Circuit Design For Reliability

Ricardo Reis · Yu Cao
Gilson Wirth Editors
Circuit
Design for
Reliability
Circuit Design for Reliability
Ricardo Reis • Yu Cao • Gilson Wirth
Editors
Circuit Design for Reliability
123
Editors
Ricardo Reis Yu Cao
Instituto de Informática School of ECEE
Universidade Federal do Rio Arizona State University
Grande do Sul (UFRGS) Tempe, USA
Porto Alegre, Rio Grande do Sul, Brazil
Gilson Wirth
Electrical Engineering Department
Universidade Federal do Rio
Grande do Sul (UFRGS)
Porto Alegre
Rio Grande do Sul
Brazil
ISBN 978-1-4614-4077-2 ISBN 978-1-4614-4078-9 (eBook)

DOI 10.1007/978-1-4614-4078-9
Springer New York Heidelberg Dordrecht London
Library of Congress Control Number: 2014952728
© Springer Science+Business Media New York 2015

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of
this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publisher’s location, in its current version, and permission for use must always be obtained from Springer.
Permissions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations
are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for
any errors or omissions that may be made. The publisher makes no warranty, express or implied, with
respect to the material contained herein.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)

Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Ricardo Reis, Yu Cao, and Gilson Wirth
2 Recent Trends in Bias Temperature Instability . . . . . . . . . . . . . . . . . . . . . . . . . 5
B. Kaczer, T. Grasser, J. Franco, M. Toledano-Luque,
J. Roussel, M. Cho, E. Simoen, and G. Groeseneken
3 Charge Trapping Phenomena in MOSFETS: From Noise
to Bias Temperature Instability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Gilson Wirth and Roberto da Silva
4 Atomistic Simulations on Reliability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
Dragica Vasileska and Nabil Ashraf
5 On-Chip Characterization of Statistical Device Degradation . . . . . . . . . 69
Takashi Sato and Hiromitsu Awano
6 Compact Modeling of BTI for Circuit Reliability Analysis . . . . . . . . . . . . 93
Ketul B. Sutaria, Jyothi B. Velamala, Athul Ramkumar,
and Yu Cao
7 Circuit Resilience Roadmap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Veit B. Kleeberger, Christian Weis, Ulf Schlichtmann,
and Norbert Wehn
8 Layout Aware Electromigration Analysis
of Power/Ground Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
Di-an Li, Malgorzata Marek-Sadowska, and Sani R. Nassif
9 Power-Gating for Leakage Control and Beyond . . . . . . . . . . . . . . . . . . . . . . . . 175
Andrea Calimera, Alberto Macii, Enrico Macii,
and Massimo Poncino
v
vi Contents
10 Soft Error Rate and Fault Tolerance Techniques for FPGAs. . . . . . . . . . 207
Fernanda Kastensmidt and Ricardo Reis
11 Low Power Robust FinFET-Based SRAM Design in Scaled
Technologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Sumeet Kumar Gupta and Kaushik Roy
12 Variability-Aware Clock Design. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
Matthew R. Guthaus and Gustavo Wilke
Chapter 1
Introduction
Ricardo Reis, Yu Cao, and Gilson Wirth
The scaling of CMOS technology to the nanometer regime inevitability increases

reliability concerns, profoundly impacting all aspects of circuit performance and
posing a fundamental challenge to future IC design. These reliability concerns arise
from many different sources, and become more severe with continuous scaling.
VLSI design in the late CMOS era is then driven by an increasing challenge to
cope with unreliable components at the device and circuit levels. Early approaches
focused only on technology improvement, but ignoring these effects in the design
process causes an excessive amount of over-margining, with negative impact in
both cost and performance of the final product. At the device and circuit levels,
physical understanding, modeling, detection, and successful design techniques for
leading reliability mechanisms are vitally important, not only to current robust
design practice, but also to the prediction and management of system reliability over
the full product life cycle. It is critical to understand, simulate, and mitigate their
impact during the stage of physical and circuit design. This book presents physical
understanding, modeling and simulation, on-chip characterization, layout solutions,
and design techniques that are effective to enhance the reliability of various circuit
units.
We start by addressing the physical mechanisms of reliability, focusing on
aging and noise. Aging is among the critical reliability issues facing present and
future deeply downscaled CMOS devices, and the major concern is the so-called
Bias Temperature Instability (BTI). The elementary definitions and experimental
R. Reis • G. Wirth
Electrical Engineering Department, Universidade Federal do Rio Grande do Sul (UFRGS),
Y. Cao ()
School of ECEE, Arizona State University, Tempe, AZ, USA
e-mail: ycao@asu.edu
© Springer ScienceCBusiness Media New York 2015 1

R. Reis et al. (eds.), Circuit Design for Reliability,
DOI 10.1007/978-1-4614-4078-9__1
2 R. Reis et al.
observations of BTI are first briefly reviewed. Afterwards it is shown how the
understanding of gate oxide defect properties can be used to explain BTI and its
variability, and possible technological solutions for both nFETs and pFETs are
discussed. The charge trapping phenomena that plays a dominant role in BTI also
dominates low-frequency noise. A modeling and circuit simulation approach that
focuses on operation conditions relevant for digital and analog design, including
large signal AC operation, is presented. The modeling and simulation approach aims
at helping circuit designers to cope with this reliability issue.
Next, the modeling and characterization is addressed. Any modeling effort
needs a solid base to be built on. Models are based on accurate measurement
methodologies, which contribute to understanding the mechanisms behind the
device degradation. In this book, authors survey characterization methods for
the purpose of device modeling that includes the effects of temporal parameter
degradation, such as BTI and noise. Particular focus is placed on the approaches that
utilize on-chip circuits to enhance the accuracy and fidelity of the measurements.
Modeling must focus on real circuit operation conditions, such as Dynamic Voltage
Scaling (DVS), or biasing found in mixed signal designs where the stress waveform
is much more random. Modeling solutions suitable to allow aging simulation under
all dynamic stress conditions are presented.
Variability in electrical parameters, such as variability in threshold voltage, is
also known to be a major reliability concern. Among the sources of variability
is random dopant fluctuation (RDF). We discuss how RDF can induce significant
variation in saturation (on) current and transconductance degradation—two key
metrics for benchmark performance of digital and analog integrated circuits.
An analysis of ongoing scaling related trends of integrated circuits is also
presented. For this purpose, authors discuss major trends observed in process
technology, as well as in memory and logic designs. While newer technology styles,
such as FinFETs provide necessary advances to enable further technology scaling,
we can observe that there exists also a trend for reduced resilience with ongoing scal-
ing. Our hope is that with the roadmap presented in this book, which concentrates on
FinFET-based designs, we are able to sensitize the research community to present
and near future problems associated with this technology. Additionally, different
architectures for SRAM memories are analyzed, to demonstrate the different design
choices available and tradeoffs that have to be made.
Finally, the last chapters present design techniques for reliability. FinFETs have
emerged as alternatives to conventional bulk MOSFETs in scaled technologies
due to superior gate control of the channel, lower short channel effects and
higher scalability. However, width quantization in FinFETs constrains the design
space of FinFET-based circuits, especially SRAMs in which transistor sizing is
critical for the circuit robustness. The adverse effects of width quantization can
be mitigated by appropriate device-circuit co-design of FinFET-based memories.
Some of such techniques are described, with an emphasis on the device-circuit
interactions associated with each methodology. The impact of different technology
1 Introduction 3
options in FinFETs like gate-underlap, fin orientation, fin height, gate workfunction
and independent control of the gates on the stability, power and performance of 6T
SRAMs is discussed.
Electromigration is an aging effect that affects metal wires (interconnections). In
this book, authors briefly introduce physical foundations of electromigration (EM).
We discuss physical parameters affecting EM wire lifetime and we introduce some
background related to the existing EM physical simulators. In the work presented
in this book for EM physical simulation, authors adopt the atomic concentration
balance-based model. We discuss the simulation setup and results. We present
VEMA, a variation-aware electromigration (EM) analysis tool for power grid wires.
The tool considers process variations caused by the chemical–mechanical polishing
(CMP) and edge placement error (EPE). VEMA is a full-chip EM analysis tool.
It performs EM lifetime calculation, and analyzes process variation effects on EM
reliability and reports variation tolerances of EM-sensitive power grid wires.
Integrated circuits subject to ionizing radiation are sensitive to transient faults
caused by the interaction of ionizing particles with the semiconductor. At ground
level, the main source of radiation are neutrons that interact with the material
generating secondary particles that can ionize the silicon, provoking transient upset
in circuits fabricated in nanometer technology. The interaction of the ionizing
radiation with the transistor may provoke transient and permanent effects. Different
fault tolerance techniques can be applied to FPGAs according to their type of
configuration technology, architecture and target operating environment. A chapter
is dedicated to present a set of fault mitigation techniques for SRAM, FLASH and
ANTIFUSE-based FPGAs. A test methodology to characterize those FPGA under
radiation is also presented. Results from neutron-induced faults are presented and
discussed.
The relevance of power consumption and leakage currents is also increasing.
Although several Design-for-Low-Power and Design-for-Variability options are
already available in modern EDA suites, the contrasting nature of the two metrics
makes their integration extremely challenging. Most of the approaches used to
compensate and/or mitigate circuit variability (e.g., Dynamic Voltage Scaling and
Adaptive Body Biasing) are, in fact, intrinsically power inefficient, as they exploit
the concept of redundancy, which is known to originate power overhead. In this
book, we introduce possible solutions for concurrent leakage minimization and
variability compensation. More specifically, authors propose Power-Gating as a
mean for simultaneously controlling static power consumption and mitigating the
effects induced by two of the most insidious sources of variability, namely, electrical
parameter variations and Aging due to BTI.
High-performance clock network design has been a challenge for many years
due to the drastically increasing effect of process variability. In addition, tight power
budgets have lowered supply voltage levels, which make designs more sensitive to
noise. Together, variability and noise present a colossal challenge to clock designers
in order to meet timing, yield, and power simultaneously. This chapter discusses the
different strategies that designers use to ameliorate variability and noise problems
in clock network design.
4 R. Reis et al.
The project of this book was born during the successful IEEE Circuits and
Systems Society Summer School on Physical Design of Reliable Circuits that was
held in Porto Alegre, Brazil, from January 12 to 15, 2010. Some chapters of the
book are based on the talks realized during the school and other well-known invited
authors wrote other chapters.
It is our hope that this book will provide the vitally important physical under-
standing and design techniques for state of the art and future technologies, ranging
from technology modeling, fault detection and analysis, circuit hardening, and
reliability management in different circuit units.
Chapter 2
Recent Trends in Bias Temperature Instability
B. Kaczer, T. Grasser, J. Franco, M. Toledano-Luque, J. Roussel, M. Cho,

E. Simoen, and G. Groeseneken
Abstract The paradigm shifts occurring in the past few years in our understanding
of BTI are reviewed. Among the most significant ones is the shift from perceiving
NBTI in terms of the Reaction-Diffusion model to analyzing BTI with the tools
originally developed for describing low-frequency noise. This includes the interpre-
tation of the time, temperature, voltage, and duty cycle dependences of BTI. It is
further demonstrated that a wealth of information about defect properties can be
obtained from deeply-scaled devices, and that this information can allow projection
of variability issues of future deeply downscaled CMOS devices. The chapter is
concluded by showing the most promising technological solutions to alleviate both
PBTI and NBTI.
1 Introduction
Among the critical reliability issues facing present and future deeply downscaled
CMOS devices is the so-called Bias Temperature Instability (BTI). While BTI in
nFET devices was generally ascribed to charge trapping in the high-k portion of
the gate oxide, the interpretation of BTI in pFET devices still generates controversy
[1–7]. This phenomenon in pFET devices has been previously described by
diffusion of hydrogen from and back into substrate/gate oxide interface states [8].
The so-called Reaction–Diffusion model based on this assumption is still popular,
especially in the design community [9], despite being inconsistent with some crucial
observations [10].
In this chapter, the elementary definitions and experimental observations of BTI
are first briefly reviewed. One of the most intriguing properties of BTI—the lack
B. Kaczer () • J. Franco • M. Toledano-Luque • J. Roussel • M. Cho

• E. Simoen • G. Groeseneken
imec, Leuven, Belgium
e-mail: kaczer@imec.be
T. Grasser
TU Wien, Wien, Austria
G. Groeseneken
KU, Leuven, Belgium

DOI 10.1007/978-1-4614-4078-9__2
6 B. Kaczer et al.
of characteristic time scale, especially in pFETs—is then argued to point to a dis-

persion in the underlying mechanism [5, 7, 11]. In CMOS technologies, a response
on many time scales is typical for low-frequency noise [and its manifestation as
Random Telegraph Noise (RTN) in deeply-scaled devices], suggesting that (the
principal component of) BTI is in fact caused by the same defects [12, 13]. The
link between BTI and low-frequency noise is then further developed—it is shown
that many properties of gate oxide defects can be directly extracted from BTI
relaxation measurements in deeply-scaled devices [14] and the noise-inspired model
capable to fully describe these properties is reviewed. Afterwards it is shown how
the understanding of gate oxide defect properties can be used to explain variability
of BTI in deeply-scaled technologies, as well as possible technological solutions for
both nFETs and pFETs.
2 Brief Overview of BTI
BTI is a consequence of charging of defect states in the gate oxide and at its
interface [2]. The defects could be both pre-existing or generated during device
operation. The trapped charge results in a shift of the device parameters, such as its
threshold voltage Vth , channel mobility, transconductance, and subthreshold slope,
and generally a decrease of the FET’s drive current. The name is derived from the
phenomenon being strongly accelerated by temperature T and gate bias VG . BTI
in n-channel FET devices, which are typically biased in circuits at positive VG ,
is referred to as Positive BTI (PBTI), while negative BTI (NBTI) takes place in
p-channel FETs. Constant VG stress bias is often referred to as static, or “DC” BTI,
while periodically interrupted VG stress is called “AC”, or dynamic BTI.
3 Static BTI
Figure 2.1 illustrates the typical gradual shift of pFET threshold voltage Vth during
accelerated stress at elevated T [5]. The stress data are typically measured at several
VG ’s and Vth is extrapolated to 10 years at the circuit operating voltage VDD (or
VDD C 10 %). The extrapolated Vth must be below a given value (typically 30 or
50 mV) for the technology to qualify.
This simple extrapolation procedure is, however, complicated by Vth decreas-
ing immediately after the stress bias is removed, as illustrated in Fig. 2.1 [15]. As
will be discussed henceforth, this recovery, or relaxation, component R typically
proceeds simultaneously on many time scales, making it difficult to determine
its beginning or end and thus separating it from the final non-recoverable, or
permanent, component P [2, 6]. This Vth relaxation is thus a crucial problem for
BTI measurement, interpretation, and extrapolation.
2 Recent Trends in Bias Temperature Instability 7
Fig. 2.1 Shift in pFET 25

threshold voltage is observed 1.4 nm SiON pFET
during negative gate bias
stress. When the stress bias is 20
removed, a recovery of the
-ΔVth (mV)
effect is seen (note: Vth 0 V stress relaxation
15
for this device). Inset
illustrates the biases applied 0V
at FET terminals during the 10
BTI measurement T = 125ⴗC
VG 0V VG,stress = -2 V
5
VG,relax = 0 V
-0.05V
0
0 500 1000 1500 2000
time (s)
50
40
ΔVth (mV)
30
~2x
20
10
T = 125 C tstress = 1000 s VS = VD = VB = VCC = 2 V
0 0
DC 10 102 104 106 108 1010
frequency (Hz)
Fig. 2.2 The Vth shifts due to 50 % AC unipolar NBTI stress in pFETs are seen independent of
frequency in the entire frequency range of 1 Hz–2 GHz. The Vth shift of the corresponding DC
NBTI stress obtained on an identical device is shown for comparison. Inset: Micrograph of the
on-chip circuit for the DC and AC BTI measurements consisting of a ring oscillator, a frequency
divider, a buffer, a pass-gate-based multiplexer, and the device under test [17]
4 Dynamic BTI
In many CMOS applications, such as logic, the majority of the FETs are constantly
switched and thus exposed to dynamic stress [16]. Figure 2.2 documents that
NBTI is present at frequencies up to the GHz range, i.e., there does not appear
to be any “cut-off” time constant of the degradation mechanism above 1 ns [17].
Furthermore, the AC bias signal reduces BTI with respect to the DC stress. This
provides some additional reliability margin, which can be factored in during the
application design phase [9].
8 B. Kaczer et al.
Fig. 2.3 Total degradation 100

Vth after 6000 S of unipolar Vstress = -2.0V
NBTI stress shows a
DC
fstress = 10 kHz
80
ΔVth = R(trelax) + P (mV)

distinctive dependence on the tstress = 6000 s
duty factor DF. In particular, trelax = 10-3( ), 10-2, … , 104 s
a weak dependence, or a 60 T = 125⬚C
“plateau” between 10 and
90 % is observed,
complemented by rapid Vth 40
increase for the outermost DF
values. Data at different
20
relaxation times are shown relaxation
0
0 20 40 60 80 100
Duty Factor
In an arbitrary FET of an arbitrary digital circuit, the average probability of a

signal being high can vary between 0 and 100 %. The dependence of BTI on the
duty cycle (called duty factor or DF here) thus needs to be studied. A NBTI Vth -
DF dependence with an inflection point around DF 50 %, first reported in [17], is
shown in Fig. 2.3 [12].
5 Similarity Between BTI Relaxation and Low-Frequency

Noise
Long, log(t)-like behavior of Vth without a characteristic time scale is typically

observed in both the initial portion of NBTI degradation [13, 18] and the recovery
phase. Figure 2.4 illustrates that the rate of degradation dVth /dtrelax [7] extracted
from the log(trelax )-like Vth NBTI relaxation transient after even a very short, 0.1 s
stress, follows 1/trelax for over seven decades. Such behavior is a signature of states
with discharging time constants covering as many decades [19].
Incidentally, superposition of states with widely distributed time scales is the
standard explanation of the 1/f noise spectra [20], which are clearly observed in our
pFETs (Fig. 2.4). This obvious similarity leads us to argue that the same states with
widely distributed time scales in fact play a fundamental role in both NBTI and
noise measurements.
6 Semi-quantitative Model for BTI Relaxation
In order to visualize this common property it is beneficial to consider an equivalent

circuit representing states with widely distributed time scales [19]. Note that in
either NBTI relaxation or 1/f noise measurements, no maximum or minimum “cut-
Fig. 2.4 (a) A characteristic a 70 101

long, log-like Vth relaxation
SiON
trace is observed after even 100
short (pulse-like) NBTI
60 tstress = 0.1 s
stress. The rate of recovery Vstress = -2 V 10-1
-dΔVth /dt (Vs-1)

dVth /dtrelax following
50 T = 125⬚C 10-2
ΔVth (mV)
1/trelax for 7 decades is a
40 10-3
signature of states with
discharging time constants 30 10-4
covering as many decades.
(b) Gate-referred noise 10-5
20
spectra measured on the same 10-6
(unstressed) devices show 10
clear 1/f dependence, 10-7
routinely explained by a 0 10-8
superposition of states with 10-4 10-3 10-2 10-1 100 101 102 103 104 105
widely distributed time scales trelax (s)
b 10-9
SiON
10-10 T = 25⬚C
unstressed
10-11
SVG (V2 Hz-1)
10-12 VG = -0.6 V
10-13
VG = -0.3 V
10-14
10-15
10-16 0
10 101 102 103 104 105 106
frequency (Hz)
off” times are typically observed [12]. For the sake of simplicity it is therefore
assumed here that the time constants are log-uniformly distributed from times
much shorter than the switching time of a pFET to very long, corresponding to
the lifetime of a CMOS application. Such states with widely distributed time scales
are then represented by “RC” elements in Fig. 2.5 with the total FET Vth being
proportional to the sum of voltages (“occupancies”) on all capacitors. For the sake
of simplicity, it is assumed that all RC elements have the same weight and can
be partially occupied, which emulates the behavior of a large-area device. Most
properties of the recoverable component can be reproduced when the ohmic resistors
in Fig. 2.5 are replaced with a non-linear component (simulated by two diodes with
different parameters, see Fig. 2.5), which emulates different charging (i.e., capture)
and discharging (i.e., emission) time constants of each defect [21]. Such a circuit
correctly reproduces DF (Fig. 2.6, cf. Fig. 2.3) and also the log-like relaxation
(Fig. 2.4a) and the log-like initial phase of stress (not shown) [19].
10 B. Kaczer et al.
a b
… …
C = 100 ... 10N C = 100 ... 10N
Fig. 2.5 (a) An equivalent circuit with exponentially increasing capacitances used to emulate
defect states with widely distributed time scales, such as those active in low-frequency noise.
(b) The same circuit modified to account for charging (i.e., capture) and discharging (i.e., emission)
time constants being voltage dependent, represented by asymmetric diodes. The sum of voltages
on capacitors is assumed to be proportional to FET Vth
Fig. 2.6 The plateau in DF 7

Vstress = 1 a.u.
R(trelax , tstress = 6000) (a.u.)
dependence of R is also
qualitatively well reproduced 6
trelax = 10-3, 10-2, … , 104a.u.
by the equivalent circuit in
Fig. 2.5, as is the decrease 5
with increasing relaxation
4 relaxation
time (cf. Fig. 2.3, which,
however, shows the sum of R 3
and P)
2
0
0 20 40 60 80 100
Duty Factor
7 Properties of Individual Defects
Figure 2.7 shows two typical Vth relaxation transients following positive VG stress
on a single 70 90 nm2 nMOSFET (i.e., corresponding to PBTI). Conversely to the
continuous relaxation curves obtained on large devices, a quantized Vth transient
is observed in the deeply-scaled devices. In such devices, the relaxation is observed
to proceed in discrete voltage steps, with each step corresponding to discharging of
a single oxide defect [12, 22, 23]. Upon repeated perturbation, each defect shows up
in the relaxation trace with a characteristic “fingerprint” consisting of its discharge,
or emission time, and its voltage step [14].
Figure 2.8 shows the two-dimensional histogram of the heights and the emission
times of the steps when the experiment was repeated 70 times at the same stressing
and relaxing condition as in Fig. 2.7 [14]. In Fig. 2.8, four clusters are clearly formed
that correspond to four active defects in the time window of the experimental setup.
The emission times of each defect are stochastically distributed and follow an
exponential distribution. This allows us to determine the average emission time e .
The capture time of each trap can be obtained by varying the stress (i.e., charging)
time from 240 ms down to 2 ms. The intensity of the cluster decreases with reducing
Fig. 2.7 Characteristic Vth
Threshold voltage shift x10-3 (V)

transients of a single 8
70 90 nm2
~2.5mV
1 nm-SiO2 /1.8 nm-HfSiO ~1.5mV
nMOSFET device stressed at 6
~3mV
25 ı C and VG D 2.8 V for
~10-3s
184 ms. Four discrete drops
are observed indicating the 4
~0.5s
existence of four active traps
at this stress condition ~20s
2 Temp = 25C
VSTRESS ~ 2.7V, ~1.5mV
0 tSTRESS = 189 ms ~100s
0.001 0.01 0.1 1 10 100

Relaxation Time (s)
Fig. 2.8 Two-dimensional

histograms (TDDS spectra) of
the heights and emission
times of the steps extracted
from 70 Vth transients of the
particular device of Fig. 2.7 at
(a) 25 ı C and (b) 50 ı C. Four
clusters are formed that shift
horizontally to shorter
emission times with
increasing temperature. Note
that trap #3 disappears from
the experimental window at
50 ı C
12 B. Kaczer et al.
stress time when the characteristic capture time is in the range of the stress time. The
fit of the intensity to Pc D 1 exp(tstress / c ) lets us calculate the average capture
time c . This technique is known as Time Dependent Defect Spectroscopy (TDDS)
[14].
In Fig. 2.8, an identical experiment was repeated at 50 ı C on the same device.
Note the large horizontal shift of the clusters to shorter emission times with only a
25 ı C temperature increase. The Arrhenius plots of the emission and capture times
obtained at T from 10 to 50 ı C (not shown) provide activation energies of 0.48 eV
for emission and 0.25 eV for capture. Similarly thermally activated capture and
emission times are also observed in both nFET and pFET (i.e., corresponding to
NBTI) with conventional SiO2 gate oxide [14, 23, 24]. One can therefore conclude
for all these cases that both emission and capture in both electron and hole gate
oxide traps are without a doubt thermally activated processes. This experimental
fact is incompatible with direct elastic tunneling theories widely used in different
oxide trap characterization techniques and calculations. Consequently, a new model
that takes into account this thermal dependence has to be considered.
8 Modeling Properties of Individual Defects
A model of the above-described properties of individual gate oxide defects can be

constructed by drawing on the above similarities with low-frequency and Random
Telegraph Noise (RTN) [25]. An example of the configuration coordinate diagram of
the model is shown in the inset of Fig. 2.9. Four different configurations of the defect
are considered [14]. Two of the states are electrically neutral while two of them
correspond to the singly positively charged state. In each charge state the defect is
represented by a double well, with the first of the two states being the equilibrium
state and the other a secondary (metastable) minimum. The time dynamics of the
defect can be described by a simple stochastic Markov process. Broadly, transition
rates between states involving charge transfer assume (1) tunneling between the
substrate and the defect, and (2) nonradiative multiphonon (NMP) theory, which
has been often applied to explain RTN [26, 27]. Introduction of the NMP theory
naturally explains the temperature dependence of both capture and emission time
constants observed in the previous section. The wide distribution of time scales is
then readily described by a distribution of the overlaps of the potential wells (i.e., a
distribution of “potential barriers”) [14].
The crucial extension of the NMP theory is the assumption of the relative position
of the potential wells changing with gate bias [14], quite naturally introducing the
required strong VG dependence. As documented in Fig. 2.9, the model successfully
describes the bias as well as the temperature dependences of the characteristic time
constants. It is also noted that, contrary to techniques for the analysis of RTN, which
only allow monitoring the defect behavior in a rather narrow time window, TDDS
can be used to study the defects capture and emission times over an extremely
wide range.
Fig. 2.9 Simulated capture and emission time constants (lines) compared with to the experimental
TDDS values obtained on SiO2 pFETs (symbols) during NBTI stressing at 125 and 175 ı C and
varying VG . The experimental occupation probability of the charged state fp is also indicated. The
configuration coordinate diagram is shown in the inset (dashed line: neutral defect state; solid line:
charged state potential)
Fig. 2.10 Simulated RTN, stress, and recovery behavior of a nano-scale device using a stochastic
solution algorithm of the proposed model. (a) At the threshold voltage (Vg1), the RTN is dominated
by defect #5 with the occasional contribution from defect #3. Defects #1, #2, and #4 remain
positively charged within the ‘simulation/experimental’ window. (b) During stress (Vg2), the
capture times are dramatically reduced by the higher (more negative) gate voltage and the defects
#3 and #5 become predominantly positively charged (£c £e). Defects #1, #2, and #4 start
producing RTN. (c) During recovery (back at Vg1), trapped charge is subsequently lost and the
dynamic equilibrium behavior is gradually restored
It has been previously argued that the phenomenon called NBTI relaxation in
pFET devices is in fact just a different facet of the well-known low-frequency noise
in these devices. While the low-frequency noise corresponds to the channel/gate
dielectrics system being in the state of dynamic equilibrium, NBTI relaxation
corresponds to the perturbed system returning to this equilibrium [28]. Figure 2.10
14 B. Kaczer et al.
Fig. 2.11 Histogram of

NBTI transient individual
step heights measured on 72
devices shows a clear
exponential distribution. The
average Vth shift
corresponding to a single
carrier discharge is
4.75 ˙ 0.30 mV in the pFETs
with metallurgic length
L D 35 nm, width
W D 90 nm, and HfO2
dielectrics with
EOT D 0.8 nm
then illustrates this concept on a simulated example of a deeply scaled pFET

containing only five active defects [23]. In particular it shows that the same defects
can be responsible both for RTN as well as the NBTI relaxation and the (initial
phase of) NBTI stress.
9 BTI Distribution in Deeply-Scaled FETs
As CMOS devices scale toward atomic dimensions, device parameters become

statistically distributed. Similarly, parameter shifts during device operation, once
studied in terms of the average value only, will have to be described in terms
of their distribution functions. The understanding of the properties of individual
defects helps us to explain this distribution. Namely, much like in the case of random
telegraph noise (RTN) [29, 30], it is observed the distribution of down-steps Vth
due to individual discharging events to be exponentially distributed (Fig. 2.11). The
exponential distribution of single-charge Vth can be understood if non-uniformities
in the pFET channel due to random dopant fluctuations (RDF) are considered
[28–30]. A single discharging event in many devices routinely exceeded 15 mV,
and in several devices exceeded 30 mV, the NBTI lifetime criterion presently used
by some groups. For comparison, Vth of less than 2 mV would be expected based
on a simple charge sheet approximation. The large observed step height amplitude
is due to the aggressively scaled dimensions of the pFETs used [28, 31].
Assuming the lateral locations of the trapped charges are uncorrelated, the
overall Vth distribution can be readily expressed as a convolution of individual
exponential distributions [28, 31]. An actual population of stressed devices will
4
99.99
3 99.9
99
2 97
90
1
percentile
probit(FN)
70
0 N = 1, 2, .., 20 50
30
-1
10
-2 N=6 η ≈ 4 mV 3
1
N=9
-3 0.1
N = 14
0.01
0 40 80 120 160
-ΔVth (mV)
Fig. 2.12 Equation (2.1) in a probit plot rescaled to fit experimental distributions from Fig. 2.10
of Ref. [22], with the corresponding values of N and readily extracted
consist of devices with a different number n of oxide defects in each device, with n
being Poisson distributed [12, 22, 28]. The total Vth distribution can be therefore
obtained as
X1
e N N n n .n; Vth =n/
FN .Vth ; / D 1 ; (2.1)
nD0
nŠ nŠ
where N is the mean number of defects in the FET gate oxide and is related to the
oxide trap (surface) density Not as N D W L Not . The CDF is plotted in Fig. 2.12
for several values of N. For comparison, measured total Vth distributions for three
different stress times from Ref. [22] are excellently fitted by the derived analytical
description.
The advantage of describing the total Vth distribution in terms of Eq. (2.1) is its
relative simplicity and tangibility of the variables. The analytical description allows,
among other things, to calculate NBTI threshold voltage shifts in an unlimited
population of devices, a feat practically impossible through device simulations.
10 Technological Solutions
Once the underlying BTI mechanisms are understood, the defect properties can be
modified to beneficial ends. Below, two possible technological solutions for both
PBTI and NBTI are discussed.
16 B. Kaczer et al.
100
10-1
ΔVth (V) ref
10-2 La
ref T = 125⬚C
10-3 trelax = 10-3s
lines: tstress = 10 years
La
symbols: tstress = 6x103s
10-4
1 2 5 10
Eox (MV cm-1)
Fig. 2.13 A significant reduction of PBTI threshold voltage shift is observed in planar nFETs
with La passivation (“La”) over the reference stack (“ref”) without passivation. Simplified power-
law projection to 10 years shows passivated stack having sufficient reliability (Vth < 30 mV) at
5 MV/cm operating field
11 Improving PBTI with Rare-Earth Incorporation
PBTI was considered a minor problem in technologies based on SiO2 . It arose

as a reliability issue when high-k materials were incorporated into the gate stack.
However, when rare earths were introduced to adjust the nFET initial threshold
voltage, this issue was mitigated, as can be seen in Fig. 2.13. A significant reduction
of PBTI is observed in planar nFETs with lanthanum with respect to a lanthanum-
free reference [32].
Positive BTI in nFETs with high-k materials like HfO2 has been linked to oxygen
vacancies, which produce a defect level in the upper part of the oxide band gap.
Group III elements compensate unpaired electrons around the oxygen vacancy in
HfO2 and the defects are “passivated” by being pushed up toward the conduction
band minimum [33]. Such states are not easily accessible to nFET channel electrons,
resulting in the significant reduction of negative charge capture in the stack and
hence the reduction of PBTI.
12 Improving NBTI in High-Mobility SiGe pFETs
Reduction of gate stack EOT, which is one of the most efficient ways to improve
FET performance, enhances NBTI due to increased oxide electric field. As a
consequence, 10 year lifetime can be guaranteed for sub-1 nm EOT Si pFETs only
at gate overdrive voltages far below the expected operating voltages (Fig. 2.14).
Another way to improve FET performance is the use of high-mobility substrates,
such as buried-channel SiGe [34]. Because of the valence band offset between the
Fig. 2.14 Plot of the operating overdrive voltage jVG Vth j for 10 year lifetime assuming a 30 mV
threshold voltage shift criterion vs. the inversion capacitance equivalent thickness Tinv , for Si
channel devices with different processing used as a reference, and for SiGe pFETs. For low Tinv ,
Si devices jVG Vth j is below the expected operating voltage. In contrast to that, optimized SiGe
devices show improved lifetime. Inset: gate-stack band diagram in inversion. Si cap acts as a barrier
EV for holes
SiGe and the Si cap (see inset of Fig. 2.14), inversion channel holes are confined in
the SiGe layer, which therefore acts as a quantum well (QW) for holes. The Si cap
lowers the inversion capacitance as compared to the accumulation capacitance. For
these devices it is therefore necessary to report the capacitance-equivalent thickness
in inversion (Tinv , evaluated at VG D Vth 0.6 V) which will be affected by the
thickness of the Si cap [35].
As can be seen from Fig. 2.14, SiGe-based device gate stacks significantly
increase operating gate overdrive while still guaranteeing 10 year device lifetime
and at the moment seem to be the only solution to the NBTI issue for sub-1 nm
EOT devices. It has been recently observed that both increasing the Ge content
in the channel as well as increasing the SiGe QW thickness reduces NBTI. Most
intriguingly, a reduction of Si cap thickness also diminishes NBTI [35]. The most
likely hypothesis explaining all three trends appears to be the energetic decoupling
of the buried channel and the gate oxide defects [36].
Conclusions
In this chapter some of the shifts occurring in the past few years in our
understanding of BTI were reviewed. Among the most significant ones is the
shift from perceiving NBTI in terms of the Reaction–Diffusion model to ana-
lyzing BTI with the tools originally developed for describing low-frequency
(continued)
18 B. Kaczer et al.
noise. This includes the interpretation of the time, temperature, voltage, and
duty cycle dependences of BTI. It was further demonstrated that a wealth
of information about defect properties can be obtained from deeply-scaled
devices, and that this information can allow interpretation of variability issues
of future deeply downscaled CMOS devices. This theme was complemented
by showing the most promising technological solutions to alleviate both PBTI
and NBTI.
References
1. J. H. Stathis and S. Zafar, Microelectronics Reliab. 46, 270 (2006).

