Isat2 are the saturation currents of Ml and M2 respectively.
338 Chapter 11

� 0.' 0. .�
� I

(a) (b)

Figure 11.15
Noise current analysis. (a) Noise current analysis for Mt. (b) Circuit transformation for a floating
current source .

Substituting the relationship 9m UT/ into Eq. 11.5.5:

I sat = K

= (
V,;,o 2kT 9ml 9m2 (rod/ro2)2 : ) (11.5.6)

The voltage gain of the inverter is 9 ml (rod/ro2)' By dividing the output­

referred noise by the gain, we get the noise voltage referred to the gate of the

(_9m1_l
input:

V;i ,
= 2kT
K
+
)
9�2 .
9 ml
(11.5.7)

Thus, to reduce the noise in this inverter circuit, we should increase the
transconductance of the input transistor M 1 and reduce the transconductance
of the current source M2. Note, however that, in subthreshold 9 m2/9ml = 1
Noise in MOS Transistors and Resistors 339

because the transconductance is invariant with geometry and depends only on

bias current, which is equal for both devices. However, the same ideas may be
used in an above threshold noise analysis to show that a large W /L for tran­
sistor Ml and a small W/ L for transistor M2 lead to a reduced gm2/ gm l ratio
that helps attenuate the noise.
Flicker noise is also part of the output noise current. The constant K in
Eq. 11.2.7 is different for nFETs and pFETS. Because we operate the circuit
in subthreshold with small bias currents thermal noise dominates over flicker
noise.

Noise in a Transconductance Amplifier

The circuit for the transconductance amplifier is shown in Fig. 11.13(a). The
noise sources associated with each transistor are included with the correspond­
ing transistor in Fig. 11.13(b). We next compute the ac transfer function be­
tween each current source and the differential output current. The noise source
Inb flows through two symmetrical pathways; one through M 1 and the second
through M2 • These currents cancel at the output node. Hence this noise current
does not contribute to the output noise. Let us consider next the noise current,
In3 as shown in Fig. 11.14(a). The current through M 4 is then � - In3 so that
the total current sums to�. This current is mirrored to the output through M 4.
Therefore, In o = -In3. The noise current, In4, flows directly to the output so
In o = In4. A similar analysis can be performed for the noise current I n4 in
Fig. 11.14(b).
The effect of the noise current from M1 on the output node is more
difficult to analyze. The current I nl flows partially through Ml and partially
through M2 (Fig. 11.15(a». To simplify the analysis, we make use of a circuit
transformation technique where we split the floating current source I nl into
2 currents (Fig. 11.15(b». This technique is used when we want to transform
some portion of a circuit so that we can apply Thevenin's theorem or Norton's
theorem to it (Kelly, 1970; Van Valkenburg and Kinariwala, 1982; Chua et al.,
1987). The resulting circuit is shown in Fig. 11.16. The noise at the output is
again In1• A similar analysis can be performed for the noise source I n2 . The
total output current noise I�,o is the sum of the squares of the current noise for
the remaining four noise sources:

(11.5.8)

The input-referred voltage noise per unit bandwidth is given by the total
340 Chapter 11

Figure 11.16
Modified circuit for noise current analysis for MJ .

output current noise divided by the transconductance of the amplifier, 9 m =

K h/2UT. Just as in the noise example in the inverter, we have ignored the
flicker noise.
 12 Layout Masks and Design Techniques

Now that we have seen how to design simple circuits, we describe the differ­
ent steps that lead to fabrication of the circuit. To fabricate a circuit, we need
to generate layout which describes the geometries required for the successive
layers of fabrication. The fabrication house uses these layers to generate masks
that, for example, determine the different areas on the chip where the polysil­
icon gates will be deposited; and which areas will be doped p-type or n-type;
and where to place the wiring between the nodes on the two-dimensional Si
substrate. The fabrication steps are described in Chapter 13. In the first half of
the chapter, we describe how to generate the different fabrication layers from
a circuit diagram. In the second half of the chapter, we describe some layout
techniques that reduce the sources of noise and errors in the fabricated circuit.
Successful layout of analog circuits is a difficult task, and a detailed coverage
of all the issues involved fills entire books (Hastings, 2001).

12.1 Mask Layout for CMOS Fabrication

Different CMOS processes are similar in that they produce the same types
of physical structures. However, each process has its own parameters, such
as the thicknesses and different dopings of the different layers. Processes
may also differ in some of the processing steps that are used to optimize
the performances of the fabricated devices. Furthermore, some processes have
more layers and options than others.
For a given fabrication process, a set of physical structures are available.
The designer has full control over the placement of some of these structures
as projected onto a plane parallel to the semiconductor surface. The position,
size and composition of the structures along the dimension perpendicular to
the semiconductor surface is specified by the manufacturer in the process
parameters. These parameters include layer thicknesses and doping profiles.
The designer-specified structures are defined by a set of binary masks,
where each mask determines the location of a layer in the projection of the
circuit onto a plane parallel to the semiconductor surface. These masks corre­
spond to some of the masks that are used for the fabrication of the circuit. Other
masks used for processing are dependent on the masks defined by the designer
and are generated from them by the manufacturer. The foundry makes avail­
able guidelines for the layout of the masks to be generated by the designer.
342 Chapter 12

These guidelines are called layout rules or design rules. These rules specify
the minimum and maximum dimensions, and spacings for the different masks
that are recommended to ensure that the circuit conforms to the manufacturer's
specifications.
The mask set is used to fabricate electrical circuits near the surface of a
uniformly doped wafer of semiconductor crystal, which is referred to as the
substrate or bulk. Nowadays, most CMOS processes use a p-type substrate.
Chapter 13 describes how the mask set is used for the processing of the wafer.

Mask Layers

For a typical process, the designer has to specify a total of about ten masks.
The most commonly specified masks are the following:

Well

Wells serve as local substrates (or local bulks) for MOSFETs. Wells with
opposite doping from the substrate may also be used as resistors or to design
diodes to the substrate (e.g. photodiodes). Some processes provide a well for
only one MOSFET type, while the other MOSFET type is located directly in
the substrate (single-well or single-tub processes). Other processes provide a
well for each MOSFET type (twin-well or twin-tub processes). For example,
in a p-substrate process, pFETs are fabricated in n-wells and nFETs are either
fabricated in the substrate or in p-wells.

Active

Active regions are all the regions that do not have a thick insulator layer
(field oxide on the surface of the semiconductor crystal. These are the regions
where the structures in the semiconductor are supposed to interact electrically
with the ones that are deposited above it. Active regions include the sources,
channels, and drains of the MOSFETs, and the contact regions to the wells and
substrate. In some processes, active regions can also be used as resistors if they
are surrounded by regions of the opposite doping type.

Select

Active regions are implanted by dopants, except where they are blocked by
gates. The select mask determines which doping type a given active region
receives. The boundaries of the masking regions are usually drawn outside
Layout Masks and Design Techniques 343

the areas that are implanted. These areas are defined by the boundaries of the
active mask and poly mask (see below). Select typically specifies one doping
type, and the active regions that are not covered by select automatically receive
doping of the other type, because the manufacturer uses a mask for it that is
complementary to the select mask. In some processes, two select masks can be
specified; one for each doping type. Active regions without select are then not
implanted by either dopant. Some manufacturers do not require the definition
of separate active and select masks, but use an n + active and a p+ active mask
instead.

Poly

Poly is a conductor layer that is electrically insulated from the substrate and is
used for the gates of the MOSFETs and for interconnects. In some processes,
poly may also be used to implement resistors. In modern silicon CMOS pro­
cesses this layer is made from doped polycrystalline silicon, hence the name.

Poly2

Poly 2 is a second conductor layer separated by a thin insulator layer from the
poly layer below it and is provided in some processes used for analog circuits.
It has the same applications as the poly layer (except that in some processes
it may not be used for MOSFET gates). However, poly2 is mainly used to
provide the option of designing poly-poly2 capacitors, which have a relatively
large capacitance per unit area and have a lower common-mode sensitivity
than poly-active capacitors. The poly2 layer is also useful for the fabrication of
charge-coupled devices, which typically have overlapping electrodes. Standard
CMOS processes used for logic circuits do not usually provide a poly2 layer,
because it significantly increases fabrication cost!.

Metal

The metal layer provides the electrical interconnections between the different
circuit structures. It has a higher conductivity than the poly layer and is there­
fore preferable for longer-range interconnects.

Contact

The contact mask specifies holes in the insulator layer, where the metal layer
is to be electrically contacted to the active, poly, or poly2 region beneath it. In

1 Typically by about 5% to 10%.

344 Chapter 12

some cases, different contact masks are specified, depending on which layer is
to be contacted.

Metal2

Some processes provide one or more (metaI3, metal4, etc.) additional metal
layers for electrical interconnections. They are separated by reasonably thick
insulator layers from the structures beneath them.

Via

Vias are electrical contacts between metal layers, that is, holes in the separating
insulator layers. Processes providing more than two metal layers have more
than one via (via2 connects metal2 with metal3, etc.).

Overglass

The entire circuit is covered by a thick insulator layer to prevent the degrada­
tion of the fabricated structure by the interaction with its environment. This in­
sulator layer is called a passivation layer and in most circuits has holes only at
the locations of the electrical contacts of the circuit to the outside world, which
are called bonding pads. The overglass layer is usually drawn where the holes
in the passivation layer are intended. It should then actually be called overglass
cut. The passivation layer absorbs electromagnetic radiation and additionally
has a strongly wavelength-dependent optical transmission due to interference
effects. It may therefore be desirable to use overglass cuts at the locations of
photodiodes and of areas where exposure to ultraviolet radiation (for example,
to facilitate electron tunneling) will occur.
The above masks are sufficient for the designer to implement a circuit
in a standard CMOS process2• The other masks needed for processing are
generated from these masks by the manufacturer according to process-specific
rules. Figure 12. 1 shows the mask layout of an inverter in a single-well process
using an active mask and a select mask that specifies the same doping type
that is used for the well. A few additional options may be available. Some
processes that do not provide a poly2 layer use a capacitor implant3 into the
semiconductor bulk as a capacitor plate to fonn linear capacitors with poly.
Some silicided processes (see Chapter 13) provide a silicide block mask. This

2 If you are lucky. Some processes require the specification of additional layers.
3 A capacitor implant is an impurity doping implant, but it differs from the source/drain implants
by the fact that it happens before the poly deposition and can therefore be located underneath the
poly.
Layout Masks and Design Techniques 345

[J Well D Select � Metal

EJ Active � Poly • Contact

Well Source Gate Drain Gate Source Substrate

contact diffusion oxide diffusion oxide diffusion contact
diffusion diffusion

Figure 12.1
Mask layout and cross-section of an inverter in a single-well process. The mask layout uses a single
select mask for the active doping of the same type as the well doping. The shown cross-section
horizontally cuts through the center of the circuit. The inverter consists of two MOSFETs of
opposite types, that have connected gates and drains. The source of the well MOSFET is connected
to the well and the source of the MOSFET in the substrate is connected to the substrate.

feature is useful for the design of photosensors, since silicided layers may
block most of the incoming light.
There is a class of processes, called BieMOS (Bipolar CMOS) processes,
which offer vertical bipolar junction transistors (BJTs) in addition to MOS­
FETs, in the same substrate. These processes provide a base implant with the
same doping type as the substrate. In order to construct a BJT in such a tech­
nology, a source/drain implant is used as the emitter; the base implant as the
base; and a well as the collector. Only one type of BJT is obtainable by this
346 Chapter 12

simple enhancement and the resulting BITs do not exhibit a very good perfor­
mance, given that the emitter and collector layers are not optimized for BITs.

A few processes also provide the option of implementing buried-channel

charge-coupled devices (BCCDs). They offer a lightly doped BCCD implant
for the buried layer (see Section 10.5). Such CMOS-compatible BCCDs are of
a lower quality than commercial ones, which are produced with dedicated CCD
processes, and are therefore not suitable for high-resolution imagers. However,
CMOS-compatible BCCDs are much better than the surface-channel CCDs
that are fabricated using standard CMOS technology, because the density
of imperfections at the semiconductor-insulator interface of a typical CMOS
process is too large for efficient multi-stage charge transfer along the interface.
The software packages used for circuit layout allow the designer to draw
the required masks with a two-dimensional graphical interface that represents
the projection of the circuit onto a plane parallel to the semiconductor surface.
The different masks are represented by polygons of different colors and/or dif­
ferent stipple patterns. A polygon marks the location where the corresponding
binary mask assumes one of its two values (for example, where it is trans­
parent to the processing step), while the absence of such a polygon denotes
the other value of the mask. In addition to the mask layers, layout programs
use a host of symbolic layers, which are ignored by the manufacturer. These
layers have different functions. Some of them are logical combinations of the
mask layers and are either used for automatic design rule checks (DRCs) or
for circuit extraction, that is, the generation of a symbolic description of the
circuit on the device level (for example, in SPICE format). Circuit extraction
is used mainly to verify if the functionality of a circuit layout matches that
of the circuit schematic that it implements. Other symbolic layers are used by
the graphical user interface, for example, to mark the locations of design rule
violations.

12.2 Layout Techniques for Better Performance

The proper layout techniques should be employed to ensure good device

matching and circuit operation on a chip. Fabricated circuits frequently do
not work exactly as in simulation. The layout of these circuits is important
to minimize parasitic effects and undesirable coupling effects. Every node on
a circuit has parasitic capacitances and resistances which should be taken into
Layout Masks and Design Techniques 347

account during design. The values of these parasitic elements depend on the
layout of the circuit. Additionally, if devices are to be matched in a design, the
layout of these devices is important in minimizing the effect of mismatches.
Processing variations across a wafer can create mismatches in devices that are
to be matched.

RESISTORS CAPACITORS TRANSISTORS

.~. .. . .
·
·
·
·
·
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
REFERENCE

1�ljgjl.
·
·
·
·
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
GOOD

��j. � Poly1 Diffusion

BAD

�
Figure 12.2
Layout elements. The first row shows the reference elements. The second row shows elements that
are matched well to the reference set and the third row shows elements that are badly matched to
the reference set.

Another major source of problems comes from mixed analog and digital
circuits that are placed on the same chip. Analog circuits are sensitive to small
changes in voltage; digital circuits typically swing over the entire power supply
range when they switch. The different operating ranges of analog and digital
circuits should be taken in consideration when planning the layout of the chip.
The switching of a digital circuit can couple into an analog circuit and affect
its operation. Digital circuits should be have their own power supply lines and
348 Chapter 12

analog circuits should be placed on a different part of the chip, preferably, as

far as possible from the digital circuits, to minimize coupling effects.

Figure 12.3
Resistors to be matched should be placed closed to one another.

In the following section, we describe some layout techniques that ensure

good device matching for better circuit performance. We also describe some
suggestions for layout techniques that apply to mixed-mode circuits on the
same chip. The techniques described in this chapter have been developed by
veteran circuit designers (Vittoz, 1985; Tsividis, 1996).

Device Matching

Device Size Matching Devices to be matched should have the same dimen­
sions. For example, two transistors that should have the same width/length
ratio should have the same lengths and same widths (Fig. 12.2). The matching
in each dimension is necessary because the dimensions of a fabricated tran­
sistor are different from the drawn dimensions. The parasitic capacitances at
each node are proportional to the area, so even if two transistors have the same
dimension ratio, the parasitic capacitances will be unequal if one transistor has
twice the dimensions of the other. The same criterion applies to capacitors and
resistors. Capacitors are usually made of a bottom poly 1 plate and a top poly2
plate with the interpoly oxide as the insulator. Resistors can be made out of
polysilicon or well. Two resistors with unequal dimensions will have different
parasitic capacitances. Jogs in resistors are not desirable because etching at
sharp corners is not isotropic. Capacitors to be matched should not be drawn
with minimum dimensions as the actual dimensions will change after fabri­
cation. To reduce the effects of the variation in the dimensions, two matched
capacitors should be of reasonable width and length and have the same di-
Layout Masks and Design Techniques 349

mensions. If one capacitor has the same area but a different perimeter, the
capacitance values might be unequal after fabrication.

(a)

GOOD

(b)

Figure 12.4
Transistors in a current-mirror circuit should be placed closed to one another. (a) Bad layout. (b)
Good layout.

Minimum Distance Because of the spatial variation in the process parame­

ters, devices to be matched should be placed next to one another. For example,
long resistors should be interleaved together (see Fig. 12.3) and the input tran­
sistors in a current mirror circuit should be placed closed together (shown in
Fig. 12.4).

Same Orientation Devices should be placed in the same orientation so that

they will be subjected to similar effects due to anisotropic process steps, found
for example, in mask alignment and anisotropic doping. Other asymmetries
come from stress after packaging. For example, transistors in a differential
350 Chapter 12

pair should be matched closely in orientation (Fig. 12.5(c)). Figures 12.5(a)

and (b) show some examples of poor layout for a differential pair.

BAD

(a) (b)

GOOD

(c)

Figure 12.5
To ensure good matching. the input transistors in a differential pair should be placed as close as
possible to one another. In addition. the transistors should be in the same orientation so that they
are subjected to same alignment mismatch. (a) Poor layout of the transistors. The orientation of
these transistors is orthogonal to each other. (b) Better but still bad layout; the current in the two
transistors flows in opposite directions. (c) Good layout; the orientation and the current flow are
the same for both transistors.

Common-Centroid Geometry Preferably, transistors in a differential pair

circuit should be placed in a common centroid arrangement (see Fig. 12.6).
However, to satisfy this requirement, the layout cost is large. This type of
arrangement is reasonable only for small numbers of transistors.
Layout Masks and Design Techniques 35 1

Common Surroundings Matched devices should have the same surround­

ings (Fig. 12.7). The edge effects, for example, on capacitors will be different
if the capacitors are surrounded by different circuitry. It is best to put dummy
capacitors on the edges as shown in the figure.

Figure 12.6
Layout example to show the common-centroid arrangement of the input transistors in a differential
pair.

Common Temperature If there is a source of heat on the chip, place

matched devices on the same isotherm line.

12.3 Short List of Matching Techniques

A summary of these layout techniques for matching devices based on ratios of

values is as follows:

1. They should have the same dimensions.

2. They should have the same structure. For example, we should not try to
352 Chapter 12

dummy device dummy device

(a)

dummy device dummy device

.,'-- (b)

Figure 12.7
To ensure good matching, the matched transistors and capacitors should be surrounded by identical
elements.

match nFETs to pFETs.

3. They should not be of minimum size because minimum-sized devices are
affected more by random factors in processing.
4. They should be placed close to one another.
5. They should have the same orientation, for example, the current flow should
be in the same direction.
6. They should be preferably be laid out in a common-centroid arrangement.
This arrangement is especially important for large transistors.
7. The surrounding circuitry should be similar.
8. They should be at the same temperature.
Layout Masks and Design Techniques 35 3

12.4 Parasitic Effects

The capacitances and resistances of the drawn circuits are supplemented by

parasitic elements from the circuit layout. The parasitic capacitances come
from the inherent capacitance between the different layers of the fabricated
circuit. The parasitic resistors exist because of the finite resistivity of the dif­
ferent materials of the layers. These numbers are provided by the fabrication
house. Other parasitic effects arise because of the increasingly small geome­
tries in modem processes. For example, the sheet resistances of the drain and
source nodes of a transistor cannot be ignored in a 0.25 /Lm CMOS technol­
ogy. These parasitic effects can reduce the speed and bandwidth of the circuit,
distort the circuit characteristics, reduce the circuit precision, create noise, and
add to signal loss. Particular care should be paid to critical nodes to reduce
or eliminate the parasitic effects. These parasitic capacitances and resistances
should be taken into account when simulating the circuit performance.
One important parasitic capacitance is the capacitance from any node to
the substrate because the substrate is common to all circuits on the chip.
Any coupling from a node to the substrate can affect many circuits. These
coupling effects are especially crucial when considering mixed-mode analog
digital circuits. Digital circuits typically switch over the entire power supply
range. If care is not taken, these transients can couple into analog circuits
which are normally sensitive to small voltage changes. The following are some
suggestions to reduce the effects of coupling.

Avoid Common Supply Lines Digital circuits and analog circuits should
have separate power supply and ground lines. A bad practice is to run a
common supply line to both analog and digital circuits. Large transient spikes
can result because digital lines switch quickly over the power supply range.
These transient currents create a voltage drop along the power supply line due
to the resistance of the line and the inductance of the packaging wires. The
power supply to the analog circuits will vary due to the switching in the digital
circuits. This variability can be reduced by bringing different lines from the
same common pad to the analog and digital circuits. Then, the switching noise
from the on-chip digital circuit can be separated from the analog circuit. An
even better solution is to bring the supply lines to separate pads on the chip, so
that the interference due to the inductance of the bonding wire can be reduced.
Remember, that the capacitances of the supply lines and the pads; and the
inductance of the wires, can form a resonant LC circuit. This impedance will
354 Chapter 12

increase if the frequency of the digital clock matches the resonant frequency
of the LC circuit.
The ground line also creates a problem. The switching current of a digital
circuit leads to large voltage spikes on the ground line. These voltage spikes
affect the operation of circuits with low power-supply rejection ratio, or if the
ground line is used to bias an input terminal. It is better to group circuits with
common functionality together and have separate ground lines running to each
group.

(a)

--
----�--��-- �s

(b)

Figure 12.8
An n-well CMOS structure showing the parasitic resistances and bipolar junction transistors that
contribute to the latchup mechanism. (a) Cross-section of the CMOS structure. (b) Equivalent
circuit that causes latchup.
Layout Masks and Design Techniques 355

The power supply and ground lines should be wide so that these lines have
low resistance. If there are large current drivers in the circuit, we can estimate
the voltage transient by measuring the total transient current and multiplying
by the resistance of the supply lines. A side effect of wide lines is that their
capacitance increases and can cause undesirable effects like parasitic coupling
to the substrate.