2. V. Huard, M. Denais, C. Parthasarathy, Microelectronics Reliab.46, 1 (2006).
3. D. K. Schroder, Microelectronics Reliab. 47, 841 (2007).
4. B. Kaczer, R. Degraeve, V. Arkhipov, N. Collaert, G. Groeseneken, M. Goodwin, as discussed
at SISC, San Diego, CA, 2006.
5. B. Kaczer, V. Arkhipov, R. Degraeve, N. Collaert, G. Groeseneken, and M. Goodwin, Proc.
Int. Reliab. Phys. Symp., 381 (2005); Appl. Phys. Lett.86, 143506 (2005).
6. T. Grasser, B. Kaczer, P. Hehenberger, W. Goes, R. O’Connor, H. Reisinger, W. Gustin, and
C. Schlunder, Int. Electron Devices Meeting Tech. Dig., 801 (2007).
7. A. Kerber, K. Maitra, A. Majumdar, M. Hargrove, R. J. Carter, and E. A. Cartier, IEEE T.
Electron Dev.55, 3175 (2008).
8. M. A. Alam, Int. Electron Devices Meeting Tech. Dig., 345 (2003).
9. S. V. Kumar, C. H. Kim, S. S. Sapatnekar, IEEE/ACM International Conference on Computer-
Aided Design ICCAD’06, 493 (2006).
10. T. Grasser, B. Kaczer, W. Goes, H. Reisinger, Th. Aichinger, Ph. Hehenberger, P.-J. Wagner,
F. Schanovsky, J. Franco, Ph. Roussel, and M. Nelhiebel, to be presented at Int. Electron
Devices Meeting 2010.
11. T. Grasser, B. Kaczer, and W. Goes, Proc. Int. Reliab. Phys. Symp, 28 (2008).
12. B. Kaczer, T. Grasser, J. Martin-Martinez, E. Simoen, M. Aoulaiche, Ph. J. Roussel, and
G. Groeseneken, Proc. Int. Rel. Phys. Symp., 55, 2009.
13. T. Grasser, B. Kaczer, W. Goes, Th. Aichinger, Ph. Hehenberger, and M. Nelhiebel, “A Two-
Stage Model for Negative Bias Temperature Instability”, Proc. Int. Reliab. Phys. Symp., 2009.
14. T. Grasser, H. Reisinger, P.-J. Wagner, F. Schanovsky, W. Goes, and B. Kaczer, Proc. Int. Rel.
Phys. Symp., 16 (2010).
15. S. Rangan, N. Mielke, and E. C. C. Yeh, Int. Electron Devices Meeting Tech. Digest, 341
(2003).
16. G. Chen, K. Y. Chuah, M. F. Li, D. S. H. Chan, C. H. Ang, J. Z. Zheng, Y. Jin, and
D. L. Kwong,Proc. Int. Reliab. Phys. Symp., 196 (2003).
17. R. Fernández, B. Kaczer, A. Nackaerts, S. Demuynck, R. Rodríguez, M. Nafría, G. Groe-
seneken, Int. Electron Devices Meeting Tech. Dig.. 1 (2006).
18. H. Reisinger, O. Blank, W. Heinrigs, A. Mühlhoff, W. Gustin, and C. Schlünder, Proc. Int.
Reliab. Phys. Symp., 448 (2006).
19. B. Kaczer, T. Grasser, Ph. J. Rousse, J. Martin-Martinez, R. O’Connor, B. J. O’Sullivan,
G. Groeseneken, Proc. Int. Reliab. Phys. Symp., 20 (2008).
20. E. Milotti, arXiv:physics/0204033v1.
21. H. Reisinger, T. Grasser, W. Gustin, and C. Schlünder, Proc. Int. Reliab. Phys. Symp., 1 (2010).
22. V. Huard, C. Parthasarathy, C. Guerin, T. Valentin, E. Pion, M. Mammasse, N. Planes and

L. Camus, Proc. Int. Reliab. Phys. Symp, 289 (2008).
23. T. Grasser, H. Reisinger, W. Goes, Th. Aichinger, Ph. Hehenberger, P.-J. Wagner, M. Nelhiebel,
J. Franco, and B. Kaczer, Int. Electron Devices Meeting Tech. Dig., 729 (2009).
24. M. Toledano-Luque, B. Kaczer, Ph. Roussel, M.J. Cho, T. Grasser, and G. Groeseneken,
presented at WoDiM 2010.
25. M. J. Uren, M. J. Kirton, and S. Collins, Phys. Rev. B37, 8346 (1988).
26. M. J. Kirton and M. J. Uren, Adv. Phys.38, 367 (1989).
27. A. Palma et al., Phys. Rev. B56, 9565 (1997).
28. B. Kaczer, T. Grasser, Ph. J. Roussel, J. Franco, R. Degraeve, L.-A. Ragnarsson, E. Simoen,
G. Groeseneken, H. Reisinger, Proc. Int. Reliab. Phys. Symp., 26 (2010).
29. A. Asenov, R. Balasubramaniam, A. R. Brown, J. H. Davies, IEEE T. Electron Dev.50, 839
(2003).
30. A. Ghetti, C. M. Compagnoni, A. S. Spinelli, A. Visconti, IEEE T. Electron Dev.56, 1746
(2009).
31. B. Kaczer, Ph. J. Roussel, T. Grasser, and G. Groeseneken, IEEE Electron Dev. Lett.31, 411
(2010).
32. B. Kaczer, A. Veloso, M. Aoulaiche, and G. Groeseneken, Microelectronic Engineering,
86(7-9), 1894 (2009).
33. D. Liu and J. Robertson, Appl. Phys. Lett. 94, 042904 (2009).
34. N. Collaert, P. Verheyen, K. De Meyer, R. Loo and M. Caymax, IEEE T. Nanotech1, 190
(2002).
35. J. Franco, B. Kaczer, M. Cho, G. Eneman, G. Groeseneken, and T. Grasser, Proc. Int. Rel.
Phys. Symp., 1082 (2010).
36. J. Franco, B. Kaczer, G. Eneman, J. Mitard, A. Stesmans, V. Afanas’ev, T. Kauerauf, Ph. J.
Roussel, M. Toledano-Luque, M. Cho, R. Degraeve, T. Grasser, L.-Å. Ragnarsson, L. Witters,
J. Tseng, S. Takeoka, W.-E. Wang, T.Y. Hoffmann, and G. Groeseneken, Int. Electron Devices
Meeting 2010.
Chapter 3
Charge Trapping Phenomena in MOSFETS:
From Noise to Bias Temperature Instability
Gilson Wirth and Roberto da Silva
Abstract Charge trapping phenomena is known to be a major reliability concern

in modern MOSFETS, dominating low-frequency noise behavior and playing a
significant role in aging effects such as Bias Temperature Instability (BTI). In this
chapter we address this reliability issue.
We start by discussing MOSFET low-frequency (LF) noise, which is known to

be dominated by charge capture and emission by defects (traps) close to the
semiconductor–dielectric interface. Standard (LF) noise models used today (e.g.
BSIM and PSP) do not properly model noise behavior under large signal excitation.
A circuit level modeling and simulation approach, valid at both DC and large signal
(AC) biasing, is presented.
The role of charge trapping and de-trapping in BTI (Bias Temperature Instability)
is also discussed and modeled.
Mutual relation between the different reliability phenomena (LF noise, BTI
and RDF) is also studied. For instance, random dopant fluctuations (RDF) may
exacerbate the impact of BTI and LF noise on circuit performance.
1 Introduction
The major goal of this chapter is to compile and critically discuss recent work per-
formed by the authors on modeling of charge trapping and de-trapping phenomena
in nanometer scale CMOS devices [1–6, 16–18]. The role of charge trapping and
de-trapping in both low-frequency noise and bias temperature instability (BTI) is
discussed.
G. Wirth ()
Electrical Engineering Department, Universidade Federal do Rio Grande do Sul (UFRGS),
e-mail: wirth@ece.ufrgs.br
R. da Silva
Physics Institute, UFRGS, Porto Alegre, Brazil

DOI 10.1007/978-1-4614-4078-9__3
22 G. Wirth and R. da Silva
Fig. 3.1 Traps within a few

kT from the Fermi Level
contribute to noise, by
switching their state between
occupied (by a charge carrier)
and empty
Fig. 3.2 If a transistor is abruptly turned on, the Fermi level abruptly changes and trap occupation
probability follows the Fermi level (EF ) change. Charge trapping/de-trapping is not an instanta-
neous event. It is governed by characteristic time constants, and the trap occupation reflects the
new Fermi level after an elapsed time. This figure depicts the situation for electron trapping in the
channel of a NMOS transistor
Charge trapping and de-trapping at localized states (charge traps) at the interface
or in the gate dielectric is a significant reliability issue for CMOS applications.
It is known to be a source of low-frequency noise, besides playing a role in Bias
Temperature Instability (BTI).
Traps that contribute to noise are the ones that keep switching their state between
occupied and empty, as depicted in Fig. 3.1. These are the traps with occupation
probability close to 50 %, which means that their capture and emission times are
similar.
Traps that contribute to BTI are the ones that have a high probability to stay
occupied after a charge trapping event, as depicted in Fig. 3.2. These are the traps
with occupation probability close to 100 %, which means that their capture time is
much shorter than the emission time.
Figure 3.3 depicts the mechanisms leading to BTI and low-frequency noise.
If a constant bias is applied to a Metal–Insulator–Semiconductor system, only
the localized states (traps) with energy close to the Fermi level show significant
activity, i.e., change their states between occupied and empty, by capture and
subsequent emission of charge carriers from the channel region. This may be
considered a steady state condition, where the occupation probability of a trap is
time independent, i.e., a constant that does not change with time. This mechanism
is known to originate low-frequency noise.
However, if the bias point changes abruptly, as is common in e.g. digital CMOS
applications, the trap occupation probability abruptly changes. For instance, if the
3 Charge Trapping Phenomena in MOSFETS: From Noise to Bias. . . 23
Fig. 3.3 (Left hand side) Traps that contribute to NBTI are the ones that stay occupied after
an capture event occurs. This leads to a degradation of transistor on current over time. (Right
hand side) Traps that contribute to noise are the ones that keep switching their state over time,
exchanging charge carriers with the channel inversion layer. This leads to Random Telegraph Noise
in the device current
gate voltage abruptly changes the Fermi level in such a way that trap occupation
probability is increased, the rate at which charge carries are captured abruptly
becomes larger than the rate at which carriers are emitted, and the number of trapped
charge increases over time. Traps change their occupation state according to their
characteristic time constant, meaning that the number of trapped charge does not
instantaneously reflect the new occupation probability. The faster traps (with shorter
capture time constants) become filled first, while the slowest traps take longer to
become filled.
Each trap that becomes occupied degrades the channel conductivity, decreasing
the device current. For nanometer scale devices, this current decrease is seen to
occur in discrete steps, each step being related to the capture of a single channel
carrier (in large area device it becomes hard to clearly identify the discrete steps).
Since the dynamics of this occupation depends on the bias point and temperature, it
may lead to bias temperature instability (BTI).
The first sections of this chapter cover the modeling of low-frequency noise,
while the last section covers the modeling of BTI.
2 Charge Trapping Events as a Source of Noise
In nanometer scale MOSFETs the alternate capture and emission of carriers at

individual defect sites (traps) generates discrete fluctuations in the device conduc-
tance. These fluctuations, also called Random Telegraph Noise (RTN), are the main
source of Low-Frequency Noise in deep-submicron MOSFETs. This work covers
measurement, analytical analysis and Monte Carlo simulation of these fluctuations.
The low-frequency noise model is based on device physics parameters which
cause statistical variation in low-frequency noise behavior of individual devices.
It includes detailed consideration of statistical effects for distribution of number
of traps per device, the trap energy distribution, trap location and its device bias
dependent noise contributions [1, 2]. Microscopic discrete quantities are used
in model derivation, and analytical equations for the statistical parameters are
provided.
In many practical applications the MOS device is not biased at steady state, but
periodically switched. This operation regime is called cyclo-stationary excitation.
The modeling approach adopted allows analysis of noise both at steady state
operation as well as under cyclo-stationary excitation. Noise behavior under cyclo-
stationary excitation is the main focus of this work.
First the autocorrelation of the RTN signal is calculated, and then the Wiener–
Khinchin formula applied, leading to an analytical formulation for the RTN
spectrum due to a single trap. After the evaluation of the power spectrum due to a
single trap, the noise behavior resulting from the combined effect of all traps found
in a device is derived. The modeling approach is valid for steady state as well as
for any periodic excitation signal, and allows the derivation of relevant statistical
parameters. Square wave excitation is used as a case study, to explore noise behavior
in detail.
It is shown that since RTN-amplitude depends on the bias point strong variations
of noise performance may appear not only between devices, but also for a single
device operated under different bias conditions.
If the trap energy distribution in the band-gap is a convex curve (e.g. “U”-
shaped), the model here presented yields a reduction in LF-noise under cyclo-
stationary excitation when the device is biased between strong and weak inversion
or even slight accumulation, which is in agreement to our experimental results [3, 7],
and experimental results found in the literature [8–11]. However, while the average
noise power is seen to decrease, the variability (normalized standard deviation) of
noise power increases, as shown by the mathematical derivation, and experimental
observations provided here.
3 Power Spectrum of the RTN Noise due to a Single Trap
The origin of RTN noise is the alternate capture and emission of charge carriers at
discrete trap levels near the Si–SiO2 interface. Figure 3.4 depicts the cross section of
an n-channel MOSFET through the location of the interface trap. The influence of
the traps on the electrical current flowing through the channel is twofold. On the one
hand, the occupation of a trap changes the number of free carriers in the inversion
layer. On the other hand, a charged trap state has an influence on the local mobility
near to its position due to Coulomb scattering. If the MOSFET biasing is kept
constant, a stationary RTN is observed at the terminals of the device as a discrete
fluctuation in electrical current, ıId being the amplitude of the current fluctuation,
as shown in the inset of Fig. 3.5. The average high current time corresponds to the
electron capture time constant ( c ), and the average low current time corresponds to
the emission time constant ( e ).
The power spectrum of a RTN fluctuation can be evaluated by calculating the
auto-covariance of the signal, and then applying the Wiener–Khinchin formula [12].
The power spectrum is then the Fourier transform of the auto-covariance, and is a
Lorentzian, as given by Eq. (3.1) and depicted in Fig. 3.5.
ı2 ˇ 1 1
S .!/ D (3.1)
.1 C ˇ/2 !0 1 C .!=!0 /2
where ı is the RTN amplitude, ¨0 is the angular corner frequency and “ is equal
to £c /£e . The amplitude of the RTN fluctuation induced by a trap depends on trap
position along the channel. It may be exacerbated by factors such as random dopant
fluctuations (RDF). For nanometer scale channel lengths (L), traps near the source
end of the channel can cause significant mobility fluctuations and surface potential
Trap
Gate Oxide
Inversion Layer
Drain Current ID
Fig. 3.4 Schematic cross section of the inversion layer of a MOS transistor through the location
of an interface trap. If the trap is electrically charged the inversion layer is disturbed by the trap,
affecting the drain current ID . The trap not only affects the number of free carriers in the inversion
layer but is also a source of electrical charge carrier scattering
S(log)
Current
τc
δΙd
τe
Time
f(log)
Fig. 3.5 Time and frequency domain representation of a stationary random telegraph noise (RTN).
In frequency domain, the power spectrum of a RTN is a Lorentzian. In time domain discrete
fluctuations are observed in the drain current, where c is the average time in the high current
state, which corresponds to the state where the trap is electrically neutral (empty). e is the average
time in the low current state, which corresponds to the state where the trap is electrically charged.
ıId is the amplitude of the current fluctuation
fluctuations, hindering electrical current flow and enhancing RTN amplitude (ı).
A trap positioned near the source junction may be able to create a repulsive Coulomb
blockade well surrounding the trap, hindering trap injection from the source into
the channel. Thus, a significant mobility fluctuation will be added to carrier number
fluctuation for carrier electrons trapped near the source side [17]. Another parameter
that may impact RTN amplitude is trap position into the dielectric.
Equation (3.1) assumes that the capture and emission time constants ( c and
e ) are time independent, constant values. This is true for constant (time invariant)
biasing. However, under cyclo-stationary excitation the applied bias voltage is a
periodic function of time, and c (t) and e (t) become periodic functions of time.
Cyclo-stationary excitation is depicted in Fig. 3.6.
In order to derive the power spectrum under cyclo-stationary excitation, we did
follow the methodology originally proposed by Machlup [12] for stationary RTN.
A RTN is considered to be a purely random signal, which may be in one of two
states, called 1 and 0. If the signal is in state 1, the probability of making a transition
to 0 in a short time dt is assumed to be dt/ c (t). If the signal is in state 0, the
probability of making a transition to 1 is assumed to be dt/ e (t). In this form, the
state 1 is related to the empty trap, i.e., high current state in Fig. 3.5, while the state
0 is related to the occupied trap, i.e., low current state.
Fig. 3.6 Energy band Gate voltage applied to the device

diagram of MOSFET with Vgs,on
noise relevant traps during
two different phases (dashed Vgs,off
and dotted) of square wave
gate voltage biasing as shown
on top of the figure. Note that E
traps close to Fermi level both
in ‘on’ and ‘off ’ state can
contribute to the noise. gt (E) Ef
shows a U-shaped trap ET
density gt (E)
Gate Oxide Substrate
In order to derive the low frequency noise spectrum of cyclo-stationary RTN, we

first calculate the autocorrelation of the RTN, and then apply the Wiener–Khinchin
formula to obtain the spectrum. Let the RTN signal be x(t).
The autocorrelation is then given by
A .s/ D< x .t/ x .t C s/ >average D P .x .t/ D 1/ P11 .s/ (3.2)
where P11 (s) is the probability of an even number of transitions in time s, given we
start in state 1. P(x(t) D 1), the probability of being in state 1 at time t, is given by
Z T
1 1
T
dt < £e1.t/ >
0 e .t/
£
P .x .t/ D 1/ D Z T Z T D (3.3)
1 1 1 1 < £e1.t/ > C < £c1.t/ >
T
d tC T dt
0 £e .t/ 0 £c .t/
where T is the periodR of the cyclo-stationary excitation, and symbol <•> is an

abbreviation for (1/T) 0 T •dt, the time average value.
The autocorrelation can then be calculated as
1
< £e .t/
>
A .s/ D :
< > C < £c1.t/ >
1
0 £e .t/Z x 1 Z s
Z s 1 1
< >C< > dy
B £c .y/ £e .y/ C
: @1 C e 0 1
£e .x/
d xA e 0
0

< 1
£c .y/
>C< 1
£e .y/
> dy
(3.4)
This formulation for the autocorrelation is a generalization of the Machlup formula,
and is valid for any kind of periodic excitation and any frequency. If c and e
become constant (independent of time), Eq. (3.4) becomes equal to (7) in [12].
4 Approximation for Excitation Frequencies Higher than the

Noise Frequency
The case of interest for most practical applications is for the noise at frequencies
below the frequency of the cyclo-stationary excitation signal.
If the probability of a trap to switch state during one period T of the cyclo-
stationary excitation signal is very small, a simplification may be done in the
calculation of the autocorrelation. This case corresponds to the limit where c (t)
and e (t) are much larger than the period T, leading to small transition probabilities
T/ c (t) and T/ e (t).
In this case we can write, without loss of generality, that s D nT, where n is a
positive integer (1, 2, 3, : : : ). In this situation we have
Z s n1 Z .iC1/T
X Xn1 Z .iC1/T
1 1 1 1
C dy D d yC dy D
0 £e .y/ £c .y/ iD0 iT
£e .y/ £c .y/
iD0 iT
D nT < £e .t/ > C < £c1.t/ >

1
(3.5)
leading to
" #
< £ 1.t/ > T< £ 1.t/ > -s < £ 1.t/ >C< £ 1.t/ >
A .s/ D e
< £ 1.t/ >C< £ 1.t/ >
1 e e e c C
e c T < 1.t/ >C< 1.t/ >
£e £c
e 1
(3.6)
< £ 1.t/ > T< £ 1.t/ >
C < £ 1.t/ >C< £ 1.t/ >
e e
e c T < 1.t/ >C< 1.t/ >
£e £c
e 1
For small values of T a Taylor expansion may be employed, leading to

< £ 1.t/ > < £ 1.t/ > -s < £ 1.t/ >C< £ 1.t/ >
A .s/ D < £ 1.t/ >C< £ 1.t/
e
>
1 e
< £ 1.t/ >C< £ 1.t/ >
e e c C
e c e c
(3.7)
< £ 1.t/ >2
C e
2
< £ 1.t/ >C< £ 1.t/ >
e c
This means that, if c (t) and e (t) are much larger than the period T of the excitation
signal, the values of 1/£c (t) and 1/£e (t) in the integrals of (3.4) are equivalent to their
time averages <1/£c (t)> and <1/£e (t)>.
The power spectrum Si (!), due to a single trap (the i-th trap), is the calculated as
the Fourier transform of the autocorrelation, leading to
Z 1
1
Si .!/ D A.s/e i!s ds (3.8)
2 1
Fig. 3.7 Noise Power

Spectral Density as a function
of “eq . For “eq D 1, i.e.,
<1/£e > D <1/£c >, the noise
power reaches its maximum
Noise PSD
10-1
10-2
10-2 100 102
βeq
which is evaluated as being
ıi2 ˇeq 1 1
Si .!/ D (3.9)
1 C ˇeq 2 !i 1 C .!=!i /2
Here •i determines the trap’s voltage amplitude and ¨ i the angular frequency, and
“eq D <1=£e .t/ >=<1=£c .t/ > (3.10)
Figure 3.7 depicts the behavior of noise power as a function of “eq .

The cyclo-stationary noise spectrum is still Lorentzian, with angular corner
frequency ¨i given by
1 1
¨i D < >C< > (3.11)
£c .t/ £e .t/
This takes us to the conclusion that making a RTN signal cyclo-stationary leads to
a Lorentzian spectrum with corner frequency equal to the sum of the inverse time
average values of the capture and emission times. Please note that for stationary
RTN the corner frequency is equal to the sum of the inverse values of the constant
capture and emission times. Please refer to equation 9 in [12].
The result for this limit is valid for any kind of periodic excitation. This is the
limit studied in [11, 13]. However, for this limit, we obtain the same result in a much
simpler derivation than in [11, 13], and without making any further assumption or
simplification. The single assumption is that the transition probabilities dt/£e (t) and
dt/£c (t) are much smaller than the excitation period T.
Equations (3.11) and (3.9) are a generalization of the Machlup formulation for
cyclo-stationary RTN with excitation frequency higher than the noise frequency.
The Machlup equations for the auto-correlation and power spectrum are recovered
if we consider <1/£e (t)> D 1/£e and <1/£c (t)> D 1/£c , i.e., constant, not time
dependent values.
5 Average Power Spectrum of the RTN Noise due

to the Ensemble of Traps
The noise behavior of a device results from the combined effect of all traps found in
the device. The noise power spectrum may then be written as the summation of the
contribution of each one of the Ntr traps found in the device. The observation that
traps are Poisson distributed results in an average noise spectral density given by
1
X X
Nt r
e N NtNr
<S > D <Si > D N <Si > (3.12)
Nt r D0 iD1
Nt r Š
Here <Si > is the average noise contribution of a trap, Ntr is actual number of traps
in a particular device, and N is average number of traps in an ensemble of devices.
As depicted in Fig. 3.8, if the corner frequencies of the Lorentzians correspond-
ing to power spectral density (PSD) of different traps are equally spaced on a log
scale, the summation of the power spectrum due to all traps leads to 1/f noise. In
time domain, this corresponds to the situation where the charge trap characteristic
time constants are uniformly distributed on a log scale.
The equations above are valid for any waveform of the periodic excitation signal.
In order to explore the noise behavior in detail and allow comparison to experimental
results, a case of particular interest for practical applications will be studied. It is
square wave excitation.
6 Square Wave Excitation
Square wave excitation is chosen as case study, because of its interest in practical
applications and because of the availability of experimental data. Comparison of
model results to relevant experimental data is performed in the last section of this
work.
Under square wave excitation, the bias voltage abruptly alternates between two
states, called on and off. Please note that the state names on and off do not imply that
the device has necessarily to be periodically turned on and off. The names refer to
two distinct states, with different gate bias. With periodically changing gate bias, the
Fermi level becomes a periodic function of time EF (t). This implies that the Fermi
level alternates between two levels: Eon being the Fermi level during the on state
and Eoff being the Fermi level during the off state. The duty cycle ˛ is the fraction
102
101
100
10−1
10−2
S(f)
10−3
10−4
10−5
10−6
10−7
10−2 10−1 100 101 102 103 104 105 106
f(Hz)
Fig. 3.8 If the corner frequencies of the Lorentzians (red lines) corresponding to different traps
are equally spaced on a log scale, the summation of the power spectrum due to all traps leads to
1/f noise (blue line)
of the period T in which the device is in the on state. The capture and emission time
constants of a trap are affected by the Fermi level.
For square wave cyclo-stationary excitation with duty cycle ’, the time averaged
capture and emission time constants may be written as

<1=c > D ˛=c;on C .1 ˛/ =c;off (3.13)

<1=e > D ˛=e;on C .1 ˛/ =e;off (3.14)

ˇeq D Eon ; Eoff ; ˛ e 2Et =kB T (3.15)
with
˛e Eon =kB T C .1 ˛/ e Eoff =kB T

Eon ; Eoff ; ˛ D (3.16)
˛e Eon =kB T C .1 ˛/ e Eoff =kB T
Assuming statistical independence of the random variables, the average noise of a

transistor <S> is then given by:
Z Ec
Ndec < ı 2 > e 2Et =kB T
<S > D 2 g .Et / dEt (3.17)
W L! Ev 1 C e 2Et =kB T
where W is the channel width and L the channel length of the transistor, and Ndec is
the trap density per unit area and frequency decade. Please see (3.1) for a detailed
explanation of these parameters. This equation again looks similar to the DC noise
behavior (see [1]). In case the cyclo-stationary period T is short compared to the
trap’s time constants their time averages and the trap numbers close to the two
Fermi levels determine the noise level. At frequencies significantly higher than the
excitation frequency the noise is given by DC theory.
Equation (3.17) above clearly relates the noise reduction to the distribution of
traps over energy. From the above equation, it follows that if the distribution of traps
in energy over the bandgap is uniform, <S> is expected to remain approximately
constant under cyclo-stationary excitation. In this case, the peak of the noise
contribution of a trap always probes the same trap density, since the trap density
around to the Fermi level is always the same. Note that for the peak of the noise
contribution of a trap occurs for ET D EF . If the trap density is uniform over the
bandgap, g(Et ) corresponding to the peak of “eq is always the same.
Monte Carlo (MC) simulations are performed in time domain. The RTN of each
trap is evaluated, and then the power spectral density is calculated. The Monte Carlo
simulation show very good agreement to the analytical results, as well as to our
experimental results, as shown in the last section of this work.
In Fig. 3.9 the results of Eq. 3.17 show noise reduction as a function of the Fermi
levels in the ‘on’ and ‘off’ state at frequencies lower than the excitation period.
0.02
S
0.015
[a.u]
0.01
0.005
1
0.8 Eoff=f(VGS,off)
0.6
0.4 1
Eon=f(VGS,on)
Fig. 3.9 Noise reduction as a function of the Fermi level in the ‘on’ and ‘off’ states. The noise
power S is evaluated according to Eq. (3.17), considering the U-shaped trap density given by
Eq. (3.18). The same parameter a D 11 is used in all figures and evaluations performed in this
work. Note that the noise reduction is larger if the biasing levels are symmetrical in relation to the
center of the U-shaped trap density
It can be seen that a higher noise reduction can be expected if the biasing levels are
symmetrical in relation to the center of the U-shaped trap density. In this case, the
traps that contribute most to the noise power are the ones close to the center of the
bandgap, where the density is lowest.
In good agreement to the works in [11, 13], the trap energy distribution g(Et )
is key for explaining experimentally observed findings. A parabolic U-shape trap
density function is assumed here:
g .Et / D aEt2 a .Ec Ev / Et C k (3.18)
where a is a fitting parameter. The integral of g(Et ) from Ev to Ec is normalized to

one with
1 a a 2
kD Ec3 Ev3 C Ec Ev Ev2 Ec C 1 (3.19)
Ec Ev 6 2
Figure 3.10 depicts a U-shape trap density.
7 Variability in the Power Spectrum of the RTN Noise due

to the Ensemble of Traps
Since the microscopic approach maintains the statistically relevant parameters, the
model proposed here also allows the modeling of the statistically relevant param-
eters. As derived in [2], the normalized standard deviation of noise performance
under constant biasing is
2.5
2
g(ET) [a.u.]
1.5
Fig. 3.10 U-shaped trap 0.5

density. The trap density is
higher close to the valance
0
(EV ) and conduction (EC ) Ev (Ec+Ev)/2 Ec
bands, and lower in the center
ET
of the bangap
r
¢np 2 1 <A4 > b
D p (3.20)
<npBW > Ndec WL <A2 >2 fH c
fL
Here fH and fL are the lower and upper boundaries of the bandwidth of interest in
a given circuit design, respectively, and b and c are constants, with b D 0.74 and
c D 0.05 [2].
A similar formulation can also be derived for cyclo-stationary operation. The
standard deviation of noise power is given by:
p
S D <S 2 > <S >2 (3.21)
Hence, in addition to <S>2 based on Eq. (3.17), it is necessary to evaluate <S2 >.
Again following a similar approach to [12], the resulting normalized standard
deviation under square wave excitation is given by:
Z 4Et =kB T
! 12
Ec 2
q e
<ı 4 > 4 g .Et / dEt
S <ı 2 >2 Ev 1 C e 2Et =kB T
Dp Z Ec
(3.22)
<S > Ndec W L e 2Et =kB T
2 g .Et / dEt
Ev 1 C e 2Et =kB T
Figure 3.11 shows the normalized standard deviation as a function of the Fermi
levels in the ‘on’ and ‘off’ state for the U-shaped trap density given by (3.18). The
normalized standard deviation becomes higher when the traps that contribute most
to noise are the ones close to the center of the band gap. In this case the normalized
standard deviation increases even though the average noise power decreases.
relative error of S
10
8
6
4 0.4
0.6
1 0.8 Eoff=f(VGS,off)
0.8 0.6 1
0.4
Eon=f(VGS,on)
Fig. 3.11 Normalized standard deviation of noise performance as given by Eq. (3.22). An increase
in normalized standard deviation is seen under cyclo-stationary excitation, especially for conditions
where the average cyclo-stationary noise reduction is high
8 Experimental Results
The low frequency noise was experimentally investigated using short and long
channel NMOS devices from a 45 nm technology. Technology details are described
in [15]. The DC and large signal noise was measured using a semi-automatic prober
setup using a BTA noise system, and the results are described in [3]. A square wave
with 50 % duty cycle (’ D 0.5) was applied to the gate using a waveform generator
HP33120A for measuring cyclo-stationary noise. Two types of cyclo-stationary
measurements were performed, one with VGS,OFF D 0 V and VGS,ON D VT , and the
other with VGS,OFF D VT and VGS, ON D VDD . Note, that the effect of modulation for
a signal modulated by a square wave with 50 % duty cycle is a reduction of noise
power (in baseband) by a factor of four.
After the standard modulation theory, for square wave excitation, switching
operation can be represented by the multiplication of the noise with a square-wave
signal with 50 % duty cycle, m(t), as follows [19]:
1 2 2 2
m .t/ D C sin !sw t C sin 3!sw t C sin 5!sw t C : : :
2 3 5
In frequency domain this corresponds to a convolution of the noise PSD with a
spectrum with delta functions at dc, ¨sw , 3¨sw , 5¨sw , and so forth. The dc-term
determines the resulting noise power in baseband, which is (1/2)2 or 6 dB, if
compared to the original noise power.
However, the measurements results did show a much larger noise reduction. The
noise reduction was larger than a factor of ten, and the mean value of the LF-noise
under DC conditions was seen to follow 1/f behavior, but with deviations at some
frequencies, showing Lorentzian components [3]. For cyclo-stationary conditions
these deviations were more pronounced in agreement to our theory. Also in good
agreement to our theory, the noise reduction factor was found to be almost constant
if only measured results below the excitation frequency are considered.
To compare theory and experiment, the average number of traps for all possible
energetic levels in the bandgap were extracted from average noise behavior. The
model comparisons are based on the total number of traps and the assumption of a
U-shaped trap density with a D 11 (see Eq. 3.18).
Experimental results did show that whereas the average LF noise is decreasing
under cyclo-stationary conditions, the normalized standard deviation of noise power
increases. As shown in Fig. 3.12, this again is in good agreement to the model
presented here. In Fig. 3.12, the model values for mean and upper spec in cyclo-
stationary conditions are given by Eqs. 3.17 and 3.22, respectively. Here the upper
boundary for the DC noise is evaluated using the model presented in [1].
MEAN (MEAS)
UPPER SPEC (MEAS)
MEAN (SIM )
UPPER SPEC (SIM)
10-17
WL SID [A2 mm2 /Hz]
10-18
10-19
DC CYCLO DC CYCLO
L = 0.04 mm L = 1 mm
CHANNEL LENGTH
Fig. 3.12 Area normalized average and C3¢ worst case noise behavior under DC and CS
conditions. All data are taken at 10 Hz noise frequency. Open symbols are data from simulation
and closed symbols are measured data reported in [3]
9 The Charge Trapping Component of Bias

Temperature Instability
Above we discussed the low-frequency noise originated from trapping and de-
trapping at localized states (charge traps) at the interface or in the gate dielectric
of transistors. Here we discuss the role of charge trapping in Bias Temperature
Instability (BTI). It is widely accepted that charge trapping plays a role in the
threshold voltage shifts (VT ) produced by Bias Temperature stress. It was reported
that a significant fraction of the threshold voltage shift is recovered spontaneously
once the Bias Temperature stress is removed [20–22]. Although first observed
decades ago, the phenomenon still remains controversial in both experimental and
theoretical terms.
There is a vast literature on Negative Bias Temperature Instability (NBTI) in
PMOSFETs, where most models are based on reaction diffusion. Reaction diffusion
models involve breaking of Hydrogen-Silicon bonds at the Silicon-Gate Dielectric
interface, related to the trapping of inversion layer holes, with the release and
diffusion of a hydrogenic species. Recovery (relaxation) is assumed to occur with
re-bonding of the Hydrogen-Silicon bonds, i.e., anneal out the interface traps [20].
Although reaction diffusion models have been very useful and successful, some
aspects of NBTI are difficult to be fully explained in a reaction diffusion framework,
as for instance the fast recovery which occurs if bias stress is removed [20]. Other
phenomena, such as charge trapping, therefore have to be also considered in the

NBTI mechanism [22, 23]. Clear steps caused by single trapping or de-trapping
events were seen in experimental works, showing the discrete nature of VT shifts,
[22–24].
The discussion here presented will not focus on the origin and nature of the
charge traps. Detailing lattice dynamics, tunneling mechanisms, or determining the
physical location of trapping centers is also not the focus of this work. This work
focuses instead on the charge trapping statistics (stochastic capture and emission
events) that contribute to degradation of transistor on-current over time (BTI). The
basic assumptions made in the modeling of BTI here presented are the same ones as
in our work presented above, on the modeling of low-frequency noise: 1) charge
trapping and de-trapping are stochastic events, governed by characteristic time
constants, which are uniformly distributed on a log-scale; 2) the number of traps is
assumed to be Poisson distributed, and the parameter of the Poisson distribution is
assumed to be constant over the time interval of interest; 3) trap energy distribution
is assumed to be U-shaped (this last assumption is key to explain the AC-Behavior),
and 4) the amplitude of the fluctuation induced by a single trap is a random variable.
In this section we develop a theoretical analysis to describe the density of
occupied traps as a function of bias, temperature and time, aiming to understand
and model the charge trapping component of the aging (degradation) process that
occurs in MOSFETs.
The same equation (model) applies to the degradation process, i.e., increase in
number of occupied traps, as well as to the recovery process, i.e., decrease in number
of occupied traps.
The capture and emission of charge carriers by a trap are described as simple
Poisson processes governed by rates £c and £e , where the capture occurs with
probability p01 (dt) D dt/£c and emission occurs with probability p10 (dt) D dt/ e .
State 1 stands for the occupied trap, while state 0 stands for the empty trap. £c
and £e are then the average residence time in states 0 and 1, respectively.
We start by evaluating the device degradation process, which is described by the
average number of occupied traps at time t, which we denote <n(t)>. First we write
the equation for the probability of a particular trap, which is initially empty (state
0), to remain in the same state after an elapsed time t. We denote this probability as
P00 (t). This probability can be calculated observing that:
P01 .t C dt / D P01 .t /p11 .dt / C P00 .t /p01 .dt / (3.23)
where p11 (dt) D 1p10 (dt) D 1dt/£e and P00 (t) D 1P01 (t). This leads to a simple
differential equation. The solution of this differential equation is [18]:

P01 .t / D 1 – exp t=£eq £e = .£c C £e / (3.24)
where 1/£eq D 1/£c C 1/£e . A similar evaluation can be performed for P11 (t),
leading to

P11 .t / D £e C £c exp t=£eq = .£c C £e / (3.25)
For a device which has Ntr traps, we can write the number of occupied traps at time
X
Ntr
t, n(t), as being n.t / D i .t /, where ™i (t) can assume the values 0 or 1. We can
iD1
then evaluate the average number of traps occupied at time t as being:
X1 N N t r e N XN t r
hn.t /i D k P .k; t/ (3.26)
N t rD0 Nt rŠ kD0
where NNtr e-N /Ntr ! is the probability that Ntr traps are found in a device, i.e., the
number of traps is assumed to be Poisson distributed, with parameter <Ntr > D N.
While P(k, t) is the probability that k traps are occupied at time t, with k D 0 : : :
Ntr . The averaging must be performed over the successive stochastic capture and
emission events of a trap, and over the number of traps of the ensemble of devices.
The average threshold voltage fluctuation is then obtained by multiplying (3.26)
by <•>, which is the average fluctuation due to a single trap. This leads to:
< VT .t / >D< ı > hn.t /i (3.27)
Since traps have different time constants, which are statistically independent and
identically distributed random variables, for considering P(k, t) in Eq. (3.26) we
evaluate the average probability over different traps by writing a binomial. In this
case

Ntr k Ntr-k
P .k; t/ D P01 .£c ; £e ; t/ P00 .£c ; £e ; t/ (3.28)
k
“
where P01 .£c ; £e ; t/ D d £c d £e P01 .£c ; £e ; t/ f .£c ; £e / and P00 .£c ; £e ; t/ D
“
d £c d £e P00 .£c ; £e ; t/ f .£c ; £e /, with f(£c , £e ) being the joint probability function
of time constants £c and £e over all traps.
£c and £e are random variables, whose values depend on temperature and bias
point, and which we assume to follow [14]:
h i
£c D 10p 1 C exp q (3.29)
£e D 10p Œ1 C exp .Cq/

(3.30)
where p 2 [pmin , pmax ]. Please note that pmin and pmax define the times constants
of the fastest and the slowest trap, respectively. This also limits the time interval
in which the model here proposed is valid, in a similar way as in the analysis of
low-frequency noise made above. Since p is assumed to be uniformly distributed,
the charge trap characteristic time constants are uniformly distributed on a log
scale. Again, this is the same assumption done in the low-frequency noise analysis
above. The variable q is given by q D (ET –EF )/kB T 2 [(EV –EF )/kB T, (EC –EF )/kB T],
where EC is the conduction band edge, while EV is the valence band edge. ET is
the energy level of the trap. Consequently, £c and £e are temperature dependent.
The assumption of the existence of individual defects with a very wide distribution
of time constants is also in line with recent NBTI data [20–24]. The assumption
of a single p, governing both£c and £e , may be restrictive and deserves deeper
investigation in future works, both theoretical and experimental. Nevertheless, we
did run numerical analysis (Monte Carlo simulations) assuming independent p
values for capture and emission. The BTI behavior is observed to stay essentially
the same (no qualitative deviation on the BTI behavior is observed), as long as the
p values remain uniformly distributed.
For the evaluation of Eq. (3.26), we consider Eqs. (3.28), (3.29) and (3.30). In
evaluating the average over the different random variables, we separate the average
over number of traps
X1 N N t r e N
h. /i D . /
N t rD0 Nt r Š
and time constants

“
. / D d £c d £e . / f .£c ; £e /
Please remember that the time constants £c and £e depend on Fermi level, trap
energy level, and temperature, as given by (3.29) and (3.30). Evaluation the averages
in this manner facilitates the analytical analysis.
This leads to:
X1 N N t r e N
hn.t /i D P01 .£c ; £e ; t/ Nt r D N P01 .£c ; £e ; t/
N t rD0 Nt r Š
Z Z !
10pmax t
N Ec
g .ET / dET .e u 1/
D du
ln10 .pmax pmi n / Ev 1 C e .ET EF /=kB T 10pmi n t u
(3.31)
where g(ET ) describes the trap energy distribution in the band-gap, and in the second
integral a change of variable was made, p D log (u/t), dp D du / (u ln 10). N is
the average number of traps found in a device.
In Eq. (3.31), the first integral contains the Fermi level and temperature depen-
dence, while the second integral contains the time dependence. This means that this
equation has the time dependence separated from the Fermi level (i.e., bias point,
since the Fermi level is defined by the bias point) and temperature dependence.
Hence, the model predicts that for measurements carried out at different tempera-
tures, there are scaling factors that can be used as a multiplicative coefficients for
the threshold voltage shift, making the curves for different temperatures to overlap
(at all measured times). The same applies for measurements carried out at different
Fermi levels, i.e., different stress voltages (bias points). Note that the Fermi level
is a function of the applied voltage. Furthermore, since Eq. (3.31) is valid for both
stress and recovery phases, it implies that the temperature and voltage dependence
(scaling factor) during stress and during recovery is the same. This is very relevant
and in agreement to experimental data, as discussed below.
Equation (3.31) can be evaluated numerically, leading to:
hn.t /i ' .T; EF / .A C B log.t // (3.32)
where A and B are constants, and the last term clearly shows that the time evolution
of number of occupied traps shows a log(t) behavior. The term ®(T, EF ) describes the
temperature and Fermi level dependence. Please note that (3.32) above is of the same
form as equation (4) in [23], which was empirically written, as an approximation for
the experimentally observed behavior.
The actual form of the term ®(T, EF ) depends on the trap energy distribution in the
band-gap g(ET ). g(ET ) is usually found to be a convex curve (e.g. U-shaped); see, for
instance, [14, 15]. In this work, a U-shape trap density function is investigated. As
in the case of low-frequency noise analysis for abruptly changing gate bias, the trap
energy distribution g(ET ) is key for explaining experimentally observed findings.
Equation (3.32) models both stress and recovery phases. Hence, both stress and
recovery phases of BTI are described by the same Eq. (3.32). If the density of
initially occupied traps is lower than the value expected for the bias point, the
number of occupied traps increases logarithmically. This corresponds to stress phase
of BTI. On the other hand, if the density of initially occupied traps is higher than
the value expected for the bias point, the density of occupied states decreases
logarithmically. This corresponds to the recovery phase of BTI. Besides analytical
analysis and evaluation, we did run Monte Carlo simulations to confirm this
behavior (Monte Carlo simulations were performed starting from different numbers
of initially occupied traps). Some of the results are presented in Fig. 3.13.
The model here presented considers that the number of traps is constant, i.e.,
there is no trap generation or annihilation. For stress or recovery over long time
intervals, there may be generation of traps during the stress phase, or annihilation
of traps (annealing) during the recovery phase. In this case the average number of
traps <Ntr > in Eq. (3.26) becomes a function of time, <Ntr (t)>. If the number
Fig. 3.13 The full line

depicts the evolution of the
density of occupied traps
obtained by numerical
integration of Eq. (3.31). The
points correspond to Monte
Carlo simulations performed
under the same conditions
or traps increases with time during the stress phase, the evolution of number of
occupied traps may become faster than log(t) as time evolves, as experimentally
observed in many works, where the time dependence is found to follow a power
law for long stress times [20, 21]. If a proper equation for <Ntr (t)> is available, the
modeling approach here presented can be used to evaluate this behavior. Modeling
Ntr as a function of time is out of the scope of this work. This work focus on the
log(t) behavior observed for recovery and short stress times, as discussed in the
experimental session below.
For numerical analysis and Monte Carlo simulations performed in this work
(pmax pmin ) was chosen to be 7, i.e., it is assumed that p is uniformly distributed
over seven decades. The behavior does not depend on pmax and pmin , as long as the
time being considered is much longer than the shortest time constant, and much
shorter than longest time constant.
Now we discuss the situation that we call cyclo-stationary excitation, where the
gate bias of the device is periodically switched.
In square wave cyclo-stationary excitation, the bias voltage abruptly alternates
between two states, called on and off. In the context of this work, the on state refers
to the state where stress is applied, and off refers to the state where no stress is
applied. The fraction of the period T in which the device is in the on state is ’ (i.e.,
the duty cycle). Figure 3.14 depicts this situation.
With periodically changing gate bias, the Fermi level becomes a periodic function
of time EF (t). As seen from Eqs. (3.23) to (3.28), the parameters that govern
the charge trapping behavior are the capture and emission time constants, c and
e , respectively. The behavior of these time constants under cyclo-stationary was
already discussed above in the context of Low-Frequency Noise. It has been shown
that as long as the cyclo-stationary period time T is short compared to the trap’s
time constants, the behavior of each trap is governed by an effective capture ( c,eff )
and emission ( e,eff ) time constant. The effective values are evaluated as being the
time average values over one excitation period:
Fig. 3.14 This figure depicts the behavior of slow traps under switched bias. If the gate bias is
periodically alternated between on and off, the Fermi Level (EF ) changes accordingly. If trap time
constants are much longer than the on (stress) and off (recovery) cycle, traps are not able to follow
device bias, and act according to equivalent time constants. For these traps, which are too slow to
follow bias point change, the behavior is described by the equivalent time constants given by Eqs.
(3.13) and (3.14)
Z T
1 1 1
D dt (3.33)
c;eff T 0 c
Z T
1 1 1
D dt
e;eff T 0 e
Under square wave excitation the effective time constants become a simple function
of the duty cycle ’, as given by Eqs (3.13) and (3.14) above.
This formulation shows that there is no dependence of the effective time con-
stants on the frequency of the cyclo-stationary excitation, as long as the excitation
period T is short compared to the trap time constants.
Using the formulation for the equivalent time constants, Eq. (3.28) becomes

Ntr k Ntr-k
P .k; t/ D P01 .£c; eff ; £e;eff ; t/ P00 .£c;eff ; £e;eff ; t/ (3.34)
k
Because the Fermi level is now a function of time, it is not possible to derive an
analytical closed form expression as Eq. (3.31). However, numerical analysis, as
well as Monte Carlo simulations, show that also under cyclo-stationary excitation
the behavior has the form of Eq. (3.32), i.e., there is one term that describes the log(t)
behavior, multiplied by another term that describes the temperature and Fermi level
dependences. Furthermore, it can be shown below that a simple equation to write
an effective Fermi level may be used, leading to a simple model, suitable for circuit
simulation purposes.
Figure 3.15 shows the results of MC simulation where the switching bias is fast
if compared to the trap time constants of all traps considered in the simulation.
In this simulation, time constants of all traps are at least one order of magnitude
longer than the period of the square wave excitation. In this figure, the dotted
12
10
8
ΔVT (a.u)
0
102 103
Time (a.u.)
Fig. 3.15 Behavior of slow traps under square wave excitation. Steady increase of VT is seem,
causing a BTI degradation which follows a log(t) behavior. Dotted (blue) line is detailed Monte
Carlo simulation, considering usual time constants. Solid (red) line is Monte Carlo simulation
using equivalent time constants, as given by Eqs. (3.13) and (3.14)
(blue) line is the result of the detailed MC simulation of the square wave excitation,
applying the usual time constants during the stress (on) and recovery (off ) phases.
The solid (red) line is MC simulation using equivalent time constants, as given by
Eqs. (3.13) and (3.14). In this simulation the same equivalent time constants are
applied during the whole simulation. In other words, it is similar to a DC simulation
(no switching bias), where the equivalent time constants are applied during the
whole simulation time. Slow traps lead to a degradation component which steadily
increases with time, with a log(t) behavior. In this MC simulation all traps are
assumed to be initially empty. Note that the slow traps do not follow the switching
bias. Steady increase in VT is seen. These MC simulations confirm that the behavior
is determined by the equivalent time constants, as given by (3.13) and (3.14).
Equation (3.34) can be solved numerically, but this approach is not practical
for circuit simulation purposes. A simpler approach, considering and equivalent
Fermi level EF,eff , which is a function of duty cycle, can fit the experimental data.
In this approach, empirically obtained after insights from extensive Monte Carlo
simulations and numerical analysis, EF,eff is written as the time average of the Fermi
level. For the case of square wave excitation with duty cycle ’ we have:
EF;eff .’/ D ’EF;on C .1 ’/ EF;off (3.35)
With this empirical approach, the same Eq. (3.31) can model both steady bias
and cyclo-stationary excitation. Monte Carlo numerical analysis confirms that the
empirical formulation fits the numerical solution of (3.34).
Conclusion
This chapter discussed the role of charge trapping in low-frequency noise
and bias temperature instability (BTI). A statistical model for the charge
trapping phenomena in MOS devices was presented. The model is based
on discrete microscopic device physics parameters. Besides evaluating the
average behavior, the model here proposed allows the derivation of statistical
relevant parameters. The model is applied to study the random telegraph noise
(RTN) and Bias Temperature Instability (BTI).
The impact of charge trapping is modeled as momentary changes in
threshold voltage (or drain current). In analog circuits, this may lead to low-
frequency noise and be a source of jitter in oscillators. In digital circuits it
may lead to aging effects and transient effects, since circuit behavior may
change due to transient VT fluctuations between two logic operations of a
digital circuit.
The modeling approach focuses on operation conditions relevant for digital
and analog design, including large signal AC operation.
It is shown that if the density of traps in the band-gap is a convex curve
(“U”-shaped) a reduction in noise power may be achieved under cyclo-
stationary excitation. However, the variability in noise behavior (normalized
standard deviation of noise power) is shown to increase. This increase in
variability will be another challenge to analog and RF circuit designs in deep
sub-micron CMOS technologies.
Regarding BTI, an analytical model for both stress and recovery phases
of BTI is presented. Furthermore, the model properly describes device
behavior under periodic switching, also called AC-BTI or cyclo-stationary
operation. It is shown that a universal logarithmic law describes the time
dependence of charge trapping in both stress and recovery phases, and that
the time dependence may be separated from the temperature and bias point
dependence.
Acknowledgments Gilson Wirth and Roberto da Silva thank CNPq and FAPERGS for research
grants that partially support the work reported here.
References
1. G Wirth, J Koh, R da Silva, R Thewes and R Brederlow, “Modeling of Statistical

Low-Frequency Noise of Deep-Submicron MOSFETs.”, IEEE Trans. Electron Dev., 52,
p. 1576–1588 (2005).
2. G Wirth, R da Silva and R Brederlow. “Statistical Model for the Circuit Bandwidth Dependence
of Low-Frequency Noise in Deep-Submicrometer MOSFETs”, IEEE Trans. Electron Dev., 54,
p. 340–345 ( 2007).
3. G Wirth, R da Silva, P Srinivasan, J Krick and R Brederlow. “Statistical model for MOSFET
low-frequency noise under cyclo-stationary conditions” Int Electron Dev Meeting - IEDM
2009, p. 30.5.1-4, (2009).
4. G Wirth and R da Silva, “Low-Frequency Noise Spectrum of Cyclo-Stationary Random
Telegraph Signals”, Electrical Eng., 90, p. 435–41 (2008).
5. R da Silva, G Wirth and L Brusamarello. “An appropriate model for the noise power spectrum
produced by traps at the Si SiO interface: a study of the influence of a time-dependent
Fermi level. Journal of Statistical Mechanics” Journal of Statistical Mechanics. Theory and
Experiment, 2008, p. P10015 (2008).
6. R Brederlow, J Koh, G Wirth, R da Silva, M Tiebout, R Thewes. “Low Frequency Noise
Considerations for CMOS Analog Circuit Design” Proceedings of the 2005 International Conf
on Noise and Fluctuations (ICNF). p. 703–708 (2005).
7. R Brederlow, J Koh and R Thewes. “A physics-based low frequency noise model for MOSFETs
under periodic large signal excitation” Solid-State El., 50, p. 668–73 (2006).
8. I Bloom and Y Nemirovsky. “1/f noise reduction of metal-oxide-semiconductor transistors by
cycling from inversion to accumulation” Appl Phys Lett, 58, p.1664–6 (1991).
9. B Dierickx and E Simoen. “The decrease of ”random telegraph signal”noise in metal-oxide-
semiconductor field-effect transistors when cycled from inversion to accumulation”. J Appl
Phys, 71, p. 2028–2029 (1992).
10. M Ertürk, T Xia and W Clark. “Gate voltage dependence of MOSFET 1/f noise statistics”.
IEEE Electron Dev Let., 28, p. 812–814 (2007).
11. A van der Wel, E Klumperink, E Hoekstra and B Nauta. “Relating random telegraph signal
noise in metal-oxide-semiconductor transistors to interface trap energy distribution” Appl Phys
Lett., 87, p. 183507 (2005).
12. S Machlup, “Noise in Semiconductors: Spectrum of a Two-Parameter Random Signal” J Appl
Phys, 35, p. 341–343 (1954).
13. A Roy and C Enz, “Analytical Modeling of Large-Signal Cyclo-Stationary Low-Frequency
Noise with Arbitrary Periodic Input” IEEE Trans. Electron Dev., 54, p.2537-2545 (2007).
14. M Kirton and M Uren, “Noise in solid-state microstructures: A new perspective on individual
defects, interface states and low-frequency (1/f) noise” Adv. in Physics, 38, p. 367–468 (1989).
15. Ekbote, S.; Benaissa, K.; Obradovic, B.; Liu, S.; Shichijo, H.; Hou, F.; Blythe, T.; Houston,
T.W.; Martin, S.; Taylor, R.; Singh, A.; Yang, H.; Baldwin, G, “45nm Low-Power CMOS
SoC Technology with Aggressive Reduction of Random Variation for SRAM and Analog
Transistors” 2008 VLSI Tech. Symp., p. 160–161 (2008).
16. Wirth, Gilson I. ; da Silva, Roberto ; Kaczer, Ben. “Statistical Model for MOSFET Bias
Temperature Instability Component Due to Charge Trapping”. IEEE Transactions on Electron
Devices, p. 2743–2751 (2011).
17. Ashraf, Nabil ; Vasileska, Dragica ; Wirth, Gilson ; Srinivasan, P. “Accurate Model for
the Threshold Voltage Fluctuation Estimation in 45-nm Channel Length MOSFET Devices
in the Presence of Random Traps and Random Dopants”. IEEE Electron Device Letters,
p. 1044–1046 (2011).
18. Camargo, Vinícius V. A. ; Ashraf, Nabil ; BRUSAMARELLO, Lucas ; Vasileska, Dragica ;
Wirth, Gilson . “Impact of RDF and RTS on the performance of SRAM cells”. Journal of
Computational Electronics, p. 122–127, (2010).
19. E Klumperink, S Gierkink, A van der Wel, and B Nauta; “Reducing MOSFET 1/f noise and
power consumption by switched biasing ,” Solid-State Circuits, IEEE Journal of , vol.35, no.7,
pp.994-1001 (2000).
20. D. K. Schroder, “Negative bias temperature instability: What do we understand?”, Microelec-
tron. Reliab., vol. 47 no. 6, pp. 841–852, (2007).
21. S. E. Rauch, “Review and reexamination of reliability effects related to NBTI statistical
variations,” IEEE Trans. Device Mater. Rel., vol. 7, no. 4, pp. 524–530, (2007).
22. B. Kaczer, T. Grasser, J. Martin-Martinez, E. Simoen, M. Aoulaiche, Ph. J. Roussel, and

G. Groeseneken, “NBTI from the Perspective of Defect States with Widely Distributed Times,”
Proc. Int. Rel. Phys. Symp., p. 55 (2009).
23. T. Grasser and B. Kaczer, “Evidence That Two Tightly Coupled Mechanisms Are Responsible
for Negative Bias Temperature Instability in Oxynitride MOSFETs”, IEEE Trans. on Electron
Dev., vol. 56, pp. 1056–1062 (2009).
24. Ang, D.; Teo, Z.; Ho, T.; Ng, C.; , “Reassessing the mechanisms of negative-bias temperature
instability by repetitive stress/relaxation experiments,” Device and Materials Reliability, IEEE
Transactions on, vol. 11 , no. 1, pp. 19 – 34 (2011).
Chapter 4
Atomistic Simulations on Reliability
Dragica Vasileska and Nabil Ashraf
Abstract Discrete impurity effects in terms of their statistical variations in number

and position in the inversion and depletion region of a MOSFET, as the gate length
is aggressively scaled, have recently been researched as a major cause of reliability
degradation observed in intra-die and die-to-die threshold voltage variation on the
same chip resulting in significant variation in saturation drive (on) current and
transconductance degradation—two key metrics for benchmark performance of
digital and analog integrated circuits. In the following chapter, the authors have
highlighted the random dopant fluctuation (RDF) based Ensemble Monte Carlo
(EMC) device simulation study conducted by the Computational Electronics (COM-
PUTEL) research group of Arizona State University. In addition to RDF, random
number and position of interface traps lying close to Si:SiO2 interface engender
additional concerns leading to enhanced experimentally observed fluctuations in
drain current and threshold voltage. In this context, the authors of this chapter
present novel EMC based simulation studies on trap induced random telegraph
noise (RTN) responsible for statistical fluctuation pattern observed in threshold
voltage, its standard deviation and drive current in saturation for 45 nm gate length
MOSFET device. From the observed simulation results and their analysis, it can
be cogently projected that with continued scaling in gate length and width, RTN
effect will eventually supersede as a major reliability bottleneck over the typical
RDF phenomenon. The fluctuation patterns observed by EMC simulation outcomes
for both drain current and threshold voltage have been analyzed and explained from
analytical device physics perspectives.
1 Introduction
The transistor mismatch due to random variations in process parameters has become
one of the major issues in deep-submicrometer technology. The term transistor
mismatch refers to the fact that supposedly identical transistors at the design phase
D. Vasileska () • N. Ashraf

Department of Electrical, Computer and Energy Engineering, Arizona State University,
Tempe, AZ 85287-5706, USA
e-mail: vasileska@asu.edu; nashraf@asu.edu

DOI 10.1007/978-1-4614-4078-9__4
48 D. Vasileska and N. Ashraf
come out as distinct devices after manufacturing due to process variations. The
extrinsic factors include variation in the implantation dose and energy, oxidation
and annealing temperatures, etc. On the other hand, the intrinsic factors include
variations due to channel dopant number, interface, and fixed oxide charge. Both
the extrinsic and the intrinsic factors influence the transistor output performance
and yield. The transistor mismatch effect will become worse in the future due
to demanding requirement on process tolerance. Hence, there is a direct need to
develop a very good understanding on the impact of variation in a given unit process
on resulting variation in a given circuit parameter. This will enable the process
engineering and manufacturing team to define appropriate process monitor and
control criteria for all the unit processes involved. This will also help during the
technology development phase to select a particular process integration scheme
that could minimize the mismatch among various available options. In this book
chapter we first discuss the impact of discrete impurity effects on critical device
parameters (Sect. 2). In Sect. 3 we elaborate on the influence of the random telegraph
noise fluctuations on the variation in the threshold voltage and device on current.
Conclusions from this work are presented in “Conclusions” section.
2 Discrete Impurity Effects
2.1 Some General Considerations
Statistical fluctuations of the channel dopant number were predicted by Keyes [1]
as a fundamental physical limitation of MOSFETs down-scaling. Entering into
the nanometer regime results in a decreasing number of channel impurities whose
random distribution leads to significant fluctuations of the threshold voltage and
off-state leakage current. These effects are likely to induce serious problems on the
operation and performances of logical and analog circuits.
It has been experimentally verified by Mizuno and co-workers [2] that threshold
voltage fluctuations are mainly caused by random fluctuations of the number of
dopant atoms (RDF) and other contributions such as fluctuations of the oxide
thickness are comparably very small. It follows from these remarks that impurities
cannot be considered anymore using the continuum doping model in advanced
semiconductor device modeling but the precise location of each individual impurity
within a full Coulomb interaction picture must be taken into account.
2.2 Drift-Diffusion Simulations of Discrete Impurity Effects
To illustrate the importance of the discrete impurity effects on the threshold voltage
and the off-state current, in 1996 the group from ASU simulated a number of MOS-
FET device structures [3]. The corresponding potential profile under equilibrium
4 Atomistic Simulations on Reliability 49
Fig. 4.1 The role of discrete impurities in large devices (left panel) and in mesoscopic devices
(right panel)
Fig. 4.2 Top panel: Potential profile in a cut through the depth of the device. Bottom panel:
Potential profile in the semiconductor in a plane parallel to the semiconductor/oxide interface
conditions along the depth and in the plane parallel to the semiconductor/oxide
interface is shown in the top and bottom panels of Fig. 4.2. In our analysis we have
found that it is not only the total number of atoms within the discrete doping region
that matters, but the location of these atoms plays very important role.
Illustrated in Fig. 4.3, are the potential profiles and the current stream lines for
two impurity distributions. Note that the significant current crowding in the upper
panel near the critical source end of the channel leads to smaller drain current for
VD D 50 mV and a range of VG values. This, in turn, results in larger threshold
voltage for this device.
Fig. 4.3 Potential profile and current stream lines in a 50 nm gate-length MOSFET device. Shown
here are two impurity distributions in the device active region
With the down-scaling of MOSFET devices, electrons in the conducting channel

are in ever closer proximity to the high-density electron gases present in the source
and drain regions—separated from each other by as little as tens of nanometers—
and in the polycrystalline Si gate—separated from the channel by as little as 1.5 nm
of SiO2 . As studied in [4], the role of these long-range Coulomb interactions
is twofold: (1) The interaction between electrons in the channel and the high-
density electron gases in the source and drain regions can be pictured classically
as a reshaping of the electron distribution in the channel caused by the potential-
fluctuations, associated with plasma oscillations in the source and drain regions,
leaking into the low-density channel. Quantum-mechanically, this corresponds to
emission and absorption of plasmons by the channel electrons. While these pro-
cesses do not subtract directly momentum from the electron gas, their net effect is a
thermalization of the hot electrons energy distribution in the channel, the resulting
higher energy tail being affected by additional momentum-relaxation processes
(phonons, ionized impurities). This causes, indirectly, a reduction of the effective
electron velocity in the channel, and so, a depression of the transconductance as the
channel length is reduced below about 4 nm. On the other hand, the interaction
between channel-electrons and electrons in the gate (Coulomb drag across the
very thin insulator) results in a direct loss of momentum of the electrons in
the channel. Semi-classically, this interaction—also plasmon-mediated—has been

studied by a group at IBM [5] predicting a significant depression of the electron
velocity for SiO2 layers thinner than about 2–3 nm. This behavior has also been
observed experimentally [6], recent results being in quantitative agreement with
early theoretical estimates.
2.3 Monte Carlo Device Simulations of Discrete