12.5 Latchup

CMOS structures are susceptible to a mechanism called latchup. This mech­

anism is aggravated by the parasitic resistances in the different regions of the
silicon bulk. There are many structures in the bulk that can trigger latchup.
When it is triggered, hugh amounts of current will flow through the chip. If the
power supply to the chip is not current limited, then the chip will be irreversibly
damaged.
Figure 12.8(a) shows the parasitic elements in a cross-section of a CMOS
n-well structure that result in latchup. We show only one source/drain region
of an nFET and one source/drain region of a pFET. There are two parasitic
bipolar junction transistors (BITs) in the structure. One is a lateral npn bipolar
junction transistor and is fonned by the n + source of an nFET, the p- substrate,
and the n-well. The other is a vertical pnp bipolar junction transistor and is
fonned by the p+ source of a pFET, the n-well, and the p- substrate. The
bulk and well have parasitic resistances, denoted by Rps and Rnw respectively.
The equivalent circuit consisting of the BITs and the resistances are shown
in Fig. 12.8(b). This npnp structure is also called a thyristor. The parasitic
resistances are critical for turning on this structure. For example, if there is a
big enough transient current on Vdd, this current causes a voltage drop across
Rnw and turns on the pnp BIT. The collector current from this BIT charges up
the base of the npn BIT through Rp and Rps, thus also turning on this BIT.
The collector current from the npn BIT further lowers the base voltage of the
pnp BJT so creating a positive feedback condition and also creating a runaway
process: Latchup. Huge currents flow in this structure and are limited only by
the resistances. If the power supply is not current-limited, the circuits on chip
will be destroyed. The same mechanism will occur if there is a transient on
Vss. Remember that there are many structures of this type all over the chip.
There are several solutions to combat this problem. First, we can make the
parasitic resistances small by separating the n + and p+ regions so that the f3
356 Chapter 12

V Switching
current

p substrate

Figure 12.9
Resistive coupling through substrate. Current can be injected into the substrate from different
sources. For example, if node Vi represents the place into which the switching current from a
digital transistor flows into the substrate, then node \2 which represents the substrate of a voltage­
sensitive transistor will be affected by this current through resistors, R and R.!. This coupling can
be reduced by placing a substrate contact next to the node l{ so that the substrate resistance, RI
is small.

of the npn BIT will be small (due to the large base width). Second, substrate
contacts should be placed frequently and close to the transistors so that R ps
and Rnw will be small. Layout rules from the foundry include a minimum
spacing between the n+ well contact and the n+ region in the substrate to
prevent latchup.
Different techniques are used in industry to minimise latchup. One tech­
nique is to grow a layer of lightly doped p-type material on top of heavily
p-doped starting material. This lightly doped material is called an epitaxial (or
epi) layer . The transistors and the n-well are formed within the epi layer. The
implanted well is adjacent to the the heavily-doped p+ layer so that the Rp
resistance will be small. Another technique is called the trench isolation tech­
nique. A vertical trench filled with polysilicon is placed around the well. The
trench reaches the heavily-doped p+ area. Some technologies like the silicon­
on-isulator (SOl) technology is completely immune to latchup.

12.6 Substrate Coupling

The substrate is common to all the devices on the chip so any non-desirable
coupling into the substrate can create problems. There are different ways that
signals can couple into the substrate: Resistive coupling (for example through
an n+ -n junction), capacitive coupling, and coupling through a bipolar tran­
sistor.
Layout Masks and Design Techniques 357

Shielding from Substrate Noise A node of a digital transistor affects the

substrate through the junction capacitor between the drain/source of the tran­
sistor and the substrate. If the digital transistor is close to an analog transistor,
the switching on the node of the digital transistor is coupled into the substrate.
Since the substrate is common to both types of transistors, this switching can
couple into any sensitive nodes including resistors and capacitors, or into the
substrate beneath the analog transistor. The switching affects the transistor cur­
rent through the body effect. It also couples into the transistor through other
parasitic capacitors of the transistor (shown in Figure 12.9). There are sev­
eral ways of reducing the impact of the coupling from the digital circuit. One
method is to separate the analog circuits and the digital circuits so that the re­
sistance R between them is large (see Fig. 12.9). However, this technique is
not effective when the process has an epitaxial layer. The heavily doped sub­
strate underneath the epitaxial layer has a low resistance. An option is to use
an insulating substrate like SOL
An effective way of reducing substrate noise from digital ground is to
separate the source and substrate wires to the digital transistor so that the high
switching currents at the source node do not affect the substrate node.

Reducing Substrate Resistances Another way of reducing the impact of the

coupling from the substrate is to reduce Rl and R2 in Fig. 12.9. This can be
done by placing guard rings around the circuit that is sensitive to interference
or that generates interference. The guard rings consist of a ring of diffusion
with substrate contacts. Make sure that the ring is tied to a ground that is
either brought out separately to the pad or use digital ground or analog ground,
whichever one least affects the circuit.

Shielding from Minority Carriers Analog switches can generate minor­

ity carriers into the substrate when the switches are turned off. Circuits with
phototransistors and photodiodes are also susceptible to minority carriers be­
cause the incoming photons generate carriers in the substrate that can diffuse
to neighboring circuits if they are not collected in the transducer area. One way
of reducing the effects of minority carriers is to build a guard ring around the
interfering circuit or sensitive circuit so that minority carriers will be soaked
up by the ring. This scenario is shown in Fig. 12.1O(a).
A second well ring can also be used to reduce substrate currents, for ex­
ample in photodiodes. The majority carriers in a p-type substrate are holes: In
358 Chapter 12

an n-well to p-substrate junction, electrons that do not go to the well, generate

substrate currents. An n-well guard ring collects these minority carriers in the
substrate.
Minority carriers from analog switches can be reduced by putting a guard
ring between a sensitive circuit or by placing the switches and sensitive circuits
in separate wells (see Fig. 12. 10).

Photodiode Positive power supply

Electrons : >17 :we!l: p substrate

(a)

Sensitive node

Separate wells

(b)

Figure 12.10
Minority carrier shielding. (a) Minority carriers generated at the junction of the photodiode can
migrate into the substrate. To reduce this current from affecting other transistors. a well guard ring
is placed around the photodiode area. (b) Sensitive nodes can be shielded from analog switches by
placing them in separate wells.

Capacitance Shielding Both the top and bottom plates of the capacitor
should be shielded from any node with high variance. The shielding of the
bottom plate is especially important because the coupling into the substrate
from any interfering circuits can cause excursions on the voltage on the bottom
plate. One way of doing this is to place a grounded n-well under the capacitor.
Shielding of resistors is also important. For example, a poly resistor can be
Layout Masks and Design Techniques 359

shielded from the top by the first-layer metal in a two-layer metal process.
Transistors that are sensitive to noise should be placed in wells if possible
and the wells tied to clean supplies. Remember that there is also capacitive
coupling through the air so sensitive nodes should be far away from interfering
nodes. Shield analog circuits from digital circuits by placing them in separate
wells.

Other Important Tips

Bonding Pads Connect nodes that will brought to the outside to the closest
bonding pads. Ensure that the bonding wires from pads to the package pins are
short as possible.

Avoid proximity Do not place sensitive lines close to one another or cross
them over each other or over any interfering lines. Avoid crossing digital
input/output lines over analog lines.

Sensitive Nodes Make lines to sensitive nodes as short as possible to avoid

parasitic coupling. Make lines to and from interfering circuits as short as
possible: for example, the outputs of digital circuits.

Differential Circuits Make circuits differential where possible. Differential

circuits are useful for reducing effects due to parasitic coupling. Make the
layout of differential circuits as symmetrical as possible.

Floorplan of Chip Plan the chip before doing the layout. Place analog
circuits together and digital circuits together.

12.7 Device Matching Measurements

Models to characterize mismatch in MOS transistors for precision analog de­

sign were first developed in the 1980s (Lakshmikumar et aI., 1986; Pelgrom
et al., 1989). The mismatch between transistors can be characterized by mea­
suring the standard deviation of the current, a( Of). This deviation can be com­
puted from the standard deviation in the f3 parameter4, a( 0: ), and the stan­
dard deviation in the threshold voltage, a(V T) across transistors5• Mismatches
4 fJ = J1.Co",W/L.
5 Variations in parameters like K, CO"" W, L are included in these two parameters.
360 Chapter 12

arise because of fabrication process variations (for example, changes in width

and length of transistors, oxide thickness, doping) and spatial "white noise"
in the device parameters; for example, local fluctuations in electron mobil­
ity (Pavasovic et al., 1994). Careless layout of a circuit can create unnecessary
mismatches. The standard deviation in VT varies between 2 and 20 mV and the
standard deviation in f3 varies between 2 and 20% (Enz and Vittoz, 1997). Both
analysis and measurements from various sizes of transistors show that both �
and a(VT) vary inversely with v'W L where W and L are the width and the
length of the transistor respectively (Lakshmikumar et al., 1986; Pelgrom et al.,
1989). Hence, one way of reducing mismatch is to make the transistor area
large. This also applies to other devices like capacitors. The results in Pelgrom
et aI. (1989) also show that the current mismatch in a current-mirror transistor
pair is largest in subthreshold and decreases when the transistors are operated
above threshold. This can be understood by computing how much the variation
in the threshold voltage 8VT contributes to a( Of). To do this, we multiply 8VT
by sr. The multiplicand sr is larger in subthreshold than in above threshold.
Hence the mismatch is larger for a current-mirror pair in subthreshold.
The current mismatches measured for arrays of different-sized transistors
operating in subthreshold (Pavasovic, 1991; Pavasovic et al., 1994) show that
that mismatches are caused by three main factors: edge effects, spatial stria­
tions, and random process effects. Edge effects occur because the transistor
characteristics depend on the surrounding devices. The variance in the cur­
rents among equal size transistors that are biased with the same gate-to-source
voltage vary anywhere from 5% to 5 0% depending on the size, type, and sur­
roundings of the transistor (Pavasovic, 1991; Pavasovic et al., 1994; Serrano­
Gotarredona and Linares-Barranco, 1999).
The mismatch due to spatial variations in the process also apply for other
devices, for example, capacitors. Measurements from capacitors of different
sizes (McCreary, 1981; Shyu et aI., 1982, 1984; Vittoz, 1985; Gregorian and
Ternes, 1986; Tuinhout et al., 1996; Minch et aI., 1996; Minch, 2000a) show
that the relative-capacitance distributions of larger size capacitors across the
chip is lower than the smaller size capacitors. Generally, capacitors are very
well matched compared to transistors or resistors 6• Designers who think that
all devices work alike are shooting themselves in the foot.

6 That's why switched-capacitor circuits rule!!

 13 A Millennium Silicon Process Technology

This chapter is a short description of typical commodity silicon process tech­

nology as we enter the new millennium. We describe a state of the art process
flow for 0.25 /Lm CMOS 1 by explaining the different process modules. We con­
clude with a brief discussion of the scaling of process technology to smaller
dimensions. Process technology is extremely complex and very highly devel­
oped; we can only touch on some important aspects of it. The interested reader
may wish to consult the excellent book (Plummer et al., 2000).

13.1 A typical 0.25 /Lm CMOS Process Flow

The lithography technology used for critical layers (usually active, poly gate,
contact and metall) is deep UV (DUV2) optical lithography. All other layers
use I-line (365 nm) lithography.

1. Field Oxide (Isolation) Formation The process begins with prepared

wafers that must be patterned with the field oxide transistor isolation structures.
The starting material is either p+/p- epi3 or p- bulk4. The process starts with the
formation of the isolation oxides. Two different isolation approaches are used
in 0.25 /Lm CMOS processes: The first is called LOCOS 5 where a nitride61ayer
locally inhibits the growth of a thick isolation oxide layer (Fig. 13.1) (Kooi
et al., 1976; Wolf, 1995).
The major drawback of this traditional approach is the oxidant diffusion
under the edges of the nitride layer which results in a smaller than expected
active width. The difference between the original active width on the mask
and the final active width is named ''bird's beak". To accommodate this bird's
beak, it is necessary to oversize the active region layer by the amount of
two times the bird's beak, leading to a loss in real estate. High-pressure

1 0.25 p,m processes entered mass production around 1998.

2 DUV is <248 mn, and generally uses 193 mn KrF laser light sources.
3 epi=epitaxially grown. Epi means that lightly doped silicon of a few microns is grown on a
heavily doped substrate. The lightly doped epi allows formation of n and p wells and the heavily
doped substrate reduces latchup problems.
4 Bulk material is sometimes used for commodity products because it is cheaper.
5 LOCOS=LOCal Oxidation of Silicon.
6 Nitride=Silicon Nitride=SiaN4. Nitride is an oxidation barrier, among other useful characteris­
tics.
362 Chapter 13

F; ��� �i02
Field implant Si3N4
' ' ' '
'' '' ''''''
. . .. . .
� . . . . r""'''�I·!
High temperature steam oxidation

! """", 9'>----
\t':=:::.�=

___ LOCOS
..& ! \S.. ...............
>,>-r------,-« Si02

Gate oxide growth

High temperature 02

gate oxide Field oxide (FOX)

Figure 13.1
Steps in a LOCOS process. A thin layer of sacrificial oxide is grown on top of the Si substrate. Next
a layer of silicon nitride is grown on top of the oxide in areas where the field oxide formation is not
desired. A shallow field implant is done to inhibit parasitic transistor action between source/drain
regions belonging to different transistors which are separated by field oxide. The nFET regions get
a p-type field implant, and the pFET regions get a n-type implant. After the oxide has been grown,
the silicon nitride and the sacrificial oxide are removed. A layer of clean gate oxide is then grown
on top of this area. The field implant diffuses vertically under the field oxide and laterally into the
channel. The encroachment of oxide under the edge of the gate is called the bird's beak; together
with the lateral diffusion of the field implant, it causes the transistor channel width to effectively
be narrower than drawn by a distance a on each side. In most present-day process flows, the field
implants are done after field oxide formation, not before, as shown here. Figure adapted from Kooi
et aI. (1976). Reprinted with permission of The Electrochemical Society, Inc.

oxidation has been shown to reduce the bird's beak, but is not usually used
in CMOS processes. At least 100 different complex LOCOS concepts have
been developed to minimize the bird's beak: These include semi or fully
recessed LOCOS; Poly Buffer LOCOS (PBL) (Han and Ma, 1984); Sealed
Interface Local Oxidation (SILO) (Bergemont et aI., 1989); and Sidewall
Masked Isolation (SWAMI) (Chiu et aI., 1982). Details of these different
LOCOS concepts can be found in Wolf (1995).
A Millennium Silicon Process Technology 363

(3) Stack & trench etch (b) Pad oxide undercut (e) Liner oxidation

(d) CVD oxide gaplill (e.g. HOP, TEOS-02) (e) eMP & HF cip (f) HSP04 Nitride strip

Figure 13.2
Different steps in a Shallow Trench Isolation (STI) process. Figure adapted from Nandakumar
et aI. (1998), Shallow trench isolation for advanced ULSI CMOS Technologies, International
Electron Devices Meeting Technical Digest. © 1998 IEEE.

The second process is called Shallow Trench Isolation (ST/), and is the
more commonly and recently used approach. In this process, the bird's beak
is reduced to near zero and it has the advantage of providing a better planar
structure for further processing. The formation of STI (Fig. 13.2) starts with
the growth of a thin pedestal oxide, followed by the deposition of a silicon
nitride layer (Holloway et aI., 1997; Nandakumar et aI., 1998; Matsuda et al.,
1998; Kuroi et aI., 1998). The active mask is used to pattern the STI openings,
followed by the etching of nitride/oxide/silicon. The trench etch into silicon
is usually around 0.4-0.5 f..lm. It is important to realize 80 0 angle sidewalls

Figure 13.3
Cross section of STI. The transistor and isolation structure are stained to delineate rt /p+ junctions
and gates. Figure from Holloway et al. (1997), 0.18 /.lm CMOS Technology for High·Performance,
low-power and RF applications, 1997 Symposium on VLSI Technology: Digest of Technical
Papers. © 1997 IEEE.
364 Chapter 13

as well as rounded bottom comers to minimize stress induced defects into the
silicon. After photoresist strip, a thin thermal oxide is grown on the trench
sidewalls to recover any defects and to produce a small round comer at the
top of the trench. Then, a dielectric film is deposited to fill the trench. Confor­
mality7 of deposition is important to avoid the formation of voids inside the
trench. High-density plasma (HDP) chemical vapor deposition (CVD) tools
are frequently used because they can adequately fill the trench. Then a chemi­
cal/mechanical polishing (CMP) step is used to polish back the HDP film until
the nitride layer is reached (nitride acts as a CMP stop layer). The dielectric is

10.0 r·..···..,..........·,............,..........··,............·,....·......·r···..........r........·..·r....·........

8.0
�
Q)
0>
� 6.0
(5
>
c
:;
0 4.0
"0
.:s:.
'"
Q)
cD
2.0

0.2 0.4 0.6 0.8 1.0

Isolation space (11m)

Figure 13.4
Isolation characteristics of STI technology. Figure adapted from Kuroi et al. (1998), Stress analysis
of shallow trench isolation for 256M DRAM and beyond, International Electron Devices Meeting
Technical Digest. © 1998 IEEE.

then densified at high temperature (1000-1100 0c) to suppress any HDP stress
induced defects into the silicon (Kuroi et aI., 1998). Then nitride and pedestal
oxide are removed, resulting in the final structure shown in Fig. 13.3 (Holloway
et aI., 1997). Figure 13.4 (Kuroi et aI., 1998) shows that STI can maintain good
isolation characteristics down to the 0.1 ILm range.

7 Conformality means uniformity of thickness of coverage.

A Millennium Silicon Process Technology 365

NMOS PMOS
pwell, VTN implants

I I I I I I I
t t t oxide
t t t t Resist
Screen (not to scale)

, .
I
I'
' p well' .... . ' � well'
I •
I.
I
I
I
I
I
I
I
----------------------------------------------------------

Figure 13.5
Well implantation.

2. Formation of Wells The next step is the implantation of the wells through
the field oxide using high energy implants (500 keV for n-Well Ph 8 implant
and 300 keY for p-Well 89 Implant.). Since the dopants are placed deeply
enough by the implant itself, there is no need for high temperature annealing
to drive the dopants 10 as in a 0.5 {Lm CMOS process. Usually the threshold
adjust and punchthrough implants 11 are performed at the same time (with the
same p- and N- masks). Figure 13.5 shows a process cross section after the
wells are formed.

3. Gate oxide formation Figure 13.6 shows the wafer cross section after
formation of gate oxide, including the next step of polysilicon deposition and
patterning. Gate oxide is continuing to scale down: The typical thickness for
2.5 V, 0.25 {Lm technology is 45-50 A. Not only does the oxide need to be
reliable from a lifetime point of view, but the oxide also needs to prevent the

8 Ph = Phosphorous
9 B = Boron
10 Annealing means temperature cycling to heal crystal defects. and driving means heating the
wafer to cause rapid diffusion of the dopants.
11 The threshold adjust/punchthrough implants increase both nFET and pFET thresholds to
prevent leakage current when the transistor is off.
366 Chapter 13

p well:' . '
.
'
.
'
.
.
. : '
.
.
, n well
I
I
I
I
I
I
I
I
L ______________________________________________________________ �

Figure 13.6
Gate oxide and poly formation.

diffusion of dopants (especially B from the p+ gate). The addition of nitrogen

in the gate oxide has been proven to be a solution for Boron penetration with
oxides grown in NO or N20 ambient.

4. Gate (Polysilicon) Formation Polysilicon deposition follows the gate

oxide growth. Figure 13.6 shows a process cross section after gate patterning
and plasma etching. The amorphous or polycrystalline polysilicon is undoped,
with a typical thickness in the range of 300-400 nm. The doping of the gate
is done later, simultaneously with the source/drain implants, allowing surface
channel fonnation for both n and p channel transistors.

LDD implant

/\1\1\/\1\ Resist I\
p w�1I' . ' . . . . .
, n well
I
I
I
I
LDD implants I
I
I
I
L ______________________________________________________________ �

Figure 13.7
Source/drain extension. The first step in full source/drain formation.
A Millennium Silicon Process Technology 367

5. Source/Drain Extension (SDE) Formation After gate formation, lightly

doped source/drain implants (LDD) are realized, self-aligned with the edge

Sidewall spacer

Field
oxide

, n well
I
I
I
I
I
I
I
I
�--------------------------------------------------------------�

Figure 13.8
Lightly-doped drain spacer formation.

of the gate. The primary functions of the SDE are to (l) form a region that
reduces the local electric field, to minimize hot carriers effects; (2) realize
extremely shallow junctions to minimize short channel effects (Drain-induced

BF2 implant

Resist
11111.-------.
p+ source drains

p we) l : . . .. , n well

I
I I
I I
I I
I I
I I
I I
1 ______------------ ---------------------------______1

Figure 13.9
Heavily-doped source/drain p+ implant.
368 Chapter 13

Arsenic implant

1 1 1 1 11 1 1 n+ source drains
--
Resist
11
, n �ell

I
I
I
I
I
I I
1 _______________________________________________________________ �

Figure 13.10
Heavily-doped source/drain n+ implant.

barrier lowering). Arsenic (As) implant is used for NLDD and BF 2 for PLDD
regions 12. Some processes also add halo implants 13 to further reduce the short

Field
oxide

p we.l l' , , n �ell

I
I
I
I
I
I
I
: _______________________________________________________________ J
Figure 13.11
Salicide formation.

12 As = Arsenic (slow diffusing donor), B = Boron (acceptor), Ph = Phosphorus (fast diffusing

donor).
13 Halo implants are tilted implants to introduce dopants under the gate edges. They help prevent
punchthrough for minimum length transistors.
A Millennium Silicon Process Technology 369

channel effects (Ph for pFETs and BF 2 for nFETs). Figure 13.7 shows a cross
section after LDD implants.

6. Spacer Formation An oxide or nitride layer is then deposited and

anisotropically etched 14 back, to form dielectric spacers (Fig. 13.8). Then
the heavily doped source/drain areas are implanted. These regions have deeper
junction depths than SDEs; the poly gates are also simultaneously doped with
those implants (Figs. 13.9, 13.10). Then a high temperature (1050 0c) rapid
thermal anneal is used to activate 15 the dopants.

Doped SAPSG Undoped TEOS

P w�II'
.
.
.
.
.
.
.
.
.
.
.
.
.
, n w�11
I
I
I
I
I
I
I
I I
L _______________________________________________________________ J

Figure 13.12
Interlevel dielectric deposition and polishing. SAPSG (Sub Atmospheric Phophorous doped
Glass) is a silicon phosphate covering material that prevents sodium contamination,TEOS
(tetraethooxysilane) is a liquid containing Si and 0 that is decomposed to form SiQ, and is un­
doped to prevent phosphorous contamination.

7. Salicide Formation The next step is the formation of metalized polysil­

icon called salicidel6, to reduce the sheet resistance of the polysilicon gate
and interconnects as well as the source/drain regions 17. The spacers formed
previously On the sides of the poly gates inhibit the formation of the salicide
over the spacer region, to avoid source/drain to gate shorts. The sequence to

14 Anisotropic etching etches vertically much more than laterally.

15 Activate means to anneal the silicon structure after implantation to make the dopants part of
the single crystal structure.
16 Salicide means self aligned silicidation; Silicide is usually CoS� or TiS b
17 Section 12.1 discusses the effect of silicide on photosensitivity.
370 Chapter 13

p we n : .. '" '.. , n well

I
I
I
I
I
I
I
I I
1 _______________________________________________________________ J

Figure 13.13
Contact openings between active and first metal.

form salicide starts with the removal of any oxide on the source/drain and poly
lines. Then the refractory metal 18 (Ti or Co) is sputtered onto the wafer. A low
temperature anneal is next used to react the metal and the silicon exposed on
the source/drains and poly lines. The refractory metal that lies on the spacer

Cobalt silicide Junction

, ' , ' .
p we!l : .... .' ..... , n well
I
I
I
I
I
I
I
I I
1 _______________________________________________________________ J

Figure 13.14
Contact filling.

or the field oxide is not converted into salicide during that step, allowing the
removal of that unreacted metal in those regions. Then a second anneal is per­
formed to convert the salicide into a more stable and lower resistivity material.
Figure 13.11 shows a cross section after salicide formation.

18 Refractory means it has a high melting temperature.

A Millennium Silicon Process Technology 37 1

Passivation

Figure 13.15
Final cross section after passivation.

8. ILD Formation The Interlevel dielectric (ILD) between poly and metall
is then deposited and polished by CMp19 to provide a planarized flat surface
(Fig. 13.12).

9. Contact Formation A contact mask is used to define openings into the

ILD (Fig. 13.13). The contacts connect metal to active or poly areas and also
interconnect metal layers. The interconnect itself is generally Al (although Cu
is becoming more popular for high speed logic), but Al cannot be directly
connected to Si when the junctions are very shallow because spikes of Al
(wormholes) form, and they short the contacts to the bulk. Hence contact
holes are generally filled with another metal, which is then alloyed to the
overlying AI. Filling the contacts can be tricky. A large aspect ratio 20 is

19 In CMP (chemical mechanical polishing), the wafers are polished by rotating buffing wheels
in a chemical slurry.
20 Aspect ratio = interlevel spacing / contact width.
372 Chapter 13

M-5-

Via-4_

M-4_

Via-3_

M-3 -

Via-2_

M-2_

Via-1_

M-1_
Contact_

Figure 13.16
Scanning electron microscope image of stacked vias. Figure adapted from Sun et al. (1998),
Foundry technology for the next decade, International Electron Devices Meeting Technical Digest.
@1998 IEEE.

desirable because it reduces interlayer capacitance while still maintammg

feature density. However, contacts with large aspect ratio are difficult to fill
and deep contact cuts are difficult to etch. For these reasonsW2 1 plugs (where
W is deposited and etched back) or W CMP (where W is deposited and
polished) have been developed (Fig. 13.14). Use ofW forces the use of barrier
metal material22 (such as TiffiN) before W deposition. The conformality of
those barriers inside the contact hole and the method of deposition are other
important issues for reliability.
The contact fill scheme is repeated for the filling of vias connecting
metal layers. Figure 13.15 shows a final cross section after passivation 23 . Fig-

21 W=tungsten.
22 A barrier metal blocks reaction of WftJ (the gas used to carry W) with silicon. This reaction
would have bad effect of forming pipes (also named wormholes) into the silicon substrate.
23 Passivation means covering the wafer with a tough, moisture and sodium ion resistant material
like phosphorous doped Si�. Sodium ions contaminate silicon by making mid-band traps that
greatly decrease minority carrier lifetime and make charge stick in the channel. Only the bonding
pads are opened up to allow connection.
A Millennium Silicon Process Technology 373

ure 13.16 shows the use of stacked vias (Sun et aI., 1998), which are now
commonly available and which greatly simplify logic layout.

13.2 Scaling Limits for Conventional Planar CMOS Architectures

Process scaling is largely driven by the demands for faster and denser logic, not
by demands for more compact, lower power, and more precise analog. Analog
designers are usually compelled to use the logic scaling results whether they
want to or not. It will be clear from the following discussion of scaling how
logic scaling is focused on three objectives: Density, speed, and lower power.
A more general discussion of process scaling that focuses on long-term trends
is given in Chapter 14.

Gate stack Gate stack

-Dual workfunction -Very thin to improve
-Low sheet resistance SeE and current drive
-No Boron penetration -Limitations: Defect density,
-Tight dimensional control tunneling current, reliability

p-

Source/drain p+
Shallow trench isolation -Shallow extension Non-uniform channel
-Litho limited dimensions to reduce SCE -Improve SCE
-Thickness independent of size -Profile optimized -Halo to counter Vi
-Lower capacitance for reliability and rolioff
-No extended thermal performance -Reduce junction
oxidation -Low sheet rho capacitance

Figure 13.17
State of the art CMOS and some of the features to consider for scaling. Figure adapted from Davari
et al. (1995), CMOS scaling for high performance and low power - the next ten years, Proceedings
of the IEEE. © 1995 IEEE.

As gate length scales down and below 0.25 j,tm, short channel effects
determine the scaling limits. These limits including finite gate leakage currents
in the transistor off state and the increased resistance of sources and drains.
All of these limits lead to degraded device performance. Figure 13.17 (Davari
et aI., 1995) shows a cross-section of a state of the art CMOS process and
some of the important features to consider for continual scaling down of
 374 Chapter 13

feature sizes. The details of how scaling is limited by the different technology
parameters are discussed in Chapter 14. In Fig. 13.18 (Thompson, 1999), we

Gate length
( ,
Source/drain Dimension(linear) A. k
extension
(SDE) Potential k k

Impurity Con. klA.2 11k

Electric Field klA. 1

Cox A. k
Source Drain
Metallurgical Current k2/A. k
k2
spacing
Power k3/A.

Power delay
* A.3/k k
k= voltage scaling (-O.7X)
A. = linear dimension scaling (-O.7X)

Figure 13.18
The historical scaling scenario. The table on the right shows how different circuit parameters are
affected when the voltage supply and all dimensions are scaled down by k and by ).. respectively.
The right column of the table shows the scenario for constant field scaling. Figure adapted
from Thompson (1999), Sub 100 nm CMOS: Technology performances, trends and challenges,
International Electron Devices Meeting (lEDM) short course, Washington D.C.@1999 IEEE.

see the historical scaling scenario of transistors. These scenarios have been
used over the years as guidelines for each new process generation. The left
column of the table shows the changes in circuit parameters as the dimensions
are scaled down by A and the voltage supply by k. For example, if the voltage
supply is not scaled down, (k=l), the electric field increases as 1/ A. The right
column shows the constant-field scaling scenario where the voltage supply is
scaled down by the same factor (A = k) as the dimensions.

Threshold Drift

For analog and mixed-signal applications in advanced CMOS technologies, the

main limitation for MOSFET scaling is the threshold shift of the pFET due to
negative bias temperature instability (NBTI) and hot-carrier injection (HCI).
The nitridation of gate oxide which is done to prevent B penetration from the
pFET gate enhances NBTI. The threshold drift due to NBTI and HCI is shown
in Fig. 13.19 in a 0.18 J.1m process (Chaparala et aI., 2(00). A surface channel
pFET was stressed under NBTI conditions (Vg = Vstress and Vd=Vs=Vsub=O)
A Millennium Silicon Process Technology 375

and HCI conditions (Vg = Vd=Vstress and Vs=Vsub=O). The threshold drift is
as much as 80 mV after 500 hours of stress @ 150°C (Chaparala et aI., 2000;
Kimizuka et aI., 20(0). This drift is due to the generation of interface traps and
the formation of fixed positive charges in the oxide under stress conditions.

0.1
Stress: T =150°C, Vg = -4.25V
0.09
0.08 ....... Vd=-4.25V
....... Vd=OV
0.07 -+- Vd=-2.0V
2:
5: 0.06
2l
Qj 0.05
0

�
0.04
0.03
0.02
0.01
1000 10000 100000
Stress time (sec)

Figure 13.19
Threshold voltage drift over stress time during NBTI and Hel stress conditions. The devices were
stressed at 150 0 C. Figure adapted from ehaparala et al. (2000), Threshold voltage drift in pFETs
due to NBTI and Hel, IEEE International Integrated Reliability Workshop,©2000 IEEE.

Gate Current Limitations due to Gate Oxide Scaling

As the MOS feature size is scaled down, the gate oxide thickness scales down
along with the feature size. However, the oxide thickness can only be scaled
down to about 2 nm before the gate leakage current becomes unacceptably
large (around 1 A/cm2 with Vdd=1 V) for device performance. This limit comes
about not because of the feasibility of manufacturing oxide thickness that is
smaller; but rather, it is set by the gate-to-channel tunneling leakage current.
The gate leakage current is due to direct tunneling through the oxide. This
tunneling occurs from the inversion layers and the accumulation layers (SDE
to gate overlap). The increase in gate current density as a function of the gate
voltage for different oxide thickness can be seen in Fig. 13.20 (Lo et aI., 1997;
Thompson, 1999). The oxide thickness limit is considered to be reached when
376 Chapter 13

104
Toxide(A)
103
102
N
E 10'
u
� 10°
�
"(ii
c:
10-'
Q)
"0
10-2
C
� 10-3
:;
10-4
u
2
'"
CJ 10-5
10-6
10-7

Gate voltage (V)

Figure 13.20
Simulated gate current density versus gate voltage from a nFET in inversion. Figure adapted from
Lo et al. (1997). Quantum-mechanical modeling of electron tunneling current from the inversion
layer of ultra-thin-oxide nMOSFET's. IEEE Electron Device Letters. © 1997 IEEE.

the tunneling current is equal to the off-state drain-to-source subthreshold

leakage (around 1 nA/j1m). This limit on the gate oxide thickness sets the limit

poly-Si

Rsificide

Figure 13.21
Components of series resistance at the source/drain of a MOS transistor.

on the channel length of the transistor, because making the channel shorter
without decreasing the gate oxide thickness eventually results in a transistor
whose current is controlled more by the source and drain nodes than by the
gate.
A Millennium Silicon Process Technology 377

Alternatives High dielectric constant materials like Ta205 are currently be­
ing considered as a method of reducing the tunneling current through the insu­
lator. Such materials permit thicker dielectrics to be used for the same inversion
charge. These materials unfortunately have their own problems. They need an
Si02 buffer between the dielectric and the substrate. They also need a metal
gate (TiN/W or TiN/AI) to prevent a reaction between the poly Si gate and the
dielectric.

Source/Drain Junction Scaling

0.8 0.5
IOFF = Constant = 1nAlllm
0.75 0.45

0.7 0.4
E 0.65 0.35 E
::l. ::l.
::( ::(
.s 0.6 0.3 .s
� �
(/) (/)
S' S'
0.55 0.25
PMOS
0.5 0.2

0.45 0.15

0.4 0.1
0 50 100 150 200
Junction depth (nm)

Figure 13.22
IDSAT versus SDE depth for an nFET and pFET. The offset spacer is at 0 nm. Figure adapted
from Thompson et at. (1998), Source/drain extension scaling for 0.1 JLm and below channel length
MOSFETs, Symposium on VLSI Technology: digest of technical papers, © 1998 IEEE.

As gate lengths scale down, the SDE depth and gate overlap are also scaled
down. The SDE depth is scaled down so that the drain charge does not signif­
icantly control the amount of channel charge. For 0.25 J.Lm process technolo­
gies, junction depths are around 50--100 nm. However, these shallow SDEs
gave rise to smaller IDsAT, as was predicted. The shallow SDE has a side effect
in that it leads to a larger series resistance. For a salicided LDD MOSFET, the
series resistance of the source and drain (Rs and Rd respectively) is the sum of
various components as shown in Fig. 13.21 (Asai and Wada, 1997; Thompson,
378 Chapter 13

1999); Rs, Rd = RSE + Rsilicide, where RSE is the sheet resistance due the
shallow SDE. The RSE can be of the same order of magnitude as the channel
resistance Rc especially in a pFET. Figure 13.22 (Taur et aI., 1997; Thompson

0.7 __ ---.--- 1.0X

...._
.. -...--. 0 . 7X
0.6
0.5X (Tox. Vee . LE)
E 0.5
::l.
::c 0.4
.s =
Tox 4.5nm. Vee = 1.8V. LE 0.121lm
=

�
(j) 0.3
-E'
0.2
NMOS data
0.1
0
0 10 20 30 40 50
Overlap (nm)

Figure 13.23
IDSAT versus SDE to gate overlap for an NMOS transistor. Figure adapted from Thompson
et a1. (1998). Source/drain extension scaling for 0.1 tLm and below channel length MOSFETs.
Symposium on VLSI Technology: digest of technical papers. © 1998 IEEE.

et aI., 1998) shows that the maximum IosAT for both nFETs and pFETs at a
constant off-state leakage (I nA/f..lm) is degraded when the junction depth is
below 35-40 nm because of the increased SDE resistance. On the other hand,
IosAT degrades when the junction depth increases above 35-40 nm because
of charge sharing.
The SDE to gate overlap also affects the saturation current through the
transistor. If the overlap is reduced, the current spreads out into a lower doped
region of the SDE. This increases the series resistance and also the accumula­
tion. Figure 13.23 (Taur et aI., 1997; Thompson et aI., 1998) shows how I osAT
degrades when the SDE overlap is less than 20 nm for a transistor in a 0.25 f..lm
technology and for transistors whose gate oxide, power supply, and gate length
are scaled down by both 0.7 and (0.7) 2.

Transistor Off-State Leakage Currents and VddIVT Scaling.

The dependence of V T on the gate length is stronger compared to other factors

that also cause VT fluctuation like random dopant distribution. As gate lengths
A Millennium Silicon Process Technology 379

8 Intel's
technology
7 trend
6
f- 5
>
-U
u
> 4
Gate drive rule
3 VT < (1/4) Vdd

2 Projected
.7X scaling

O L-__L-__�__�__�__�__-L__-L__-L__�____
.07 .10 .15 .22 .35 .50 .70 1.2 1.6 2.1
LGATE (�m)

Figure 13.24
Vdd/VT trend as devices scaled down in Intel's process technology. Figure adapted from Thomp·
son (1999), Sub 100 nm CMOS: Technology performances, trends and challenges, International
Electron Devices Meeting (lEDM) short course, Washington D.C. © 1999 IEEE.

scale below 0.1 f.lm, both the power supply and the threshold voltage are also
scaled down. Reducing the power supply is necessary to decrease the power
consumption. However, this leads to reduced gate overdrive for the same V T.
Hence, threshold voltages also have to scale down. A side effect of this is
that the off-state leakage increases due to the increased subthreshold leakage
currents. A general rule of thumb is to have Vdd about 4 VT, to maintain gate
overdrive at the same off-state leakage. Figure 13.24 (Thompson, 1999) shows
the gate overdrive problems that will be encountered at 0.1 f.lm by using the
historical transistor scaling scenario. At Vdd=1 V, a threshold voltage of 0.25 V
will be required to maintain good performance! Another important factor is the
limitation due to larger VT variations especially at low VT. This limitation is
illustrated in Fig. 13.25 (Sun and Tsui, 1994), which shows a ring oscillator
propagation delay varying wildly when Vdd is scaled down towards VT.

Poly-Gate Depletion

The polysilicon gate can be depleted by the voltage applied to it. The effect of
this depletion is to effectively decrease Cox, lowering the maximum current.
This effect was not a concern in the past because the drain and source junctions
380 Chapter 13

2.0e-9

I o Wafer 11
6. Wafer 12

1.5e-9
� o Wafer 13

I
x Wafer 14

+ Wafer 15

U
Q)
!!!-
,.,
'"
a; 1.0e-9

i
"0
Q)
OJ
l'l
(J)
+
x

x
5.0e-l0 +

�
+

I I I
0.0
2 3 4 5
Supply voltage (V)

Figure 13.25
Ring oscillator frequency variation with supply voltage scaling. Figure adapted from Sun and
Tsui (1994), Limitation of CMOS supply-voltage scaling by MOSFET threshold-voltage variation,
IEEE Custom Integrated Circuits Conference, © 1994 IEEE.

and the polysilicon gate were doped separately, so the gate could be doped
heavily enough to prevent its own depletion. Because the source/drain and
the gate poly are now implanted at the same time, there is a compromise
between poly depletion (tied to poly thickness), depth of the junctions (tied
to the final thermal processing used to activate the dopants) and poly thickness
(to avoid B penetration through the gate oxide). A heavy gate implant and
high temperature is needed to reduce poly-gate depletion, but the B penetration
through the gate oxide is enhanced. With the extremely thin active junctions
now common and the requirement for very small thermal cycling budgets, the
gate can be insufficiently doped to prevent its own depletion. Figure 13.26
shows how gate doping concentration affects transistor operation (Arora et aI.,
1995; Ricco et aI., 1996).
A Millennium Silicon Process Technology 38 1

8
nMOST
tox = 70A
Wm/Lm = 100.35 �m
� 6
E {} __ -O __ �
<n
:2
c'
�
:;
"
c: -- -0 ---0
.�
0

2 3
Drain Voltage, Vds(V)
(a)

8
nMOST
tox = 70A
Wm/Lm = 100.35 �m
6
�
E
<n
:2
'E
� 4
:;
"
c:
0---0----0
.�
0
2

2 3
Drain Voltage, Vds(V)
(b)

Figure 13.26
Poly-gate depletion. The drain current versus drain voltage for 3 different polysilicon gate doping
concentrations (Na=5 x 1019, I X 10 19 and 0.5 x 1019 cm - 3) and three gate to source voltages.
(a) for nMOST (nFET). (b) for pMOST (pFET). The lines are fits from a model and the symbols are
2-D simulated data. Figure adapted from Arora et aI. (1995), Modeling the polysilicon depletion
effect and its impact on submicrometer CMOS circuit performance, IEEE Transactions on Electron
Devices, © 1995 IEEE.

A Steep Subthreshold Slope

A steep subthreshold slope (which is good for analog circuits because it makes
for large gm in subthreshold) will be difficult to maintain with further process
382 Chapter 13

scaling, unless the starting material is switched from bulk Si to silicon on

insulator (SOl). The ever-increasing bulk doping means that K, will tend to
decrease.

1000 r-------�--__,
\
\
� --- --- -- \
Ii> O.BOum - -- \
> - ----\ -_ - Eeff-3
N­ 0.60um _ - -- - -
,,
E 0.35um--""
�
0.25um / \\
--"" - --
-
�
:0 0.1Bum \\
� 100
c
o
Technology 0.13um
I "\ Eeff-2
is node
Q) \
w \
\
\
\
Source: Intel Technology
\
10 �----��-����---�
0.1 10
Eeff (MV/cm)

Figure 13.27
Mobility degradation with channel length scaling. Figure adapted from Thompson (1999). Sub
100 nm CMOS: Technology performances. trends and challenges. International Electron Devices
Meeting (lEDM) short course, Washington D.C. © 1999 IEEE.

Mobility Degradation

The electron mobility degrades with channel length scaling as shown in

Fig. 13.27 (Thompson, 1999). This degradation is the result of impurity scat­
tering due to the extremely highly doped channel region. The degradation can
be reduced by growing undoped thin epitaxial layers on top of the highly doped
substrate.

13.3 Conclusions and Guidelines for New Generations

New devices and process technologies are constantly being developed, some
radically different than the conventional (year 2000) flow described here.
When the time is right for new concepts and architectures, everything is
A Millennium Silicon Process Technology 383

possible. A good cook is innovative by definition! For a new concept to be

successful, it has to be

1. Done yesterday or face the "I need a job!" dilemma (Fig. 13.28).
2. Cheap.
3. Done right and easy to manufacture: Tastes good first time, every time!
4. Done by a good cook (kitchen engineer): Learn how to be one of those, start
with the taste of freedom!

You're gonna ask: what else?! This is the art of cooking. Bon appetit!

r (.jANT �()() TO ITACI\ TO TRULY UNDER5TN'<D

I'\E E'IERYT"!NG YOU OffICE POLITICS YOU
KNOW ABOUT COI\FIJAAIT I'\U5T WEAR A
POLITICS 50 I CAN GET WASTE BASKET ON IT WORKS
TO YOUR YOUR HEAD fOR ONE fOR ME.
LEVEL. fUll. DA'i .
./
\

\1.a..tJ 0..\1 ,� h>..�� +" lo�<!.-

0.// e.:3' "e.u� .s�dt.s

Figure 13.28
Staying sane Adams (l996b,a, 1998b,a). Specified text excerpts from pages 12 and 64 from THE
DILBERT PRINCIPLE by SCOTT ADAMS. Copyright (c) 1996 by United Features Syndicate,
Inc. Reprinted by permission of HarperCollins Publishers, Inc.
 14 Scaling of MOS Technology to Submicrometer Feature
Sizes

It is always difficult to predict the future; few attempts to do so have met with
resounding success. One remarkable example of successful prediction is the
exponential increase in complexity of integrated circuits, first noted by Gor­
don E. Moore!. As we contemplate the ongoing evolution of this great tech­
nology, many questions arise: Can the trend continue? Will single-chip sys­
tems attain levels of complexity that render present system architectures un­
workable (Mead and Conway, 1980)? Will digital techniques completely re­
place analog methods (Mead, 1989)? The answers to these questions depend
critically on the properties of the individual transistors that provide the essen­
tial active functions, without which no interesting system behavior is possible.
Integrated-circuit density is increased by a reduction in the size of elementary
features of the underlying structures. Therefore, any discussion of the capabili­
ties of future technologies must rely on an understanding of how the properties
of transistors evolve as the transistors' dimensions are made smaller2
Elsewhere (Hoeneisen and Mead, 1972b), we described the factors that
limit how small an MOS transistor can be and still operate properly. That
discussion will not be repeated here, but we will outline the major issues:

1. For the device current to be primarily controlled by the gate, the device
should not be punched through; that is, the sum of the source and drain
depletion layers should be less than the geometric channel length. As a direct
consequence of this requirement, the bulk doping must increase as dimensions
are decreased.
2. Increasing the bulk doping has two important consequences: (a) Junction
breakdown voltage is lowered; and (b) a larger electric field is required in the
gate oxide to obtain a given change in surface potential.