Impurity Effects
In the past, the effect of discrete dopant random distribution in MOSFET channel
has been assessed by analytical or drift-diffusion (DD) approaches. The first DD
study consisted in using a stochastically fluctuating dopant distribution obeying
Poisson statistics [7]. 3D atomistic simulators have also been developed for studying
threshold voltage fluctuations [8, 9] Even though the DD/HD (hydrodynamic)
methods are very useful because of their simplicity and fast computing times, it
is not at all clear whether such macroscopic simulation schemes can be exploited
into the atomistic regime. In fact, it is not at all clear how such discrete electrons
and impurities are modeled in macroscopic device simulations due to the long-range
nature of the Coulomb potential. 3D DD/HD macroscopic models may be accurate
for modeling the threshold voltage fluctuations (since the device is in the off-state)
but they are definitely not accurate when examining the on-state current fluctuations.
Three-dimensional Monte Carlo (MC) simulations should provide a more realis-
tic transport description in ultra-short MOSFETs, in particular in the on-state. The
MC procedure gives an exact solution of the Boltzmann transport equation. It, thus,
correctly describes the non-stationary transport conditions. Even if microscopic
simulations such as the MC method are concerned, the treatment of the electrons
and impurities is not straightforward due to, again, the long-range nature of the
Coulomb potential. The incorporation of the long-range Coulomb potential in the
MC method has been a long-standing issue [10]. This problem is, in general,
avoided by assuming that the electrons and the impurities are always screened by the
other carriers so that the long-range part of the Coulomb interaction is effectively
suppressed. The complexity of the MC simulation increases as one takes into
account more complicated screening processes by using the dynamical and wave-
vector dependent dielectric function obtained from, for example, the random phase
approximation. Indeed, screening is a very complicated many-body matter [11].
A novel approach has been introduced by the ASU group [12], in which the MC
method is supplemented by a molecular dynamics (MD) routine. In this approach,
the mutual Coulomb interaction among electrons and impurities is treated in the
drift part of the MC transport kernel. Indeed, the various aspects associated with
the Coulomb interaction, such as dynamical screening and multiple scatterings,
are automatically taken into account. Very recently, the MC/MD method has
been extended for spatially inhomogeneous systems. Since a part of the Coulomb
interaction is already taken into account by the solution of the Poisson equation,
the MD treatment of the Coulomb interaction is restricted only to the limited area
near the charged particles. It is claimed that the full incorporation of the Coulomb
interaction is indispensable to reproduce the correct electron mobility in highly
doped silicon samples.
Although real space treatments eliminate the problem of double counting of the
force, a drawback is that the 3D Poisson equation must be solved repeatedly to
properly describe the self-consistent fields which consume over 80 % of the total
simulation time. To further speed up simulations, the ASU team has, for the first
time, utilized a 3D Fast Multi-pole Method (FMM) [13–16] instead. The FMM
allows calculation of the field and the potential in a system of n particles connected
by a central force within O(n) operations given certain prescribed accuracy. The
FMM is based on the idea of condensing the information of the potential generated
by point sources in truncated series expansions. After calculating suitable expan-
sions, the long range part of the potential is obtained by evaluating the truncated
series at the point in question and the short range part is calculated by direct
summation. The field due to the applied boundary biases is obtained at the beginning
of the simulation by solving the Laplace equation. Hence the total field acting on
each electron is the sum of this constant field and the contribution from the electron–
electron and electron–impurity interactions handled by the FMM calculations. The
image charges, which arise because of the dielectric discontinuity, are handled by
the method of images [17].
Next we present several success stories on the use of our 3D particle-based device
simulation code that properly takes into account the short range and the long range
Coulomb forces (Fig. 4.4). We begin with the example of energy relaxation of the
with e-e and e-i

400
mesh force only
Electron energy [meV]
VD =1 V, VG = 1 V
300
200
100
channel drain
0
100 110 120 130 140 150 160 170 180
Length [nm]
Fig. 4.4 Average energy of the electrons coming to the drain from the channel. The applied bias
equals VD D VG D 1 V. Filled (open) circles correspond to the case when the short-range e–e and
e–i interactions are included (omitted) in the simulations
carriers at the drain end of the MOSFET channel. The simulated device has channel
length LG D 80 nm, channel width WG D 80 nm and oxide thickness Tox D 3 nm.
The lateral extension of the source and drain regions is 50 nm. The channel doping
equals 3 1018 cm3 . The applied bias is VG D VD D 1 V. Only those electrons that
entered the channel region from the source side were “tagged” and their energy
and position was monitored and used in the average energy calculation. From the
average velocity simulation results, it follows that the short-range electron–electron
(e–e) and electron–impurity (e–i) interaction terms damp the velocity overshoot
effect, thus increasing the transit time of the carriers through the device; thus
reducing its cut-off frequency. It is also quite clear that when we use the mesh
force only, i.e. we skip the MD loop that allows us to correct for the short-range
e–e and e–i interactions, those electrons that enter the drain end of the device from
the channel never reach equilibrium. Their average energy is more than 60 meV
far into the drain region. Also, the average energy peaks past the drain junction.
The addition of the short-range Coulomb forces to the mesh force via the MD loop,
leads to rapid thermalization of the carriers once they enter the drain region. The
characteristic distance over which carriers thermalize is on the order of a few nm.
Second important example we have chosen in this paper is the quantitative
prediction of the threshold voltage fluctuations versus device gate width, channel
doping and oxide thickness, that are shown in Fig. 4.5. Also shown in this figure
are the analytical model predictions derived from experimental studies by Mizuno,
Okamura, and Toriumi [2], according to which
p ! p
4
q 3 "s b Tox 4 N
vt D p p ; (4.1)
2"ox Leff Weff
where N is the average channel doping density, ¥b is the built-in potential, Tox is the
oxide thickness, Leff and Weff are the effective channel length and width, and "s and
"ox are the semiconductor and oxide permittivity, respectively. Stolk, Widdershoven
and Klaassen [18] generalized the analytical result by Mizuno and his co-workers,
by taking into account the finite thickness of the inversion layer, depth-distribution
of charges in the depletion layer and the influence of the source and drain dopant
distributions and depletion regions. For a uniform channel dopant distribution, the
analytical expression for the threshold voltage standard deviation simplifies to
p ! p
4 4
q 3 "s b kb T 1 Tox N
vt D p p C p : (4.2)
3 q 4"s b Na "ox Leff Weff
In Eq. (4.2), the first term in the square brackets represents the surface potential
fluctuations whereas the second term represents the fluctuations in the electric field.
The decrease of the threshold voltage fluctuations with increasing the width of the
gate is due to the averaging effects, in agreement with the experimental findings by
Horstmann et al. [19]. We want to point out that we still observed significant spread
Fig. 4.5 Variation of the a

threshold voltage with 70
(a) gate width, (b) channel
60 σVth => Equation (1)
doping, and (c) oxide
thickness σVth => Equation (2)
50
σVth => Simulation results
σVth [mV]
40
30
20
10 NA =5x1018 cm-3, Tox =2 nm, LG =50 nm

0
20 40 60 80 100 120
Gate width [nm]
b 40
35 LG=WG=50 nm, Tox=2 nm
30
σVth [mV]
25
20
5 σVth => Simulation results
0
1x1018 3x1018 5x1018 7x1018
Doping density NA [cm-3]
c 100
σVth => Equation (1)

σVth => Simulation results
σVth [mV]
60
40
LG =WG =50 nm
20
NA =5x1018 cm-3
0
0 1 2 3 4 5
Oxide thickness Tox [nm]
of the device transfer characteristics along the gate voltage axis even for devices
with WG D 100 nm. This is due to the nonuniformity of the potential barrier, which
allows for early turn-on of some parts of the channel. As expected, the increase in
the channel doping leads to larger threshold voltage standard deviation ¢VTH . These
results also imply that the fluctuations in the threshold voltage can be even larger in
devices in which counter ion implantation is used for threshold voltage adjustments.
Similarly, the increase in the oxide thickness leads to linear increase in the threshold
voltage standard deviation. The results shown in Fig. 4.5(a–c) also suggest that
reconstruction of the established scaling laws is needed to reduce the fluctuations
in the threshold voltage. In other words, within some new scaling methodology, Tox
should become much thinner, or NA much lower that what the conventional scaling
laws give.
To further elaborate on the role of Coulomb interactions, in Example 3, as
suggested by the results presented in Example 2, we consider 50 nm channel
length narrow-width SOI device with undoped channel. The results presented in
Fig. 4.6 also suggest that there are fluctuations in the device threshold voltage
for devices fabricated on the same chip due to unintentional doping and random
positioning of the impurity atoms. This can also be deduced from the scatter of the
experimental data from [20]. The simulation results of the transfer characteristics
with a single impurity present in different regions in the channel of the device,
shown in the top panel of Fig. 4.6, clearly demonstrate the origin of the threshold
voltage shifts for devices with 10 and 5 nm channel width. The width dependence
of the threshold voltage for the case of a uniform (undoped) and a discrete impurity
model is shown in the bottom panel of Fig. 4.6. These results suggest that both size-
quantization effects and unintentional doping must be concurrently considered to
explain threshold voltage variation in small devises.
3 Random Telegraph Signal
3.1 Importance of Random Trap Fluctuations
The single most important interface in semiconductor technology is that between

silicon and its thermally grown oxide. This interface with its propensity to surface
micro-roughness after in-situ fabrication plays a crucial role in the performance
of today’s high speed MOSFET devices. The degree of perfection of the interface
has been stipulated to be really exacting in terms of process integrity where a
typical device-quality interface has defect densities on the interfacial plane of the
order of 108 –1010 cm2 eV1 resulting in defect densities of the order of 1–100
defects per square micron assuming the defects are located within 1 eV of energy
distribution from the Fermi energy. As the device area is shrunk to aggressively
scaled sub-m2 size with scaling-preserved process tolerances, the number of defect
densities do show an upward trend and considering that 1011 cm2 eV1 values are
b 0.6
Experimental: Lg = 250nm
0.5 Simulation: Lg = 50 nm
Simulation: Vt roll-off adjusted
0.4 Discrete single dopants
Threshold Voltage [V]
0.3
0.2
0.1
-0.1
-0.2
2 4 6 8 10 12 14 16
Channel Width [nm]
Fig. 4.6 Top Panel: Transfer characteristics of the device with 10 and 5 nm channel widths and
different location of the impurity atoms. Bottom Panel: Width dependence of the threshold voltage
for the case of a uniform and a discrete impurity model. Clearly seen in this figure are two trends:
(a) Threshold voltage increase with decreasing channel width due to quantum-mechanical size
quantization effects, and (b) Scatter in the threshold voltage data due to unintentional doping
at least readily encountered, the number of defects seem to reduce to less than 3
in number per square micron within an eV energy distribution for a device size
of W L D 50 nm 50 nm. The reason reliability concern did not arise in wide
gate area technology generation is because with a good number of traps lying
within a few eV of Fermi energy, the spatial distribution of energy levels is more
tighter making the energy barrier values 4EB a fraction of an eV. Hence the carriers
trapped in traps easily reemit to inversion layers making RTS amplitude variation
almost nonexistent. But in today’s aggressively scaled device size, even though the
trap numbers are countable and sparse, these traps can be located deep within the
oxide with higher 4EB differential and a carrier once trapped in a trapped site,
may stay there for a prolonged period of time and never get reemitted to inversion
layer causing severe RTS amplitude drawbacks. The surface level trap lying close
to the interface can block the carrier flow in the channel by causing local potential
fluctuations where significant spread of RTS variation can be observed for even a
single trap closer to source side to mid-channel zone impeding the carrier flow. Also
one way carrier gets trapped and get detrapped is through tunneling from inversion
layer to a trap location and in earlier technological generation with thicker gate
oxide, tunneling was not as significant as it is today with the gate-oxide reaching
nanometric thickness. The conclusion is with the channel electrons being random,
presence of very few defects will suffice to cause notable RTS related device
operation failure for current ongoing technological generation of MOSFETs,
Figure 4.7 (top panel) illustrates the measurement data from a 90 nm SRAM
design [21]. The minimum supply voltage (Vccmin ), which is highly sensitive to
device threshold voltage, exhibits a similar pattern in the time domain as that of
RTF. The impact of RTF in this case is more than 200 mV, which is catastrophic to
the yield and low-power design of SRAM. Therefore, accurate and physical models
of RTF are essential to predict and optimize circuit performance during the design
stage. Currently, such models are not available for circuit simulation. The compound
between RTF and other sources of variation, such as RDF, further complicates the
Fig. 4.7 Top panel: The fluctuation of SRAM Vccmin due to RTF. Bottom panel: A random
telegraph signal (namely x(t)) produced by a carrier trap
situation especially in extremely scaled CMOS design. Figure 4.7 (bottom panel)
shows a typical two level RTS signal time domain characteristics.
3.2 Monte Carlo Device Simulations of Random Traps

at the Semiconductor/Oxide Interface
In this research work the integration of random defects positioned across the channel
at the Si/SiO2 interface from source end to the drain end in the presence of different
random channel and bulk dopant distributions are used to conduct Ensemble
Monte-Carlo (EMC) based numerical simulation of key device performance metrics
for 45 nm gate length MOSFET device. The two main performance parameters
that affect RTF based reliability measurements are percentage change in relative
drain current fluctuation, particularly in the saturation region where most digital
circuits operate and percentage change in threshold voltage that affects the device
transconductance and on-current drive.
The simulator described in the previous section is presently being used in the
investigation of the random trap fluctuations in 45 nm technology node MOSFET
device where, in addition to the randomness of the position and the actual number of
the impurity atoms in the whole simulated domain of the device, a random double-
charged trap is introduced in the middle section of the channel and moved from the
source-end to the drain end of the channel. An example of a discrete impurity pattern
and a double trap located at the source-end of the channel is shown in Fig. 4.8. The
effective channel length of 45 nm technology node is taken to be 35 nm.
We consider ensemble of 20 devices with different random dopant distribution.
The threshold voltage of each of these 20 devices without the presence of the trap is
shown in Fig. 4.9.
The total variation of the threshold voltage as a function of the double trap
position in the middle portion of the channel, when moved from the source end to
the drain end of the channel, is shown in Fig. 4.10. We see that the threshold voltage
increases from its average value when the double trap is located at the source end of
the channel. This is due to the fact that carriers see additional large potential barrier
due to the presence of the charged double trap and are reflected back in the source
contact. The threshold voltage reduces when the double trap is moved away from the
source injection barrier because when the electrons are injected in the channel, even
though the electric field is small (due to small drain bias applied when measuring
threshold voltage), they slowly drift towards the drain contact.
In Fig. 4.11 we depict the threshold voltage standard deviation variation as a
function of the double trap position when the double trap is being moved from the
middle of the source end of the channel to the middle of the drain end of the channel.
An explanation of the results given in Fig. 4.11 is schematically shown in Fig. 4.12.
At threshold voltage, the sheet electron density in the channel is small, therefore
screening is not important. Traps near the source end of the channel have the largest
Trap present in the middle at the source end of the channel

(doping profile is in the plane at the SC/oxide interface)
50
Effective
45 channel length
is 35 nm
40
35
Distance [nm]
30
25
20
15
10
0
0 5 10 15 20 25 30 35 40 45
Distance [nm]
Trap position
Fig. 4.8 Random dopant distribution and a double trap located in the middle of the source end of
the channel
Threshold voltage Vt [V]
0.25
0.2
0.15
0.1
0.05
0
0 5 10 15 20
Dopant Configuration
Fig. 4.9 Threshold voltage fluctuations due to random dopant fluctuations (without traps) for a
statistical ensemble of 20 devices with different number and different distribution of the impurity
atoms
0.19
0.18 Average threshold voltage

cubic interpolation
0.17
Threshold voltage Vt [V]
0.16
0.15
0.14
0.13
0.12
0.11
0.1
0 5 10 15 20 25 30 35 40
distance along the channel [nm]
Fig. 4.10 Threshold voltage variation with trap position variation. These results are averages over
the ensemble of 20 devices
45
Simulated data
cubic interpolation
40
percentage Vt variation
35
30
25
Trap position (b)
20
Trap position (a)

15
0 5 10 15 20 25 30 35 40
Fig. 4.11 Percentage threshold voltage fluctuation due to a double trap located at the semiconduc-
tor/oxide interface and different position along the middle section of the channel. Twenty devices
with different random dopant distributions have been averaged out. Detailed description of the
trend of the results shown in this figure is given in Fig. 4.12
a b
Sheet density Sheet density
distance distance
VD small VD small
Fig. 4.12 Schematic explanation of the results from Fig. 4.11
10
Barrier dominated Screening not

8 current degradation effective due to
pinch-off of the
delta ID/ID [%]
channel
6
2
VDs = 0.7 V, VGs = 0.8 V
0
0 10 20 30 40
Fig. 4.13 ON-current degradation vs. trap position. The statistical ensemble used here consists of
the first seven devices from Fig. 4.9 with different number and different distribution of the impurity
atoms
influence since they are major obstacles to the electrons because of the large input
barrier depicted in case (a) shown on the left panel of Fig. 4.12. Traps near the
drain end of the channel have smaller influence since electrons are accelerated by
the small electric field—case (b) shown on the right panel of Fig. 4.12.
Figure 4.13 shows the ON-current degradation as a function of the double trap
position. As depicted on the figure, near the source end of the channel the current
degradation due to the presence of a negatively charged double trap is large because
the double trap introduces additional barrier for the current flow. When the double
trap is in the middle section of the channel, the current degradation is smaller.
Traps near the drain contact, where the electron density is pinched off for the bias
Sheet Sheet
density density
distance distance
(a)
(a)
(b)
(b)
VD small
VD large
Fig. 4.14 Under small gate and drain bias, we have the situation depicted in Fig. 4.12. For large
gate bias: (a) the sheet density increases which means that screening increases; (b) traps near
the drain are surrounded with large number of electrons for small VD , therefore screening of the
Coulomb potential is large and there is smaller degradation of the current; (c) for large VD traps
near the drain are surrounded with smaller number of electrons, therefore screening is smaller and
we have slightly larger ID degradation
conditions used, are not effectively screened and a notable increase of the current
degradation is observed. At the drain contact, the degradation drops practically to
zero because there traps are effectively screened by the electrons. The expected
current trends under small and large drain voltages are explained in Fig. 4.14.
4 Conclusions
Unlike most of the analytical revelation of published articles, where it has been
reported that fluctuation in drain current amplitude variation and threshold voltage
variation tend to diminish at strong inversion and saturation conditions imposing
higher gate and drain bias owing to the screened out potential due to high inversion
charge density at the surface with associated improvement of Coulombic-scattering
related mobility, our simulations conducted at saturation bias conditions on a 45 nm
MOSFET reveal that for different random dopant distributions in the channel
and bulk, the fluctuation pattern exhibited by drain current amplitude variation
and threshold voltage variation are truly statistical and random in nature, i.e., for
some specific random dopant distributions, the fluctuation nature is well controlled
whereas for some other random dopants, the fluctuation pattern shows significant
transitions between local peaks and valleys. From our EMC simulation, it has been
demonstrated that fluctuations in the drain current amplitude and the threshold
voltage have been dependent on particular random dopant distribution type, i.e., its
number within the channel area and its position, in addition to having strong
correlation on strategically positioned interface traps along the channel from source
to drain.
In order to truly represent the amplitude variation to show more dependence on
spatial positioning of trap than specific random dopant type, the expectation value
of the statistic (drain current or threshold voltage amplitude change) or the average
term needs to be studied out of a significant number of possible random channel
dopant distributions. These averaged-out values of drain current and threshold
voltage fluctuations as extracted by the simulations can be analytically rationalized
through the concept of device physics of scaled MOSFETs. We propose herein the
following analytical physical explanations supporting the simulated results obtained
for drain current fluctuations and threshold voltage fluctuations with associated
fluctuations in its standard deviation induced by trap’s occurrences in the channel
close to the interface spatially located between source and drain.
4.1 Local Surface Potential Fluctuations
It has long been known that random dopant fluctuations cause fluctuations in the
surface potential along the channel from source to drain and also from interface to
some depth into the bulk. The fact that at strong inversion and saturation condition,
the ready availability of inversion charge density smears out the surface potential
variation, thus possible fluctuation in drain current and threshold voltage, needs
a second look. With considerable scaling of gate length along with narrow width
MOSFET devices, there are only few 100 atoms available per device area of scaled
MOSFET and considering their random positioning along the channel and also
along the depth from interface towards bulk, based on the fact that the depletion
region available underneath the channel itself has a spatial random distribution
owing to the random doping of the bulk, the inversion charge density formed
by surface band bending is not continuous along the interface from source to
drain, rather it is discrete and spatially random and nonuniform. Therefore the
very possibility of screened out or smeared out surface potential at the interface
is never a reality and depending upon the local fluctuation enhanced inhomogeneity
of inversion charge density even at high gate bias, there will be sizeable spatial
variation of surface potential at the interface from source towards drain. In addition
to experiencing fluctuation due to random discreteness of channel dopant position
and its number, the local surface potential fluctuation is also enhanced by the
strategic location of trap or a number of traps interspersed between one another
and distributed along the channel from source to drain. A repulsive trap center
originating from the presence of acceptor like traps will enhance the spike in local
potential amplitude at the trap’s position and if this trap happens to be situated
right at the source side or even a few nm away from the source side, the spike
in potential barrier is going to significantly impede the carrier flow and thereby
reduce the number of carriers that reach the drain side and contribute to current
value. A trap’s positioning at the source side being repulsive type consequently
causes significant negative shift in current value and significant rise in fluctuation
percentage value of relative drain current amplitude. This phenomenon is not
uncommon in modern process controlled MOSFET fabrication as most RTF related
phenomenon are caused by acceptor like traps being repulsive in nature. As the
drain bias increases ensuring saturation conditions, the spatial conduction energy
band gradually decreases towards the drain, enhancing drift related carrier transport
where the carriers either enter the condition of velocity saturation or velocity
overshoot near the drain zone and easily surmount any local barrier constituted
surface potential fluctuation. Hence if the traps are near the drain side, the fluctuation
pattern reduces to a steady and stable value. This physical aspect is corroborated
by all the simulation outputs of present research work regarding drain current
fluctuation. When the double-trap is located at the pinch-off region there is slight
increase in the current degradation.
For the threshold voltage fluctuations and fluctuation in standard deviation of
threshold voltage as extracted by our simulation results, the analytical explanation
holds and since a trap’s presence at the source side of the channel is causing
maximum resistance to carrier flow by showing a peak of local surface potential
at trap’s position, therefore this will affect the inversion charge density available
at threshold and shift the threshold voltage by altering the gate bias conditions at
threshold. Therefore relative threshold voltage amplitude reaches a high peak near
source side trap positions and due to presence of high enough energetic carriers
along the drain region, a trap’s positioning there would not alter the inversion
condition to the extent that a considerable shift in threshold voltage amplitude will
be expected. Hence, the threshold voltage shift for drain side-positioned traps will be
minimal. This physical aspect is also vindicated by the available data statistics that
have been shown in the plots shown previously for threshold voltage variation with
its fluctuation in standard deviation of threshold voltage. The explanation presented
in this subsection from device physics standpoint is attributed to the condition that
a large number of channel random dopant distributions are considered and have
been averaged to arrive at the values of final drain current fluctuation statistic and
threshold voltage fluctuation statistic. Expectedly, the randomness of the distribution
of channel and bulk dopants will induce comprehensive randomness in local surface
potential fluctuations making the after effect on the drain current fluctuations
and threshold voltage fluctuations to be purely statistical. In essence, the plots
shown with different random dopant distributions related drain current amplitude
fluctuation and threshold voltage variation reveal a very important observation that
strategically devised random channel dopant distributions through ion implantation
can more successfully suppress the variability in drain current and threshold voltage
and excel against RTF related device failure.
4.2 Fluctuation in Carrier Mobility
Each trap acts as a potential scattering center and impacts the scattering related
transport efficiency. When the traps are closely apart and randomly positioned along
the channel from source to drain, trap’s interaction with carrier electron in the inver-
sion channel and fixed ions in the underlying depletion region modifies the short
range trap-to-electron–electron and trap-to-electron–ion Coulombic potential giving
rise to significant Coulomb scattering related mobility reduction. Since our EMC
transport kernel is equipped with properly accounting for this Coulomb effect, the
results of the simulations perfectly reveal the fact that due to the random reduction
of transport mobility of carrier due to trap’s position and interactions with nearby
electrons and ions, drain current amplitude reduction and hence larger fluctuation
of drain current amplitude is expected which will be spatially nonuniform. The
Coulomb related short range and long range e–e and e–ion interaction with traps will
also result in significant threshold voltage shift at inversion condition on the gate,
hence is responsible for trap related threshold voltage fluctuation. The modification
of Coulomb potential seen by a trap surrounding its vicinity where a significant or no
carrier may reside and significant or no ion may reside stemming from their mutual
random and discrete nature, is also severely impacted by a trap’s proximity to a
nearby trap. Since for all our simulations, the traps were very closely positioned to
one another within 1 nm separation, this will induce more Coulomb potential change
and changes in local short range trap-electron and trap-ion interaction affecting
carrier mobility. The rapid screening of Coulomb potential only occurs at the drain
side with high bias on drain and hence at those locations (except the pinch-off
region), mobility reduction is minimal and the fluctuation pattern for both relative
drain current and threshold voltage is minimized. In addition, depending on the
discrete positioning and number density of channel dopants, the particular trap-to-
electron and trap-to-ion interaction for both short and long range Coulomb forces
will be fully distinctive and random causing randomness in mobility reduction
spreading over different random channel and bulk dopant distributions. Larger
spikes and surges in local surface potential values expected near the source side,
will contribute to associated variation in short range and long range Coulomb force
related scattering probability of carrier transport for a trap positioned there.
4.3 Interface Conditions
From the EMC simulation results of our present research work on RTF related drain
current and threshold voltage fluctuation, a close observation reveals the important
viewpoint that for interface trap occurrences, only traps within a thousandth fraction
of a nm from the interface in the channel are considered for study. It has been shown
by the exhaustive work of Professor Dragica Vasileska and her co-researchers that
both the drain current and threshold voltage are strongly correlated to within a few
nm from the interface to the vertical channel depth. Since the traps designed in our
simulation are almost within a considerable fraction of a nm from the interface,
it is vindicated from previous simulation study with present EMC code that both
drain current fluctuation pattern and threshold voltage fluctuation pattern will show
strong positive correlation with trap’s proximity to the interface. This feature has
been found to be another potent factor for reasoning out wide fluctuation patterns
observed for both drain current amplitude and threshold voltage variation with its
standard deviation. In addition, the interface roughness and chemical imperfections
due to trap’s positioning as scattering centers at random locations from source to
drain are also a significant source for interface roughness scattering related mobility
reduction which from theory is found to be one of most dominant causes for
mobility reduction at high gate and drain bias, i.e. saturation bias condition of the
MOSFET which we employed in our simulation for on current fluctuations study.
The interface roughness will reduce the overall effective mobility with the traps
acting as scattering sites randomly distributed within the inversion channel, being in
conjunction with Coulomb-force related mobility reduction and will impact spatial
random and statistical variation in both relative drain current amplitude.
References
1. Keyes R. W., The effect of randomness in the distribution of impurity atoms on FET thresholds,
Appl. Phys. vol. 8, 251–259 (1975).
2. Mizuno T., Okumtura J. and Toriumi A., Experimental study of threshold voltage fluctuation
due to statistical variation of channel dopant number in MOSFET’s, IEEE Trans. Electron
Devices 41, 2216–2221 (1994).
3. Vasileska D., Gross W. J., Kafedziski V. and Ferry D. K., Convergence properties of the
Bi-CGSTAB method for the solution of the 3D Poisson and 3D electron current continuity
equations for scaled Si MOSFETs, VLSI Design 8, Nos. 1-4, 301–305, (1998).
4. Fischetti M. V., Towards Fully Quantum Mechanical 3D Device Simulations, Journal of
Computational Electronics 1, 81–85 (2002).
5. Fischetti M. V., Effect of the electron–plasmon interaction on the electron mo-bility in silicon,
Phys. Rev. B 44, 5527–5534 (1991).
6. Lilly M. P., Eisenstein J. P., Pfeiffer L. N., and West K. W., Coulomb Drag in the Extreme
Quantum Limit, Phys. Rev Lett. 80, pp. 1714–1717, (1998).
7. Wong H. S. and Taur Y., Three-dimensional “atomistic” simulation of 50 nm FETs, in Proc.
IEDM, 29.2.1 (1993).
8. Gross W. J., Vasileska D. and Ferry D. K., 3D Simulations of Ultra-Small MOSFETs with
Real-Space Treatment of the Electron–Electron and Electron–Ion Interactions, VLSI Design,
Vol. 10, 437–452 (2000).
9. Asenov A., Random Dopant Induced Threshold Voltage Lowering and Fluctuations in Sub-
0.1 um MOSFETs: A 3-D Atomistic Simulation Study, IEEE Trans. Electron Devices 45,
2505–2513 (1998).
10. Sano N., Matsuzawa K., Mukai M., and Nakayama N., On discrete random dopant modeling
in drift-diffusion simulations: physical meaning of ‘atom istic dopants, Microelectronics
Reliability, Volume 42, Issue 2, 189–199 Feb. (2002).
11. Ferry D. K., Kriman A. M., Kann M. J., and Joshi R. P., Molecular dynamics extensions
of Monte Carlo simulation in semiconductor modeling, Computer Physics Comm., vol. 67,
119–134 (1991).
12. Gross W. J., Vasileska D. and Ferry D. K., A Novel Approach for Introducing the Electron–
Electron and Electron–Impurity Interactions in Particle-Based Simulations, IEEE Electron
Device Lett. 20, No. 9, 463–465 (1999).
13. Greengard L. and Rokhlin V., A fast algorithm for particle simulations, J. Comput. Phys. 135,
280–292 (1997).
14. Beatson R. and L. Greengard, A short course on fast multipole methods, in Wavelets, Multilevel
Methods and Elliptic PDEs (Leicester, 1996), ser. Nu mer.Math. Sci. Comput. New York:
Oxford Univ. Press, 1–37 (1997).
15. Cheng H., Greengard L., and Rokhlin V., A Fast Adaptive Multipole Algo-rithm in Three
Dimensions, J. Comput. Phys. 155, 468–498 (1999).
16. FMMPART3D user’s guide, version 1.0 ed., MadMax Optics, Hamden, CT, USA.
17. Khan H. R.,Vasileska D., Ahmed S. S., Ringhofer C. and Heitzinger C., Mod-eling of FinFETs:
3D MC Simulation Using FMM and Unintentional Doping Effects on Device Operation,
Journal of Computational Electronics, Vol. 3, Nos. 3-4, 337–340 (2005).
18. Stolk P. A., Widdershoven F. P. and Klaassen D. B. M., IEEE Trans. Electron Devices 45,
1960–1971 (1998).
19. Horstmann J. T., Hilleringmann U. and Goser K. F., IEEE Trans. Electron Devices 45, 299–306
(1998).
20. Majima H., Ishikuro H., and Hiramoto T., IEEE Electron Dev. Lett. 21, 396–398 (2000).
21. Agostinelli T., Arca M., Caironi M., Ferrero V., Natali D. and Sampietro M., Trapping effects
on the frequency response of dithiolane-based planar photodetectors, ICSM0 International
Conference on Science and Technology of Synthetic Metals.
Chapter 5
On-Chip Characterization of Statistical Device
Degradation
Takashi Sato and Hiromitsu Awano
Abstract Bias temperature instability (BTI) is one of the most critical degradation
mechanisms that occur in modern semiconductor devices. The degradation due
to BTI is transient, and known to be greatly influenced by bias voltages and
temperature, making it very difficult to detect possible BTI-related failures during
manufacturing test. Characterization and modeling of BTI is hence extremely
important to protect a chip from BTI-related failures. In this chapter, an array
structure that accelerates the statistical characterization of BTI is described. By over-
lapping the stress-application period for each device, measurements on hundreds or
thousands of devices can be conducted concurrently. Test chip measurement results
that provides a statistical insight on the parameters of BTI-related degradation
process are also presented.
1 Introduction
The quality of the semiconductor–insulator interface in silicon devices has been

a major topic of research. Defects in the interface between silicon (Si) and
silicon dioxide (SiO2 ) not only lower device performance of the metal–oxide–
semiconductor (MOS) transistor but may also lead to severe reliability degradation.
The lifetime of a device decreases when the quality of the interface film is
insufficient. Owing to the appropriate treatments of the interface, such as optimizing
the annealing process using hydrogen (H) gas, which passivates the Si-dangling
bond by H, modern silicon devices have become considerably more stable and are
thus durable. Enhanced device reliability has allowed semiconductor circuits to be
adapted in a broad range of applications.
Despite the significant quality improvement, unsaturated bonds are still consid-
ered to exist. The interface states and the associated change in the surface potential
is again becoming an increasing concern for circuit designers who utilize devices
fabricated using nanometer-scale technology. As device dimensions become even
smaller, a single charge that is trapped in the interface region can greatly distort
T. Sato () • H. Awano