Because of 2(a), the operating voltage must be reduced. So that sufficient

electric field can be obtained with a lower operating voltage, the gate oxide
must be made thinner. Thus, it is inevitable that, as the minification process
is continued, both drain depletion layer and gate oxide will become thin
enough that electron tunneling through them will become comparable with

1 According to Moore's Law, the number of transistors on a chip doubles every 18 months.
2 This chapter is slightly modified from a paper by Mead (1994), Scaling of MOS technology to
submicrometer feature sizes, Journal of VLSI Signal Processing. Reprinted with permission from
Kluwer Academic/Plenum Publishers.
386 Chapter 14

other device currents. In 1971, when our original study (Hoeneisen and Mead,
1972b) was written, we described a device of 0.15 micrometers (JLm) channel
length, having a 50 Angstrom (A) gate oxide. Although we were confident that
a device of this size could be made to work, we were not at all sure that smaller
devices could be made viable.
Over the ensuing 22 years, feature sizes have evolved from 6 JLm to 0.6 JLm,
and the trend shows no sign of abating (Nagata, 1992; Davari et al., 1992;
Chang et al., 1992; Bryant et al., 1992; Yan et al., 1992; Yamaguchi et al., 1993;
Iwase et al., 1993). In this chapter, we shall examine what we have learned
from the past 22 years of technology evolution, and shall discuss to what extent
these same trends may continue into the future. We conclude that at least
one more order of magnitude of scaling can be obtained with a concomitant
increase in both density and performance. Several of the conclusions of this
study were reached independently by Hu (1993).

14.1 Scaling Approach

The historic trend of gate-oxide thickness tox as a function of I, the minimum

feature size of the process, is plotted in Fig. 14.1. The trend can be expressed
accurately as
tox = 2101°.77,

where the feature size is in JLm, and the gate-oxide thickness is in A. This
observation suggests that it may be fruitful to express all important process
parameters as powers of the feature size, and to determine whether there is a
scaling of this form that allows sensible process evolution to dimensions well
below 0.1 JLm. To prevent the gate-oxide thickness from becoming thinner than
a single atomic layer, we have chosen a scaling of the form

(14.1.1)

This expression is plotted as the solid line in Fig. 14.1. In reviewing the
historic trend, it is clear that we previously expressed (Hoeneisen and Mead,
1972b) more concern with gate-oxide tunneling than has been justified by the
experience accumulated through the intervening years. It is conceivable that
the same bit of paranoia occurs here. In any case, if oxide thickness continues
to decrease at the present rate, the resulting devices will be somewhat more
capable than those we present.
Scaling of MOS Technology to Submicrometer Feature Sizes 387

• 0.77

E
e
1ii •
Cl
c
::s ...
•
(/)
(/) 102
Q) •
c
�
u
£ •
Q)
"0
'x
0

0.55

Feature size (11)

Figure 14.1
Gate-oxide thickness as a function of feature size. The solid circles are production processes in
silicon-gate technology, starting in 1970. Triangles are processes reported in the literature. Solid
squares are the two most advanced devices described in our previous study (Hoeneisen and Mead,
1972b). The solid line is the analytic expression used in this chapter (Eq. 14.1.1).

The oxide thickness and feature size together determine the gate-oxide
capacitance C9 of a minimum-sized device:
[2
Cg = €ox-'
tox
The historic trend in supply voltage V is shown in Fig. 14.2. This trend
is not as smooth as the trend in oxide thickness, due to the long period of
standardization at 5 volts (V). It is clear, however, that modem submicrometer
devices operate better on lower voltages (Bryant et aI., 1992; Lyon, 1993), and
that this trend to lower voltages must continue. The scaling used here is

(14.1.2)

This expression is plotted as the solid line in Fig. 14.2.

Once the gate-oxide capacitance and supply voltage are known, the energy
Wg stored on the gate of a minimum-sized transistor at any given feature size
388 Chapter 14

102

•
•

• • • • • • •
•
•

0.75

10 1 +----+--+-+-�++�----���---+��
10-2 10-1 100 101
Feature size (!l)

Figure 14.2
Power-supply voltage as a function of feature size. The solid line is the analytic expression used
in this chapter (Eq. 14.1.2).

can be estimated. The stored energy is slightly overestimated as

1
Wg = 2CgV 2 . (14.1.3)

For the scaling laws given here, the stored energy (in Joules) is

Wg = 2.2 X 10-14[2.75. (14.1.4)

This expression is plotted as the long solid line in Fig. 14.3. Even with the
slight "kink" introduced by Eq. 14.1.1, this expression is a good abstraction
of the actual energy over the entire range of the plot. In the central section of
historic data, however, the constant 5 V power-supply voltage has established
a trend with much less dependence on feature size.
This shorter trend is well-represented by the expression

W5 = 2 X 10-14[1.22.

Also shown for reference on Fig. 14.3 is the thermal energy kT, and the
Scaling of MOS Technology to Submicrometer Feature Sizes 389

spacing of levels in the channel with momenta in the direction of current flow.
It is clear that the stored energy is more than 10 kT even at feature sizes of
O.oI f.lm.

10-11
•

10-12 •

10-13

10-14

10-15
s:
>-
e> 10-16
Q)
c
W
10-17

10-18

10-19
llE

10-20 kT

10-21
10-2 10-1 100 101
Feature size (11.)

Figure 14.3
Energy stored on the gate of a minimum-sized transistor as a function of feature size. We compute
the points from Eq. 14.1.3 using oxide thickness values from Fig. 14.1 and the supply-voltage
values from Fig. 14.2. The solid line is the analytic expression used in this chapter (Eq. 14.1.4).
Also shown for reference are the thermal energy kT at room temperature, and the quantum-level
spacing for electrons in the channel with momenta in the direction of current flow.

The minimum stored energy is an interesting quantity because it sets the

scale for the switching energy dissipated in a digital system. The energy per
operation of compute-intensive digital chips is compared with the minimum
stored energy in Fig. 14.4. The system energy per operation is 4 to 6 orders of
magnitude higher than the minimum stored energy, and can be bounded by the
two solid trend lines
Wmax = 1. 15 X 1O-8Z3.4 (14.1.5)

Wmin = 2.5 X 10-10 Z3.25. (14.1.6)

390 Chapter 14

The overall system trend is steeper than that for minimum stored energy,
presumably because designers have become more skilled over the years, and
processes have an ever increasing set of features on which designers can draw
(multiple levels of metal, for example). A 5 V sub-trend is clearly discernible
in the system data as well.

10-3

10-6

10-9

�
OJ
10-12
c
W

10-18

10-21 +---�--�����----�--�+-�-+-�
10-2 10-1 100 101
Feature size (Jl)

Figure 14.4
Energy dissipated per operation at the chip level. Filled triangles are data taken from the literature
and from manufacturers' data sheets. Examples are all compute-intensive single chips, such as
multipliers, digital signal processors, and similar devices. So that the data could be plotted on a
single scale, all values were normalized to 8 x 8 multiply-add operations, assuming that the energy
is proportional to the product of the word lengths of the multiplicand and multiplier. Minimum and
maximum trend lines shown are Eqs. 14.1.5 and 14.1.6. Also shown for reference are the data of
Fig. 14.3.

With the information on hand, we can determine the tunneling current

density Jox through the gate oxide (Lenzlinger and Snow, 1969; Hori et aI.,
1992; Sum! et aI., 1992), making the worst-case assumption that the entire
supply voltage appears across the entire gate area:

(14.1.7)
Scaling of MOS Technology to Submicrometer Feature Sizes 39 1

where Jo = 6.5 X 1010 A/(V /cm)2 was adjusted to match experimental data,
as shown in Fig. 14.5. The imaginary part of the wave vector k is given by

2ko </>
k = ""3
V
[ 1- ( 1- mm 1, . ( ¢))V 3/2
] .

101

100 •

10-1

10-2

E 10-3 ...
u
r:r
� 10-4 ...

�10-5 0 ...

0
10-6 ...

0
10-7 ...

0
10-6 ...

0
10-9
0.6 0.7 0.8 0.9 1.0 1.1 1.2 1.3 1.4
Eox (V/cm) (107)

Figure 14.5
Oxide tunneling current as a function of electric field. The open circles are from the original work
of Lenzlinger and Snow (1969). Filled circles are from the recent work of Sum! et al. (1992).
Filled triangles are from Hori et al. (1992). The solid line is the analytical expression used in
this chapter (Eq. 14.1.7). The filled square is inferred from Iwase et al. (1993), but is not directly
comparable with the other data because it was taken from a transistor drain characteristic, and may
be corrupted with other effects such as gate-enhanced drain tunneling. The gate current was not
reported separately, so this value shown represents a worst-case estimate.

These expressions are valid for voltages both above and below the barrier
potential </>, which was taken to be 3.2 V. The pre-exponential constant k 0 =
1. 2 A-I was used. It is comforting to note that oxide tunneling data are
available over the entire range of electric fields that will be encountered down
to the smallest dimensions studied here. It will be helpful, however, to have
392 Chapter 14

actual experimental data in the 10 A range. For these extremely thin oxides, it
will be essential to take into account the quantum corrections discussed in SUnt�
et ai. (1992).

1020

E1019
u
u
-1.6
:0
::::J
U
CD
2:
g'1018 •
"Ci
o
""0
Q) •

� •
Ui ...
.0

Jl 1017

101 6 +-----��--+__+��_+�++----�---+--+-��+-�
10-2 10-1 100
Feature size (Il)

Figure 14.6
Substrate doping as a function of feature size. The solid line is the analytical expression used in
this chapter (Eq. 14.1.8). Filled triangles represent processes reported in the literature. The two
solid squares are the two smallest transistor designs shown our earlier work (Hoeneisen and Mead,
1972b).

The other major source of parasitic current is tunneling through the drain
junction. The junction-tunneling current density J j is critically dependent
on the substrate acceptor concentration n, which must be increased to avoid
punch-through as device dimensions are decreased (Chynoweth et aI., 1960;
Logan and Chynoweth, 1963; Krieger, 1966; Fair and Wivell, 1976; Stork and
Isaac, 1983; Hackbarth and Tang, 1988; Reisch, 1990). The scaling law used
here is plotted in Fig. 14.6:

(14.1.8)

Given the doping density n, we can compute the depletion-layer thickness x

Scaling of MOS Technology to Submicrometer Feature Sizes 393

for any potential1/J relative to substrate using the usual step-junction approxi­
mation

x
=
J 2€si1/J
qn
•
(14.1. 9)

The corresponding depletion-layer capacitance C is given by

€si.
C=
X

We can determine the maximum electric field in the drain junction, from the
junction voltage, which in the worst case will be the supply voltage plus the
built-in voltage:

2qn(V + Vb)
€si

We could alternatively use a graded-junction approximation, such as that

used by Fair and Wivell (1976). For our purposes, the two approaches are
nearly equivalent, so the simpler step-junction expression with the junction
built-in voltage Vb = 1.1 V is used. In either case, the tunneling current density
is a function of the maximum electric field:

(14.1. 10)

The constant Eo = 2.9 X107 V fern was taken from Fair and Wivell (1976),
and the pre-exponential factor Go = 3 X 109 AfVem2 was chosen to fit the
experimental data plotted in Fig. 14.7. It is significant that experimental data
exist that allow us to predict the tunneling currents in junctions of devices down
to 0. 03 j.tm feature sizes. Previously (Hoeneisen and Mead, 1972b), we pointed
out that the "drain corner" tunneling occurs at lower voltage than that across the
junction area. a fact that has received considerable attention (Li et al. , 1988).
For the present study, we use Eq. 14.1. 10 for area tunneling, both for simplicity
and because considerable cleverness on the part of process designers as this
phenomenon becomes limiting can be expected. Caution, however, that corner
effects may significantly increase the drain tunneling over the values shown in
the following figures.
394 Chapter 14

102
•

1
E 10
u
•
CJ
(J)
�
-',- 100 •

10-1 •

10-2 +---L---�----�--+--��-
1.2 1.4 1.6 1.8 2.0 2.2 2.4
Ej (V/em) (106)

Figure 14.7
Junction·tunneling current density as a function of peak electric field in the junction. The filled
triangles are from alloy tunnel diodes, which were reported as step junctions by Chynoweth et al.
(1960). The filled circles are from diffused emitter-base junctions reported as graded junctions
by Fair and Wivell (1976). These were the only references that we were able to locate for electric
fields in the range encountered in the finest feature sizes considered in this chapter. Some data are
shown by Reisch (1990), but not enough information is given to allow direct comparison with the
other data. For reference, the solid square represents the parameters encountered in the 0.03 JLm
device described in this chapter. The solid line is the analytical expression used here (Eq. 14.1.10).

14.2 Threshold Scaling

To determine the detailed properties of small devices, we must take into

account the short-channel properties, most notable of which are carrier­
velocity saturation and drain-induced barrier lowering (the precursor to punch­
through). Previously (Mead, 1989), we developed a model that gives closed­
form expressions for the current in short-channel devices, including the effects
of velocity saturation . To apply the model, we need some abstraction of the
vertical doping profile under the gate. The most widely used such abstrac­
tion is the threshold voltage VT. We therefore proceed by choosing a nominal
threshold voltage of the form
Scaling of MOS Technology to Submicrometer Feature Sizes 395

VT = 0.551°.23• (14.2. 1)

The actual threshold voltage will be lower than the nominal one by the amount
of drain-induced barrier lowering (DIBL) (Troutman, 1979; Bakker, 1991;
Deen and Yan, 1992; Van der Tol and Chamberlain, 1993). The expression

(Xs)
given by Fjeldly and Shur (1993 ) is

Xc .
smh -
DIBL = V
>.
cosh
(l-Xd) (Xs)
--
>.

- cosh -:f
(14.2. 2)

>.

where Xs Xd
and are the classical depletion-layer thicknesses of the source

to computeXc,
and drain junctions. We have used a surface potential of 0.5 V in Eq. 14. 1. 9
the thickness of the depletion layer under the channel. The
distance scale>. is given by

where the depletion-layer capacitance per unit area C c from channel to sub­

c- Xfsci
strate is

and the oxide capacitance per unit area C ox from gate to channel is

fox
Cox =-
tox
.

The nominal threshold voltage; the actual threshold voltage, including DIBL;
and the supply voltage are plotted as a function of feature size in Fig. 14. 8. For
the scaling parameters used here, DIBL does not become a serious problem
until feature sizes are less than 0. 03 /Lm.

14.3 Device Characteristics

Threshold is defined as the gate voltage at which mobile charge Q at the

source end of the channel changes the surface potential by kT/ q (Mead, 1989).
s
The channel charge at threshold is

Qt =-
kT
q
(Cox + C ) . c (14. 3 . 1)
396 Chapter 14

101

100

Q)
0>
i9
g

10-1 DIBL

10-2 �----��-----+--�--+---��
10-2 10-1 100
Feature size (�)

Figure 14.8
Threshold voltage. The middle curve is the nominal threshold voltage, given by Eq. 14.2.1. The
bottom curve is the actual threshold voltage, which is lowered from the nominal value by drain­
induced barrier lowering (DIBL), given by Eq. 14.2.2. The top curve is the nominal supply voltage
from Eq. 14.1.2.

For higher gate voltages, essentially all charge on the gate attracts equal and
opposite counter-charge of mobile carriers in the channel. Thus, we can form
an excellent estimate of the channel charge Q s at the source end of the channel:

Qs = Cox (V - VT). (14. 3 . 2)

For gate voltages below VT, channel current decreases exponentially with
decreasing gate voltage. At zero gate voltage, the channel charge is

(14. 3 . 3 )

where
Cox
K,- ----
- Cc + Cox '
Given Qt and Qs, we can compute the saturated channel current for a
minimum-sized transistor of any given channel length using Eq. B. 28 from Mead
Scaling of MOS Technology to Submicrometer Feature Sizes 397

(1989):

I 'M � Q, VQ +Q,VQ C: ) (
+ I 1- 1+ ��;: C: f ) + I (14.3.4)

where vo, the saturated velocity of electrons in silicon, is taken to be

107 cm/s (Noor Mohammad, 1992), and to = D/vo can be thought of as
the mean free path of the carrier, which is taken to be 0. 007 /-tm (Mead, 1989).

10-3

10-4

10-5

10-6
t= ------
=� -- ___
__ �T�h�r e=s�
hOld

� 10-7
C
�
:; 10-8
()

10-9
Off
10-10
Tunneling
10-11 Gate Drain
10- 12+---L-r--�-�+-���+----+--�-r-��+-�
10-2 10-1 100
Feature size (11)
Figure 14.9

threshold current It by substituting Qs =

Currents characteristic of minimum-sized devices as a function of feature size. We obtain the
Q t from Eq. 14.3.1 into Eq. 14.3.4. We obtain the on
current Ion by substituting Qs from Eq. 14.3.2 into Eq. 14.3.4, using the threshold voltage lowered
only by the built-in junction voltage, rather than by the total junction voltage. We obtain the off
current IoJ J by substituting Qs from Eq. 14.3.3 into Eq. 14.3.4, using the threshold voltage as
lowered by the full supply voltage. The junction tunneling current was computed from Eq. 14.1.10,
assuming the drain area is the square of the feature size. The gate-oxide tunneling current was
computed from Eq. 14.1.7, assuming that the full supply voltage is present across the full gate area
(the square of the feature size).

We obtain the threshold current It by substituting Qs = Qt from

Eq. 14.3.1 into Eq. 14.3.4. We obtain the on current Ion by substituting Qs
398 Chapter 14

from Eq. 14. 3.2 into Eq. 14. 3 . 4, using the threshold voltage lowered only by
the built-in junction voltage, rather than by the total junction voltage. We ob­
tain the off current Ioff by substituting Q 8 from Eq. 14. 3 . 3 into Eq. 14. 3 . 4,
using the threshold voltage as lowered by DIBL. These expressions thus repre­
sent a conservative characterization of the transistor performance, because the
on current will be somewhat underestimated.
The several currents associated with a minimum-sized transistor are shown
as a function of feature size in Fig. 14.9. The tradeoffs mentioned in the
introduction are immediately apparent in this plot. As features become smaller,
substrate doping must increase to prevent punch-through. The increase in
substrate doping increases the junction electric field, thereby increasing drain­
junction tunneling current into the substrate. To limit the tunneling current to
a reasonable value, we reduce the supply voltage, thereby reducing the ratio of
channel on current to channel off current. The most remarkable conclusion
from Fig. 14. 9 is that transistors of 0. 03 /-Lm channel length still function
essentially as do present-day devices. With proper scaling of all parameters
of the process, device miniaturization is alive and well. Many issues will arise
in the development of ever finer-scale fabrication, but, in the end, the endeavor
will prevail.
Given that devices at least 1 order of magnitude smaller than today's are
feasible, we may enquire what their characteristics may be. Figure 14.10 shows
several quantities of interest. It is clear that discreteness of all quantities will
become increasingly important at smaller feature sizes-particularly that of
doping ions in the substrate. We have given elsewhere a simple discussion of
the effects of discrete substrate charge (Hoeneisen and Mead, 1972b); a recent
analysis is presented by Nishinohara et al. (1992).
Perhaps the single most important aspect of device performance is the
speed of logic fabricated from any particular technology. We can estimate the
time T required for an elementary logic element to drive another like it:

VCtot
T =--­ (14. 3 . 5 )
Ion
where the total capacitance C tot is taken to be three times the sum of the
oxide and drain junction capacitances. This delay should correspond rather
directly to the delay per stage measured for ring oscillators in any given
process, and is plotted along with several experimental points in Fig. 14.11. It is
remarkable that, despite the reduction in supply voltage at small feature sizes,
logic performance continues to improve. Several authors have emphasized the
Scaling of MOS Technology to Submicrometer Feature Sizes 399

Quantum levels
Electrons

103

Voltage levels above thermal noise

100+--L--�----�����--+-��
10-2 10-1 10°
Feature size (Il)
Figure 14.10
Number of signal levels resolvable by a minimum·sized device according to the scaling laws
used in this chapter. Thermal noise limits the analog depth representable by a single voltage. The
number of voltage levels above thermal noise was taken to be the square root of the minimum
stored energy shown in Fig. 14.3, expressed in units of kT. The quantum-level separation was
taken to be the energy spacing of states in a one-dimensional box of length I - J;; - Xd. The
number of electrons under the gate was taken to be the on-value of Q. multiplied by the gate area
(a slight overestimate). The number of depletion ions was taken to be the doping density n given
by Eq. 14.1.8, multiplied by the gate area and the depletion depth x from Eq. 14.1.9, using I V for
'IjJ. As the number of depletion ions becomes smaller, the range of threshold voltages encountered
across a single chip increases. In analog systems, adaptation techniques can mitigate or eliminate
the variation among transistors.

improvement in speed that we can make available by reducing threshold and

power-supply voltages (Burr and Peterson, 1991; Liu and Svensson, 1993;
Murphy, 1993 ; Lyon, 1993).
The primary effect behind this somewhat counter-intuitive trend is velocity
saturation, an excellent recent account of which can be found in Noor Moham­
mad (1992). We gave an early treatment of the effect of velocity saturation
on device characteristics (Hoeneisen and Mead, 1972a); an extended analysis
appears in Appendix B of a previous work (Mead, 1989).
The supply voltage V affects the performance of standard CMOS digital
logic in three ways: The channel charge is proportional to V - VT; the electric
400 Chapter 14

field in the channel is proportional to V; and the logic swing is proportional to

V. For long-channel devices, the carrier velocity is proportional to the electric
field in the channel. The channel current is the product of the channel charge
and the carrier velocity. Therefore, the device current has a quadratic depen­
dence on the supply voltage. This current must charge the load capacitance to
approximately one-half of the supply voltage to achieve a logic transition. This
factor cancels one of the V terms in the current, leaving the circuit speed linear
in the supply voltage.

10-9

Minimum inverter

10-10

10-11

10-12 +------<>---+--+--+-+-++--+-+-<-+-1
10-2 10-1 100
Feature size (J.!)
Figure 14.11
Delay of minimally loaded inverter as a function of feature size. Filled triangles are experimen­
tal results from ring oscillators reported in the literature. Solid line is the expression given in
Eq. 14.3.5.

Once the carrier velocity is saturated, however, increasing the electric field
in the channel no longer increases the channel current. Both the charge in
transit and the voltage to be traversed by the output are increased by the same
factor. In this regime, the only effect of increased supply voltage is an increase
in the switching energy, with virtually no increase in performance. Just how
close devices of the present day come to this limit can be seen in the delay-
Scaling of MOS Technology to Submicrometer Feature Sizes 40 1

versus-voltage plots in the recent literature; see, for example, (Hori et aI. , 1992;
Chang et aI. , 1992; Iwase et aI. , 1993).
Because we have at our disposal the currents associated with all termi­
nals of the transistor, we can evaluate the conductances associated with these
currents. For logic devices to function properly, it is necessary that an elemen­
tary logic circuit have a gain greater than unity, which in tum requires that the
transconductance gm of the transistor be larger than the sum of all contribu­
tions to the drain conductance. As feature size decreases below 0.1 f.lm, both
DIBL and drain-junction tunneling make rapidly increasing contributions to
the drain conductance, as can be seen in Fig. 14.12. Despite these parasitic ef­
fects, the device is still capable of providing greater than unity gain down to
the smallest feature sizes investigated.

10-3

10-4

�OJ
U
� 10-5
u
::J
"0
c
o
()

10-6

10- 7 +_------�--���+-+-��+_--�+-+-��
10-2 10-1 100
Feature size (�)

Figure 14.12
Several conductances associated with minimum-sized transistors, as a function of feature size.
The top curve is the transconductance. The filled triangles are experimental values given in the
literature, normalized to a minimum-sized device at the reported dimension. The second curve is
the drain conductance due to DIBL, computed by evaluating Eq. 14.3.4 at a drain voltage equal to
V and at 0.9 V, and dividing the difference by 0.1 V. The current through this conductance flows
from drain to source. The bottom curve is the drain conductance due to drain-junction tunneling.
Current through this conductance flows from drain to substrate.
402 Chapter 14

14.4 System Properties

The enormous effect of device scaling on computational capability becomes

apparent only when viewed from the system level. We can estimate the system­
level capabilities of digital chips fabricated with advanced processes by extrap­
olation from present-day systems. The first such extrapolation is the number
of devices per unit area. If every transistor in a modem digital chip were to
be shrunk to minimum size, the entire active area would cover approximately
2 % of the chip area. If we assume that this coverage factor can be maintained
in future designs, the density of active elements scales with feature size, as
shown in Fig. 14.13. The system-clock period in today's processors is approx­
imately 100 T. Even today, it is becoming more economical to break each chip
into several processors that can operate in parallel, than it is to merely build
larger "dinosaur" processors. For purposes of extrapolation, we can assume
that each processor contains 106 transistors. The computation available un­
der these clearly oversimplified assumptions is plotted versus feature size in
Fig. 14.14. If we further assume that all devices are in fact of minimum size,
and that they are clocked at the system-clock frequency, we can estimate the
power that will be dissipated by chips built in these advanced technologies.
The power attributable to useful switching, and the dissipations of various par­
asitic currents that do not depend on clock speed, are shown in Fig. 14.15.
Down to about 0.03 /Lm feature size, most of the energy supplied to the chip
is dissipated in real, useful computation. Only below this scale do the parasitic
currents overwhelm the energy consumed in performing real computation.

14.5 Conclusions

The MOS transistor has become the workhorse of modem microelectronics;

it has survived many generations of process scaling to finer feature sizes. In
this chapter, we have explored the extent to which the MOS device, as we
know it today, can be scaled to still smaller dimensions. There is no data
available that provides experimental support for the tunneling currents that
will be encountered in the heavily doped source and drain junctions of devices
down to 0.03 /Lm. There is no comparable data to support the theory for
oxides in the loA range, nor is there direct experimental verification of the
effect of statistical fluctuations on very small structures built in heavily doped
material. As such data become available, we will be better able to chart the
Scaling of MOS Technology to Submicrometer Feature Sizes 403

1011

1010

107

10 6 +-----�----�-+��_+��--�--�+-+-��
10-2 10-1 100
Feature size (�)

Figure 14.13
Assumed number of active devices per square centimeter of chip area. If all devices are of
minimum size, active (transistor channel) area is 2 % of total area.

course of future minification, of which the present study is only an outline.

It is already clear that MOS circuits will be integrated to upward of 10 9
devices per square centimeter merely by scaling, without any major change
in the conceptual framework that we use today. There are many challenges
involved in this technology evolution (Nagata, 1992), but show-stoppers are
not expected. The prospect of very high levels of integration was daunting in
1971 when our earlier study was written (Hoeneisen and Mead, 1972b), and is
far more daunting today. Whereas massive parallelism is possible in present­
day technology, it will clearly become mandatory if we are to realize even
a fraction of the potential of more highly evolved technology. Even as this
chapter was written, there was far more potential in a square centimeter of
silicon than we have developed the paradigms to use, as has often been the
case in periods of rapid technological evolution.
We should clarify the "limits" considered here. It is clear that devices much
smaller than those treated here can be made to show useful characteristics.
Conventional MOS devices can be fabricated on insulating substrates (SOI-
404 Chapter 14

1024

1021

1018
E
u
cr
�
&l1015
�Cl.
Q.
1 012

109
Clock freq felk

10 6 +-------�--�--�+_+_��+_--+--�+-�+-�
10-2 10-1 100
Feature size (�.)
Figure 14.14
Several measures of computation capability per unit area as a function of feature size. The bottom
curve is a typical processor clock frequency, the clock period assumed to be 100 times the inverter
delay shown in Fig. 14.11. The second curve is the number of systems (of 1 (j3 transistors each) per
square centimeter multiplied by the clock frequency. The third curve is the number of transistors
per square centimeter shown in Fig. 14.13 multiplied by the clock frequency. The top curve is
the number of transistors per square centimeter multiplied by the reciprocal of the inverter delay
shown in Fig. 14.11.

SOS), thereby removing the constraint imposed by substrate tunneling. Much

smaller devices are possible at molecular scale. The most obvious example
of an extremely small device is an electron-transfer reaction occurring along
an amino-acid path, the potential of which is determined by the charge on a
nearby atomic site. Such arrangements are thought to occur in many biological
systems. The physics of such a transfer corresponds directly to that of an MOS
transistor operating in weak inversion (below threshold). Imagining a device
that functions is easy; building a device that works is much harder; and having
a process by which billions of devices can be constructed in a single physical
structure is many orders of magnitude harder still. This study is limited to the
consideration of direct extensions to existing technology.
Finally, we emphasize that only the properties of transistors themselves
have been considered, and we have not even touched many other important
Scaling of MOS Technology to Submicrometer Feature Sizes 405

1 3
0

1 2
0
Switching

1 1
0
E
u
C"
.!!! Off current
� lO-{)
iii
� Tunneling
1 -1
0

1 -2
Gate
0

1 � +------L+---�--+-��+-�+---+-��-���
0
1 -2 1 -1 1 0
0 0 0
Feature size (�)
I<'igure 14.15
Several contributions to the power dissipated by typical digital systems as a function of feature
size. The curve labeled Switching was obtained by multiplying the number of transistors per unit
area shown in Fig. 14.13 by the switching energy shown in Fig. 14.3 and by the clock frequency
shown in Fig. 14.14. This power contributes to the performance of computation: It scales directly
with clock frequency. In addition to the switching power, there are several parasitic mechanisms
by which power is wasted, each being the result of one of the parasitic currents shown in Fig. 14.9.
These parasitic mechanisms are present even at zero clock frequency, and perform no useful work.
The values shown assume that all devices are of minimum size, and have the full voltage V on
their drains. All values depend critically on the assumptions embodied in the scaling laws of
Eqs. 14.1.1, 14.1.2, 14.1.8, and 14.2.1. Even slightly different scaling can lead to substantially
different results for the smallest feature sizes. The particular laws discussed in this chapter were
fine tuned to produce reasonable results down to 0.02 /.lm. For example, a slight increase in doping
density markedly decreases the off current by reducing DIBL, while dramatically increasing the
drain-junction tunneling current. Similar tradeoffs can be made with other parameters.

aspects of the technology. Of the latter, interconnect (both within a single

chip and across chip boundaries) is obviously a key concern. We have given
elsewhere a preliminary discussion of the global scaling properties of a single­
chip interconnect network for ultra-dense technology (Mead and Conway,
1980). The topic of interconnect, along with many other issues, such as the
fabrication technology itself, deserve a great deal of consideration as the
technology evolves. W hatever complications arise, however, it is clear that
the technology will evolve. It will evolve because that evolution is possible,
406 Chapter 14

because there is so much to be gained at the system level by that evolution,

and because the same energy and will on the part of bright, energetic, devoted
people that has overcome enormous obstacles in the past will overcome those
that lie ahead.
A
Appendix:
Units and symbols

The SI units

Basic units
Quantity Unit Sym.
Length Meter m
Mass Kilogram kg
Time Second s
Therm. temp. Kelvin K
Electrical current Ampere A
Luminous intensity Candela cd
Amount of subst. Mol mol

Derived units with special names

Quantity Unit Sym. Derivation
Frequency Hertz Hz S -1
Force Newton N kg m S-2· ·

Pressure Pascal Pa N m 2
·
-

Energy Joule J N m '

Power Watt W J. S-1

Charge Coulomb C A s ·

Electrical Potential Volt V W A 1 ·

-

Electrical Capacitance Farad F C V 1

·
-

Electrical Resistance Ohm n V A 1

·
-

Electrical Conductance Siemens S A V 1

·
-

Magnetic flux Weber Wb V · s

Magnetic flux density Tesla T Wb m 2 ·

-

Inductance Henry H Wb A 1 ·
-

Luminous flux Lumen 1m cd sr ·

Illuminance Lux Ix 1m m 2 ·
-
408 Appendix A

Physical Constants

Name Symbol Value Unit

Angstrom Unit A 10-10 m
Number 7r 7r 3.141592653589793238
Number e e 2.718281828459
Speed of light in vacuum c 2. 9 9792458. 10 8 mls (def)
Permittivity of the vacuum co 8. 8541 87. 10-12 F/m
Permeability of the vacuum J.to 47r·1O-7 HIm

Planck's constant h 6.6260755. 10-34 Js

Dirac's constant n = h/27r 1.0 545727. 10-34 Js
Molar gasconstant R 8.31441 J/mol
Avogadro's constant NA 6.0221 367. 1023 mol-1
Boltzmann's constant k 1. 3806 58. 10-23 JIK
Elementary charge q 1.60217733 . 10-19 C
Electron mass me 9.10 9 3897. 10-31 kg
Electron volt eV 1.6021 8·10-19 J eV
Units and symbols 409

Properties of Si

Name Symbol Value Unit

Breakdown field EB 3 . 105 V/cm
Diffusion coefficient:
Electron Dn 34.91 cm2/s
Hole Dp 12.41 cm2/s
Effective mass:
Electron m�/mo 1.18
Hole m�/mo 0.81
Electron affinity qX 4.05 eV
Energy gap Eg 1.124 eV
Intrinsic carrier concentration ni 1. 45 . 1010 cm-3
Intrinsic resistivity p 3.16 . 10 5 O-cm
Minority carrier lifetime T 2. 5 . 10-3 s
Mobility:
Electron JLn 1350 cm2N-s
Hole JLp 480 cm2N-s
Relative permittivity €r 11.9
Effective density of states:
Conduction band Nc 2. 8 . 1019 cm-3
Valence band Nv 1.0 4 . 1019 cm-3
410 Appendix A

List of Symbols

Symbol Description Unit

a Lattice Constant A
c Speed of light in vacuum rn/s
C Capacitance F
D Diffusion coefficient m2/s
E Energy eV
Ec Bottom of conduction band eV
Ep Fermi level energy eV
Ev Top of valence band eV
£ Electric Field Vim
£c Critical Field Vim
£m Maximum Field Vim
f Frequency Hz
F(E) Fermi-Dirac distribution function
I Current A
J Current density Nm2
L Transistor's length m
mo Electron rest mass kg
m* Effective mass kg
n Density of free electrons m-3
ni Intrinsic density m-3
N Doping concentration m-3
NA Acceptor impurity density m-3
Nc Effective density of states in conduction band m-3
ND Donor impurity density m-3
Nv Effective density of states in valence band m-3
p Density of free holes m-3
q Magnitude of electronic charge C
R Resistance n
T Absolute temperature K
v Carrier velocity rn/s
Vs Saturation velocity rn/s
Vth Thermal velocity rn/s
continued on next page
Units and symbols 411

Symbol Description Unit

Vbi Built-in potential V
VEB Emitter-base voltage V
VB Breakdown voltage V
W Transistor width m
Wb Base thickness m
€o Permettivity in vacuum F/m
€s Semiconductor permettivity F/m
€i Insulator permettivity F/m
€s/€o Dielectric constant
€i/€O Dielectric constant
f..to Permeability in vacuum Him
f..tn Electron mobility m2N-s
f..tp Hole mobility m2N-s
p Resistivity O-m
UT Thermal voltage V
w Angular frequency (27rj) Hz

Prefixes

Name Symbol Value Name Symbol Value

exa E 1018 deci d 10-1
peta P 1015 centi c 10-2
tera T 1012 milli m 10-3
giga G 109 micro f..t 10-6
mega M 106 nano n 10-9
kilo k 103 pico P 10-12
hecto h 102 femto f 10-15
deca da 10 atto a 10-18
412 Appendix A

The � (nabla) symbol

The symbol � represents a vector differential operator:

t'7 a ... a ... a ...
x+ y+ z
e e aze
v
ax ay
=

where ex, ey, and ez are the base vectors of unit length directed along the
positive directions of the x, y, and z axes respectively.

The symbol �2 is called the Laplacian operator:

a2f a2f a2f
�2f
ax2 + ay2 + az2
=

If fO =
f(x, y, z) represents a scalar field with continuous first partial
derivatives, then

t'7 af ... af ... af ...

f gradf x y z
axe + ay e + az e
v = =

is called the gradient of the scalar function f. The gradient operator defines a
vector.

If v( x, y, z) is a differentiable vector function, with components v 1, V 2,

and v 3 , then
t'7 ... d' ... aVl aV aV3
v V lV V - + -2 + -
ax ay
• = =

az
is called the divergence of the vector field V. The divergence operator defines
a scalar.

If v( x, y, z) is a differentiable vector function defined as

then

t'7 ...
v xv
'"
curI v
(-
aaz aay ) ... ( aax aaz ) ... ( aay aax ) ...
- - ex+ - - - ey+ - - - ez
ay az az ax ax ay
= =

is called the curl of the vector function V. The curl operator curl V, also denoted
by rot V, defines a vector.
Units and symbols 413

Basic relations for the mathematical operators

If f and 9 are two scalar functions, and v and u are two vector functions, the
following relations hold:

V( fg) = fVg + gVf, V( f/g) = ( 1/g2)( gVf- fVg)

V2f = div( Vf), V2( fg) = gV2f + 2Vf· Vg + fV2g
div( fV) = fdiv v + v· V , div( fVg) = fV2g + Vf· Vg
curl( fV) = Vf x v + fcurl ( V), div( u x V) = v· curl- u· curl v
curl( Vf) = 0, div(curl v) = 0
References

Adams, R. W. (1979). Filtering in the log domain, Preprint 1470. In Audio Engineering Society
Convention 63.
Adams, S. (1996a). The Dilbert Principle. Harper Business, New York.
Adams, S. (1996b). Fugitivefrom the cubicle police. Andrews and McMeel, Kansas City, MO.
Adams, S. (1998a). The Dilbert Future, Thriving on stupidity in the 21st century. Harper
Business, New York.
Adams, S. (1998b). The Joy of Work. Harper Business, New York.
Allen, P. E. and Holberg, D. R. (2002). CMOS Analog Circuit Design. Oxford University Press,
2nd edition.
Aispector, J. and Allen, R. B. (1987). A neuromorphic VLSI learning system. In Losleben, P.,
editor, Proceedings of the 1987 Stanford Conference on Advanced Research in VLSI, pages
313-349, Cambridge, MA. MIT Press.
Amelio, G. F., Bertram, Jr., W. J., and Tompsett, M. F. (1971). Charge-coupled imaging devices.
IEEE Transactions on Electron Devices, ED-18:986-992.
Andreou, A. G. and Boahen, K. A. (1994). Neuromorphic information processing II. In Ismail,
M. and Fiez, T., editors, Analog VLSI : Signal and Information Processing, chapter 8, pages
358-413. McGraw-Hill, New York.

Andreou, A. G. and Boahen, K. A. (1996). Translinear circuits in subthreshold MOS. Analog

Integrated Circuits and Signal Processing, 9(2):141-166.
Andreou, A. G., Boahen, K. A., Pouliquen, P. O., Pavasovic, A., Jenkins, R. E., and Strohbehn,
K. (1991). Current-mode subthreshold MOS circuits for analog VLSI neural systems. IEEE
Transactions on Neural Networks, 2(2):205-213.
Aritome, S., Shirota, R., Hetnink, G., Endoh, T., and Masuoka, F. (1993). Reliability issues of
flash memory cells. Proceedings of the IEEE, 81(5):776-788.
Arnold, M. G., Bailey, T. A., Cowles, J. R., and Cupal, J. J. (1990). Redundant logaritbtnic
arithmetic. IEEE Transactions on Computers, 39(8):1077-1086.
Atora, N. D., Rios, R., and Huang, c.-L. (1995). Modeling the polysilicon depletion effect and
its impact on submicrometer CMOS circuit performance. IEEE Transactions on Electron
Devices, 42(5):935-943.
Asai, S. and Wada, Y. (1997). Technology challenges for integration neat and below 0.1 11m.
Proceedings of the IEEE, 85(4):505-520.
Ashok, S. (1976a). Integrable sinusoidal frequency doubler. IEEE Journal of Solid-State
Circuits, SC-ll(2):341-343.
Ashok, S. (1976b). Translinear root-difference-of-squares circuit. Electronics Letters,
12(8):194-195.
Bakker, J. G. C. (1991). Simple analytical expressions for the fringing field and
fringing-field-induced transfer time in charge-coupled devices. IEEE Transactions on Electron
Devices, 38(5):1152-1161.
Barker, R. W. J. and Hart, B. L. (1974). Root-law circuit using monolithic bipolar-transistor
arrays. Electronics Letters, 10(21):439-440.
Ben-Yishai, R., Lev Bar-Or, R., and Sompolinsky, H. (1995). Theory of orientation tuning in
visual cortex. Proceedings of the National Academy of Sciences of the USA, 92(9):3844-3848.
Bergemont, A., Deleonibus, S., Guegan, G., Guillaumot, B., Laurens, M., and Martin, F. (1989).
A high performance CMOS process for subtnicron 16Mb EPROM. In International Electron
Devices Meeting, pages 591-594.
Bertsekas, D. P. (1982). Constrained optimization and Lagrange multiplier methods. Acadetnic
Press, New York.
Boahen, K. A. (1997). Retinomorphic Vision Systems: Reverse Engineering the Vertebrate
Retina. Ph.D. thesis, California Institute of Technology, Pasadena, CA.
416 References

Boahen, K. A. (1998). Communicating neuronal ensembles between neuromorphic chips. In

Lande, T. S., editor, Neuromorphic Systems Engineering, pages 229-259. Kluwer, Boston, MA.
Boahen, K. A. and Andreou, A. G. (1992). A contrast sensitive silicon retina with reciprocal
synapses. In Moody, J. E., Hanson, S. J., and Lippman, R. P., editors, Advances in neural
information processing systems, volume 4, pages 764-772, San Mateo, CA. Morgan Kaufmann.
Boyle, W. S. and Smith, G. E. (1970). Charge coupled semiconductor devices. The Bell System
Technical Journal, 49(4):587-593.
Brajovic, V. and Kanade, T. (1998). Computational sensor for visual tracking with attention.
IEEE Journal of Solid-State Circuits, 33(8):1199-1207.
Briiggemann, H. (1970). Feedback stabilized four-quadrant analog multiplier. IEEE Journal of
Solid-State Circuits, SC-5(4):150-159.
Bryant, A., El-Kareh, B., Furukawa, T., Noble, W. P., Nowak, E. J., Schwittek, W., and Tonti, W.
(1992). A fundamental performance limit of optimized 3.3-V sub-quarter-micrometer fully
overlapped LDD MOSFET's. IEEE Transactions on Electron Devices, 39(5):1208-1215.
Burr, J. B. and Peterson, A. M. (1991). Energy considerations in multichip-module based
multiprocessors. In IEEE International Conference on Computer Design, pages 593-600.
Carlson, A. B. (1986). Communication Systems. McGraw-Hill, New York, 3rd edition.
Cauwenberghs, G. and Bayoumi, M. A., editors (1999). Learning on Silicon: Adaptive VLSI
Neural Systems. Kluwer, Boston, MA.
Chamberlain, S. G. and Lee , J. P. Y. (1983). A novel wide dynamic range silicon photodetector
and linear imaging array. In Proceedings of the IEEE Custom Integrated Circuits Conference,
pages 441-445.
Chang, W.-H., Davari, B., Wordeman, M. R., Taur, Y., Hsu, C. C.-H., and Rodriguez, M. D.
(1992). A high-performance 0.25-Jtm CMOS technology. I. Design and characterization. IEEE
Transactions on Electron Devices, 39(4):959-966.
Chaparala, P., Shibley, J., and Lim, P. (2000). Threshold voltage drift in PMOSFETs due to
NBT I and HCI. In IEEE International Integrated Reliability Workshop.
Chiu, K. Y., Moll, J. L., and Manoliu, J. (1982). A bird's beak free local oxidation technology
feasible for VLSI circuits fabrication. IEEE Transactions on Electron Devices,
ED-29(4):536-540.
Choi, J. and Sheu, B. J. (1993). A high-precision VLSI winner-lake-all circuit for self-orgauizing
neural networks. IEEE Journal of Solid-State Circuits, 28(5):576-584.
Chua. L. 0., Desoer, C. A., and Kuh, E. S. (1987). Linear and nonlinear circuits, pages 696-700.
McGraw-Hill, New York.
Chung, S. S., Kuo, S. N., Yih, C. M., and Chao, T. S. (1997). Performance and reliability
evaluations of p-channel flash memories with different programming schemes. In International
Electron Devices Meeting technical digest, pages 295-298.
Chute, C. (2000). CMOS Versus CCD: Battle for Market Share. IDC, Framingham, MA.
Chynoweth, A. G., Feldmann, W. L., Lee, C. A., Logan, R. A., and Pearson, G. L. (1960).
Internal field emission at narrow silicon and germauium p-n junctions. Physical Review,
118(2):425-434.
Cohen, M. H. and Andreou, A. G. (1992). Current-mode subthreshold MOS implementation of
the Herault-Jutten autoadaptive network. IEEE Journal of Solid-State Circuits, 27(5):714-727.
Davari, B., Chang, W. H., Petrillo, K. E., Wong, C. Y., Moy, D., Taur, Y., Wordeman, M. R., Sun,
J. Y.-c., Hsu, C. C.-H., and Polcari, M. R. (1992). A high-performance 0.25-Jtm CMOS
technology: II. T echnology. IEEE Transactions on Electron Devices, 39(4):967-975.
Davari, B., Dennard, R. H., and Shahidi, G. G. (1995). CMOS scaling for high performance and
low power - the next ten years. Proceedings of the IEEE, 83(4):595-606.
Deen, M. J. and Yan, Z. X. (1992). DIBL in short-channel NMOS devices at 17K. IEEE
References 417

Transactions on Electron Devices, 39(4):908-915.