Kyoto University, 36-1 Yoshida-Hommachi, Sakyo-ku, Kyoto 606-8501, Japan
e-mail: takashi@i.kyoto-u.ac.jp

DOI 10.1007/978-1-4614-4078-9__5
70 T. Sato and H. Awano
channel formation. The temporal change in the charges in the channel region will
cause a drift in circuit performance or, in the worst case, will result in a malfunction
of the circuit.
As semiconductor circuits are utilized in various applications, diverse levels
of reliability are demanded. Certain applications impose very stringent reliability
requirements. Examples include transportation systems such as automated car-
driving systems or medical applications. On the other hand, in another application,
cost lowering may be pursued as the highest priority, requiring a certain level of
reliability to guarantee equal or longer than product lifespans. It has now become
a common practice to design for the reliability of a circuit in addition to its
performance.
A device model that incorporates a transient change in the device parameters is a
key enabler to facilitate the design of device and circuit reliability. As a solid basis
for building such a model, it is important to establish an accurate measurement
methodology that will contribute to understanding the mechanisms behind the
device degradation. In particular, the phenomena of bias temperature instability
(BTI) are drawing increasing attention.
In this chapter, we survey characterization methods for the purpose of device
modeling that includes the effects of temporal parameter degradation. We begin
by quickly reviewing the physical mechanisms of temporal threshold voltage
(Vth) changes: random telegraph noise (RTN), negative bias temperature instability
(NBTI), and positive bias temperature instability (PBTI), which are considered to
be interrelated and important in miniaturized devices. Then, we review the work to
accurately capture the RTN, NBTI, and PBTI phenomena. Particular focus is placed
on the approaches that utilize on-chip circuits to enhance the accuracy and fidelity
of the measurements. Finally, the characterization of statistical degradation will be
reviewed with examples.
1.1 Transient Change in Device Parameters
BTI is considered to be one of the major limiting factors for long-term performance
and reliability of semiconductor devices [25, 30].
NBTI and PBTI degradations are observed in pMOS and nMOS transistors,
respectively. These phenomena are observed as a gradual increase in threshold
voltage during circuit operation. The threshold voltage increase is accelerated under
a strong vertical electric field and under an elevated temperature. It is also known
that the increased threshold voltage will be partially recovered once the stress
condition of the electric field or temperature is alleviated. Hence, their models
naturally become dependent on history. The prediction of a device parameter at a
certain time in the future has been found to be very difficult.
Physical mechanisms of threshold voltage shifts due to NBTI have long been
under debate. As device dimensions decreases, there is an urgent need for develop-
ing a model that is physically correct and thus predictable for both degradation and
5 On-Chip Characterization of Statistical Device Degradation 71
recovery phenomena. One promising hypothesis is the formation of a positive oxide

charge due to interface traps. The other explanation is the hydrogen-ion diffusion
generated by broken Si-H bonds.
Another developing concern is PBTI, which is the degradation in nMOS
transistors. PBTI is attributed to the application of new materials such as high-
k gate insulators and metal gates. In addition to the interface state degradation,
the capture of carriers in the intrinsic defects in high-k insulator stacks causes a
threshold voltage shift. For example, the void formulated in an HfO2 dielectric gives
a rise to the threshold voltage [32].
The degradation processes in pMOS and nMOS transistors originate in discrete
charges. The effects of BTI are expected to become more discrete and larger in
devices fabricated by using scaled technologies [17]. Hence, the characterization of
device-dependent degradation has gathered keen interest.
1.2 Measurement Requirements
It is essential to incorporate BTI effects into device models. Good experimental

data is critically important to gain a full understanding of the BTI phenomena and
to improve the reliability of a circuit.
Variation in the NBTI parameter values has not been fully investigated despite its
importance. One major difficulty involved in the measurement of BTI phenomena
is the inclusion of very slow components in the threshold voltage change. Even
if the degradation is accelerated under high-temperature and high-stress-voltage
conditions, hours or even days are required to observe threshold voltage change
because of slow components. This has made it almost impossible to collect statistical
data on a large number of devices in a practical time frame.
It is also known for the BTI phenomena that the threshold voltage recovery
incorporates a very rapid component as well. In the rapid recovery of small devices,
stair-like changes in threshold voltage have been observed. This suggests the release
of a charge from the interface trap [9]. Such a wide-spread time response of the
threshold voltage recovery makes measurement even more difficult.
In most of the existing measurement structures, large-channel-area devices have
been used. Even when individual small-area devices are used, the equivalent device
area is considered to be large. An example is a ring oscillator structure whose
frequency change reflects the device degradation [12]. Although the oscillation
frequency indicates an average degradation of the devices, it is more important to
separate the degradation of a single transistor to create an accurate model.
Device size, i.e., the channel area of a device, has a strong impact on the
temporal threshold voltage changes. In particular, in the context of trap-based
models [3, 8, 9, 20], device size is a critical parameter because it is closely related
to the expected number of defects in a device. The amplitude of threshold voltage
change caused by the capture and release of a charge in an interface trap is also
severely affected. Understanding the variability in NBTI parameters for devices that
are approximately the sizes of logic LSIs is thus critically important. Additionally,
note that as device size decreases, the measurements become more difficult because
small parasitic elements in the measurement system may significantly alter the
measurement results.
2 Circuit Structures for the Measurement

of Device Degradation
Reflecting the importance and difficulty of characterizing device degradation on the

basis of the underlying physics, various circuit structures have been proposed [5, 7,
11, 12, 14].
2.1 Measurements Using Off-Chip Equipment
In this section, we review the measurement techniques for the characterization of

BTI of a single transistor using off-chip equipment. The off-chip equipment refers
to parametric analyzers with source-measurement units (SMUs).
The most straightforward and widely used method is to measure I–V curves
through direct probing because BTI is observed as a change in threshold voltage.
Once the full I–V curve is obtained before and after stress application, most of the
device parameters, such as threshold voltage or carrier mobility, can be extracted to
analyze their temporal changes. The drawback of this measurement is a relatively
long interruption of the bias voltages. The bias conditions during voltage sweeping
for I–V curve measurement are, in most cases, different from the stress or recovery
bias in the BTI measurement.
The threshold voltage measurement using the constant current method defines
the threshold voltage to be the gate-to-source voltage at which a particular current,
e.g., 107 A, in which the transistor is in the linear region, is observed [16].
Because this measurement focuses on observing the threshold voltage only, this
measurement evades full I–V curve measurement. However, the constant current
method still requires a device to depart from a strong stress or recovery bias
condition. Though the interruption is short, partial recovery is still unavoidable
during the measurement. This may lead to under-estimation of the degradation.
In [19], the use of an operational-amplifier-based on-chip current generator is
proposed. This type of circuit can potentially shorten the settling time of the
output node voltage as compared to what can be achieved using off-chip parametric
analyzers.
On-the-fly (OTF) measurement [4] attempts to eliminate the bias interruptions
associated with the threshold voltage measurement. In this method, instead of
measuring I-V curves, the change in the drain current is measured while the stress
voltage is continuously applied to the device. Hence, the degradation of a device can
be observed without recovery associated with the stress interruption. An ultra-fast
on-the-fly (UFOTF) method improved timing resolution to 1 s [13]. This method
utilizes a current-to-voltage converter to isolate the device of interest from an oscil-
loscope probe. However, these methods require the transconductance of the device
in order to convert the measured current to the corresponding threshold voltage.
Because transconductance may also be subject to variation between devices, error
associated with the conversion may be involved. In addition, the sensitivity of the
threshold voltage to the drain current tends to be very weak in the saturation region,
in which a strong stress bias condition for the device is achieved. In this case, it is
difficult to obtain high accuracy.
In [14], the drain current measurement of a transistor in the weak inversion
region, instead of the saturation region, is proposed to utilize the exponential
sensitivity. In order to overcome a slow settling time due to the extremely small
charging and discharging current of a device in the weak-inversion region, the device
of interest has to be large, and an assist circuit is required. In the literature, it is
reported that a measurement delay as fast as 400 ns has been achieved. Because this
method also relies on the drain current measurement, a conversion is also required
to obtain the threshold voltage by using separately acquired transconductance.
2.2 On-Chip Measurements
A ring oscillator is a widely used circuit structure for device characterization

because it is relatively easy to implement and measure. Ring oscillators are also
applied for BTI measurements. In [12], Kim et al. proposed a ring-oscillator-based
degradation sensor, which is known as a silicon odometer. It accurately measures
device degradation by monitoring the beat frequencies of two ring oscillators,
one that is fresh and another that has experienced stress periods. This sensor is
suitable for monitoring dynamic degradation behavior, such as on-chip path-delay
degradation. The advantages of utilizing ring-oscillators are twofold: they do not
require additional off-chip measurement equipment, and they can be placed in
environments where the circuits are in operational condition. The drawback is that
this technique cannot measure the degradation of a single transistor because all
transistors that compose the ring oscillator are involved in the degradation of the
oscillation frequency.
In an attempt to enhance the sensitivity of several selected transistors in the ring-
oscillator sensor, an inhomogeneous ring oscillator sensor was proposed [6]. In this
type of sensor, an inverter stage is designed to have significantly larger delay than
the other stages by connecting a pass gate transistor—pMOSFET or nMOSFET—
between the output of the preceding inverter and the inverter stage of interest.
The device type is selected such that the gate voltage of the inverter is limited by
the threshold voltage, making its delay very sensitive to the threshold voltage of the
pass gate transistor. To keep track of the degradation of SRAM cells, a ring oscillator
composed of embedded SRAM was proposed in [26].
3 BTIarray for Statistical Characterization

of Device Degradation
As device dimensions continue to shrink, the variability of device parameters is

becoming an increasing concern. Similarly, the variability of the parameter degrada-
tion may differ between devices. Even if two identical devices have experienced the
same BTI stress, their degradation behaviour differs because of inherent variability.
In order to accurately account for the impact of degradation during circuit designs,
the modeling of degradation variability, in addition to the average degradation trend,
is very important [28, 29]. For this purpose, measurements for a large number
of devices are vital in characterizing the statistical properties of BTI phenomena.
However, the methods that were previously reviewed target the accurate capture of
the degradation of a single transistor. Considering the long measurement period,
they are difficult to apply for characterizing a statistically significant number of
devices.
One promising approach is to construct an array circuit on a chip to facilitate
overlap of a long biasing time when collecting degradation data on a large number
of devices. In the following subsections, the circuit design and experimental results
of an array-based on-chip approach will be reviewed.
3.1 Stress and Recovery Time Overlapping
When measuring BTI phenomena, bias voltages have to be applied for a long time
to observe a detectable degradation in the threshold voltage. The largest obstacle for
statistical BTI characterization is the very long time required for the measurement
of multiple devices.
Upon close look examination of the BTI measurements, it is observed that each
threshold voltage measurement takes up only a very short time as compared to the
stress or recovery period. Most of the measurement period is used for applying
the bias voltages to the device. In such a condition, which is typical in BTI
characterizations, the total period for the measurement of multiple devices will be
significantly shortened by the overlapping stress periods of different devices.
A straightforward approach is to conduct I–V curve measurements for multiple
devices using a high-pin-count probe. In that case, all electrodes of the devices
need to be equipped with individual pad electrodes for probing. The use of a probe
card with a very high pin-count correspondingly requires a large number of SMUs,
which is impractical. Signal synchronizations become more complicated and costly
Operating
mode MS SA MS
measurement Time
DUT0 stress
DUT1 stress
DUT2 stress
DUT3 stress
DUT4 stress
DUT5 stress
DUT(n−1) stress
A B
Fig. 5.1 Stress period overlapping on n devices. Stress periods with equal durations can be
achieved for all devices
as a higher number of parametric analyzers are used. Another drawback of this

type of approach is area overhead. A set of four probing pads is necessary for
the measurement of a device. Because a probe needle has to physically contact a
probing pad, the pad area cannot be arbitrarily small. A probing pad occupies a
significantly larger area than the device of interest does, which significantly reduces
the area efficiency. This area gap between the device and probing pad will increase
as transistors are miniaturized.
To resolve the above inefficiency, an array-based approach that simultaneously
measures the degradation of many devices has been proposed, which is called
BTIarray [22]. The concept of the bias overlapping in the BTIarray is illustrated
in Fig. 5.1. In [22], the measurement results of 128 transistors are also presented,
which validates the stress overlapping concept. A similar idea has been proposed
in [24] with only simulation results.
The bold lines in Fig. 5.1 represent the stress periods in which devices under test
(DUTs) are in a stress condition. The small box represents a period in which the
threshold voltage of a selected DUT is measured. A single voltage source is shared,
and the voltage is distributed for all DUTs in the array to apply an equal stress
voltage for an exactly equal period, while the threshold voltage measurements are
performed individually. With the aid of on-chip circuitry, probing pads and hence
the SMUs are shared by the all DUTs in the array.
The application of stress or recovery bias voltages is controlled by an on-chip
circuit that is implemented with the DUT. In contrast to the parallel probing,
a single supply voltage is shared among all the DUTs on the chip. The circuit
structure and measurement process will be reviewed in detail in the next section.
It is also crucially important to completely automate the measurement process in
the statistical device measurements. The measurement environment of the BTIarray
will also be described in the following section.
3.2 Circuit Structure of BTIarray
BTIarray provides several modes to realize a series of measurement procedures of

many DUTs. Figure 5.1 depicts example modes and procedures. In time slot A, a
single DUT is selected for threshold voltage measurement, and all other DUTs are in
a stress condition. During time slot B, a stress bias voltage is applied for all DUTs.
When observed from a DUT, the DUT is always in one of three conditions—stress,
recovery, or measurement. Considering the typical measurement procedure of BTI
degradation, the BTIarray should provide at least four operating modes:
• stress all (SA): apply stress bias for all DUTs simultaneously,
• recovery all (RA): apply recovery bias for all DUTs simultaneously,
• measurement stress (MS): measure threshold voltage of a selected DUT, and
simultaneously apply stress bias for other DUTs,
• measurement recovery (MR): measure threshold voltage of a selected DUT, and
simultaneously apply recovery bias for other DUTs.
In [22], to realize all of the above operational modes, the unit structure of the
BTIarray is defined as shown in Fig. 5.2a. This circuit is called a DUT unit. The
DUT unit consists of a DUT cell and a control logic circuit. The DUT unit is
designed so that it can be arranged adjacent to the other DUT units in the horizontal
and vertical directions to form a device array similar to [23]. The array structure
facilitates sharing of the supply voltage and control signal networks among many
DUTs. Forming an array also maximizes DUT density by sharing probing pad
electrodes, realizing a compact test structure. Control logic generates switch control
signals according to the combination of the control and address-select signals:
MEAS, STRS, and MSEL. The internal signal MSEL becomes logical ‘1’ when
the DUT unit is selected by the address signal.
A simplified schematic of a DUT cell that implements a pMOS-transistor as
the DUT is shown in Fig. 5.2b. In the DUT cell, three terminals of the DUT
are connected to different voltage sources through pass-gate switches to realize a
specified operating mode. Voltages of VSS , VSTR , VREC , and VDD are shared among
all DUTs in the circuit, each of which is connected to a constant voltage source. VDD
and VSS are nominal supply and ground voltages, respectively. VCC is the terminal
to connect a current source for the threshold voltage measurement using the constant
current method. All switches are realized by CMOS pass-gate transistors.
An example construction of a BTIarray is shown in Fig. 5.2c [22]. In this
example, an array of 128 DUTs is formed. In order to select a DUT in the array,
7-bit address signal is decoded into 8-bit and 16-bit select signals, using X- and
Y-decoders, respectively.
a b
VSS
MEAS sw0
STRS XSEL
VSTR
VSS sw1 DUT
VSTR
YSEL sw4
sw2
DUT
MSEL Control sw sw3
logic cell VREC
cntl
VM
VREC
VM sw6 sw5
VDD VDD
VCC sw7
VCC
DUT unit DUT cell
c
YSEL<7:0>
DUT unit
Y−decoder (3b)
W/L=705/60 W/L=705/120
32 DUTs 32 DUTs
W/L=360/60 W/L=360/120
32 DUTs 32 DUTs
XSEL<15:0>
Address X−decoder (4b)
BTIarray
Fig. 5.2 Circuit structure of BTIarray for a pMOS DUT. (a) The DUT unit consists of “control
logic” and a “DUT cell.” All signals pass through the DUT unit, allowing an array circuit to be
easily formed. Control logic generates switch control signals for the DUT cell. (b) Schematic
diagram of a DUT cell. Switches are realized by pass-gate transistors. (c) Placement of DUT units
in an implementation of a BTIarray [22]. An array of 128 DUTs is formed. Power supply and
control signals are omitted for clarity except for XSEL and YSEL
Switch Control
In the SA mode, stress voltage is applied for a DUT in the DUT cell by closing
switches sw1, sw4, and sw6. Other switches are controlled to be open. The stress
voltage is VSTR minus VDD , where VSTR is more negative than VDD . In the RA
mode, switches sw0, sw3, and sw6 will be closed, and others are left open. VREC is
a recovery voltage and nominally equals VDD , which turns devices off.
In the two measurement modes, MS and MR, a measurement is carried out for a
DUT that is selected by an address signal. Other DUTs that are not selected are all
in either a stress or a recovery condition. DUT selection is specified by XSEL and
YSEL signals, both of which are generated as one-hot signals by the decoders in
Table 5.1 An example truth MEAS STRS MSEL Mode

table of control logic
0 0 * Recovery
0 1 * Stress
1 0 0 Recovery (in MR)
1 0 1 Measurement (in MR)
1 1 0 Stress (in MS)
1 1 1 Measurement (in MS)
An asterisk (*) in the table indicates “don’t care”
Fig. 5.2c. Threshold voltage is measured by the constant-current method [16]. The
voltage that appears at node VM is captured during a constant current is applied to
Vcc. Hence, the switches sw0, sw2, sw5, and sw7 are closed in the selected DUT
cell. The switches of other DUT cells that are not selected are equal to those of
the SA mode or RA mode, depending on whether the current mode is MS or MR,
respectively.
Control Logic Design
The control logic circuit in the DUT unit in Fig. 5.2a can be realized according to
the truth table in Table 5.1. In the BTIarray realization in [22], mode switching
is achieved by the combination of three control signals: “MEAS,” “STRS,” and
“MSEL.” A DUT is selected for the measurement only when both MEAS and MSEL
are “1,” i.e., the array is in one of the measurement modes, and the address of the
DUT is selected. Otherwise, the constant bias voltages that correspond to either the
stress or recovery mode is applied to all the DUTs in parallel.
Advantages of BTIarray
BTIarray achieves substantial reduction during the measurement period. Because

the stress period is usually several or many orders of magnitude longer than the time
required for threshold voltage measurement, the total measurement period becomes
very close to that of a single DUT, regardless of the number of DUTs in the array. In
addition, with the array structure, the time for the mode transitions can be made short
and equal for all DUTs. This is a preferable characteristic when collecting statistical
data over many DUTs. During the entire measurement period, all voltage sources
and a current source are maintained at constant values so that the change in bias
voltages is realized by the on-chip pass-gate switches. Quick current and voltage
steering is achieve by this construction because parasitic capacitance existing in the
array side of the switches is significantly smaller than that of an off-chip parameter
analyzer. Hence, fast transitions between the operational modes are achieved, which
contributes to precise equalization of the stress and recovery periods.
Table 5.2 Chip specifications [1]

pMOS array nMOS array
Technology 65-nm, 11-metal-layer CMOS
Circuit area 225.2 m 62.3 m (DUT array), 489.2 m 332.8 m (incl. pads)
Implemented DUTs 128 DUTs, 32 each size
width(nm)/length(nm) 705/60, 705/120, 360/60, 360/120 360/60, 360/120, 180/60, 180/120
Precision
0.07 mV rms 0.05 mV rms
at 10 ms/acq.
a b
DUT
Control
332.8 μm
128 DUT array
7.4 μm
logic
225μm x 58.6μm
AND
Switches
10.8 μm 489.2 μm
Layout of the DUT unit. Chip microphotograph.
Fig. 5.3 An example implementation of the BTIarray. (a) Layout of the DUT unit, showing
selected layers up to the second level wiring. (b) A total of 128 DUTs are implemented sharing
only 16 probing pads [22]
3.3 Implementation Example
The implementation and measurement results of the BTIarray for pMOS transistors
were first presented in [22]. Then, in [1], measurement results of BTIarrays for both
pMOS and nMOS transistors were presented. The nMOS array was designed as a
separate array from the pMOS array, but the circuit structure of the two arrays is
similar. The only difference between the two arrays is the switched polarities of the
stress and recovery voltages, and the connection of the DUT.
The specifications of the array design using a 65-nm process are summarized
in Table 5.2. In both arrays, four transistor sizes were implemented as DUTs in
combination with two channel lengths and two channel widths.
The layout diagram of the pMOS DUT unit including the pass-gate switches
and control logic circuit is shown in Fig. 5.3a. The layout size of a DUT unit is
10.8 m 7.4 m. Maximum layout regularities have been pursued by sharing the
single layout skeleton in Fig. 5.3a for four different DUT sizes. As a result, the
layout designs of the switches and the control logic were completely identical except
for the sizes of transistors that were located at the center of the DUT unit. This
regularity minimized layout-dependent device parameter variation. Additionally, it
greatly eased the layout design of the arrays.
As observed Fig. 5.2c, the number of DUT units with an equally sized DUT is
32. In these designs, the DUT units are regularly placed to form a 4 8 array. Four
4 8 arrays are again arranged to construct an entire array; thus, 128 DUTs are
implemented in total. A microphotograph of a test chip is presented in Fig. 5.3b.
The layout size of the array including decoders and periphery circuit is 225.2 m
62.3 m, and the total test circuit area including the probing pads is 489.2 m
332.8 m.
Considering the circuit structure of the DUT unit and formation of the entire array
structure, the number of DUTs in an array can be easily scaled. More recently, it has
been reported that 3,996 DUTs have been successfully implemented in an array
circuit by extending the concept of BTIarray without compromising measurement
accuracy [2]. A slight circuit modification was made to suppress leakage current
caused by a large number of off-state pass-gate switches to maintain measurement
accuracy. For that purpose, the concept of a dummy supply voltage [23] is utilized
to eliminate the leakage current. If the dummy supply voltage is not used, the
constant current, which should flow through a selected DUT to develop its threshold
voltage, becomes inaccurate because the current will be reduced significantly by the
leakages.
3.4 Measurement Automation Through Scripting
When measuring of device degradation for many transistors, it is important to

precisely control stress and recovery periods. The signal timing is particularly
critical because a BTI-induced threshold voltage shift has components that quickly
respond to the change in the bias voltages. Even a short stress interruption can lead
to a partial recovery of the threshold voltages. Other conditions such as equalizing
the temperature of the transistors are also important when statistical degradation
results for a large number of DUTs are to be collected.
When measuring statistical degradation, it is also important to make the exper-
iments completely reproducible. Because of the long measurement period, com-
plicated commands have to be sent to the circuit many times in order to observe
the response to the changing bias voltages. The ability to repeat exactly the same
measurement procedure for multiple arrays is also required to further increase the
number of samples.
In order to fulfill all these requirements, a scripting language interface for
the BTIarray measurement is proposed [21]. The automation of setting mea-
surement parameters and of issuing commands to the instruments is realized by
the instrumentation as depicted in Fig. 5.4. A scripting language, BTIscript, has
been developed in order to specify signal switch timing, voltage levels, chuck
temperature, and other parameters, at a higher abstraction level. The BTIscript
was realized on the basis of the Python scripting language. When writing the
measurement scenario using the BTIscript, all types of powerful features equipped
Fig. 5.4 Instrumentation for the measurement of BTIarray. Stress, recovery, and measurements
can be entirely written in a macro language, BTIscript
Table 5.3 Subset of macro commands realized as functions in Python used in BTIscript
Macro Description
stress_all(duration) All DUTs are biased to stress for duration seconds.
recovery_all(duration) All DUTs are biased to recover for duration seconds.
measure_stress_all(burst) Measure all DUTs in ascending order. Measurement
of a DUT is conducted for burst times while all other
DUTs are biased to stress.
measure_stress_one(adrs, burst) Measure single DUT specified by address adrs. Mea-
surement of a DUT is conducted for burst times while
all other DUTs are biased to stress.
measure_recovery_all(burst) Measure all DUTs in ascending order. Measurement
of a DUT is conducted by burst times while all other
DUTs are biased to recover.
measure_recovery_one(adrs, Measure single DUT specified by address adrs. Mea-
burst) surement of a DUT is conducted by burst times while
all other DUTs are biased to recover.
Parameters in parentheses may have a default value as in a regular Python function. The addition
of new commands that are dedicated to a specific measurement is easy [21]
with the Python language, such as predefined and user-defined classes, subroutine
calls, and flow controls, can be fully utilized.
The scripting language is useful to efficiently develop and debug measurement
scripts because of its user-friendliness. Without such an environment based on
a programming language, lengthy, complicated, and thus error-prone commands
would have to be hand-written. In particular, iterations and procedure calls with
parameters are very useful to represent measurement procedures both formally and
compactly. Macro commands covering typical measurement procedures for BTI
degradation are listed in Table 5.3.
To understand the versatility of the script, let us proceed through an example
script in Fig. 5.8 by comparing the commands to the measurement results, such as
in Fig. 5.9. Other measurement examples can also be found in [21, 27, 29].
Figure 5.5 shows a block diagram of the measurement system. A hand-written
BTIscript is first converted into a pattern file and GPIB commands. The pattern file
defines the transition timings of the control signals, and the commands from GPIB
and triggers from the FPGA-based pattern generator control the parametric analyzer.
The generated pattern file is transferred to the FPGA via a USB interface.
Sequences of bit patterns and output timings are transferred from the host PC via
Fig. 5.5 Block diagram of the FPGA-based pattern generator. A 24-bit command output is
generated to control the BTIarray and a parametric analyzer
USB. Each atomic command is formatted in a 64 bit command word. The upper
40 bits and lower 24 bits correspond to the count value and the state of the control
signals, respectively. The count value determines the duration for which the control
signals of a mode are applied. It is used as an initial value of the 40-bit countdown
counter. When the counter reaches zero, the next command word is fetched from the
FIFO queue to reload the counter and the output register.
The GPIB commands setup output voltages, output current, and the measurement
parameters such as integration time. Once the measurement is started, threshold
voltage measurement is conducted by the parametric analyzer according to the
trigger signal generated in the FPGA.
3.5 Measurement Accuracy
The timing accuracy of the FPGA-based measurement environment was evaluated