Delbriick, T. (1989). A chip that focuses an image on itself. In Mead, C. and Ismail, M., editors,
Analog VLSl lmplementation of Neural Systems, pages 171-188. Kluwer, Boston, MA.
Delbriick, T. (1993). Investigations of analog VLSI visual transduction and motion processing.
Ph.D. thesis, California Institute of Technology, Pasadena, CA.
Delbriick, T. and Mead, C. A. (1994). Adaptive photoreceptor with wide dynamic range. In
1994 IEEE International Symposium On Circuits and Systems, volume 4, pages 339-342.
ISCAS '94: London, England, 30 May-2 June.
Delbriick, T. and Mead, C. A. (1995). Analog VLSI phototransduction by continuous-time,
adaptive, logarithmic photoreceptor circuits. In Koch, C. and Li, H., editors, Vision Chips:
Implementing Vision Algorithms with Analog VLSI Circuits, pages 139-161. IEEE Computer
Society Press, Los Alamitos, CA.
Demosthenous, A., Smedley, S., and Taylor, J. (1998). A CMOS analog winner-take-all network
for large-scale applications. IEEE Transactions on Circuits and Systems I: Fundamental Theory
and Applications, 45(3):300-304.
Diorio, C. (1997). Neurally Inspired Silicon Learning: From Synapse Transistors to Learning
Arrays. Ph.D. thesis, California Institute of Technology, Pasadena, CA.
Diorio, C. (2000). A p-channel MOS synapse transistor with self-convergent memory writes.
IEEE Transactions on Electron Devices, 47(2):464-472.
Diorio, C., Hasler, P., Minch, B. A., and Mead, C. (1998a). Floating-gate MOS synapse
transistors. In Lande, T. S., editor, Neuromorphic Systems Engineering: Neural Networks in
Silicon, pages 315-338. Kluwer, Boston, MA.
Diorio, c., Hasler, P., Minch, B. A., and Mead, C. A. (1996). A single-transistor silicon synapse.
IEEE Transactions on Electron Devices, 43(11):1972-1980.
Diorio, C., Hasler, P., Minch, B. A., and Mead, C. A. (1997a). A complementary pair of
four-tenninal silicon synapses. Analog Integrated Circuits and Signal Processing,
13(1-2): 153-166.
Diorio, c., Hasler, P., Minch, B. A., and Mead, C. A. (1997b). A floating-gate MOS learning
array with locally computed weight updates. IEEE Transactions on Electron Devices,
44(12):2281-2289.
Diorio, C. J., Hasler, P. E., Minch, B. A., and Mead, C. A. (1997c). Semiconductor structure for
long term learning. U.S. Patent No. 5,627,392, Issued May 6.
Diorio, C. J., Hasler, P. E., Minch, B. A., and Mead, C. A. (1998b). Three-terminal silicon
synaptic device. U.S. Patent No. 5,825,063, Issued October 20.
Doorenbosch, F. and Goinga, Y. (1976). Integrable, wideband, automatic volume control
(A.V.C.) using Pythagoras's Law for amplitude detection. Electronics Letters, 12(16):41�20.
Drakakis, E. M., Payne, A. J., and Toumazou, C. (1997). Bernoulli operator: A low-level
approach to log-domain processing. Electronics Letters, 33(12):1008-1009.
Drakakis, E. M., Payne, A. J., and Toumazou, C. (1998). Multiple feedback log-domain filters.
In Proceedings of the 1998 IEEE International Symposium on Circuits and Systems, volume I,
pages 317-320. ISCAS '98: Monterey, CA, 31 May-3 June.
Drakakis, E. M., Payne, A. J., and Toumazou, C. (1999). "Log-domain state-space": A
systematic transistor-level approach for log-domain filtering. IEEE Transactions on Circuits and
Systems II: Analog and Digital Signal Processing, 46(3):290-305.
Dunlap, Jr., W. C. (1957). An Introduction to Semiconductors. Wiley, New York.
Edgar, A. D. and Lee, S. C. (1979). FOCUS microcomputer number system. Communications of
the ACM, 22(3):166-177.
Enz, C. and Punzenberger, M. (1999). I-V Log-domain filters. In Huijsing, J., van de Plassche,
Analog Circuit Design: Volt Electronics; Mixed-Mode Systems;
R., and Sansen, W., editors,
Low-Noise and RF Power Amplifiers for Telecommunication, pages 33-67. Kluwer, Boston, MA.
418 References

Enz, C. C. (1989). High Precision CMOS Micropower Amplifiers. Ph.D. thesis, Ecole
polytechnique fMerale de Lausanne, Lausanne, Switzerland. No. 802.
Enz, C. c., Krummenacher, E, and Vittoz, E. A. (1995). An analytical MOS transistor model
valid in all regions of operation and dedicated to low-voltage and low-current applications.
Analog Integrated Circuits and Signal Processing, 8(1):83-114.
Enz, C. C. and Vittoz, E. A. (1997). MOS transistor modeling for low-voltage and low-power
analog IC design. Microelectronic Engineering, 39:59-76.
Etienne-Cummings, R., Van der Spiegel, J., and Mueller, P. (1996). VLSI model of Primate
visual smooth pursuit. In Touretzky, D. S., Mozer, M. c., and Hasselmo, M. E., editors,
Advances in Neural Information Processing Systems, volume 8, pages 706-712, Cambridge, MA.
MIT Press.
Fabre, A. (1984). An integrable multiple output translinear current converter. International
Journal of Electronics, 57(5):713-717.
Fabre, A. (1985). Translinear current conveyors implementation. International Journal of
Electronics, 59(5):619-623.
Fabre, A. (1986). The translinear operational current amplifier: A new building block.
International Journal of Electronics, 60(2):275-279.
Fabre, A. (1988). Translinear Current-Controlled Current Amplifier. Electronics Letters,
24(9):548-549.
Fabre, A. and Rochegude, P. (1987). Current processing circuits with translinear operational
current amplifiers. International Journal of Electronics, 63(1):9-28.
Fair, R. B. and Wivell, H. W. (1976). Zener and avalanche breakdown in As-implanted
low-voltage Si n-p junctions. IEEE Transactions on Electron Devices, ED-23(5):512-518.
Fjeldly, T. A. and Shur, M. (1993). Threshold voltage modeling and the subthreshold regime of
operation of short-channel MOSFET's. IEEE Transactions on Electron Devices, 40(1):137-145.
Fossum, E. R. (1989). Architectures for focal plane image processing. Optical Engineering,
28(8):865-871.
Fossum, E. R. (1993). Active pixel sensors: Are CCDs dinosaurs? In Blouke, M. M., editor,
Charge-Coupled Devices and Solid State Optical Sensors 11/, Proceedings of the SPIE, volume
1900, pages 2-14.
Fossum, E. R. (1997). CMOS image sensors: Electronic camera-on-a-chip. IEEE Transactions
on Electron Devices, 44(10): 1689-1698.
Frey, D. R. (1993). Log-domain filtering: An approach to current-mode filtering. lEE
Proceedings G: Circuits, Devices and Systems, 140(6):406-416.
Frey, D. R. (1996a). Explicit Log Domain Root-Mean-Square Detector. U.S. Patent No.
5,585,757, Issued December 17.
Frey, D. R. (1996b). Exponential state space filters: A generic current-mode design strategy.
IEEE Transactions on Circuits and Systems I: Fundamental Theory and Applications,
43(1):34-42.
Fried, R. and Enz, C. C. (1996). CMOS parametric current amplifier. Electronics Letters,
32(14):1249-1250.
Frohman-Bentchkowsky, D. (1971). Memory behavior in a floating-gate avalanche-injection
MOS (FAMOS) structure. Applied Physics Letters, 18(8):332-334.
Fujita, O. and Amerniya, Y. (1993). A floating-gate analog memory device for neural networks.
IEEE Transactions on Electron Devices, 40(11):2029-2035.
Genin, R. and Konn, R. (1979). Sinusoidal frequency doubler. Electronics Letters, 15(2):47-48.
Gilbert, B. (1968a). A DC-500 MHz Amplifier/Multiplier Principle. In Raper, J. A. A., editor,
1968 International Solid-State Circuits Conference Digest of Technical Papers, volume XI, pages
114-115, New York. L. Winner. Philadelphia, PA, 14-16 February.
References 419

Gilbert, B. (1968b). A new wide-band amplifier technique. IEEE Journal of Solid-State Circuits.
SC-3(4):353-365.
Gilbert, B. (1968c). A precise four-quadrant multiplier with subnanosecond response. IEEE
Journal of Solid-State Circuits. SC-3(4):365-373.
Gilbert, B. (1974). A high-performance monolithic multiplier using active feedback. IEEE
Journal of Solid-State Circuits. SC-9(6):364-373.
Gilbert, B. (1975). Translinear circuits: A proposed classification. Electronics Letters.
11(1):14-16. See also Errata. 11(6):136. March 1975.
Gilbert, B. (1976). High-accuracy vector-difference and vector-sum circuits. Electronics Letters.
12(11):293-294.
Gilbert, B. (1983). A four-quadrant analog divider/multiplier with 0.01 % distortion. In Digest of
Technical Papers of the 1983 IEEE International Solid-State Circuits Conference. pages
248-249. Philadelphia. PA.
Gilbert, B. (1990). Current-mode circuits from a translinear viewpoint: A tutorial. In Tomazou.
c.. Lidgey. F. J and Haigh. D. G editors. Analogue IC design: the current-mode approach.
.• .•

chapter 2. pages 11-91. Peregrinus. Stevenage. Herts UK.

.•

Gilbert, B. (1993). Translinear circuits - 25 years on. Part I: The foundations. Electronic
Engineering. 65(800):21-24.
Gilbert, B. (1996). Translinear circuits: An historical review. Analog Integrated Circuits and
Signal Processing. 9(2):95-118.
Gilbert, B. and Counts. L. W. (1976). A monolithic RMS-DC converter with Crest-Factor
compensation. In Digest of Technical Papers of the 1976 IEEE International Solid-State Circuits
Conference. pages 110-111. Philadelphia. PA.
Gilbert, B. and Holloway. P. (1980). A wideband two-quadrant analog multiplier. In Digest of
Technical Papers of the IEEE International Solid-State Circuits Conference. pages 200-201. San
Francisco. CA.
Godfrey. M. D. (1992). CMOS device modeling for subthreshold circuits. IEEE Transactions on
Circuits and Systems II: Analog and Digital Signal Processing. 39(8):532-539.
Gray. P. R.. Hurst, P. J Lewis. S. H and Meyer. R. G. (2001). Analysis and Design of Analog
.• .•

Integrated Circuits. Wiley. New York. 4th edition.

Gray. P. R. and Meyer. R. G. (1993). Analysis and design of analog integrated circuits. pages
716-717. Wiley. New York. 3rd edition.
Gregorian. R. (1999). Introduction to CMOS OP-AMPs and Comparators. Wiley. New York.
Gregorian. R. and Ternes. G. C. (1986). Analog MOS Integrated Circuits for Signal Processing.
Wiley. New York.
Grossberg. S. (1978). Competition. decision. and consensus. Journal of Mathematical Analysis
and Applications. 66(2):470-493.
Grove. A. S. (1967). Physics and Technology of Semiconductor Devices. Wiley. New York.
Hackbarth. E. and Tang. D. D.-L. (1988). Inherent and stress-induced leakage in heavily doped
silicon junctions. IEEE Transactions on Electron Devices. 35(12):2108-2118.
Hlifliger. P. and Mahowald. M. (1998). Weight vector normalization in an analog VLSI artificial
neuron using a backpropagating action potential. In Neuromorphic Systems: Engineering Silicon
from Neurobiology. chapter 16. pages 191-196. World Scientific.
Hahnloser. R. H. R.. Sarpesbkar. R.. Mahowald. M. A.. Douglas. R. J and Seung. H. S. (2000).
.•

Digital selection and aualogue amplification coexist in a cortex-inspired silicon circuit Nature.
405(6789):947-951.
Hall. E. L.. Lynch. D. D and Dwyer. III. S. J. (1970). Generation of products and quotients
.•

using approximate binary logarithms for digital filtering applications. IEEE Transactions on
Computers. C-19(2):97-105.
420 References

Han. Y. P. and Ma. B. (1984). Isolation process using polysilicon buffer layer for scaled
MOSNLSI. Journal of the Electrochemical Society. 131(3):85C. Abstract 67. Cincinnati. Ohio
meeting. May 6-11.
Hasler. P Andreou. A. G Diorio. C Minch. B. A.. and Mead. C. A. (1998). Impact ionization
.• .• .•

and hot-electron injection derived consistently from Boltzmann transport. VLSI Design.
8(1-4):455-461.
Hasler. P. E. (1997). Foundations of Learning in Analog VLSI. Ph.D. thesis. California Institute
of Technology. Pasadena. CA.
Hastings. A. (2001). The Art of Analog Layout. Prentice Hall. Upper Saddle River. NJ.
Hawkins. G. A. (1985). Lateral profiling of interface states along the sidewalls of channel-stop
isolation. Solid State Electronics. 28(9):945-956.
Hertz. J Krogh. A.. and Palmer. R. G. (1991). Introduction to the Theory of Neural
.•

Computation. Addison-Wesley. Reading. MA.

Hoeneisen. B. and Mead. C. A. (1972a). Current-voltage characteristics of small size MOS
transistors. IEEE Transactions on Electron Devices. ED-19(3):382-383.
Hoeneisen. B. and Mead. C. A. (1972b). Fundamental limitations in microelectronics - I. MOS
technology. Solid-state Electronics. 15:819-829.
Holler. M Tam. S Castro. H and Benson. R. (1989). An electrically trainable artificial neural
.• .• .•

network (ETANN) with 10240 floating gate synapses. In Proceedings of the 1989 IEEE INNS
International Joint Conference on Neural Networks. volume 2. pages 191-196. Washington.
D.C.
Holloway. T. C Dixit. G. A.. Grider. D. T Ashburn. S. P Aggarwal. R.. Shih. A.. Zhang. X
.• .• .• .•

Misium. G Esquivel. A. L.. Jain. M Madan. S Breedijk. T Singh. A Thakar. G Shinn. G

.• .• .• .• .• .• .•

Riemenschneider. B O·Brien. S Frystak. D Kittl. J Amerasekera, A.. Aur. S Nicollian. P

.• .• .• .• .• .•

Aldrich. D Eklund. B.. Appel. A Bowles. C and Parrill. T. (1997). 0.18 J-Lm CMOS
.• .• .•

technology for high-performance. low-power and RF applications. In 1997 Symposium on VLSI

Technology: digest of technical papers. pages 13-14.
Hori. T Akamatsu. S and Odake. Y. (1992). Deep-submicrometer CMOS technology with
.• .•

reoxidized or annealed nitrided-oxide gate dielectrics prepared by rapid thermal processing.

IEEE Transactions on Electron Devices. 39(1):118-126.
Horiuchi. T. K. and Koch. C. (1999). Analog VLSI-based modeling of the primate oculomotor
system. Neural Computation. 11(1):243-265.
Horowitz. P. and Hill. W. (1989). The Art of Electronics. Cambridge University Press. 2nd
edition.
Hu. C. (1993). Future CMOS scaling and reliability. Proceedings of the IEEE. 81(5):682-689.
Huijsing. J. H. (2001). Operational Amplifiers: Theory and Design. Kluwer. Dordrecht. The
Netherlands.
Huijsing. J. H Lucas. P and de Bruin. B. (1982). Monolithic analog multiplier-divider. IEEE
.• .•

Journal of Solid-State Circuits. SC-17(1):9-15.

Indiveri. G. (2000). Modeling selective attention using a neuromorphic analog VLSI device.
Neural Computation. 12(12):2857-2880.
Indiveri. G. (2001a). A current-mode hysteretic winner-take-all network. with excitatory and
inhibitory coupling. Analog Integrated Circuits and Signal Processing. 28(3):279-291.
Indiveri. G. (2001b). A neuromorphic VLSI device for implementing 2-D selective attention
systems. IEEE Transactions on Neural Networks. 12(6):1455-1463.
Indiveri. G Kramer. J and Koch. C. (1996). System implementations of analog VLSI velocity
.• .•

sensors. IEEE Micro. 16(5):40-49.

Ismail. M. and Fiez. T. (1994). Analog VLSI : Signal and Information Processing. McGraw-Hill.
New York.
References 421

Iwase, M., Mizuno, T., Takahashi, M., Niiyama, H., Fukumoto, M., Ishida, K., Inaba, S.,
Takigami, Y., Sanda, A., Toriumi, A., and Yoshimi, M. (1993). High-performance O.IO-ttm
CMOS devices operating at room temperature. IEEE Electron Device Letters, 14(2):51-53.
Johns, D. A. and Martin, K. (1997). Analog integrated circuit design. Wiley, New York.
Kahng, D. (1967). Semipermanent memory using capacitor charge storage and IGFET read-out.
The Bell System Technical Journal, XLVI(6):1296-1300.
Kahng, D. and Sze, S. M. (1967). A floating-gate and its applications to memory devices. The
Bell System Technical Journal, XLVI(6):1288-1295.
Kaski, S. and Kohonen, T. (1994). Winner-take-all networks for physiological models of
competitive learning. Neural Networks, 7(6n):973-984.
Kelly, R. D. (1970). Electronic circuit analysis and design by driving-point impedance
techniques. IEEE Transactions on Education, E-13(3):154-167.
Kerns, D. A., Tanner, J. E., Sivilotti, M. A., and Luo, J. (1991). CMOS UV-writeable
non-volatile analog storage. In Sequin, C. H., editor, Advanced Research in VLSI, pages
245-261. MIT Press, Cambridge, MA.
Kirnizuka, N., Yamaguchi, K., Imai, K., Iizuka, T., Liu, C. T., Keller, R. C., and Horiuchi, T.
(2000). NBTI enhancement by nitrogen incorporation into ultrathin gate oxide for O.lOttm gate
CMOS generation. In Digest of Technical Papers / 2000 Symposium on VLSI Technology, pages
92-93, Piscataway, NJ. IEEE Electron Devices Society.
Kingsbury, N. G. and Rayner, P. J. W. (1971). Digital filtering using logarithmic arithmetic.
Electronics Letters, 7(2):56-58.
Kittel, C. (1996). Introduction to Solid State Physics. Wiley, New York, 7th edition.
Konn, R. and Genin, R. (1979). High-performance aperiodic frequency multiplying. Electronics
Letters, 15(6):187-189.
Kooi, E., van Lierop, J. G., and Appels, J. A. (1976). Formation of Silicon Nitride at a Si-Si�
interface during local oxidation of silicon and during heat-treatment of oxidized silicon in NIl
gas. Solid-State Science and Technology, Journal of the Electrochemical Society,
123(7):1117-1120.
Kramer, J., Sarpeshkar, R., and Koch, C. (1997). Pulse-based analog VLSI velocity sensors.
IEEE Transactions on Circuits and Systems II: Analog and Digital Signal Processing,
44(2):86-101.
Krieger, J. B. (1966). Theory of electron tunneling in semiconductors with degenerate band
structure. Annals of Physics, 36:1-60.
Kuroi, T., Uchida, T., Horita, K., Sakai, M., Inoue, Y., and Nishimura, T. (1998). Stress analysis
of shallow trench isolation for 256M DRAM and beyond. In International Electron Devices
Meeting technical digest, pages 141-144.
Kurokawa, T. and Mizukoshi, T. (1991). Computer Graphics Using Logarithmic Number
Systems. The Transactions of the Institute of Electronics, Information and Communication
Engineers E, 74(2):447-451.
Lai, F. (1991). A IOns hybrid number system data execution unit for digital signal processing
systems. IEEE Journal of Solid-State Circuits, 26(4):590-599.
Lai, F.-S. and Wu, C.-F. E. (1991). A hybrid number system processor with geometric and
complex arithmetic capabilities. IEEE Transactions on Computers, 40(8):952-962.
Lakshmikumar, K. R., Hadaway, R. A., and Copeland, M. A. (1986). Characterization and
modeling of mismatch in MOS transistors for precision analog design. IEEE Journal of
Solid-State Circuits, SC-21(6):1057-1066.
LaMaire, R. O. and Lang, J. H. (1986). Performance of digital linear regulators which use
logarithmic arithmetic. IEEE Transactions on Automatic Control, AC-31(5):394-400.
Lang, J. H., Zukowski, C. A., LaMaire, R. 0., and An, C. H. (1985). Integrated-circuit
logarithmic arithmetic units. IEEE Transactions on Computers, C-34(5):475-483.
422 References

Lau. K. T. and Lee . S. T. (1998). A CMOS winner-takes-all circuit for self-organizing neural
networks. International Journal oj Electronics. 84(2):131-136.
Lazzaro. J. (1990). Silicon Models oj Early Audition. Ph.D. thesis. California Institute of
Technology. Pasadena. CA.
Lazzaro. J. and Mead. C. A. (1989). A silicon model of auditory localization. Neural
Computation. 1(1):47-57.
Lazzaro. J Ryckebusch. S Mahowald. M. A.. and Mead. C. A. (1989). Winner-take-all
.• .•

networks of O( n) complexity. In Touretzky. D. S editor. Advances in Neural Information

.•

Processing Systems. volume I. pages 703-711. San Mateo. CA. Morgan Kaufmann.
Lenzlinger. M. and Snow. E. H. (1969). Fowler-Nordheim tunneling into thennally grown SiQ.
Journal oj Applied Physics. 40(1):278-283.
Lewis. D. M. (1995). 114 MFLOPS logarithmic number system arithmetic unit for DSP
applications. IEEE Journal oj Solid-State Circuits. 30(12):1547-1553.
Li. O. P Hackbarth. E and Chen. T.-C. (1988). Identification and implication of a perimeter
.• .•

tunneling current component in advanced self-aligned bipolar transistors. IEEE Transactions on

Electron Devices. 35(1):89-95.
Liu. D. and Svensson. C. (1993). Trading speed for low power by choice of supply and threshold
voltages. IEEE Journal oJ Solid-State Circuits. 28(1):10-17.
Liu. S.-C. (1996). Silicon model of motion adaptation in the fly visual system. In Proceedings oj
the Third Joint UCSDICaltech Symposium. pages 1-10.
Liu. S.-C. (1999). Silicon retina with adaptive filtering properties. Analog Integrated Circuits
and Signal Processing. 18(2/3):243-254.
Liu. S.-C. (2000). A winner-take-all circuit with controllable soft max property. In Solla. S. A..
Leen. T. K.. and Miiller. K.-R.. editors. Advances in Neural InJormation Processing Systems.
volume 12. pages 717-723. MIT Press. Cambridge. MA.
Lo. S.-H Buchanan. D. A.. Taur. Y and Wang. W. (1997). Quantum-mechanical modeling of
.• .•

electron tunneling current from the inversion layer of ultra-thin-oxide nMOSFET·s. IEEE
Electron Device Letters. 18(5):209-211.
Logan. R. A. and Chynoweth. A. O. (1963). Effect of degenerate semiconductor band structure
on current-voltage characteristics of silicon tunnel diodes. Physical Review. 131(1):89-95.
Lyon. R. F. (1993). Cost. power. and parallelism in speech signal processing. In Proceedings oj
the IEEE 1993 Custom Integrated Circuits Conference. pages 15.1.1-15.1.9. San Diego. CA.
Maher. M. A. C. (1989). A charge-controlled model Jor MOS transistors. Ph.D. thesis.
California Institute of Technology. Pasadena. CA.
Mahowald. M. (1994). An Analog VLSI System Jor Stereoscopic Vision. Kluwer. Boston. MA.
Maloberti. F. (2001). Analog Design Jor CMOS VLSI Systems. Kluwer. Dordtecht. The
Netherlands.
Matsuda. S Sato. T Yoshimura. H Takegawa. Y.. Sudo. A.. Mizushima. l.. Tsunashima. Y
.• .• .• .•

and Toyoshima. Y. (1998). Novel comer rounding process for shallow trench isolation utilizing
MSTS (Micro-Structure Transfonnation of Silicon). In International Electron Devices Meeting
technical digest. pages 137-140.
McCreary. J. L. (1981). Matching properties. and voltage and temperature dependence of MOS
capacitors. IEEE Journal oj Solid-State Circuits. SC-16(6):608-616.
Mead. C. A. (1989). Analog VLSI and Neural Systems. Addison-Wesley. Reading. MA.
Mead. C. A. (1990). Neuromorphic electronic systems. Proceedings oJ the IEEE.
78(10):1629-1636.
Mead. C. A. (1994). Scaling of MOS technology to subtnictometer feature sizes. Journal oJ
VLSI Signal Processing. 8(1):9-25.
Mead. C. A. and Conway. L. A. (1980). Introduction to VLSI Systems. Addison-Wesley.
References 423

Reading. MA.
Mendis. S. K Kemeny. S. E Gee. R. C Pain. B Staller. C. 0 Kim. Q and Fossum. E. R.
.• .• .• .• .• .•

(1997). CMOS Active Pixel Image Sensors for highly integrated imaging systems. IEEE
Journal of Solid-State Circuits. 32(2):187-197.
Minch. B. A. (1997). Analysis, Synthesis, and Implementation of Networks of Multiple-Input
Translinear Elements. Ph.D. thesis. California Institute of Technology. Pasadena. CA.
Minch. B. A. (2000a). Floating-gate techniques for assessing mismatch. In Emerging
technologies for the 21st century: Proceedings of the 2000 IEEE International Symposium on
Circuits and Systems. volume 4. pages 385-388. ISCAS 2000 Geneva. Switzerland. 28-31 May.
Minch. B. A. (2000b). Synthesis of dynamic multiple-input translinear element networks. In
Emerging technologies for the 21st century: Proceedings of the 2000 IEEE International
Symposium on Circuits and Systems. volume 1. pages 48�86. ISCAS 2000 Geneva.
Switzerland. 28-31 May.
Minch. B. A.. Diorio. C Hasler. P and Mead. C. A. (1996). Translinear circuits using
.• .•

subthreshold floating-gate MOS transistors. Analog Integrated Circuits and Signal Processing.
9(2):167-180.
Minch. B. A.. Hasler. P and Diorio. C. (1998). The multiple-input translinear element: A
.•

versatile circuit element. In Proceedings of the 1998 IEEE International Symposium on Circuits
and Systems. volume 1. pages 527-530. ISCAS '98: Monterey. CA. 31 May-3 June.
Minch. B. A.. Hasler. P and Diorio. C. (1999). Synthesis of multiple-input translinear element
.•

networks. In Proceedings of the 1999 IEEE International Symposium on Circuits and Systems.
volume 2. pages 236-239. ISCAS '99: Orlando. FL. 30 May-2 June.
Mitchell. Jr J. N. (1962). Computer multiplication and division using binary logarithms. IRE
.•

Transactions on Electronic Computers. EC-ll(4):512-517.

Montoro. C. G Schneider. M. C and Cunha. A. I. A. (1999). A current-based MOSFET model
.• .•

for integrated circuit design. In Sanchez-Sinencio. E. and Andreou. A. G editors.

.•

Low- Voltage/Low-Power Integrated Circuits and Systems. pages 7-55. IEEE Press. Piscataway.
NJ.
Moss. T. S editor (1980). Handbook on Semiconductors. volume 1-4. North-Holland.
.•

Amsterdam.
Mudra. R.. Hahnloser. R.. and Douglas. R. J. (1999). Integrating neuromorphic action-oriented
perceptual inputs to generate a navigation behaviour for a robot. International Journal of Neural
Systems. 9(5):411-416.
Mulder. J Serdijn. W. A.. van der Woerd. A. C and van Roennund. A. H. M. (1996). Dynamic
.• .•

translinear RMS-DC converter. Electronics Letters. 32(22):2067-2068.

Mulder. J Serdijn. W. A.. van der Woerd. A. C and van Roennund. A. H. M. (1997a). A
.• .•

current-mode synthesis method for translinear companding filters. In Proceedings of the Fourth
IEEE International Conference on Electronics, Circuits, and Systems. volume 3. pages
1419-1422. Cairo.
Mulder. J Serdijn. W. A van der Woerd. A. C and van Roennund. A. H. M. (1997b).
.• .• .•

Dynamic translinear circuits-An overview. In Proceedings of the 2nd IEEE-CAS Region 8

Workshop on Analog and Mixed IC Design. pages 65-72. Baveno. Italy.
Mulder. J Serdijn. W. A van der Woerd. A. C and van Roennund. A. H. M. (1999). Dynamic
.• .• .•

Translinear and Log-Domain Circuits: Analysis and Synthesis. Kluwer. Boston. MA.
Mulder. J van der Woerd. A. C Serdijn. W. A.. and van Roennund. A. H. M. (1995).
.• .•

Application of the back gate in MOS weak inversion translinear circuits. IEEE Transactions on
Circuits and Systems I: Fundamental Theory and Applications. 42(11):958-962.
Mulder. J van der Woerd. A. C Serdijn. W. A.. and van Roennund. A. H. M. (1997c). An
.• .•

RMS-DC converter based on the dynamic translinear principle. IEEE Journal of Solid-State
Circuits. 32(7):1146-1150.
424 References

Murphy, B. T. (1993). Minimization of transistor delay at a given power density. IEEE

Transactions on Electron Devices, 40(2):414-420.
Nagata, M. (1992). Limitations, innovations, and challenges of circuits and devices into a half
micrometer and beyond. IEEE Journal of Solid-State Circuits, 27(4):465-472.
Nandakumar, M., ChatteIjee, A., Sridhat, S., Joyner, K., Rodder, M., and Chen, I.-C. (1998).
Shallow trench isolation for advanced ULSI CMOS technologies. In International Electron
Devices Meeting technical digest, pages 133-136.
Nass, M. M. and Cooper, L. N. (1975). A theory for the development of feature detecting cells in
visual cortex. Biological Cybernetics, 19:1-18.
Nishinohata, K., Shigyo, N., and Wada, T. (1992). Effects of microscopic fluctuations in dopant
distributions on MOSFET threshold voltage. IEEE Transactions on Electron Devices,
39(3):63�39.
Nixon, R. H., Kemeny, S. E., Staller, C. 0., and Fossum, E. R. (1995). 128x128 CMOS
photodiode-type active pixel sensor with on-chip timing, control and signal chain electronics. In
Blouke, M. M., editor, Charge-Coupled Devices and Solid State Optical Sensors V. Proceedings
of the SPIE, volume 2415, pages 117-123.
Noble, P. J. W. (1968). Self-scanned silicon image detector arrays. IEEE Transactions on
Electron Devices, ED-15(4):202-209.
Noor Mohanunad, S. (1992). Unified model for drift velocities of electrons and holes in
semiconductors as a fnnction of temperature and electric field. Solid-State Electronics,
35(10):1391-1396.
Normand, G. (1985). Translineat current conveyors. International Journal of Electronics,
59(6):771-777.
Nyquist, H. (1928). Thermal agitation of electric charge in conductors. Physical Review,
32(1):110-113.
Oppenheim, A. V., Willsky, A. S., and Nawab, S. H. (1996). Signals & Systems. Prentice Hall,
Upper Saddle River, NJ, 2nd edition.
Paterson, W. L. (1963). Multiplication and logarithmic conversion by operational
amplifier-transistor circuits. The Review of Scientific Instruments, 34(12):1311-1316.
Pavasovic, A. (1991). Subthreshold region MOSFET mismatch analysis and modeling for analog
VLSI systems. Ph.D. thesis, The Johns Hopkins University, Baltimore, MD.
Pavasovic, A., Andreou, A. G., and Westgate, C. R. (1994). Characterization of subthreshold
MOS mismatch in transistors for VLSI systems. Journal of VLSI Signal Processing, 8(1):75-85.
Payne, A. and Thanachayanont, A. (1997). Translineat circuit for phase detection. Electronics
Letters, 33(18):1507-1509.
Payne, A., Thanachayanont, A., and Papavassilliou, C. (1998). A 150-MHz translineat
phase-locked loop. IEEE Transactions on Circuits and Systems II: Analog and Digital Signal
Processing, 45(9):1220-1231.
Pelgrom, M. J. M., Duinmaijer, A. C. J., and Welbers, A. P. G. (1989). Matching properties of
MOS transistors. IEEE Journal of Solid-State Circuits, 24(5):1433-1440.
Perry, D. and Roberts, G. W. (1996). The design of log-domain filters based on the operational
simulation of LC ladders. IEEE Transactions on Circuits and Systems II: Analog and Digital
Signal Processing, 43(11):763-774.
Plummer, J. D., Deal, M., and Griffen, P. B. (2000). Silicon VLSI Technology: Fundamentals,
Practice and Modeling. Prentice Hall, Upper Saddle River, NJ.
Pookaiyaudom, S. and Mahattanakul, J. (1995). A 3.3 Volt high-frequency capacitorless
electronically-tunable log-domain oscillator. In 1995 IEEE International Symposium on Circuits
and Systems, volume 2, pages 829-832. ISCAS '95: Seattle, WA, 30 April-3 May.
Poularikas, A. D. and Seely, S. (1994). Signals and Systems. PWS-KENT, Boston, MA, 2nd
edition.
References 425

Punzenberger, M. and Enz, C. (1996). A new 1.2 V BiCMOS log-domain integrator for
companding current-mode filters. In 1996 IEEE International Symposium on Circuits and
Systems, volume I, pages 125-128. ISCAS '96: Atlanta, GA, 12-15 May.
Razavi, B. (2001). Design of analog CMOS integrated circuits. McGraw-Hill, Boston, MA.
Reimbold, G. (1984). Modified IIf trapping noise theory and experiments in MOS transistors
biased from weak to strong inversion-influence of interface states. IEEE Transactions on
Electron Devices, 31(9):1190-1197.
Reisch, M. (1990). Tunneling-induced leakage currents in pn junctions. AID: Archiv for
Elektronik und Ubenragungstechnik, 44(5):368-376.
Ricco, B., Versari, R., and Esseni, D. (1996). Characterization of polysilicon-gate depletion in
MOS structures. IEEE Electron Device Letters, 17(3):103-105.
Robinson, F. N. H. (1974). Noise and Fluctuations in Electronic Devices and Circuits. Oxford
University Press.
Rose, A. (1973). Vision: Human and Electronic. Plenum Press, New York.
Sanchez, J. J. and DeMassa, T. A. (1991). Review of carrier injection in the
silicon/silicon-dioxide system. lEE Proceedings G: Circuits, Devices and Systems,
138(3):377-389.
Sarpeshkar, R. (1997). Efficient Precise Computation with Noisy Components: Extrapolating
from an Electronic Cochlea to the Brain. Ph.D. thesis, California Institute of Technology,
Pasadena. CA.
Sarpeshkar, R. (1998). Analog versus digital: Extrapolating from electronics to neurobiology.
Neural Computation, 10(7):1601-1638.
Sarpeshkar, R., Delbriick, T., and Mead, C. A. (1993). White noise in MOS transistors and
resistors. IEEE Circuits and Devices Magazine, 9(6):23-29.
Schlotzhauer, K. G. and Viswanathan, T. R. (1972). New bipolar analogue multiplier.
Electronics Letters, 8(16):425-427.
Schutte, C. and Rademeyer, P. (1992). Subthreshold IIf noise measurements in MOS transistors
aimed at optimizing focal plane array signal processing. Analog Integrated Circuits and Signal
Processing, 2(3):171-177.
Sedra, A. and Smith, K. C. (1970). A second generation current conveyor and its applications.
IEEE Transactions on Circuit Theory, CT-17(1):132-134.
Seevinck, E. (1981). Analysis and Synthesis of Translinear Integrated Circuits. D.Sc. thesis,
University of Pretoria, Pretoria, South Africa.
Seevinck, E. (1988). Analysis and Synthesis of Translinear Integrated Circuits. Elsevier,
Amsterdam.
Seevinck, E. (1990). Companding current-mode integrator: A new circnit principle for
continuous-time monlithic filters. Electronics Letters, 26(24):2046-2047.
Seevinck, E., Wassenaar, R. F., and Wong, H. C. K. (1984). A wide-band technique for vector
summation and RMS-DC conversion. IEEE Journal of Solid-State Circuits, SC-19(3):311-318.
Seitz, P., Leipold, D., Kramer, J., and Raynor, J. M. (1993). Smart optical and image sensors
fabricated with industrial CMOS/CCD semiconductor processes. In Blouke, M. M., editor,
Charge-Coupled Devices and Solid State Optical Sensors lll. Proceedings of the SPIE, volume
1900, pages 21-30.
Serdijn, W. A., Mulder, J., Poort, P., Kouwenhoven, M., van Staveren, A., and van Roermund, A.
H. M. (1999). Dynamic Translinear Circuits. In Huijsing, 1., van de Plassche, R., and Sansen,
W., editors, Analog Circuit Design: Volt Electronics; Mixed-Mode Systems; Low-Noise and RF
Power Amplifiers for Telecommunication, pages 3-32. Kluwer, Boston, MA.
Serdijn, W. A., Mulder, J., van der Woerd, A. C., and van Roermund, A. H. M. (1997a). Design
of wide-tunable translinear second-order oscillators. In Proceedings of the 1997 IEEE
International Symposium on Circuits and Systems, volume 2, pages 829-832. ISCAS '97: Hong
426 References

Kong, 9-12 June.

Serdijn, W. A., Mulder, J., van der Woerd, A. c., and van Roermund, A. H. M. (1998). A
wide-tunable translinear second-order oscillator. IEEE Journal of Solid-State Circuits,
33(2):195-201.
Serdijn, W. A., Mulder, J., and van Roermund, A. H. M. (1997b). Shortening the Analog Design
Trajectory by Means of the Dynamic Translinear Principle. In Proceedings of the ProRISC
Workshop on Circuits, Systems and Signal Processing, pages 483-489.
Serrano, T. and Linares-Barranco, 8. (1995). A modular current-mode high-precision
wiuner-take-all circuit IEEE Transactions on Circuits and Systems II: Analog and Digital Signal
Processing, 42(2):132-134.
Serrano-Gotarredona, T. and Linares-Barranco, B. (1999). Systematic width-and-length
dependent CMOS transistor mismatch characterization and simulation. Analog Integrated
Circuits and Signal Processing, 21(3):271-296.
Serrano-Gotarredona, T., Linares-Barranco, 8., and Andreou, A. G. (1999). A general translinear
principle for subthreshold MOS transistors. IEEE Transactions on Circuits and Systems I:
Fundamental Theory and Applications, 46(5):607-616.
Shibata, T. and Ohmi, T. (1992). A functional MOS transistor featuring gate-level weighted sum
and threshold operations. IEEE Transactions on Electron Devices, 39(6):1444-1455.
Shockley, W. (1961). Problems related to p-n junctions in silicon. Solid-State Electronics,
2(1):35-67.
Shyu, J.-B., Ternes, G. c., and Krummenacher, F. (1984). Random error effects in matched MOS
capacitors and current sources. IEEE Journal of Solid-State Circuits, SC-19(6):948-955.
Shyu, J.-8., Ternes, G. C., and Yao, K. (1982). Random errors in MOS capacitors. IEEE Journal
of Solid-State Circuits, SC-17(6):1070-1076.
Sin, C.-K., Kramer, A., Hu, V., Chu, R. R., and Ko, P. K. (1992). EEPROM as an analog storage
device, with particular applications in neural networks. IEEE Transactions on Electron Devices,
39(6):1410-1419.
Singh, J. (2001). Semiconductor Devices: Basic Principles. Wiley, New York.
Smith, K. C. and Sedra, A. (1968). The current conveyor - a new circuit bulding block.
Proceedings of the IEEE, 56(8): 1368-1369.
Smith, R. A. (1979). Semiconductors. Cambridge University Press, 2nd edition.
Starzyk, 1. A. and Fang, X. (1993). CMOS current mode wiuner-take-all circuit with both
excitatory and inhibitory feedback. Electronics Letters, 29(10):908-910.
Stork, J. M. C. and Isaac, R. D. (1983). Tunneling in base-emitter junctions. IEEE Transactions
on Electron Devices, ED-30(11):1527-1534.
Sun, J. Y. C., Chiang, S.-Y., and Liu, M. (1998). Foundry technology for the next decade. In
International Electron Devices Meeting technical digest, pages 321-324.
Sun, S. W. and Tsni, P. G. Y. (1994). Limitation of CMOS supply-voltage scaling by MOSFET
threshold-voltage variation. In 1994 IEEE Custom Integrated Circuits Conference, pages
267-270.
Sune, J., Olivo, P., and Ricco, 8. (1992). Quantum-mechanical modeling of accumulation layers
in MOS structure. IEEE Transactions on Electron Devices, 39(7):1732-1739.
Surakampontom, W., Jutaviriya, S., and Apajinda, T. (1988). Dual translinear sinusoidal
frequency doubler and full-wave rectifier. International Journal of Electronics, 65(6):1203-1208.
Swartzlander, Jr., E. E. and Alexopoulos, A. G. (1975). The sign/logarithm number system.
IEEE Transactions on Computers, C-24(12):1238-1242.
Swartzlander, Jr., E. E., Chandra, D. V. S., Nagle, Jr., H. T., and Starks, S. A. (1983).
Sign/logarithm arithmetic for FFT implementation. IEEE Transactions on Computers,
C-32(6):526-534.
References 427

Sze, S. M. (1981). Physics of Semiconductor Devices. Wiley, New York, 2nd edition.
Takeda, E., Yang, C. Y., and Miura-Hamada, A. (1995). Hot-Carrier Effects in MOS Devices.
Academic Press, San Diego, CA.
Tam, S., Ko, P. K., and Hu, C. (1984). Lucky-electron model of channel hot-electron injection in
MOSFET's. IEEE Transactions on Electron Devices, 31(9): 1116-1125.
Taur, Y., Buchanan, D. A., Chen, w., Frank, D. J., Ismail, K. E., Lo, S.-H., Sai-Halasz, G. A.,
Viswanathan, R. G., Wann, H.-J. C., Wind, S. J., and Wong, H. S. (1997). CMOS scaling into the
nanometer regime. Proceedings of the IEEE, 85(4):486-504.
Taylor, F. J., Gill, R., Joseph, J., and Radke, J. (1988). A 20 bit logarithmic number system
processor. IEEE Transactions on Computers, 37(2): 190-200.
Teranishi, N., Kohno, A., Ishihara, Y., Oda, E., and Arai, K. (1984). An interline CCD image
sensor with reduced image lag. IEEE Transactions on Electron Devices, ED-31(12):1829-1833.
Thanachayanont, A., Payne, A., and Pookaiyaudom, S. (1997). A current-mode phase-locked
loop using a log-domain oscillator. In Proceedings of the 1997 IEEE International Symposium
on Circuits and Systems, volume 1, pages 277-280. ISCAS '97: Hong Kong, 9-12 June.
Thanachayanont, A., Pookaiyaudom, S., and Toumazou, C. (1995). State-space synthesis of
log-domain oscillators. Electronics Letters, 31(21):1797-1799.
Theuwissen, A. J. P. (1995). Solid-state imaging with charge-coupled devices. Kluwer,
Dordrecht, The Netherlands.
Thompson, S. (1999). Sub 100 um CMOS: Technology performances, trends and challenges. In
International Electron Devices Meeting (IEDM) shon course, Washington D.C., pages 23, 26, 28
&29.
Thompson, S., Packan, P., Ghani, T., Stettler, M., Alavi, M., Post, I., Tyagi, S., Ahmed, S., Yang,
S., and Bohr, M. (1998). Source/drain extension scaling for 0.1 J.!m and below chaunel length
MOSFETs. In 1998 Symposium on VLSI Technology: digest of technical papers, pages 132-133.
Tomazou, C., Lidgey, F. J., and Haigh, D. G., editors (1990). Analogue IC design: the
current-mode approach. Peregrinus, Stevenage, Herts., UK.
Toumazou, C., Lidgey, F. J., and Yang, M. (1989). Translinear Class AB Current Amplifier.
Electronics Letters, 25(13):873-874.
Troutman, R. R. (1979). VLSI limitations from drain-induced barrier lowering. IEEE
Transactions on Electron Devices, ED-26(4):461-469.
Tsividis, Y. (1996). Mixed analog-digital VLSI devices and technology. McGraw-Hill, New
York.
Tsividis, Y. (1998). Operation and modeling of the MOS transistor. McGraw-Hill, New York.
Tuinhout, H. P., Elzinga, H., Brugman, J. T. H., and Postma, F. (1996). The floating gate
measurement technique for characterization of capacitor matching. IEEE Transactions on
Semiconductor Manufacturing, 9(1):2-8.
Vainio, O. and Neuvo, Y. (1986). Logarithmic arithmetic in FIR filters. IEEE Transactions on
Circuits and Systems, CAS-33(8):826-828.
van der Gevel, M. and Kuenen, 1. C. (1994). y"x circnit based on a novel, back-gate-using
multiplier. Electronics Letters, 30(3):183-184.
Van der Tol, M. J. and Chamberlain, S. G. (1993). Drain-induced barrier lowering in
buried-channel MOSFET's. IEEE Transactions on Electron Devices, 40(4):741-749.
van der Ziel, A. (1970). Noise: Sources, Characterization, Measurement, pages 171-173.
Prentice-Hall.
Van Valkenburg, M. E. and Kinariwala, B. K. (1982). Linear Circuits, pages 68-72.
Prentice-Hall, Englewood Cliffs, NJ.
Vittoz, E. (1996). Analog VLSI implementation of neural networks. In Fiesler, E. and Beale, R.,
editors, Handbook of Neural Computation, chapter E1.3. Oxford University Press and Institute of
428 References

Physics Publishing, Oxford and Bristol.