using an oscilloscope, referencing the trigger output that is generated by the
parametric analyzer at the end of each measurement operation. When the target
interval is 10 ms, the standard deviation of the transition timing was 19.0 s in the
FPGA-based pulse generator [1]. That of a PC-based command control [21, 22]
using a GPIB interface was 452 s.
In [1], the precision of the voltage measurements was also evaluated, which
was approximately 50 V rms or lower. It is reported that the voltage precision
was evaluated using a DUT that had very small fluctuations because small stair-
like fluctuations in the threshold voltage due to RTN were observed in almost all
DUTs. The RTN of a DUT can be easily measured by repeating threshold voltage
measurements in the MR mode. An example RTN waveform captured using the
BTIarray fabricated using a 65-nm process is shown in Fig. 5.6. In the figure,
multilevel RTN is clearly captured.
Time (s)
Time (s)
Fig. 5.6 Example threshold voltage measurement of a pMOS device, showing multilevel random
telegraph noise (DUT size: W/L = 360/60 nm/nm)
Ideally, the currents and voltages of the measurement paths should be switched
instantly by the on-chip pass-gate switches. However, in reality, there is a lower
limit on the transition time because of parasitic capacitances in the circuit and the
measurement system. Hence, the output voltage node requires a finite time to settle
to a final value. The settling time of the output voltage of the BTIarray has also been
evaluated in [1] using an oscilloscope. The measured settling time was distributed
between 10 and 60 s for almost all DUTs in the pMOS and nMOS arrays. The
minimum interval of the measurements can be reduced to 100 s or shorter in the
128-DUT array.
4 Characterization Example of Statistical Degradations
4.1 Measurement Scenario
Figure 5.7 depicts the scenario for a BTI degradation measurement. The BTIscript
that defines the scenario is also shown in Fig. 5.8. First, the output values of the
voltage sources and current source are set (lines 1–8). Next, measurement timings
are defined for later use (lines 10–12). Then, the threshold voltages of all DUTs
in recovery bias mode are measured to obtain initial threshold voltages for the
DUTs before applying stress (line 21). The measurement interval is set to 10 ms,
which is determined by the minimum interval of the parametric analyzer. This step
Fig. 5.7 Scenario for BTI degradation measurement
Fig. 5.8 Script that realizes measurement scenario of Fig. 5.7. Stress, recovery, and measurement
processes are written using macro commands. After voltage sources and current source are
initialized and measurement interval is defined, threshold voltages of DUTs in recovery bias mode
are measured 1,000 times. Then, all DUTs undergo stress for 20,000 s during which time threshold
voltage measurements are conducted intermittently to reduce effect of interrupting stress. Finally,
all DUTs returned to recovery mode, and threshold voltages are measured 1,000 times
requires 112810 ms D 1:28 s in total. Note that, during these measurements, the
DUTs that are not selected for measurement are kept in recovery bias mode to avoid
degradation. After that, all DUTs enter stress bias mode (lines 22–24). Stress is
continuously applied for 20,000 s. Measurements are intermittently conducted
during the stress period for all DUTs in series at 1, 2, 5, 10, 20, 50, 100, 500,
1,000, 5,000, 10,000, 12,000, 15,000, 18,000 and 20,000 s. The macro command
“measure_stress_all” is repeatedly issued by the “for” statement according to the
list of measurement intervals. Then, all DUTs return to recovery bias mode, and
measurements are conducted for all DUTs 1,000 times (lines 25–26). Repetitions
of 1,000 are realized by issuing the macro command “measure_recovery_all” in a
for loop.
The upper part of Fig. 5.7 shows the detailed timing of the control signals used
for acquiring a threshold voltage of the selected DUT in the BTIarray. First, a
“MEAS” signal is asserted. Then, the measurement trigger is issued after 100 s
have elapsed. This timing was determined according to the settling time of the output
node voltage. This ensures that the parametric analyzer starts an integration of VM
after the node voltage has been stabilized.
The threshold voltage measurement is based on the constant current method [16]
in which the gate voltage of the DUT is switched from VSTR to VSS when the DUT
is under measurement. The source voltage becomes the threshold voltage of the
selected DUT, which reduces the stress on the DUT under measurement. Because
the change in the threshold voltages contains a very fast component, partial recovery
may occur during the measurement. To minimize this recovery, the “MEAS” signal
is negated immediately after the sampling time required by the parametric analyzer
has elapsed.
4.2 Threshold Voltage Shifts for Small DUTs
Examples of the measured threshold voltage shift are shown in Fig. 5.9. The results
for two DUT sizes are presented: W/L = 360 nm/60 nm and 705 nm/120 nm. Ten
traces were randomly selected from each size. In the 20,000 s stress period, the
threshold voltage increased for all DUTs, but the traces of the increases differed
among DUTs. The variation in the threshold voltage increases was larger for DUTs
with a smaller channel area. Because the variation is expected to increase as the
transistor size decreases, statistical representations of the degradation parameters
are required.
In addition, a nonmonotonic increase in the threshold voltage was observed
for almost all of the DUTs. Although its cause is still not understood, there are
two possible hypotheses. The first hypothesis is recovery during measurement. As
described above, the measured DUTs experience less stress voltage than the other
fully stressed DUTs. This may lead to device-dependent partial recovery during
measurement.
a b
Fig. 5.9 Temporal changes in the threshold voltage (pMOS array) for ten randomly selected
DUTs. Variation in threshold voltage increases was larger for smaller-area DUTs. In the recovery
period, discrete changes in threshold voltage increases were observed for most of the DUTs. The
discrete changes are more distinct for smaller-area DUTs. (a) DUT size: W/L=360 nm/60 nm.
(b) DUT size: W/L=705 nm/120 nm
The other hypothesis is the interaction between carrier traps and emissions during
the stress period. The nonmonotonic increase may have been caused by the carrier
trap and emission process at interface defects, which is considered to be the major
mechanism of BTI degradation [10, 31]. This is a stochastic process; thus, the
fluctuation in the threshold voltage is observed at any time.
The time courses of threshold voltage recovery are magnified to the right of the
figures. The discrete recovery of the threshold voltages was more prominent for
the DUTs with a small channel area. Smoother recovery was again observed for
the DUTs with a larger channel area. This trend held true for the reminder of the
DUTs. The origin can also be explained by the averaging effect of carrier trap and
emission process. Under an assumption of uniform defect density over all areas of
the measured chip, a larger number of interface traps should be found in the DUTs
of a larger channel area than in those of a smaller channel area. The trapping and
emission of carriers occur more often in the larger DUTs, and the threshold voltage
shift caused by the trap of a single carrier is smaller than that of the smaller DUTs.
To investigate the stair-like changes in more detail, the response to the stress
and recovery bias was repeatedly measured for the DUT with a small channel area.
A stress with a duration of 1,000 s was first applied, and then, recovery mode
was applied for 1,280 s, during which 1,000 measurements were conducted for
each DUT. This set of stress-recovery measurements was repeated 100 times to
characterize the stochastic recovery process. The upper figure in Fig. 5.10 shows the
recovery period of the threshold voltage shift. The results of 30 trials, which are
again randomly selected from 100 trials, are presented.
In the recovery period, the threshold voltage shifts of an equal magnitude were
found throughout different repetitions. The magnitude of the threshold voltage
shift was determined by the defect location in the channel. Hence, the defects
could be identified by the magnitude. This characterization method is known as
time dependent defect spectroscopy (TDDS) [9]. The associated TDDS plot is also
Fig. 5.10 The threshold 40

voltage shift of a DUT
observed in repetitive
30
stress-recovery
VTH (a.u.)
measurements. The results of #1 #0
30 trials randomly selected 20
from 100 trials are presented. #0
The corresponding TDDS 10
#1 #0
plot is shown in the lower #0
graph [1]
0
0 100 200 300 400
Time (s)
10
8
ΔVTH (a.u.)
6
Defect #0
4
Defect #1
2
0 100 200 300 400
Emission time (s)
a b
DUT size: W/L=360 nm/60 nm DUT size: W/L=705 nm/120 nm
Fig. 5.11 Temporal changes in threshold voltage for ten randomly selected DUTs, measured for
an nMOS array. Both the degradation and recovery occurred in a very short time
shown in Fig. 5.10. There are two major clusters of defects. The emission time
constant of Defect #0 assumes a value of approximately 10 to 50 s, whereas that of
Defect #1 distributes wider in the range of 100 to 400 s. Characterizing the range and
distribution of the emission time is important in device lifetime extension methods,
such as a circuit that alternately uses one of two identical circuits [15].
Similarly, Fig. 5.11 shows the results of the nMOS array obtained using the same
measurement scenario as the pMOS array. The DUT set is again randomly selected.
By comparing the graphs with those in [1], readers can find the similar trend of
the temporal threshold voltage changes for nMOS DUTs. Although smaller than
that of pMOS DUTs, the threshold voltage of an nMOS DUT is increased. The
threshold voltage shift seemingly reached a plateau relatively quickly within 5,000 s,
and remained almost constant for the reminder of the stress time. In this device,
nMOS DUTs seemingly lack “slow trap” components [18]. In addition, regardless
of the device size, the threshold voltages of most of the DUTs recovered almost
completely once the stress was removed, which is different from the case of the
pMOS DUTs.
4.3 Statistical Model Parameter Extraction for NBTI
As we have observed in Fig. 5.9, there was a large difference in the threshold voltage
shift even for DUTs of equal channel-area size. The threshold voltage increase
during a stress period has been commonly modeled using a power law model in
which the stress period and threshold voltage increase are related by Vth / t n ,
where t is the stress period, and n is the power law exponent that varies between
transistors.
The distributions of the power law exponent for 720 DUTs (180 DUTs for each
size) measured using five pMOS arrays at 125 ı C are depicted in Fig. 5.12. The
a b
15 15
Sample quantile
Sample quantile
0.4 0.4
μ = -1.99 0.3 μ = -2.05 0.3
σ = 0.47 σ = 0.35
Frequency
Frequency
0.2 0.2
10 0.1
10 0.1
0.0 0.0
0.0 0.1 0.2 0.3 0.4 0.0 0.1 0.2 0.3 0.4
Model quantile Model quantile
5 5
0 0
0.0 0.1 0.2 0.3 0.4 0.0 0.1 0.2 0.3 0.4
n n
c 15 d 15
Sample quantile
0.4 0.4
Sample quantile
μ = -2.00 0.3 μ = -2.01 0.3

σ = 0.32 σ = 0.26
Frequency
Frequency
0.2 0.2
10 0.1
10 0.1
0.0 0.0
0.0 0.1 0.2 0.3 0.4 0.0 0.1 0.2 0.3 0.4
Model quantile
5 5 Model quantile
0 0
0.0 0.1 0.2 0.3 0.4 0.0 0.1 0.2 0.3 0.4
n n
Fig. 5.12 Measured distributions of power law exponent and their log-normal approximations
for four sizes of DUTs [1]. The channel-area dependency of the variation is evident.
(a) W/L D 360 nm/60 nm. (b) W/L D 360 nm/120 nm. (c) W/L D 705 nm/60 nm. (d) W/L D 705
nm/120 nm
5 W/L=360/60 5 W/L=360/60
W/L=705/60 W/L=705/60
4 W/L=360/120 4 W/L=360/120
(a.u.)
(a.u.)
W/L=705/120 W/L=705/120
3 3
2 2
1 1
0 0
0 1 2 3 4 5 Probability
Initial VTH (a.u.)
Fig. 5.13 Correlations between threshold voltages of fresh DUTs and the threshold voltage shift
after a stress with a duration of 20,000 s [1]. The approximated distributions of the threshold
voltage shift are also plotted
variance in the exponent decreased as the DUT size increased. All distributions
are approximated by a log-normal distribution, which is indicated by a solid line.
The two shape parameters of each distribution, and , are also provided. An
average of the log-normal distribution , which represents the logarithmic mean
of the power-law exponent, was approximately 2 regardless of the DUT size.1 The
inset graph shows quantile-quantile plot for the log-normal approximation.
The standard deviation of the power-law exponent increases as the channel
area of the DUTs decreases. The standard deviation of the power-law exponent
is inversely proportional to the channel area of a device. The inverse dependency
between the device channel area and exponent variance has also been reported in [2].
The exponent variation increases at a faster rate than the threshold voltage, whose
variation is inversely proportional to the square root of the channel area.
The accurate consideration of correlations between device parameters is an
effective way to avoid overdesign. Figure 5.13 shows a scatter plot between
threshold voltages of fresh DUTs and threshold voltage shifts after a stress period
of 20,000 s. The results of the four DUT sizes are shown. For the transistors of these
sizes, initial threshold voltages and the threshold voltage shifts fit normal and log-
normal distributions, respectively. As expected, no correlation was found between
the initial threshold voltage and its shift because the variation sources are different
for these variations.
1
D 2 approximately corresponds to an exponent n D 0:15.
4.4 Measurement Efficiency
The time to complete the scenario in Fig. 5.7 was approximately 5 h and 55 min
whereas that for a single DUT was approximately 5 h and 33 min. If a single DUT
measurement is serially conducted for 128 DUTs, the measurement would require
approximately 710 h, which is 29.6 days, even when the time for moving the probe
needles to the next DUT is ignored. Considering that 128 DUTs are integrated in an
array, BTIarray has thus achieved an increase in speed of a factor of 120 with this
measurement scenario, enabling month-long measurements to be completed within
a day. An ideal increase in speed of a factor of 128 could not be achieved because
the time required for threshold voltage measurements is finite. The parametric
analyzer requires 10 ms for the voltage measurement. Therefore, the acquisitions
of the threshold voltages for 128 DUTs for a total of 1,018 measurements require
approximately 22 min. The use of a faster SMU shall further shorten the total
measurement time.
5 Chapter Summary
Characterization techniques for the transient changes in device parameters have

been reviewed. Aiming to efficiently characterize degradation statistics, an on-chip
circuit structure called BTIarray has been developed. In the BTIarray, transistors
are placed in an array such that the stress bias voltages will be applied for all the
devices in parallel by sharing a single source voltage. By applying bias voltages
in parallel, the aggregate stress time required for a large number of devices is
substantially shortened. The array also offers a function to measure the threshold
voltage of a selected device while maintaining the stress voltages to other devices,
i.e., the measurement of a device can be carried out in parallel while other devices
are in stress or recovery conditions. With the array circuit structure, measurement
conditions, including timing, voltage, temperature, and other parameters, will be
equalized. In order to facilitate the description of a scenario in which trigger
timings for stress, recovery, and measurements are given, a scripting language
environment has been provided. Array-based circuit design as well as a development
and debugging environment will become even more useful to achieve accurate and
reproducible measurements.
For the proof of concept, the experimental results of BTIarray, fabricated using
a 65-nm CMOS process, are also presented. Using an array consisting of 128
devices, transient changes in the threshold voltages were measured. With the array
structure, a month-long measurement time has been shortened to less than 6 h,
while maintaining sufficient voltage and timing precision required in statistical
characterizations. The measurement results have shown the following statistical
observations. (1) Degradation and recovery slopes vary between devices even for
devices with equal channel areas. (2) The variance of the time exponent is inversely
proportional to the channel area. (3) Most of the measurement results involve a stair-
like recovery trace, which suggests the interface trap charge is playing a key role in
the threshold voltage shift.
Acknowledgements Part of this work has been supported by MEXT/JSPS KAKENHI Grant
Numbers 22360143 and 20431992.
References
1. Awano, H., Hiromoto, M., Sato, T.: BTIarray: A time-overlapping transistor array for efficient
statistical characterization of bias temperature instability. IEEE Trans. Dev. & Mat. Reliab.
(2014)
2. Awano, H., Hiromoto, M., Sato, T.: Variability in device degradations: Statistical observation
of NBTI for 3996 transistors. In: Proc. European Solid-State Dev. Res. Conf. (ESSDERC)
(2014)
3. Campbell, J., Lenahan, P., Krishnan, A., Krishnan, S.: Observations of NBTI-induced atomic-
scale defects. IEEE Trans. Dev. & Mat. Reliab. 6(2), 117–122 (2006)
4. Denais, M., Parthasarathy, C., Ribes, G., Rey-Tauriac, Y., Revil, N., Bravaix, A., Huard, V.,
Perrier, F.: On-the-fly characterization of NBTI in ultra-thin gate oxide PMOSFET’s. In: IEDM
Tech. Dig., pp. 109–112 (2004)
5. Du, G., Ang, D., Teo, Z., Hu, Y.: Ultrafast measurement on NBTI. IEEE Trans. Electron Device
Lett. 30(3), 275–277 (2009)
6. Fujimoto, S., Islam, A.M., Matsumoto, T., Onodera, H.: Inhomogeneous ring oscillator for
within-die variability and RTN characterization. IEEE Trans. Semicond. Manufact. 26(3),
296–305 (2013)
7. Ghosh, A., Rao, R.M., Brown, R.B., Chuang, C.T.: On-chip negative bias temperature
instability sensor using slew rate monitoring circuitry. In: ACM IEEE Intl. Symposium on
Low Power Electronics and Design (2009)
8. Grasser, T., Kaczer, B., Goes, W.: An energy-level perspective of bias temperature instability.
In: Proc. Intl. Reliab. Phys. Symp. (IRPS), pp. 28–38 (2008)
9. Grasser, T., Reisinger, H., Wagner, P., Schanovsky, F., Goes, W., Kaczer, B.: The time
dependent defect spectroscopy (TDDS) for the characterization of the bias temperature
instability. In: Proc. Intl. Reliab. Phys. Symp. (IRPS), pp. 16–25 (2010)
10. Huard, V., Denais, M.: Hole trapping effect on methodology for DC and AC negative bias
temperature instability measurements in pMOS transistors. In: Proc. Intl. Reliab. Phys. Symp.
(IRPS), pp. 40–45 (2004)
11. Kim, J.J., Rao, R., Schaub, J., Ghosh, A., Bansal, A., Zhao, K., Linder, B., Stathis, J.:
PBTI/NBTI monitoring ring oscillator circuits with on-chip Vt characterization and high
frequency AC stress capability. In: Proc. VLSI Technology symposium, pp. 224–225 (2011)
12. Kim, T., Persaud, R., Kim, C.: Silicon odometer: An on-chip reliability monitor for measuring
frequency degradation of digital circuits. IEEE J. Solid-State Circ. 43(4), 874–880 (2008)
13. Kumar, E., Maheta, V., Purawat, S., Islam, A., Olsen, C., Ahmed, K., Alam, M., Mahapatra, S.:
Material dependence of NBTI physical mechanism in silicon oxynitride (SiON) p-MOSFETs:
A comprehensive study by ultra-fast on-the-fly (UF-OTF) iDLIN technique. In: IEDM Tech.
Dig., pp. 809–812. IEEE (2007)
14. Matsumoto, T., Makino, H., Kobayashi, K., Onodera, H.: A 65 nm complementary metal-
oxide-semiconductor 400 ns measurement delay negative-bias-temperature-instability recov-
ery sensor with minimum assist circuit. Japan. J. Appl. Phys. 50(4), 04DE06 (2011)
15. Matsumoto, T., Makino, H., Kobayashi, K., Onodera, H.: Multicore large-scale integration
lifetime extension by negative bias temperature instability recovery-based self-healing. Japan.
J. Appl. Phys. 51, 04DE02 (2012)
16. Ortis-Conde, A., García Sánchez, F.J., Liou, J.J., Cerdeira, A., Estrada, M., Yue, Y.: A review
of recent MOSFET threshold voltage extraction methods. Microelec. Reliability 42, 583–596
(2002)
17. Rauch, S.E.: Review and reexamination of reliability effects related to NBTI-induced statistical
variations. IEEE Trans. Dev. & Mat. Reliab. 7(4), 524–529 (2007)
18. Reisinger, H., Blank, O., Heinrigs, W., Gustin, W., Schlunder, C.: A comparison of very fast
to very slow components in degradation and recovery due to NBTI and bulk hole trapping to
existing physical models. IEEE Trans. Dev. & Mat. Reliab. 7(1), 119–129 (2007)
19. Reisinger, H., Blank, O., Heinrigs, W., Muhlhoff, A., Gustin, W., Schlunder, C.: Analysis of
NBTI degradation- and recovery-behavior based on ultra fast VT-measurements. In: Proc. Intl.
Reliab. Phys. Symp. (IRPS), pp. 448–453 (2006)
20. Reisinger, H., Grasser, T., Gustin, W., Schlunder, C.: The statistical analysis of individual
defects constituting NBTI and its implications for modeling DC-and AC-stress. In: Proc. Intl.
Reliab. Phys. Symp. (IRPS), pp. 7–15 (2010)
21. Sato, T., Awano, H., Shimizu, H., Tsutsui, H., Ochi, H.: Statistical observations of NBTI-
induced threshold voltage shifts on small channel-area devices. In: Proc. Intl. Symp. Quality
Electronic Design (ISQED), pp. 306–311. IEEE (2012)
22. Sato, T., Kozaki, T., Uezono, T., Tsutsui, H., Ochi, H.: A device array for efficient bias-
temperature instability measurements. In: Proc. European Solid-State Dev. Res. Conf.
(ESSDERC), pp. 143–146 (2011)
23. Sato, T., Ueyama, H., Nakayama, N., Masu, K.: Accurate array-based measurement for
subthreshold-current of MOS transistors. IEEE J. Solid-State Circ. 44(11), 2977–2986 (2009)
24. da Silva, M., Kaczer, B., Van der Plas, G., Wirth, G., Groeseneken, G.: On-chip circuit for
massively parallel BTI characterization. In: IEEE Intl. Integrated Reliability Workshop Final
Report (IRW), pp. 90–93 (2011)
25. Stathis, J.H., Zafar, S.: The negative bias temperature instability in MOS devices: A review.
Microelec. Reliability 46(2–4), 270–286 (2006)
26. Tsai, M.C., Lin, Y.W., Yang, H.I., Tu, M.H., Shih, W.C., Lien, N.C., Lee, K.D., Jou, S.J.,
Chuang, C.T., Hwang, W.: Embedded SRAM ring oscillator for in-situ measurement of NBTI
and PBTI degradation in CMOS 6T SRAM array. In: VLSI Design, Automation, and Test
(VLSI-DAT), 2012 International Symposium on, pp. 1–4. IEEE (2012)
27. Velamala, J., Sutaria, K., Shimuzu, H., Awano, H., Sato, T., Wirth, G., Cao, Y.: Logarithmic
modeling of BTI under dynamic circuit operation: Static, dynamic and long-term prediction.
In: Proc. Intl. Reliab. Phys. Symp. (IRPS), pp. CM.3.1–CM.3.5 (2013)
28. Velamala, J.B., Sutaria, K.B., Sato, T., Cao, Y.: Aging statistics based on trapping/detrapping:
Silicon evidence, modeling and long-term prediction. In: Proc. Intl. Reliab. Phys. Symp.
(IRPS), pp. 2F2.1–2F2.5 (2012)
29. Velamala, J.B., Sutaria, K.B., Sato, T., Cao, Y.: Physics matters: statistical aging prediction
under trapping/detrapping. In: Proc. ACM/IEEE Design Automation Conf. (DAC), pp. 139–
144 (2012)
30. Wang, W., Yang, S., Bhardwaj, S., Vrudhula, S., Liu, F., Cao, Y.: The impact of NBTI effect on
combinational circuit: Modeling, simulation, and analysis. Trans. VLSI 18(2), 173–183 (2010)
31. Wirth, G., da Silva, R., Kaczer, B.: Statistical model for MOSFET bias temperature instability
component due to charge trapping. IEEE Trans. Electron Dev. 58(8), 2743–2751 (2011)
32. Zafar, S., Callegari, A., Gusev, E., Fischetti, M.V.: Charge trapping related threshold voltage
instabilities in high permittivity gate dielectric stacks. J. Appl. Phys. 93(11), 9298–9303 (2003)
Chapter 6
Compact Modeling of BTI for Circuit
Reliability Analysis
Ketul B. Sutaria, Jyothi B. Velamala, Athul Ramkumar, and Yu Cao
Abstract The aging process due to Bias Temperature Instability (BTI) is a key
limiting factor of circuit lifetime in contemporary CMOS design. Threshold voltage
shift induced by BTI is a strong function of stress voltage and temperature.
Furthermore, BTI consists of both stress and recovery phases, depending on the
dynamic stress conditions. This behavior poses a unique challenge for long-term
aging prediction for a wide range of stress patterns encountered in today’s circuits.
Traditional approaches usually resort to an average, constant stress waveform
to simplify the lifetime prediction. They are efficient, but fail to capture the
reality of circuit operation, especially under Dynamic Voltage Scaling (DVS) or
in analog/mixed signal designs where the stress waveform is much more random. In
this chapter, we present a suite of modeling solutions that enable aging simulation
under all dynamic stress conditions. The key innovation of this chapter is to develop
compact models of BTI when the stress voltage is varying. The results cover
the underlying physics of two leading mechanisms, Reaction–Diffusion (R–D)
and Trapping/Detrapping (T–D). Moreover, silicon validation of these models
is performed at 45 and 65 nm technology nodes, at both device and circuit
levels. Leveraging the newly developed BTI models under DVS and random input
waveforms, efficient aging simulation is demonstrated in representative digital and
analog circuits. Our proposed work provides a general and comprehensive solution
to circuit aging analysis under random stress patterns.
1 Introduction
Aggressive scaling of CMOS technology has been driving the evolution of elec-
tronics and the increase in the total number of transistors per chip. Moore’s law
predicts that the total number of transistors placed on a chip will be approximately
doubled every 18 months. Benefiting from the scaling of CMOS technology, system-
on-chips (SoCs) further improve the system capability by integrating both analog
K.B. Sutaria • J.B. Velamala • A. Ramkumar • Y. Cao ()

School of Electrical, Computer and Energy Engineering,
Arizona State University, Tempe, Arizona 85287, USA
e-mail: ycao@asu.edu

DOI 10.1007/978-1-4614-4078-9__6
94 K.B. Sutaria et al.
and digital circuits on a single platform. Present SoCs pack close to 800 million
transistors incorporating analog and mixed signal units such as data converters (A/D,
D/A), PLLs, etc., providing higher performance at a lower cost.
On the other side, aggressive CMOS scaling results in serious reliability threats
such as Bias Temperature Instability (BTI), Channel Hot Carrier (CHC) and Time
Dependent Dielectric Breakdown (TDDB) [1–10]. These reliability effects become
more pronounced partially due to process scaling techniques that improve device
and circuit performance. Moreover, scaling of supply voltage and threshold voltage
does not go hand in hand with scaling of device feature size. This greatly increases
vertical and lateral electrical field causing the exacerbation of reliability issues. As
an aftermath of these issues, circuit performance degrades over time and eventually
results in functionality error, which is called circuit aging.
Threshold voltage shift ( Vth ) due to negative BTI (NBTI) in PMOS devices
is the dominant aging mechanism in scaled digital and analog circuits. Accurate
characterization of the aging rate is essential to lifetime evaluation and guard
banding [11–24]. One major challenge in reliability prediction of NBTI comes
from the recovery phase, which causes the dependence on the stress pattern and
its history. The situation is more complex under Dynamic Voltage Scaling (DVS)
in low power digital design. Analog/mixed signal circuits face similar challenges
due to random input pattern. Modulation of voltage levels due to such operations
results in partial recovery of Vth and thus, the average or the worst case pattern does
not provide an accurate prediction of the lifetime. Figure 6.1 presents the operation
pattern in a 65 nm dual core Intel processor and Vth under such an operation.
DVS in a 65nm Intel Processor

1.25
VDD (V)
1.00
1.0
Simulated ΔVth
ΔVth (a.u.)
0.5
ΔVth from random DVS waveform
ΔVth using the averaged pattern
0.0
0.0 5.0x104 1.0x105
Time (s)
Fig. 6.1 4Vth predicted from the averaged voltage significantly deviates from that simulated by
dynamic aging models
6 Compact Modeling of BTI for Circuit Reliability Analysis 95
When VDD fluctuates, the degradation is strongly non-linear and varies according
to the stress condition. The average value of the stress voltage predicts a lower
or higher Vth shifts, as compared to that by the real-time model. Therefore, real-
time prediction of Vth shift allows a realistic and accurate assessment of circuit
aging, avoiding over-design to compensate the degradation. Overall, it is critical
to understand NBTI with recovery effects, simulate its effect under dynamic stress
voltages, and ensure reliable circuit operation for the desired lifetime.
1.1 BTI Aging Models: Challenges and Needs
To date, research work on aging mechanisms has mainly focused on device and
reliability physics [24–31]. Proprietary efforts also exist in leading industrial com-
panies to develop their own reliability models and tools [32–35]. Such knowledge
needs to be propagated into circuit design and CAD tools to assess the impact of
device degradation on various circuit performance metrics. As BTI is the dominant
aging mechanism, the accuracy and efficiency of BTI models directly affect the
quality of reliability analysis.
In a contemporary SoC, a wide variety of circuit operation patterns co-exists.
The diversity of circuit operation presents a unique challenge to predict the lifetime
under the BTI effect: A device operating under constant stress voltage needs a
static aging model; different from static stress, devices under AMS operation may
experience totally random stress during their lifetime. For digital circuits, they
employ DVS extensively to balance the workload and power consumption. Under
such operation, a static model is insufficient to accurately determine circuit aging.
Therefore, a model which can analyze aging under any random input stress pattern
is necessary.
Different from AMS circuits, large-scale logic designs encounter periodic inputs
with different duty cycles and frequency. Although a random stress model can
handle such situations, long-term aging simulation is not efficient and requires
expensive simulation time. To improve the efficiency, a long-term BTI model is
preferred which predicts an upper bound of threshold voltage shift for a periodic
input. In summary, a set of new BTI models are needed to improve circuit reliability
prediction in both VLSI and AMS design under dynamic operations.
In reliability physics, NBTI was first assumed to be induced by the Reaction–
Diffusion (R–D) mechanism, where holes initiate the breaking of Si–H bonds
(reaction phase) at the silicon substrate/oxide interface [4, 28, 31]. The broken
hydrogen species diffuse away from the interface into gate oxide and the poly-
silicon gate (diffusion phase) and thus, interface charges are generated. This
mechanism leads to the classical power-law model of Vth shift. More recent research
suggests a role of hole Trapping/Detrapping (T–D) for Vth shift of a device under
stress [36–51]. According to the T–D theory, if the trap gains sufficient energy it may
capture a charge carrier, thus reducing the number of available carriers in the channel
and modulating the threshold voltage of a device. A logarithmic time dependence
of Vth shift is observed under T–D as opposed to the power-law behavior. There are
research works which have suggested a combination of R–D and T–D. Whether T–
D or R–D dominates the Vth shift behavior depends on the fabrication technology.
For reliable circuit design, it is important to have an access to both R–D and T–D
based aging models in order to accurately determine the guard-band of a design. In
this chapter, we present a set of BTI models based on R–D and T–D theory which
can predict threshold voltage shift under static, random and long-term stress input.
Such a set of aging models provides an accurate and efficient solution for reliable
design.
2 Reliability Physics: Device-Level Modeling of BTI
The primary impact of BTI at the device level is the gradual increase in transistor
Vth , whereas the degradation of other device parameters are less pronounced. Since
the threshold voltage directly affects the delay of a digital gate, operating frequency
of a logic path decreases temporally. Similarly, in AMS circuits, increase in Vth
degrades the gain, trans-conductance and other performance metrics. To estimate
circuit aging rate, the fundamental step is to model device Vth shift under given
stress voltage and temperature. In this section, compact models based on both R–D
and T–D mechanisms are presented.
Two prevalent theories, Reaction–Diffusion and Trapping/Detrapping, are pro-
posed to explain BTI. Figure 6.2 shows the cross section of a PMOS device which
shows the difference between these mechanisms. R–D is a two-step process namely:
Reaction and Diffusion. According to the R–D theory, the stress voltage causes
covalent bonds (Si–H) at interface to break, which is Reaction. In the Diffusion
step the broken hydrogen atoms combine to form H2 , which diffuses towards gate.
For current thin-oxide devices, diffusion in poly-gate dominates the incremental
behavior of Vth shift. The interface states left at Si–SiO2 due to H2 diffusion increase
the threshold voltage. This leads to a power law relation (tn ) with time exponent
(n) 1/6, which should be independent of process parameters. 4Vth exponentially
depends on the stress voltage and temperature [28, 31].
According to the T–D theory shown in Fig. 6.2b, there exist a number of
defect states with different energy levels, and capture and emission time constants.
Threshold voltage increases when a trap captures a charge carrier from the channel
a b
D p+ D p+
ND
xxxxxx
Channel
G H2 Nit G
Fig. 6.2 Two mechanisms Carriers
for BTI: (a) Reaction
Diffusion (RD), and (b) S p+ S p+
Trapping/Detrapping (TD)
of a MOS device. Reduced number of channel carriers also causes drain current
to decrease. The probability of trapping and detrapping is a function of capture and
emission time constants, respectively. The gradual change in the number of occupied
traps results in a logarithmic time evolution of Vth shift, different from the power law
behavior in R–D. The stress voltage and temperature has an exponential impact on
the degradation rate, similar as that in the R–D theory [47].
2.1 Static BTI Models

2.1.1 Reaction: Diffusion Based Static BTI Model
There are two critical steps based on the Reaction–Diffusion theory. Reaction: Si–H
or Si–O bonds at the substrate/gate oxide interface are broken under the electrical
stress [16–18]. The interface charges are induced in gate oxide and further in poly-
gate, which cause the increase of Vth . Given the initial concentration of the Si–H
bonds (N0 ) at the interface and the concentration of inversion carriers (P), the
generation rate of the interface traps is given by [28]:
dNI T
D kf .N0 NI T / P kr NH NI T (6.1)
dt
where, kf and kr are forward and reverse reaction rates. The generation rate is
exponentially dependent on the stress voltage and temperature. During the initial
phase of stress period, the trap generation is slow with respect to time. Thus dNIT /dt
0 and NIT < < N0 , reducing Eq. (6.1):
kf
NH NI T D N0 P (6.2)
kr
The hydrogen atoms combine to form H2 molecules. The relationship between H

and H2 molecule is given as NH2 D kh NH 2 , where kh is the rate constant.
Diffusion: Generated hydrogen species diffuse away from Si–SiO2 interface towards
the gate. This diffusion is driven by the gradient of the density. This process is
governed by the following equation [28]:
dNH2 d 2 NH2
DC (6.3)
dt dx 2
C is the diffusion constant for hydrogen molecules in poly–Si, which depends on
the activation energy. To solve differential Eqs. (6.1) and (6.3) to derive a compact
model, an approximate profile of H2 concentration in gate oxide and poly–Si is
assumed, as shown in Fig. 6.3. NH2 is the concentration of hydrogen p molecules at
distance ‘x’ from Si–SiO2 interface at time (t), where x is given by C t. Diffusion
of H2 molecules in oxide is much faster than in poly–Si. The total number of

interface traps generated for a given stress time is solved as [31]:
Z
x.t/
NI T D 2 NH2 .x/dx (6.4)

0
H2 diffusion is divided in silicon oxide and poly-gate. Fast diffusion of H2 in

oxide and small thickness of dielectric lead to a very small difference in H2
concentration between Si–SiO2 interface and SiO2 –Poly interface. Figure 6.3 shows
the approximate profile of hydrogen concentration where a fitting parameter ı
is introduced to account for the fraction drop of H2 concentration. Using this
approximation, Eq. (6.4) can be written as [31],
p
Ztox CZt Ctox
NI T D 2 NH2 .x/dx C NH2 .x/dx (6.5)

0 tox

1 1 p
2 .1 C ı/ NH2 .0/ tox C NH2 .0/ C t (6.6)
2 2
NH2 (0) is H2 concentration at Si–SiO2 interface while ı NH2 (0) is the density at
Si–Poly interface. Finally, using NH2 D kh NH 2 , NIT can be represented as
NH2(t)
NH2(0)
Poly-Silicon
Fig. 6.3 Approximate

diffusion profile of hydrogen x
atoms under constant stress
x=0 Tox C•t
Fig. 6.4 Illustration of

statistical trapping/
de-trapping process in gate
oxide for a PMOS device,
leading to threshold voltage
shift
p ! 23
kh kf N 0 P p 13
NI T D .1 C ı/ tox C C t (6.7)
kr
where, inversion hole density P D Cox (Vgs Vth ). Based on the interface charges,
threshold voltage shift can be derived as: 4Vth D qNIT /Cox . This R–D based model
predicts the Vth shift under any given constant stress voltage, temperature and time.
The final form of the degradation is given as:
p 2n
qNI T
Vth D Cox
D A .1 C ı/ tox C C t
r (6.8)
qtox
AD "ox
1=2n
K2 Cox Vgs Vth exp 2Eox
E0
C is the diffusion constant which incorporates the temperature dependence. Time

exponent n is 1/6 when the diffusion species is H2 . The static model equation is
verified by 45 nm silicon data along with the T–D model (Fig. 6.5).
2.1.2 Trapping/De-trapping Based Static BTI Model
Several works have presented the evidence of the T–D mechanism, such as the
discrete Vth shift, especially during the recovery phase in NBTI observed with the
fast measurement techniques [52, 53]. Figure 6.4 illustrates the physical picture
of T–D: when a negative bias voltage is applied to the gate of a PMOS device,
the trap energy (relative to the Fermi energy level) is modulated. If the trap gains
sufficient energy, it may capture a charge carrier, thus reducing the number of
available carriers in the channel [38]. The charged trap state modulates the local
Vth and acts as a scattering source, reducing the effective mobility [38]. Faster traps
(with shorter time constants) having a higher probability of capturing carriers; the
occupation probability increases with voltage and temperature. Trapping and de-
trapping events are stochastic in nature and hence a compact model is based on the
statistics of trap properties.
The basic modeling assumptions based on T-D theory are the same as the ones
used in modeling of low-frequency noise, since the charge trapping dynamics
(capture and emission time statistics) that contribute to the degradation of device
performance over time is similar [40–44]. The main assumptions of the trap
properties are:
• The number of traps follows a Poisson distribution, which is common for a
discrete process.
• Capture and emission time constants are uniformly distributed on the logarithmic
scale. This microscopic assumption is critical to derive the logarithmic time
evolution at the macro scale.
• The distribution of trap energy is approximated as a U-shape, which is verified

by silicon measurement and key to the voltage and temperature dependence of
the aging effect.
Based on the T–D theory, the Vth shift at a given stress time is the result of number
of traps (n(t)) occupied by the channel carriers. The probability of a particular trap,
initially empty (0), to be occupied (1) after an elapsed time t is given by P01 (t). This
occupation probability can be calculated by observing that
P01 .t C dt / D P01 .t /p11 .dt / C P00 .t /p01 .dt / (6.9)
where p01 (dt) D 1/ c and p11 (dt) D 1 p10 (dt) D 1/ e . Integrating it from t0 to t:
eq
P01 .t C t0 / D 1 e t=eq C P01 .t0 / e t=eq (6.10)
c
where 1/ eq D 1/ c C 1/ e . c, e are random in nature, representing capture and
emission time constants respectively, and dependent on bias point and temperature.
The values are determined by [38]:
c D 10p .1 C e q / (6.11)

e D 10p 1 C e Cq (6.12)
where p 2 [pmin , pmax ]. pmin and pmax define the time constants for fastest and
slowest traps respectively (pmin 1 and pmax > 10). This assumption of the existence
of defects with wide distribution of time constants is in line with recent NBTI
data [38]. Since p is assumed to be uniformly distributed, the characteristic time
constants are uniformly distributed on logarithmic scale. The parameter q is given
by (ET EF )/kT, where ET is the trap energy and EF is the Fermi energy level.
The trap energy (relative to Fermi energy) is a function of applied electric field.
Consequently, c and e are dependent on voltage and temperature.
The occupation probability of the trap at time t, assuming that it is under
constant stress from time t0 D 0 is obtained by substituting P01 (0) D 1 P01 (0) D 0
in Eq. (6.10), integrating P01 and multiplying with the number of available traps,
the average number of occupied traps obtained by substituting the logarithmic
distribution of time constants, and the U shaped distribution of trap energies:
EZ
T max 10p
Z max t
N g .ET / dET e u 1
n.t / D EF
: du (6.13)
ln 10 .pmax pmin / 1 C exp ETkT u
0 10p min t
where g(ET ) is the trap energy distribution, and pmin and pmax represent fast and slow
traps, respectively. The trap energy, ET changes as a function of electric field (Eox ).
Assuming pmin 1 and pmax > 10, and ET 1/Eox ,
Fig. 6.5 Both R–D and T–D 1.6

based compact static model Symbols: 45nm data
matches 45 nm silicon data RD model
1.2
under constant stress for
TD model
different voltage and 1.1V, 105°C
temperature
DVth (a.u.)
0.8
1.3V, 30°C
0.4
1.1V, 30°C
4x103 8x103
Time (s)

N ˇVg E0
n.t / D exp exp ŒA C B log 10p max t
ln 10 .pmax pmin / tox kT kT

(6.14)
Equation (6.14) describes the aging under a constant stress voltage and temperature.
Similar as previous R–D model, it is an exponential function of the stress voltage,
temperature and Tox . Furthermore, it has a statistical nature with N, an index for the
number of traps per device. For the simplicity, Eq. (6.14) is written as:
Vth .t / D ŒA C B log .1 C C t /
(6.15)
Equations (6.8) and (6.15) represent Vth shift based on R–D and T–D theory
respectively. Aging prediction of these models is compared with 45 nm silicon
data in Fig. 6.5. Within the range of data fluctuations, both models match well with
measurement under the static stress. The exponential dependence on stress voltage
and temperature are correctly captured by both aging models.
2.2 BTI Models Under Random Stress Patterns
Today’s circuits typically have a reduced activity factor (or duty cycle) and dynamic
voltage scaling (DVS), to reduce power consumption. Therefore, a significant
portion of the operation is under lower supply voltage, resulting in large recovery.
Since the degradation is highly sensitive to the stress voltage, DVS leads to different
amounts of circuit aging. For a random stress pattern, it is necessary to derive
voltage dependent recovery to accurately predict Vth shift.
2.2.1 Reaction: Diffusion Based Aging Model for Random Input
The hydrogen atoms that are generated during the stress phase will recover if the
stress voltage is removed completely. Atoms close to Si–SiO2 interface anneal the
broken Si–H bond while atoms deep in Poly-gate continue to diffuse away, leading
to incomplete recovery of Vth . Approximate profile of hydrogen species is shown in
Fig. 6.6. Hydrogen atoms diffuse quickly in oxide and anneal some of Si–H bonds
in a short time. If a stress voltage is removed after time (t1 ), the interface charges
generated at the given time are NIT (t1 ). The total number of charges to be annealed
is given by NA IT (t). Interface charges at a given time t is given by
NI T .t / D NI T .t1 / NIAT .t / (6.16)
as shown in Fig. 6.6; the number of annealed traps can be divided in to two parts:
(1) recombination of H2 in oxide, and (2) back diffusion of H2 in poly-gate [31].
Thus we have

1p
NI T .t / D 2 1 te C
A
2 C .t t1 / NH2 .0/ (6.17)
2
In Eq. (6.17) above, 1 and 2 are the back diffusion constants. From Eqs. (6.7) and
(6.17), we get
p !
21 te C 2 C .t t1 /
NIAT .t / D NI T .t / p (6.18)
.1 C ı/ tox C C t
Substituting Eq. (6.18) in Eq. (6.16), simplifying equations and using 4Vth D qNIT /
Cox , we get the recovery model as:
p !
21 te C 2 C .t t1 /
Vth .t / D Vth .t1 / 1 p (6.19)
.1 C ı/ tox C C t
NH2(t)
Poly-Silicon
NAH2(0)
Fig. 6.6 Approximate

diffusion profile of hydrogen x
atoms under recovery
x = 0 Tox C•t
a NH2 NH2 NH2
NH2(0) NH2(0) NH2(0)
NH2(0)’ NH2(0)’ NH2(0)’
Tox Te Tox Te C (t−t ) C • t0 Tox Te C (t−t0) C • t0

C • t0 0
b NH2 NH2 NH2
NH2(0) NH2(0) NH2(0)
NH2(0)’ NH2(0)’ NH2(0)’
Tox Te C • t0 ToxTe C (t−t ) C•t Tox Te C (t−t0) C•t

0
Fig. 6.7 Approximate diffusion profile of hydrogen atoms in PMOS under dynamic stress voltage:
(a) from high to low and (b) from low to high
The equation above represents Vth recovery when stress voltage is completely
removed. This model is not applicable to the Vth behavior when the stress voltage
is lowered or increased. Under DVS operation, the stress voltage is changed up
and down. Diffusion profile of hydrogen atoms when the stress voltage changes
from high to low is different from that when the voltage changes from low to
high [54]. Time-dependent diffusion profiles are shown in Fig. 6.7 for a voltage
change from high (V1 ) to low (V2 ) at time (t0 ). The resulting Vth shift is divided in
to two components. The first component is the diffusion due to the lower voltage
(V2 ) which is the un-shaded region in the figure. The shaded part is the recovery
component that gradually decreases with time. Eventually, the 4Vth degradation is
driven by the lower voltage. To derive a closed form solution, we start with 4Vth
under the lower voltage:
p 2n
Vth .t / D A2 .1 C ı/ tox C C .t t0 / C s.t / (6.20)
where A2 is the function of the lower voltage (V2 ). s(t) is time-dependent initial
distance of hydrogen diffused, shown as the shaded area in first graph of Fig. 6.7a.
As time progresses, these hydrogen atoms recover, given by the equation:
p !
21 Te C 2 C .t t0 /
Vthr .t / D Vth .T / 1 p (6.21)
.1 C ı/ tox C C t
By substituting t D t0 in Eq. (6.20) and equating the results with Eq. (6.21), we get
a compact model which predicts Vth shift when stress voltage changes from a high
to low voltage:
0 v !12n
u p
p p u 2 T C C .t t /
Vth .t / D @ A2 C .t t0 / C tVth .t0 / 1 A
1 e 2 0
2n 2n
p
.1 C ı/ tox C C t
(6.22)
1 and 2 are back diffusion constants, same as those in the recovery model. This
equation is capable of predicting the non-monotonic behavior of initial recovery and
eventually converging with rising Vth shift due to the lower voltage.
When the stress voltage changes from a low (V2 ) to high (V1 ) value at time (t0 ),
the diffusion profile approximation is shown in Fig. 6.7b. In this case, there is no
recovery component. The diffusion due to low voltage continues to diffuse at the
same rate. However, the diffusion front due to higher voltage is dominant, and Vth
shift eventually converges. We start from Eq. (6.20) in the derivation. In this case,
A2 is the function of the higher voltage (V1 ). The diffusion profile of hydrogen under
V1 follows the shaded region in Fig. 6.7b. Vth shift under the lower voltage (V2 ) at
time (t0 ) is given by:
p 2n
Vth .t0 / D A1 .1 C ı/ tox C C t0 (6.23)
Substituting t D t0 in Eq. (6.20) and equating with Eq. (6.23), we arrive at the model
which describes the behavior when voltage transits from a low to a high value:
hp p p p i2n
Vth .t / D 2n
A2 .1 C ı/ tox C C .t t0 / C 2n A1 .1 C ı/ tox C C t
(6.24)
Table 6.1 summarizes the models under random stress, and static models based on
the R–D theory. Figure 6.8 validated the model prediction with 45 nm silicon data
under random stress patterns. The stress and recovery behaviors of threshold voltage
are accurately captured when device is stressed under arbitrary voltage (Fig. 6.8a).
Fig. 6.8b emphasizes the model prediction of the recovery component. In this figure,
a device is stress under 1.3 V and after 1,000 s, the stress voltage is lowered to 1.2 V.
As a result of excessive stress until 1,000 s, the device initially recovers and then
Vth shift increases under the lower voltage. This non-monotonic behavior is well
predicted by the dynamic model [54].
Table 6.1 Summary of R–D based BTI models

Constant stress Vth (t) D A tn
Random Input Stress Low to high voltage transition: Vt h .t / D
hp p p p i2n
2n
A2 .1 C ı/ tox C C .t t0 / C 2n A1 .1 C ı/ tox C C t
High to low voltage transition: Vt h .t C t0 / D
r p 2n
p p 21 te C 2 C.t /
2n
A2 C.t / C 2n Vt h .t0 / 1 .1Cı/t p
C C .t Ct /
ox 0
a b
1.0 Vgs=1.4V Vgs=1.2V

Vgs=1.3V Vgs=1.2V
1.0
Vgs=1.3V Vgs=1.0V
ΔVth (a.u.)
0.5
Symbols: 45nm data 0.5 Symbols: 45nm data

R-D model R-D model
0.0
0 20000 40000 102 103 104
Time (s) Time (s)
Fig. 6.8 Random input model validation (a) arbitrary stress pattern and (b) non-monotonic
behavior of Vth shift under a lower voltage
Fig. 6.9 Vth shift under DVS

is non-monotonic; when the V1 V2<V1
stress voltage is changed from
V1 to V2 (assuming V2 < V1 )
Δ1
ΔVth
Δ2
t0
Time (sec)
2.2.2 Trapping/De-Trapping Based Random Input Aging Model
In this section, trapping/de-trapping based models are presented which can handle
random stress waveform. Since the degradation is highly sensitive to the voltage
(Eq. (6.15)), dynamic voltage scaling leads to different amounts of circuit aging.
The voltage tuning is categorized into two cases. In case 1, stress voltage is changed
from lower voltage V1 to a higher voltage V2 and in case 2, V2 is lower than V1 at
time t0 . To handle such a voltage transition, using Eq. (6.10) and a non-zero time,
t0 , to calculate the occupation probability at time t (time elapsed after t0 ), as shown
in Fig. 6.9 [48].
To handle such a voltage transition, using a non-zero time, t0 to calculate the

occupation probability at time t (time elapsed after t0 ) using Eq. (6.10):
eq2
P01 .t C t0 / D 1 e t=eq2 C P01 .t0 / e t=eq2 (6.25)
c2
where eq2 , c2 represent the time constants under voltage V2 . Using Eqs. (6.5) and
(6.6), eq1 D eq2 , since eq depends only on parameter p, which is independent of
the voltage. Substituting this property in Eq. (6.25):
eq eq t=eq
P01 .t C t0 / D 1 e t=eq e e .tCt0 /=eq (6.26)
c1 c2
where c1 and c2 correspond to voltages V1 and V2 . Following similar steps as in

static model derivation, we arrive at a closed form solution:

1 C C .t C t0 /
Vth .t / D 2 ŒA C B log .1 C C t /
C 1 B log (6.27)
1 C Ct
where 1 corresponds to the voltage V1 and 2 corresponds to V2 . The degradation

in Eq. (6.27) is physically interpreted as a sum of two components, 1 and 2
which are proportional to 1 and 2 respectively. When the voltage is changed to a
lower voltage, traps emit some of the charge carriers, and the number of occupied
traps reaches a new equilibrium. 2 dominates initially, which contributes to the
recovery. If the operation under V2 continues for a longer time, 1 eventually takes
over and Vth increases. Such a non-monotonic behavior is correctly predicted.
Table 6.2 presents the summary of static and dynamic models based on trapping/de-
trapping theory [48]. When the voltage is increased, the degradation rate rises at the
point of voltage change. Figure 6.10 validates the dynamic model. Non monotonic
behavior when stress voltage transition to a lower value is correctly predicted by T–
D based random input model. Under this condition, the device experience recovery
period, before the stress goes back to the equilibrium condition. Eventually, the
degradation rate goes to the same as the constant stress under the lower voltage.
This behavior is predicted from Eq. (6.27), where the second component dominates
initially, resulting in the recovery; after t >> 200 s, the second component decays
down and the first component takes over, leading to the stress behavior under the
second voltage. Experimental results from the test chip well validate these non-
monotonic behaviors, as shown in Fig. 6.10, supporting further study on aging
Table 6.2 Summary of T–D based BTI models

Constant stress Vth (t) D [A C B log(1 C Ct)]
1 D ' .A C Blog .1 C C t // ;
Random Input Stress Vth (t C t0 ) D 1 C 2 kClog.1CC t /
2 D Vt h .t0 / 1 kClog.1CC .t Ct0 //
Fig. 6.10 The T–D model

predicts the dynamic behavior 1.2 Symbols: 65nm Data
under voltage tuning, such as T-D model VDD=1.8V
the transition period, and the
convergence to the constant
stress condition 0.8
ΔVth (a.u.)
VDD=1.5V VDD=1.5V
0.4
0.0
VDD=1.2V
101 102 103 104

Time (s)
Fig. 6.11 Vth shift under 1.2

voltage tuning from 1.8 to Symbols: 65nm data
1.65 V. The shift eventually T-D model
converges to the final voltage,
weakly dependent on 0.8
previous stress history
ΔVth (a.u.)
1.8V 1.65V
0.4
0.0
50s 200s
101 102 103 104

Time (s)
prediction under DVS. The two components in Eq. (6.27) play an important role
in long-term prediction under multiple cycles.
Higher recovery is seen when the device is stressed at 1.8 V compared to 1.65 V
as more traps are captured under higher stress voltage. The T–D model captures such
behavior. As stress voltage continues to be 1.2 V after 200 s, the temporary recovery
is overwhelmed by the capture of traps under lower stress voltage. Increasing stress
time causes the degradation to converge to this constant stress voltage at 1.2 V.
This behavior is predicted from Eq. (6.27), where the second component dominates
initially, resulting in the recovery, while the second component decays down and the
first component takes over, leading to the stress behavior under the second voltage.
Figure 6.11 evaluates the model prediction, with different periods under the same
voltage. In this study, the device is initially stressed under 1.8 V, for 50 or 200 s;
a b
1.6
Symbols: 65nm Data Symbols: 65nm Data
Line: T-D model Line: T-D model
1.2
1.8V- 1.2V with α =0.5 1.8V- 0V with α =0.5
ΔVth (a.u.)
0.8
0.4
0.0
3000 6000 9000 60000 63000
Time (s) Time (s)
c 1.6
d
Symbols: 65nm Data Symbols: 65nm Data
Line: T-D model Line: T-D model
1.2
1.8V- 0V with α=0.97
1.8V- 0V with α =0.03
ΔVth (a.u.)
0.8
0.4
0.0
12000 15000 105000 108000
Time (s) Time (s)
Fig. 6.12 Validation of the dynamic stress model under different voltages and duty cycles
then the voltage is switched to 1.65 V. As the stress voltage is lowered, a temporary
recovery behavior is observed due to the emission of excessive amount of trapped
charges. The T–D model captures such behavior in both cases. In the case where the
device is first stressed under 1.8 V for 200 s, the stress time is higher as compared
to the case where the initial stress is 50 s causing higher Vth shift. However, since
in both cases, the device is later stressed for a much longer time ( 10 ks) at 1.65 V,
the degradation converges to the constant stress condition at 1.65 V. This validation
helps predict the aging under various switching activities (˛) much needed for aging
evaluation of digital circuits.
Figure 6.12 presents the measurement under different patterns of voltages and
duty cycles. Figure 6.12a and b show the multiple cycle prediction stressed at 1.8,
1.2 V and 1.8, 0 V at ˛ D 0.5. When the voltage is lowered from 1.8 V, a recovery
is observed from silicon data, which is well predicted from the model. Furthermore,
if the voltage is lowered down to 0 V, the recovery is more pronounced (Fig. 6.12b).
The cycle-to-cycle model in Eq. (6.27b) is scalable with different duty cycles.
Figure 6.12c and d show the validation of our model with silicon data at very low
(˛ D 0.03) and very high (˛ D 0.97) duty cycles, respectively. A voltage transition
even for a short duration results in the sudden shift in threshold voltage, due to the
fast trapping/de-trapping (Fig. 6.12c and d).
Random input models derived in this section are thoroughly validated with 45
and 65 nm data. These models provide an accurate and efficient way to identify
aging under arbitrary stress patterns, allowing designers to adequately guard-band
their design. Such models can be more beneficial to AMS designs where the number
of transistors are less than that in a digital design and the input is highly random.
However for large scale digital circuits, using random stress models is insufficient
to predict the lifetime of a high-speed design. Thus, there is a clear need for long-
term aging models based on both R–D and T–D theories, which can estimate a tight
upper bound of aging, without tracking the behavior cycle by cycle. The following
section presents the solutions.
2.3 Long-Term BTI Model Under DVS
Under the extensive usage of DVS, accurate Vth shift can be predicted using the
random input stress models derived in previous section. However for a large scale
logic design, applying such random stress models will not be efficient in circuit
simulation. Most digital circuits operate at a regular frequency periodically with
a representative duty cycle. A long-term model will be desirable for efficient aging
prediction, which directly estimates aging at a given operation time without tracking
the stress-recovery over many cycles. This long-term model predicts a tight upper
bound after multi-cycle operations under DVS. Based on the multi-cycle model
in previous section, stress ( Vths,m ) and recovery ( Vths,mC1 ) in Fig. 6.13 can be
connected to derive a long-term model based on both R–D and T–D theories.
2.3.1 Long-Term BTI Model Based on Reaction–Diffusion
It is possible to obtain a closed form solution to predict the upper bound of Vth
for different clock cycle (Tclk ), duty cycle (˛) and stress time for a circuit oscillating
between two voltages. To derive a closed form solution, Vths,m and Vths,mC1 are
connected iteratively using random stress models as:
2 p 32n
1=2n 1=2n 1=2n
2n
A2 X 1 C ˇm C ˇm ˇm1 C : : : C
Vths;mC1 D 4 2n
p p 1=2n 1=2n 1=2n
5
A1 Y C C ˛Tclk 1 C ˇm C ˇm ˇm1 C : : :
(6.28)
Fig. 6.13 Vth shift during the ΔVth

stress and recovery cycles,
presenting the important
Recovery
parameters used in the
Stress Stress
multi-cycle model
ΔVths,m+1
ΔVths,m
ΔVthr,m+1
ΔVthr,m
Tclk
(m-1)Tclk mTclk (m+1)Tclk Time
where,
p p
X D .1 C ı/ tox C C ˛Tclk ; Y D C .1 ˛/ Tclk
p
21 te C 2 C .1 ˛/ Tclk
ˇ D1 p
.1 C ı/ tox C C mTclk
Using ˇ 1 < ˇ 2 < ˇ m-1 < ˇ m and geometric series approximation, the upper bound
of degradation is derived as:
p p p 2n
1 1
Vths;mC1 D 2n
A2 X C A1 Y C C ˛Tclk
2n
1 ˇ 1=2n 1 ˇ 1=2n
(6.29)
The new long-term model is capable of predicting the upper bound of Vth for cycle
of two non-zero voltages [54].
Figure 6.14 confirms that for multiple cycles, the long-term model predicts the
same result as the model under random input, given a constant duty cycle. Such
model predicts a tight upper bound under the periodic stress.
2.3.2 Long-Term BTI Model Based on Trapping/De-Trapping
Similar to the derivation of the R–D model, a long-term model based on the T–
D theory is obtained. Based on the multi-cycle model in the previous sub-section,
Vths,m and Vths,mC1 are connected by:
Vths;mC1 D 1 ŒA C B log .1 C C ˛TC lk /

C 2 ŒA C B log .1 C C .1 ˛/ TC lk /
ˇ1;m

C V ths;m 1 ˇ1;m 1 ˇ2;m
(6.30)
1.0
α=0.9
ΔVth (a.u.) 0.7
α=0.5
α=0.1
0.3
Dynamic R-D model

Long-term R-Dmodel
0.0
0 1x105 2x105
Time (s)
Fig. 6.14 Under constant duty cycles, the long-term model is consistent with the model for
random input
Using Eq. (6.30) and repeatedly replacing the Vths,mC1 by Vths,i for i D m, : : : ,1,
we get:
0 1
X
m Y
m
Vths;m D 1 ŒA C B log .1 C C ˛TC lk /
@1 C ˇ1;:j :ˇ2;:j A
iD1 j DmiC1
0 1
X
m Y
m
ˇ1;m @1 C ˇ1;:j 1 :ˇ2;:j A
iD1 j DmiC1
(6.31)
Since obtaining a closed-form solution for Eq. (6.31) is not straight-forward, we use
the property ˇ 1,m-1 < ˇ 1,m and ˇ 2,m-1 < ˇ 2,m ,
2
Vths;mC1 1 ŒACB log .1CC ˛TC lk /
1Cˇ1;:m :ˇ2;:m C ˇ1;:m :ˇ2;:m C : : :
2
C2 ˇ1;m ŒACB log .1CC .1˛/ TC lk /
1Cˇ1;:m :ˇ2;:m C ˇ1;:m :ˇ2;:m C : : :
(6.32)
Equation (6.32) is a geometric series and the upper bound of degradation is:
Vths;m D 1 ŒA C B log .1 C C ˛TC lk /

1ˇ1;:m1 :ˇ2;:m
ˇ (6.33)
1ˇ1;:m
1;:m
:ˇ2;:m
Fig. 6.15 Long-term model

prediction for different duty 1.0 Model
cycle V1=1.8V V2=0V
V1=1.8V V2=1.2V
0.8
ΔVth (a.u.)
0.6
0.4
Tclk=1000s
0.2
0.0 0.2 0.4 0.6 0.8 1.0

α (duty cycle)
Equation (6.33) is sensitive to the duty cycle, ˛ (ratio of time under V1 to time under
V2 ), time period (sum of operation times under V1 and V2 for a single cycle), and the
stress voltage. The long term model captures the tight upper bound of cycle-to-cycle
prediction. Furthermore, Fig. 6.15 presents the aging behavior under a wide range
of duty cycles as validated with 65 nm silicon data. The degradation rate changes
rapidly when ˛ approaches 0 or 1, but gradual for intermediate duty cycles. This
behavior is due to the sudden change in degradation at the beginning of the stress
and recovery phase, due to the voltage dependent time constants. It is well predicted
by a single equation in the long-term model [48].
3 Circuit Aging Simulation
New models are derived from basic physics and are comprehensively validated with
discrete device data in previous sections. In this section, these models are further
integrated with design tools at the circuit level. Different types of circuits have
different requirements on reliability tools, due to the diversity in circuit operation
patterns. Digital circuits usually operate at two distinct voltage levels (VDD and
GND) as opposed to continuous voltages in analog circuits. Thus long-term models
are more appropriate for digital circuits and random input stress models are better
for AMS designs. Aging models required to predict circuit lifetime under any given
stress conditions are presented in previous sections. Based on them, we present
the aging analysis of digital and AMS designs by implementing aging models in
SPICE for real time simulation under any possible stress conditions, supporting
design practice for reliability. The model prediction is validated with 45 nm ring
oscillator data.
VRING
VDIV
Frequency Output
Divider
VRING : ROVDD
Venable VDIV : VDDof peripheral circuits
Venable : OscillationEnable
Fig. 6.16 Test circuit at 45 nm used for the validation
In order to estimate the impact of aging, a sub-circuit module is implemented

in SPICE. Model equations (R–D or T–D based) described in previous sections are
coded in Verilog-A, which take the stress voltage, temperature and time as inputs
[31]. A voltage controlled voltage source then subtracts the Vth shift calculated by
the model from Vg , emulating aging of a device and resulting in decrease in drain
current. To improve the simulation speed, input voltages are quantized into discrete
levels for random input models. Thus, there exists a trade-off between simulation
speed and step size of the voltage. Stress time is the simulation time from the
SPICE environment allowing designer to track circuit performance at real time and
to closely monitor the degradation behavior.
3.1 Aging Analysis: Digital Circuits
To validate the new models at circuit level, silicon data is measured from 45 nm
ring oscillators (ROs). The frequency change (4F/F) of RO is measured as a direct
index of the degradation, which is proportional to PMOS transistor threshold voltage
change under NBTI. Figure 6.16 presents the test circuit of ROs used for aging
characterization. Frequency change in 11 stage ring oscillator is monitored during
the test. The ring oscillator is activated by the enable (Venable ) pin. Different from
traditional RO based aging test, the supply voltage of this circuit, VRING , is switched
between VDD (stress mode) and 0 (recovery mode) at different duration, in order to
emulate dynamic voltage scaling encountered in logic circuits. The data is collected
at regular intervals at multiple supply voltages and temperature. The pin, VDIV , is
implemented to control the frequency divider and the output buffer. It also helps to
eliminate the impact on frequency shift of ROs due to aging of peripheral circuits,
thus giving a clean data [54].
Vth of a device in RO has a strong dependence on the dynamic scaling of supply
voltages due to the recovery effect. Figure 6.17a shows the basic test pattern in a
ring oscillator with alternate active and sleep modes, each for 5,000 s duration. The
a b
1: active sleep active sleep 1.0 active sleep
1.0
active sleep active
0.8
0.8
ΔF/F (a.u.)
0.6
0.6
0.4
0.4
Symbols: 45nm data Symbols: 45nm data
Aging models 0.2 Aging models
0.2
0 1x104 2x104 0 1x104

Time (s) Time (s)
Fig. 6.17 (a) Model validation with circuit-level DVS data. (b) Aging under different stress
patterns well captured by the model
increase of frequency degradation in the active mode and its decrease in the sleep
mode are well predicted by the new random stress models based on R–D or T–D
theory. Figure 6.17b shows the impact of the recovery on the frequency shift of RO.
Although the stress times are same for both patterns, the recovery in PMOS device
causes over all degradation to be less. These effects are successfully captured by the
aging models described previously.
3.2 Aging Analysis: AMS Circuits
Analog/mixed signal circuits encounter more complex stress patterns. Furthermore,

device degradation changes the DC biasing conditions, which in turn change circuit
performance. Figure 6.18 shows a case study of the impact of NBTI on a PMOS
input differential amplifier. In this case, we demonstrate that the amplifier is being
used to amplify the output of a mixer circuit. The stress pattern of a device is a
combination of two different frequencies signal ! 1 and ! 2 . Conventional prediction
with an average pattern fails to capture amplitude distortion (Fig. 6.19b). Using new
random input stress models, an accurate aging analysis on the differential amplifier
can be conducted. Note that the accuracy of aging prediction has a trade-off with
simulation time in SPICE. Real-time stress and recovery analysis requires SPICE to
Vbias P
sin (ω1t)
MIXER P P Vcm
sin (ω2t) sin (ω1t) Vout

+ N N
sin (ω2t)
Fig. 6.18 A differential amplifier in aging analysis
a b
0.6
Average ΔVth
Acutal ΔVth
0.6
Output Voltage (V)
ΔVth (a.u.)
0.4
0.3
Average ΔVth
(Sim. time = 1.50s) 43mV distortion in
Acutal ΔVth 0.2 the amplitude is not
0.0
(Sim. time = 3.44s) captured by the
average aging model
200 400 600 200 400 600