Vittoz, E. A. (1983). MOS transistors operated in the lateral bipolar mode and their application
in CMOS technology. IEEE Journal oJ Solid-State Circuits, SC-18(3):273-279.
Vittoz, E. A. (1985). The design of high-performance analog circuits on digital CMOS chips.
IEEE Journal oj Solid-State Circuits, SC-20(3):657-665.
Vittoz, E. A. (1990). MOS transistor. Intensive Summer Course on CMOS VLSI Design, pages
532-539. Ecole polytechnique federale de Lausanne, Switzerland.
Vittoz, E. A. (1994). Micropower techniques. In Franca, J. E. and Tsividis, Y., editors, Design oj
Analog-Digital VLSI Circuits Jor Telecommunications and Signal Processing, chapter 3, pages
53-96. Prentice Hall, Englewood Cliffs, NJ, 2nd edition.
Vittoz, E. A. and Arreguit, X. (1993). Linear networks based on transistors. Electronics Letters,
29(3):297-299.
von der Malsburg, C. (1973). Self-organization of orientation sensitive cells in the striate cortex.
Kybernetik, 14:85-100.
Walden, R. H., Krambeck, R. H., Strain, R. J., McKenna, J., Schryer, N. L., and Smith, G. E.
(1972). The buried channel charge coupled device. The Bell System Technical Journal,
51(7):1635-1640.
Wassenaar, R. E, Seevinck, E., van Leeuwen, M. G., Speelman, C. J., and Holle, E. (1988). New
techniques for high-frequency RMS-to-DC conversion based on a multifunctional V-to-I
convertor. IEEE Journal oJSolid-State Circuits, 23(3):802-815.
Weckler, G. P. (1967). Operation of p-n junction photodetectors in a photon flux integrating
mode. IEEE Journal oJSolid-State Circuits, SC-2(3):65-73.
Weste, N. H. E. and Eshraghian, K. (1994). Principles oJCMOS VLSI design. Addison-Wesley,
Reading, MA, 2nd edition.
Wilson, B. (1990). Recent developments in current conveyors and current-mode circuits. lEE
Proceedings G: Circuits, Devices and Systems, 137(2):63-77.
Wilson, C. S., Morris, T. G., and DeWeerth, S. P. (1999). A two-dimensional, object-based
analog VLSI visual attention system. In DeWeerth, S. P., Wills, S. M., and Ishii, A. T., editors,
Proceedings oj the 20th Anniversary Conference on Advanced Research in VLSI, pages 291-308.
IEEE Computer Society Press. Los Alamitos, CA.
Wolf, S. (1995). Silicon processing Jor the VLSI era, volume 3 - The submicron MOSFET.
Lattice Press, Sunset Beach, CA. ISBN 0-9616721-5-3.
Wong, C. K., Wassenaar, R. E, and Seevinck, E. (1983). A wideband accurate vector-sum
circuit. In ESSCIRe'83, Ninth European Solid-State Circuits Conference Digest oJ Technical
Papers, pages 135-138, Lausanne, Switzerland. Presses Polytechniques Romandes.
Yamaguchi, Y., Ishibashi, A., Shimizu, M., Nishimura, T., Tsukamoto, K., Horie, K., and
Akasaka, Y. (1993). A high-speed 0.6-ttm 16K CMOS gate array on a thin SIMOX film. IEEE
Transactions on Electron Devices, 40(1):179-186.
Yan, R. H., Lee, K. E, Jeon, D. Y., Kim, Y. 0., Tennant, D. M., Westerwick, E. H., Chin, G. M.,
Morris, M. D., Early, K., and Mulgrew, P. (1992). High-performance deep-submicrometer Si
MOSFET's using vertical doping engineering. IEEE Transactions on Electron Devices,
39(11):2639.
Yang, H., Sheu, B. J., and Lee, J.-C. (1992). A nonvolatile analog neural memory using
floating-gate MOS transistors. Analog Integrated Circuits and Signal Processing, 2(1):19-25.
Yu, L. K. and Lewis, D. M. (1991). A 30-b integrated logarithmic number system processor.
IEEE Journal oJSolid-State Circuits, 26(10):1433-1440.
Yuille, A. L. and Geiger, D. (1995). Winner-take-all mechanisms. In Arbib, M. A., editor, The
Handbook oj Brain Theory and Neural Networks, pages 1056-1060. MIT Press, Cambridge, MA.
Index

Above threshold, Capacitor-resistor circuit, 240, 261

SeeMOSFET Cascode, 299
Absorption, 275 CCO (Charge-coupled device), 304-307
Acceptor, 15 buried-channel, 304, 346
Accumulation, 35, 52 frame-transfer, 304
Active mask, 342 interline-transfer, 305
Active pixel sensor, surface-channel, 304
SeeAPS Channel hot-electron injection,
Adaptation, 292 SeeCHEI
Affinity, 36 Channel-length modulation effect, 77
Amplifier, Charge carrier
See Transconductance amplifier concentration, 13-15
APS (Active pixel sensor), 303-304 majority, 18
Avalanche, 80 minority, 18
multiplication, 34 Charge-coupled device,
photodiode, 282 SeeCCO
Back gate, 68 CHEI (Channel hot-electron injection), 104,
Band edge, 11 110, 112
Bandgap, 10 Circuit extraction, 346
direct, 11 Circuit simulator
energy, 11 SPICE, 86
indirect, 11 CLBT (Compatible lateral bipolar transistor),
Barrier potential, 391 SeeBIT
Base implant, 345 CMOS (Complementary metal oxide semicon­
BCCO (Buried-channel CCO), 304, 346 ductor), 50-51
implant, 346 CMP (Chemical mechanical polishing), 364,
BiCMOS (Bipolar CMOS), 345 371
Bipolar junction transistor, Common-mode sensitivity, 134
SeeBIT Comparator, 140
Bird's beak, 79, 361-363 Composition property, 253, 256
BIT (Bipolar junction transistor), 47, 89-91, Conductance, 165
177, 283, 345, 355 pseudo, 165
compatible lateral (CLBT), 199 Conduction band, 10
Blooming, 304 Conductor, 9
Bode plot, 248, 267 Contact
Body effect of MOSFET, 68 formation, 371-372
Boltzmann holes, 371-372
constant, 14 mask, 343-344
distribution, 14 Contaminant
Bonding pad, 344 use of nitride to prevent, 366
Breakdown, 34 Contamination
Buffer, 143 use of nitride to prevent, 372, 374, 380
current, 146 Continuity equations, 23
Built-in potential, 26, 393 Continuous-time, 160
Bulk, 342 currents, 150
Bump circnit, 168-175 mode, 299
Buried photodiode, 309 Contrast, 287
Capacitance Convolution, 234-236
depletion, 42-43, 57 Covalent bonds, 7
gate-to-bulk, 82 Crystal, 7
gate-to-drain, 82 Curl operator, 412
gate-to-source, 82 Current
source-to-bulk, 82 conveyor, 145-148
Capacitor implant, 344 correlator, 168-170
430 Index

dark, 275, 280, 298, 307-310 effect, 75-77, 164

density, 19-24 voltage, 164
diffusor, 165 Einstein relation, 22, 327
leakage, 132, 284, 298 Electron tunneling, 96-98
mirror, 127-128 Electron volt (eV ), 11
normalizer, 148-150 Electrostatic potential, 11
source, 124-125 Element
Current-mirror integrator, 256-260 clockwise, 193
Current-mode circuits, 145-175, 181 counterclockwise, 193, 194
Cutoff Emitter follower, 197
frequency, 248 Energy band, 10
frequency of MOSFET, 85 diagram, 9-13
wavelength, 279 Energy gap, 10
Dark current, 275, 280, 298, 307-310 Enz-Krummenacher-Vittoz model, 86-89
importance for DRAM, 309 Epitaxial layer, 356, 361
noise due to, 310 Equipartition theorem, 330
use of HAD photodiode to reduce, 309 Fabrication of MOSFETs, 361-383
Degenerate semiconductor, 18 design rule check,
Depletion, 35, 53 SeeDRC
capacitance, 57 layout and mask generation, 342-346
deep, 44 Fermi level, 13
region, 26 Fermi-Dirac distribution, 13
Design rule check, FET (Field-effect transistor),
SeeDRC SeeMOSFET
Design rules, 342 Field
DIBL (Drain-induced barrier lowering), 77, implant, 362
368, 394-395 oxide, 342, 361-362
Differential amplifier, Fill factor, 302
See Transconductance amplifier Filter
Differential pair, 133-135 band-pass, 250
Differentiator high-pass, 250, 262, 263
capacitor, 261 ideal, 250
capacitor-resistor circuit, 240, 261 low-pass, 248
follower-differentiator circuit, 263-264 Fixed-pattern noise,
hysteretic, 270-273 SeeFPN
Diffusion, 20 Flat-band
coefficient, 21 condition, 36
current density, 21 voltage, 45
length, 32, 168 Flicker noise, 315,
potential, 26 SeeNoise
velocity, 21 Floating-gate MOSFET,
Diffusor network, 166-168 SeeMOSFET
Diode, 24-34 Focal-plane processing, 300
Diode-connected, 127 Follower-differentiator circuit, 263-264
Dirac delta function, 236 Follower-integrator circuit, 252-256
Divergence operator, 412 Forward
Donor, 15 bias, 28
Doping, 9 current, 28, 56, 161
Drain-junction tunneling, 398, 401 Fowler-Nordheim tunneling, 96-99
DRC (Design rule check), 346 FOX (Field oxide), 342, 361-362
Drift, 20 FPN (Fixed-pattern noise), 292
current density, 22 Gabor function, 232
velocity, 22 Gate
Early formation of polysilicon, 366
Index 431

oxide, 365 use of epitaxial1y grown wafer to prevent,

Gate-oxide tunneling current, 397 361
Generation, 23 Layout
Gilbert nonnalizer, 150 capacitance shielding, 358
Gradient operator, 412 for device matching, 348-352
Guard ring, 357-358 guard ring, 357-358
HAD (Hole accumulation diode), 309 mask, 342-346
Hole, 10 mininum distance between
Hot pixels, 307 resistors, 349
Hot-electron injection, 80, 96, 99, 101-102, transistors, 349
See CHEI,IHEI minority carrier shielding, 357
Hot-electron injection efficiency of bonding pads, 359
in MOSFETs, 102-104 of mixed-mode analog digital circuits, 357
Hysteretic differentiator, 270-273 parasitic effects, 353
IHEI (Impact-ionized hot-e1ectron injection), rules, 342
104, 106 substrate coupling, 353-356
llluminance, 287 substrate noise, 356
Impact ionization, 80, 104, 114, 115 transistor
Impact-ionized hot-electron injection, common-centroid geometry, 350
SeeIHEI orientation, 350
Impulse LDD (Lightly doped drain), 367
Leakage current, 132, 284, 298
definition, 236
Learning rule in synaptic array, 113
integration properties, 237
LED (Light-emitting diode), 24
response, 237, 239
Lens effect on conversion from scene illumi­
Impurity doping, 9, 15-19
nation to image illumination, 309
Insulator, 9
Lightly doped drain,
Integration mode, 300
SeeLDD
Integrator types
Linear
resistor-capacitor circuit, 251
resistor, 125-126
Integrator types
systems, 231-234
current-mirror, 256-260
threshold units, 154, 155
follower-integrator circuit, 252-256
transistor network, 165
resistor-capacitor circuit, 240-241
units, 153
Intrinsic
Linearity, 231
carrier density, 15 Lithography, 361
semiconductor, 14 Local
Inversion, 35 bulk, 342
moderate, 40 substrate, 342
strong, 39 LOCOS (local oxidation of silicon), 361-362
weak, 39 Lux (photometric unit), 308
Inverting amplifier, 132-133 Majority carrier, 18
Ionization energy, 10 Maxwell equations, 19
Irradiance, 287 Memory
Junction DRAM, 307, 309, 364
leakage, 307-310 EEPROM, 93
Junction breakdown, 34 EPROM, 93
voltage, 385 Flash EEPROM, 93
Junction-tunneling current density, 392, 394 PROM, 94
K, (Subthreshold slope factor), 44-45, 56, 57 Metal, 9
Lagrange multipliers, 158 mask, 343
Laplace transform, 243-244 Miller effect, 299
Laplacian operator, 412 Minority carrier, 18
Latchup, 355 MIS (Metal-insulator-semiconductor), 35
432 Index

Mismatch, 140, 142, 143, 359-360 weak inversion, 53-59

Mobility, 22 Multiple-input translinear element (MITE),
MOS (Metal-oxide-silicon), 35 202-205
MOSFET (Metal-oxide-silicon field-effect n-type semiconductor, 16
transistor), 35, 47 Neural synapses, 94
above threshold, 59-68 Neuron, 47, 151, 232
back gate, 68 Noise
backgate effect, 68 1/f, 182, 315, 317-320
bird's beak, 79 current, 326
body effect, 68 fixed-pattern, 292
body effect coefficient (1\;), 61, 70 flicke�315, 317-320
channel of, 49 input-referred, 335
conductance, 71-75 inverter, 336
Early effect, 75-77, 164 model of MOSFET, 335-336
Early voltage, 77, 164 per unit bandwidth, 332
electron mobility, 78 photoreceptor, 320, 332
extrinsic part of, 81 pink, 315
floating-gate (FGMOSFET), 93-96, 202- power spectral density, 314
204 power spectrurn, 315
hot-carrier effects, 79 readout, 303, 307
hot-electron injection, 80 RMS (root-mean-square), 314
impact ionization, 80 shot, 313, 316, 319
intrinsic capacitances above threshold, 84 thermal, 316, 319, 399
intrinsic capacitances in subthreshold, 83 thermal noise is shot noise, 325
intrinsic part of, 81 transconductance amplifier, 339
inversion layer, 53 white, 315-317, 320
I\; (Subthreshold slope factor), 56, 57 Normalization
measuring 1\;, 57 circuit, 109, 113-114, 148
mismatch, 359-360 signal, 148
narrow-channel effects, 79 Operational amplifier,
noise model, 335-336 See Transconductance amplifier
output conductance, 76 Optical absorption coefficient, 278
pinchoff point, 67 Optical conversion from scene illumation to
pinchoff region, 67 image illumination, 309
punch-through, 80, 392, 398 OTA,
saturation region, 67 See Transconductance amplifier
short-channel effects, 78, 79 Overglass mask, 344
small-signal model at low frequencies, 71 Oxide, 9
small-signal model at moderate frequen- field, 342
cies, 81 gate, 365
strong inversion, 53, 59-68 traps, 45
structure, 49 p-i-n photodiode, 282
subthreshold operation, 52-59 p-n junction, 24-34
surface potential, 59 abrupt, 26
threshold drift, 374 p-type semiconductor, 16
threshold drift by NBTI and HeI, 374 Parasitic current, 392
threshold voltage, 61-62, 394-398 Passivation, 372
transconductance, 71-73 layer, 344
triode regime, 62 Passive pixel sensor, 300-303
unity-gain cutoff frequency <IT), 84-86 Phonon, 12, 106
use of halo implant to reduce short channel Photo-electric effect, 275
effects, 369 Photoconductor, 275
use of LDD to reduce DIBL, 368 Photocurrent, 276
velocity saturation, 78, 394 Photodiode, 275-282
Index 433

buried, 309 Reverse

p-i-n, 282 bias, 28
pinned, 309 current, 28, 56, 162
Photogate, 284-286 Salicide, 369-370
Photometetric conversion from lux to photon source/drain resistance, 377
flux, 308 Saturation region, 59
Photon, 12 Scaling effects
Photoreceptor noise, 320, 332 gate leakage, 375-377
Photosensing mode, 278 mobility degradation, 382
Phototransistor, 283-284 off-state leakage, 378-379
Photovoltaic mode, 277 polycrysta1line gate depletion, 379
Pink noise, source/drain junctions, 377-378
See Noise subthreshold slope, 381
Pinned photodiode, 309 threshold drift, 374-375
Pixel, 299 Scaling limits, 373-374
Poisson Scaling of MOS Technology, 385-407
equation, 19, 38 SDE (Source/drain extension)
process, 321 formation of, 366
Poly mask, 343 Select mask, 342-343
Polycrystalline gate depletion, 379-380 Semiconductor, 9
Polycrystalline silicon, 9, 343 degenerate, 18
formation, 366 n-type, 16
Polysilicon, p-type, 16
See Polycrysta1line silicon Shockley
Positive feedback, 113 approximation, 33
Power spectrum, 322 equation, 32
Power-law circuits, 196, 220 Shot noise,
Principle See Noise
of additivity, 232 Silicide, 369-370
of homogeneity, 231 block, 344
of superposition, 232 effect on photosensitivity, 344
translinear, 177 Silicon, 7
Process Silicon dioxide, 9
flow, 361-373 Silicon learning array, 107
parameters, 341 Silicon nitride, 362, 364
Pseudo-conductance, 165 Single-well (or single-tub) process, 342
Pseudo-voltage, 165 Small-signal model of MOSFET at low fre-
Pnnch-through, 80, 368, 398 quencies, 71-75
Quantum efficiency, 281 Solar cell, 277
Quasi-Fermi levels, 29 Source follower, 128-132
Recombination, 24 Space-charge region, 26
Rectifier, 32 Spacer formation, 367, 369
Reflectivity, 287 Stacked vias, 373
Refractory metal, 370 Step
Renormalization, 112 function, 239
Resistance, 164 response, 239
negative, 113 Step-junction approximation, 393
pseudo, 165 STI (Shallow trench isolation), 363-364
Resistive element Storage pixel
compressive, 296 effect of dark current, 310
expansive, 296 Strong inversion, 39-43
Resistive networks, 164-168 in MOSFETs, 53, 59-68
Resistor-capacitor circuit, 240-241, 251 Substrate, 342
Responsivity, 281 Substrate tunneling, 404
434 Index

Subthreshold, See MOSFET

See MOSFET synapse,
Subthreshold slope factor, See Synapse transistor
See K, Translinear
Surface potential, 37 circuit, 177, 178, 182
Symbolic layer, 346 dynamic circuits, 178
Synapse, 260 dynamic principle, 178
four-tenninal transistor, 98-10I element, 179, 185
nFET, 101-104 loop, 184, 188, 190-196
nFET gate current, 96-98 multiplier, 201
pFET, 104-106 principle, 177, 183, 187
pFET gate current, 106-107 Trapping centers, 34
transistor, 93, 98-101 Trench isolation, 356
channel-length modulation, 114 Triode region, 57
weight-update rule, 107-109 Tunneling
Synapse transistor area, 393
channel-length modulation, 115 current density, 390
System drain comer, 393
linear, 231 gate oxide, 386
time-constant, 240 in synapse transistor, 96-98
Thennal process, 101
energy, 388 Twin-well (or twin-tub) process, 342
equilibrium, 14 Unity-gain follower, 142-143, 252, 263
noise, 316 Unsupervised learning, 116
voltage, 22 Valence
Threshold drift, 374 band, lO
NBTI (Negative bias temperature instabil- electron, 7
ity), 374 Velocity
Threshold voltage, 42, 61-62, 394-398 diffusion, 21
Time constant, 240, 251, 262 drift, 22
Transconductance amplifier, 135-142 Velocity saturation, 394
open-circuit voltage gain, 139 Via mask, 344
output conductance, 138 Voltage gain, 139
transconductance, 137 Wafer, 342
Transfer function, 153, 231, 244-246, 252, Weak inversion, 39-43
255 in MOSFETs, 53-59
composition property, 253 Weight nonnalization circuit, 111-112
non-linear, 153 Well mask, 342
Transistor White noise,
bipolar junction, See Noise
See BIT White spots, 307
floating -gate, Winner-take-all (WTA)
See MOSFET circuit, 150, 160-164
insulated-gate bipolar (IGBT), 179 network, 152-160
metal-oxide-silicon field-effect, Work function, 35