Time (s) Time (s)
Fig. 6.19 Aging analysis with the average value fails to capture the distortion. (a) Vth shift
prediction (b) output voltage of the amplifier
increase the internal time steps which take longer simulation time compared to the
average aging stress pattern in Fig. 6.19.
4 Summary
Aging issue has become increasingly important as the technology scales. BTI,
especially NBTI in a PMOS device, has emerged as a dominant mechanism in digital
as well as AMS designs, impacting circuit performance over time. The challenges
are further compounded with Dynamic Voltage Scaling (DVS) and random inputs
that are encountered in a realistic operation. In this chapter, both R–D and T–
D theories are comprehended in detail. Compact BTI models for both R–D and
Fig. 6.20 Hierarchical development of BTI models for circuit aging analysis
T–D theories are then derived and validated using 45 and 65 nm silicon data.
Figure 6.20 presents distinct levels of reliability modeling. Three different types of
models are developed for diverse design needs: Static, Random stress and Long-
term models. The models are implemented in the SPICE environment, through
Verilog-A. Depending on the operation pattern, appropriate models can be applied
to circuit-level aging analysis and help improve circuit reliability.
References
1. C. Constantinescu, “Trends and challenges in vlsi circuit reliability,” IEEE Computer Society,
pp. 14–19, 2003.
2. Semiconductor Industry Associate, International Technology Roadmap for Semiconductors,
2010. (Available at http://public.itrs.net).
3. D. K. Schroder and J. A. Babcock, “Negative bias temperature instability: Road to cross in
deep submicron silicon semiconductor manufacturing,” Journal of Applied Physics, vol. 94,
no. 1, pp. 1-18, 2003.
4. K. O. Jeppson and C. M. Svensson, “Negative bias stress of MOS devices at high electric fields
and degradation of MNOS devices,” J. Applied Physics, vol. 48, no. 5, pp. 2004-2014, 1977.
5. K. Bernstein, D. J. Frank, A. E. Gattiker, W. Haensch, B. L. Ji, S. R. Nassif, E. J. Nowak,
D. J. Pearson, N. J. Rohrer, “High-performance CMOS variability in the 65-nm regime and
beyond,” IBM Journal of Research and Development, vol. 50, no. 4/5, pp. 433-449, July 2006.
6. S. Borkar, T. Karnik, S. Narendra, J. Tschanz, A. Keshavarzi, V. De, “Parameter variations and
impact on circuits and microarchitecture,” Design Automation Conference, pp. 338-342, 2003.
7. F. Arnaud, L. Pinzelli, C. Gallon, M. Rafik, P. Mora, F. Boeuf, “Challenges and opportunity in
performance, variability and reliability in sub-45 nm CMOS technologies,” Microelectronics
Reliability, vol. 51, no. 9-11, pp. 1508-1514, Sept.-Nov. 2011.
8. B. H. Calhoun, Y. Cao, X. Li, K. Mai, L. T. Pileggi, R. A. Rutenbar, and K. L. Shepard, “Digital
circuit design challenges and opportunities in the era of nanoscale CMOS,” Proceedings of the
IEEE, vol. 96, no. 2, pp. 343-365, Feb. 2008.
9. A. Ramadan, “Compact model council’s standard circuit simulator interface for reliability
modeling,” International Reliability Physics Symposium, pp. 2A.5.1-2A.5.6, 2013.
10. J. Hicks, D. Bergstrom, M. Hattendorf, J. Jopling, J. Maiz, S. Pae, et al., “45 nm transistor
reliability,” Intel Technology Journal, vol. 12, no. 02, pp. 131-144, Jun. 2008.
11. S. V. Kumar, C. H. Kim and S. S. Sapatnekar, “Adaptive techniques for overcoming perfor-
mance degradation due to aging in digital circuits,” Asia and South Pacific Design Automation
Conference, pp. 284-289, 2009.
12. V. Reddy, “Impact of negative bias temperature instability in digital circuit reliability,”
International Reliability and Physics Symposium, pp. 248-253, 2002.
13. R. Zheng, et al., “Circuit aging prediction for low-power operation,” Customs Integrated
Circuits Conference, pp. 427-430, 2009
14. G. Chen, K. Y. Chuah, M. F. Li, Daniel SH Chan, C. H. Ang, J. Z. Zheng, Y. Jin, and D. L.
Kwong, “Dynamic NBTI of PMOS transistors and its impact on device lifetime,” International
Reliability Physics Symposium, pp. 196-202, 2003.
15. B. C. Paul, K. Kang, H. Kufluoglu, M. A. Alam, and K. Roy, “Impact of NBTI on the temporal
performance degradation of digital circuits,” IEEE Electronic Device Letters, vol. 26, no. 8,
pp. 560-562, Aug. 2005.
16. R. Vattikonda, W. Wang, and Y. Cao, “Modeling and minimization of pmos nbti effect for
robust nanometer design,” IEEE/ACM Design Automation Conference, pp. 1047-1052, Jul.
2006.
17. M. Agarwal, B. C. Paul, Ming Zhang, and S. Mitra, “Circuit failure prediction and its
application to transistor aging,” VLSI Test Symposium, pp. 277-286, 2007.
18. V. Reddy, A. T. Krishnan, A. Marshall, J. Rodriguez, S. Natarajan, T. Rost, and S. Krishnan,
“Impact of negative bias temperature instability on digital circuit reliability,” International
19. W. Wang, Z. Wei, S. Yang and Y. Cao, “An efficient method to identify critical gates under
circuit aging,” Int. Conference on Computer Aided Design, pp. 735-740, 2007.
20. A. T. Krishnan, F. Cano, C. Chancellor, V. Reddy, Q. Zhangfen, P. Jain, J. Carulli, J. Masin, S.
Zuhoski, S. Krishnan, and J. Ondrusek, “Product drift from NBTI: Guardbanding, circuit and
statistical effects,” Int. Electron Devices Meeting, pp. 4.3.1-4.3.4, 2010.
21. W. Wang, S. Yang, S. Bhardwaj, R. Vattikonda, S. Vrudhula, F. Liu, Y. Cao, “The impact of
NBTI effect on combinational circuit: Modeling, simulation, and analysis,” IEEE Transactions
on VLSI Systems, vol. 18, no. 2, pp. 173-183, 2010.
22. H. Sangwoo, K. Juho, “NBTI-aware statistical timing analysis framework,” IEEE International
SOC Conference, pp.158-163, 2010.
23. E. Maricau and G. Gielen, “Computer-aided analog circuit design for reliability in nanometer
CMOS,” IEEE Transactions on Emerging and Selected Topics in Circuits and Systems, vol. 1,
no. 1, pp. 50-58, Mar. 2011.
24. A. R. Brown, V. Huard and A. Asenov, “Statistical simulation of progressive NBTI degradation
in a 45-nm technology pMOSFET,” IEEE Transactions on Electron Devices, vol. 57, no. 9, pp.
2320-2323, Sept. 2010.
25. S. Ogawa and N. Shiono, “Generalized diffusion-reaction model for the low-field charge-
buildup instability at the Si-SiO2 interface,” Physical Review B, vol. 51, no. 7, pp. 4218-4230,
Feb. 1995.
26. G. Chen, K. Y. Chuah, M. F. Li, D. S. H. Chan, C. H. Ang, J. Z. Zheng, Y. Jin, et al., “Dynamic
NBTI of PMOS transistors and its impact on device lifetime,” International Reliability Physics
Symposium, pp. 196-202, 2003.
27. S. Chakravarthi, A. T. Krishnan, V. Reddy, C. F. Machala and S. Krishnan, “A comprehensive
framework for predictive modeling of negative bias temperature instability,” International
28. M. A. Alam, S. Mahapatra, “A comprehensive model of PMOS NBTI degradation,” Micro-
electronics Reliability, vol. 45, pp. 71-81, 2005.
29. V. Huard, C. R. Parthasarathy, C. Guerin, M. Denais, “Physical modeling of negative
bias temperature instabilities for predictive extrapolation,” International Reliability Physics
Symposium, pp. 733-734, 2006.
30. S. V. Kumar, C. H. Kim, and S. S. Sapatnekar, “An analytical model for Negative Bias
Temperature Instability (NBTI),” International Conference for Computer Aided Design, pp.
493-496, 2006.
31. W. Wang, V. Reddy, A. Krishnan, R. Vattikonda, S. Krishnan and Y. Cao, “Compact modeling
and simulation of circuit reliability for 65 nm CMOS technology,” IEEE Transactions on
Device and Materials Reliability, vol. 7, no. 4, pp. 509-517, Dec. 2007.
32. R. Tu, E. Rosenbaum, W. Chan, C. Li, E. Minami, K. Quader, et al., “Berkeley reliability tools –
BERT,” IEEE Transactions on Computer-Aided Design of Integrate Circuits and Systems, vol.
12, no. 10, pp. 1524-1534, Oct. 1993.
33. Reliability Simulation in Integrated Circuit Design, Cadence, 2003.
34. MOS Device Aging Analysis with HSPICE and CustomSim, Synopsys, 2011.
35. ELDO User’s Manual, Mentor Graphics, 2005.
36. V. Huard, M. Denais, “Hole trapping effect on methodology for DC and AC negative bias
temperature instability measurements in pMOS transistors,” International Reliability Physics
Symposium, pp. 40-45, 2004.
37. T. Grasser, B. Kaczer, W. Goes, H. Reisinger, T. Aichinger, P. Hehenberger, et al., “The
paradigm shift in understanding the bias temperature instability: from reaction-diffusion to
switching oxide traps,” IEEE Transactions on Electron Devices, vol. 58, no. 11, pp. 3652-
3666, Nov. 2011.
38. G. I. Wirth, R. da Silva and B. Kaczer, “Statistical model for MOSFET bias temperature
instability component due to charge trapping,” IEEE Transactions on Electron Devices, vol.
58, no. 8, pp. 2743-2751, Aug. 2011.
39. J. B. Velamala, K. B. Sutaria, T. Sato, Y. Cao, “Aging statistics based on trapping/detrapping:
silicon evidence, modeling and long-term prediction,” International Reliability Physics Sym-
posium, pp. 2 F.2.1-2 F.2.5, 2012.
40. T. Grasser, H. Reisinger, W. Goes, T. Aichinger, P. Hehenberger, P.-J. Wagner, M. Nelhiebel,
J. Franco, B. Kaczer, “Switching oxide traps as the missing link between negative bias
temperature instability and random telegraph noise,” International Electron Devices Meeting,
pp.1-4, 2009.
41. G. I. Wirth, J. Koh, R. da Silva, R. Thewes, and Ralf Brederlow, “Modeling of statistical low-
frequency noise of deep-submicron MOSFETs,” Trans. on Electron Dev., vol. 52, pp. 1576-
1588, 2005.
42. A.P. van der Wel, E.A.M. Klumperink, J.S. Kolhatkar, E. Hoekstra, M.S. Snoeij, C. Salm,
H. Wallinga, B. Nauta, “Low-Frequency Noise Phenomena in Switched MOSFETs,” IEEE
Journal of Solid-State Circuits, vol. 42, no. 3, pp.540-550, March 2007.
43. M. J. Kirton and M. J. Uren, “Noise in solid-state microstructures: A new perspective on
individual defects, interface states and sow-frequency (1/f) noise,” Advances in Physics, vol.
38, p. 367-468, 1989.
44. G. Wirth, R. da Silva and R. Brederlow, “Statistical model for the circuit bandwidth dependence
of low-frequency noise in deep-submicrometer MOSFETs,” Trans. on Electron Devices, vol.
54, pp.340-345, Feb. 2007.
45. G. Wirth, R. da Silva, P. Srinivasan, J. Krick and R. Brederlow. “Statistical model for
MOSFET low-frequency noise under cyclo-stationary conditions,” International Electron
Devices Meeting, p.30.5.1-4, 2009.
46. B. Kaczer, T. Grasser, Ph. J. Rousse, J. Martin-Martinez, R. O’Connor, B. J. O’Sullivan,
and G. Groeseneken, “Ubiquitous relaxation in BTI stressing—new evaluation and insights,”
International Reliability Physics Symposium, pp. 20-27, 2008.
47. T. Grasser, H. Reisinger, P.-J. Wagner, F. Schanovsky, W. Goes and B. Kaczer, “The time
dependent defect spectroscopy (TDDS) for the characterization of the bias temperature
instability,” International Reliability Physics Symposium, pp. 16-25, 2010.
48. J. B. Velamala, K. B. Sutaria, H. Shimizu, H. Awano, T. Sato, G. Wirth, Yu Cao, “Compact
modeling of statistical BTI under trapping/detrapping,” IEEE Transactions on Electron
Devices, vol.60, no.11, pp.3645-3654, Nov. 2013
49. J. B. Velamala, K. Sutaria, H. Shimizu, H. Awano, T. Sato, Y. Cao, “Statistical aging
under dynamic voltage scaling: A logarithmic model approach,” Custom Integrated Circuits
50. K. Sutaria, J. Velamala, V. Ravi, G. Wirth, T. Sato, Y. Cao, “Multilevel reliability simulation
for IC design,” pp. 719-749, in Bias Temperature Instability for Devices and Circuits, edited
by T. Grasser, Springer, 2014.
51. K. B. Sutaria, J. B. Velamala, C. Kim, T. Sato, Y. Cao, “Aging statistics based on trap-
ping/detrapping: Compact modeling and silicon validation,” IEEE Transactions on Device and
Materials Reliability, vol. 14, no. 2, pp. 607-615, June 2014.
52. M. Denais, C. R. Parthasarathy, G. Ribes, Y. Rey-Tauriac, N. Revil, A. Bravaix, V. Huard,
F. Perrier, “On-the-fly characterization of NBTI in ultra-thin gate oxide PMOSFET’s,”
International Electron Devices Meeting, pp. 109-112, 2004.
53. H. Reisinger, O. Blank, W. Heinrigs, A. Muhlhoff, W. Gustin, C. Schlunder, “Analysis of NBTI
degradation- and recovery-behavior based on ultra-fast VT-measurements,” International
Reliability Physics Symposium, 2006.
54. K. B. Sutaria, A. Ramakumar, R. Zhu, R. Rajeev, Y. Ma, Y. Cao, “BTI-induced aging under
random stress waveforms: Modeling, simulation and silicon Validation,” Design Automation
Chapter 7
Circuit Resilience Roadmap
Veit B. Kleeberger, Christian Weis, Ulf Schlichtmann, and Norbert Wehn
Abstract Technology scaling has an increasing impact on the resilience of inte-

grated circuits. This growth is the result of (a) increasing sensitivity to various
noise sources, and (b) an increase in parametric variability. This chapter examines
the issue of circuit resilience by studying ongoing trends in technology scaling.
Additional experiments with basic circuit blocks, such as memory or logic cells,
reveal insights into their behavior for future technology generations and major
threats for circuit resilience.
1 Introduction
Faults in integrated circuits can broadly result from four aspects: (a) design bugs
which originate from mistakes in the system specification or implementation, (b)
manufacturing problems which can range from variations of design parameters to
complete changes in topology, such as resistive shorts or opens, (c) aging problems
which arise over lifetime and also can range from a degradation of design parameters
to changes in the circuit topology, and (d) signal corruption which results from
intrinsic or extrinsic temporary disturbances.
All these faults may lead to errors and failures in the whole system. These we
can again categorize broadly: (a) permanent errors which relate to changes in the
topology, (b) soft errors which result from signal corruptions, such as a particle
V.B. Kleeberger ()

Infineon Technologies AG, Neubiberg, Germany
e-mail: veit.kleeberger@infineon.com; kleeberger@tum.de
C. Weis • N. Wehn
Microelectronic Systems Design Research Group, Technische Universität Kaiserslautern,
Kaiserslautern, Germany
e-mail: weis@eit.uni-kl.de; wehn@eit.uni-kl.de
U. Schlichtmann
Institute for Electronic Design Automation, Technische Universität München, Munich, Germany
e-mail: ulf.schlichtmann@tum.de

DOI 10.1007/978-1-4614-4078-9__7
122 V.B. Kleeberger et al.
strike changing the contents of a storage element, and (c) parametric errors, where
the circuit fails to meet its specification in terms of frequency of operation, power,
or similar metrics.
In this context, a circuit is more resilient if it is able to tolerate increased
occurrence of soft and parametric errors. For digital synchronous circuits, we
can easily evaluate their resilience, using standardized measures and metrics for
their performance and correctness. Consider for example a metric such as logic
functionality: we can easily state that the steady-state output voltage of an inverter
has to be the logic inverse of its input voltage. Similarly, for example a latch has to
keep its stored value until the next latching clock pulse.
Traditionally, hard and soft errors were distinguished in the following way. Hard
errors are associated with an incorrect behavior of the circuit under all conditions,
which results from a topological change of the circuit. In contrast, soft errors are
based on changes of electrical parameters of the circuit, which causes a faulty
behavior only under specific conditions. For example, a short or open inside a
CMOS inverter may cause its output to be stuck at logic one or zero, while a shift in
gate oxide thickness behaves more subtle, as it changes the time the inverter takes
to change its state given a switching input signal. This in turn, leads only to an error
if this change causes a path to violate one of its timing constraints.
However, excessive parametric variability can cause the circuit to behave in a way
which we would traditionally associate with a hard error. A prominent example for
this in current technologies can be found in static random access memory (SRAM),
where we use the smallest possible device due to high density requirements. These
small devices are extremely susceptible to various sources of variability, such as
Random Dopant Fluctuations (RDF) or Line Edge Roughness (LER) [1]. Such
excessive variability can lead to scenarios, where a particular SRAM cell cannot
be read or written under all practical operating conditions, although the underlying
topology might still be correctly manufactured [13].
In addition to parametric variability, which we discussed above, circuits—as
they get smaller—become more sensitive to various forms of noise. As the amount
of charge being held in a storage element or transported by a switching signal
decreases, the potential for a noise source, such as a striking particle, to disrupt
this charge increases.
These problems already exist in memories for several years, which has lead
to a huge variety of solutions and techniques to deal with them. Over the years,
various kinds of redundancy, parity checking, and error correction, were proposed
and are nowadays an essential part of any practical memory design. As we start to
face similar problems now with the remaining logic parts of the design, also here
several techniques have been proposed over the last years, such as RAZOR [2] or
BISER [22] to name just two prominent examples.
Our goal in this chapter is to discuss and analyze the ongoing trend of continued
technology scaling and its implications for circuit resilience. This continues earlier
work on predicting circuit resilience conducted in [10] and [14]. To demonstrate this
behavior on specific examples, we will focus on two basic circuit blocks. On the
one hand we analyze the behavior and resilience of different SRAM architectures.
7 Circuit Resilience Roadmap 123
As already mentioned above, SRAM cells represent the most susceptible parts of a
chip due to their minimal sizing. On the other hand we discuss trends of technology
scaling in logic standard cells. Standard cells in logic designs, such as inverters,
represent (compared to SRAM cells) the other extreme and are typically considered
to be the most resilient parts of the design.
In the following, Sect. 2 discusses general trends in technology, focusing on man-
ufacturing variability, aging degradation, and particle strikes. Section 3 discusses
the behavior of SRAM performances for different architectures and technologies.
Section 4 does the same for different logic standard cells. Section 5 concludes the
chapter.
2 Trends in Technology
There exist several technology related effects, which influence the final resilience of
the circuit and its individual components. For example, manufacturing variations,
which may be in their nature either random, such as process variations, or system-
atic, such as chemical-mechanical polishing, cause a static deviation, which remains
constant over the entire life-time of the chip. Aging effects cause an additional
degradation of these characteristics which increases over life-time. Finally, effects
such as particle strikes, disturb chip operation for small amounts of time, but this
possibly quite frequently.
2.1 Manufacturing Variations
Due to imperfections of the semiconductor manufacturing process multiple param-

eters deviate from their designed nominal values. These variations can be either
random or systematic and manifest in fluctuations of transistor parameters, such
as channel length, oxide thickness or number of doping atoms. Additionally these
variations may be spatially distributed, which leads to correlations of parameter
variations of two transistors, depending on their proximity.
Channel length of transistors may deviate globally (i.e., all transistors on one die
are effected in the same way) due to mask offset variations. Typically, this effect
is assumed to be Gaussian distributed with a 3 variation being equal to 10 %
deviation from its nominal value [7]. Channel width variation typically exhibits a
similar behavior. Gate oxide thickness also follows a Gaussian distribution, but this
time due to variations in film thickness. A common assumption here is 3 being
equal to 5–10 % deviation of the nominal value.
To pattern transistors subwavelength lithography is used. As the utilized wave-
length nowadays remains constant at 193 nm, there exists a lithographic gap between
wavelength and minimum feature size of the transistors, which will increase
with ongoing technology scaling until ultraviolet lithography becomes available.
Fig. 7.1 Threshold voltage
variability due to RDF
dependent on doping
concentration for a 20 nm
planar CMOS and FinFET
technology [11, 17]
This leads to line edge roughness (LER) which impacts for example the effective
gate length of all transistors. LER does not scale with device scaling and is almost
independent of the lithographic process.
Metal gates, which were introduced to overcome limitations of using polysilicon
gate electrodes, are also subject to another variability source called metal gate
granularity (MGG). Metal gate granularity refers to the random orientation of
the different metal grains and leads to variations in the gate work function. The
magnitude of this effect is primarily dependent on the number of metal grains inside
the gate. As the size of metal grains is a material-specific property [3], the number
of metal grains decreases with ongoing scaling, which increases the effect of MGG
in future technologies.
The number and position of dopant atoms in the channel is another common
source of process variability called Random Dopant Fluctuations (RDF). Due to
decreasing numbers of dopant atoms in the channel, this effects tends to have
more influence in CMOS technologies, as one dopant more or less means a large
difference in resulting device properties.
With ongoing technology scaling transistors are not anymore produced in their
common planar style, but more advanced techniques, such as FinFET evolved,
especially due to problems resulting from RDF. Figure 7.1 shows the variation in
threshold voltage dependent on doping concentration for a planar and a FinFET
transistor produced in a similar technology.
As FinFET transistors are able to tolerate low channel doping, they can be
produced with almost negligible amount of threshold voltage variation due to RDF.
Traditional planar CMOS technologies require higher doping concentrations and,
hence, impose a higher threshold variability due to RDF.
In FinFET the transistor channel is produced in a 3D fashion and the gate is
folded around the channel, which enables a better control of the channel as well
as low channel doping, but also introduces additional variation sources. Fin height
Table 7.1 Process parameter variations for different technology effects

Effect Transistor parameter Correlations Amount
Mask offsets, etc. Channel length Global 3= D 10 % [7]
Fin thickness Global 3= D 10 % [7, 17]
Fin height Global 3= D 10 % [7, 17]
Film thickness variation Oxide thickness Global 3= D 5 % [7]
Line edge roughness Channel length Local According to [7]
Fin thickness Local According to [7]
Random dopant fluctuations Threshold voltage Local According to [11, 17]
Metal gate granularity Gate work function Local According to [3]
and fin thickness are additionally subject to global variations due to mask offsets.
Additionally fin thickness also suffers from line edge roughness. In contrast, fin
height is only dependent on film thickness and not on lithography and, hence, does
not vary due to LER.
Table 7.1 summarizes the main process parameter variations which are modeled
in this work. Global variations affect each transistor on the die in the same
way, hence, each transistor pair is fully correlated in this regard. Local variations
affect each transistor on the die individually, hence, each transistor pair is fully
uncorrelated.
2.2 Aging
Aging changes transistor parameters after the manufacturing process. These changes
are the consequence of a certain stress condition, which lead to changes in the
physical structures of the transistor. For example, Time Dependent Dielectric
Breakdown (TDDB) leads to additional conducting paths in the gate oxide, which
increase the gate leakage current. TDDB can be divided into a soft breakdown phase,
where this additional current leads to a degradation of the physical properties of the
transistor. The soft breakdown is typically followed by a hard breakdown, which
leads to a complete loss of oxide dielectric properties of the transistor. Another
effect is hot carrier (HC) degradation, which is caused by a large electric field
near the drain of the transistor channel. HC leads to a degradation of threshold
voltage and carrier mobility of the transistor. Bias Temperature instability (BTI)
has attracted much attention in the last years. As gate oxides become thinner than
4 nm, BTI-induced shift in Vth has become the dominant factor limiting device
lifetime and circuit reliability [19]. BTI can be distinguished into Negative Bias
Temperature Instability (NBTI), which occurs in p-channel transistors and Positive
Bias Temperature Instability (PBTI), which occurs in high-k dielectric n-channel
transistors. BTI requires that the transistor is operated in inversion, which means
for a PMOS transistor that its gate is negatively biased with respect to its drain.
Fig. 7.2 Prediction of Vt h

degradation in future
technologies (T D 125 ı C,
tstress D 9 years)
Fig. 7.3 Temporary current

disturbance Ip .t / by an SET
in (a) an inverter and (b) a
6-T SRAM cell
The exact amount of Vth shift is a strong function of device parameters and circuit
operation conditions such as the switching activity. Figure 7.2 presents a prediction
of NBTI induced Vth degradation for future technologies based on a model from [9].
2.3 Particle Strikes
Logic upsets and value changes in high density memories can be induced by charge
collection from ion tracks [21]. Soft error transients (SETs) are momentary current
or voltage disturbances in a circuit caused by energetic particle strikes (see Fig. 7.3).
Such an SET may propagate through subsequent circuitry, eventually reaching
a memory element or latch and causing a logic failure. The sensitivity of digital
ICs to SETs is rapidly increasing with aggressive technology scaling. Due to
decreasing parasitic capacitances and operating voltage, the critical charge Qc ,
which is required to upset a storage cell also decreases. Figure 7.4 shows the
dependence of Qc on supply voltage for different technology nodes in a 6T SRAM
cell, which can be found using a binary search and the simulation setup from
Fig. 7.3. When entering the f C region for the critical charge, as in current logic
and SRAM devices, lighter particles such as alpha and proton particles become
Fig. 7.4 Critical charge

dependence on technology
node and supply voltage for a
6T SRAM cell
Fig. 7.5 Particle dominance

dependent on critical charge
(adapted from [6])
dominant (see Fig. 7.5). This increases not only error rates, but also their spread,
as the range of lighter particles is much longer compared to residual nucleus [6].
For a scaled transistor, the surroundings of the drain junction are the most
sensitive sites to SET, e.g., the drain junction of the inverter and internal storage
nodes in a SRAM cell. For alpha particles, the charge deposition is modeled by a
double-exponential current pulse (i.e., funneling current) [12]:
Ip .t / D I0 .e t=˛ e t=ˇ / (7.1)
where I0 is the maximum charge collection current; ˛ is the charge collection time
constant and ˇ is the time constant for the rising edge of the current pulse, which
is mainly determined by the energy of the particle.
Assuming the device is a bulk CMOS transistor, the dependence of I0 and ˛ on
device parameters is expressed as:
I0 D qNE0 (7.2)
and
˛ D k"0 =qN (7.3)

2.0x10-12
-5
6.0x10 3.0x10-11
1.5x10-12
-11
tβ (s)
2.5x10
4.0x10-5
I0 (A)
tα (s)
1.0x10-12
-11
2.0x10
-13
5.0x10
2.0x10-5 -11
1.5x10
0.0
12 16 22 32 45 12 16 22 32 45
Technology (nm)
Fig. 7.6 Maximum charge collection current I0 , and collection time constants ˛ and ˇ as a
function of CMOS technology
where N is the doping concentration and is the effective mobility. With the
scaling of CMOS technology, the product of N increases, even though the effective
mobility decreases due to higher channel doping concentration. Hence, I0 increases
and ˛ decreases when the device moves to a smaller technology node, as shown in
Fig. 7.6. However, ˇ is relatively independent on device parameters (Fig. 7.6). It is
also much smaller compared to ˛ [12].
3 Trends in Memory
In this section, we first analyze different design options for SRAM cells based
on a 14 nm FinFET technology extracted from TCAD simulation. Second, we
demonstrate resilience trends for FinFET-based SRAM architectures down to 7 nm
with the help of Predictive Technology Models (PTM) [16, 18]. For this purpose we
have chosen two common SRAM architectures, namely the 6-transistor (6T) and
8-transistor (8T) bit cell shown in Fig. 7.7.
3.1 CMOS SRAM
The SRAM is well known to already have high failure rates in current technologies.
Figure 7.7 shows both architectures we analyzed. For the 6T architecture we have as
design choices the number of fins for the pull-up transistors (PU), the number of fins
for the pull-down transistors (PD) and the number of fins for the access transistors
(PG). The resulting architecture choice is then depicted by 6T_(PU:PG:PD). For
Fig. 7.7 Circuit schematics for standard 6T (a) and 8T (b) SRAM bit cells
the 8T architecture we have additionally two transistors for the read access (PGR).
Hence, the corresponding architecture choice is named 8T_(PU:(PG:PGR):PD).
As any other circuit an SRAM cell can fail in many different ways, for
example:
• Read delay failure: An SRAM cell cannot be read within a specified time.
• Write Trip Voltage (WTV) failure: The voltage swing during a write is not high
enough at the SRAM cell.
• Static Voltage Noise Margin (SVNM) failure: An SRAM cell can be flipped
unintentionally, when the voltage noise margin is too low (stability).
• Soft Error/Single Event Upset (SEU) failure: If the critical charge Qcrit is low,
the susceptibility to a bit flip caused by radiation is higher.
For the design parameters we have extended this list by:
• Write delay: Write a new content into the cell within a specified time.
• Dynamic power (SW): The power measured during a new content write
(switching).
• Leakage power: Standby leakage with zero and one content averaged (word line
set to zero).
We measure the write delay in our experiments, by first initializing the SRAM
content with a one, setting the bit lines such that they will attempt to write a zero
into the cell, pulse the word line, and measure the delay until the internal cell node
takes on the new value.
For the complete SRAM array architecture we assumed 128 cells connected to
one column (nrows = 128).
b
a
Fig. 7.8 N-Curve extraction circuit (a) and resulting N-Curve (b)
3.2 Resilience Key Parameters
We identified four parameters, namely Read delay, WTV, SVNM and Qcrit as
resilience key parameters.
To measure the SVNM and WTV we used the N-Curve metric [20]. This metric
finds the minimal amount of voltage that is required to change the cell’s content.
The extraction procedure for the N-curve is as follows (see also Fig. 7.8):
1. The bitlines BL and BR and the wordline W L are clamped to VDD
2. A DC voltage source Vi n is attached to the node R
3. The voltage of Vi n is swept from VSS to VDD and the current Ii n through the
voltage source is measured.
This then results on the curve shown in Fig. 7.8b. The voltage difference between
the points A and B in the curve is the maximum tolerable DC voltage noise or the
static voltage noise margin (SVNM) [5]. Similarly, the voltage difference between
the points C and B is the voltage drop required to flip the internal node of the cell,
also called the write-trip voltage (WTV).
For measuring the read delay, we initialize the SRAM bit with a one, initialize
both bit lines to the supply voltage, leave them floating (high-Z), and pulse the
wordline. Then we measure the delay as the time the bit line node requires to drop
to 2/3 of the supply voltage.
Figure 7.9a shows the impact of the number of rows used in one column on
the read delay for various 6T and 8T architectures. The 6T and 8T one fin for all
transistors designs are very similar (in fact their read delays are equal) and show
the largest impact, while the double sized 8T cell is almost as good as the high-
performance 6T from Intel (6T_(1:2:3)) [8]. Figure 7.9b shows the impact of VDD
on the SVNM. We see that 8T cells show for both models PTM 20 nm and TCAD
14 nm a very similar trend as they have nearly identical P/N ratios. The SVNM of
the 8T cell relies only on the intrinsic feedback of the storage inverter. The 6T cell
a 60
b 500
6T_(1:1:1) 6T_1:1:1_TCAD_14nm
8T_1:(1:1):1 450 8T_1:(1:1):1_TCAD_14nm
50 6T_1:1:1_PTM_20nm
8T_2:(2:2):2 400 8T_1:(1:1):1_PTM_20nm
Read delay [ps]
6T_(1:2:3)
SVNM [mV]
40 350
30 300
250
20
200
10 150
0 100
nrows:128 nrows:256 nrows:512 500 550 600 650 700 750 800 850 900 950 1000
VDD [mV]
Fig. 7.9 Impact of number of rows on read delay for a 14 nm TCAD model (a) and SVNM
comparison between 6T and 8T cells using different models (b)
behavior differs between TCAD based model and PTM. This is due to the 100 mV
lower threshold voltage (Vth ) of the PTM models, which reduces the SVNM.
For Qcrit measurements we use the setup shown in Fig. 7.3b. Basically, we
measured the amount of charge needed to flip the cell, i.e., to corrupt the data.
3.3 Design Trends and Considerations
We observed that there is no single cell, which provides the best solution under any
requirement specification. Therefore, also Intel presented three different sizings for
the 6T cell [8]. There is a need for at least a high-density cell and a high-performance
cell. So we included in our study the three flavors of Intel’s 6T cell and several 8T
design alternatives to show what might be achievable.
First, we compared various 6T and 8T architectures by an exhaustive analysis to
quantify the design influence.
In Fig. 7.10 different SRAM design options are evaluated for the resilience
key parameters normalized to the respective values of a 6T_(1:1:1) minimal sized
cell. In general, higher values indicate a performance improvement (e.g. +10 %).
Obviously, it is not good idea to use the design options 6T_(1:2:1) and 6T_(2:2:1),
as they decrease the SVNM by 30 %. For 8 fin designs the 8T_(1:(1:1):1) cell might
be a much better solution, which offers 22 % SVNM improvement. But also the
6T_(1:1:2) used by Intel as high density low voltage design has an increased SVNM
(10 %). When it comes to high performance cell options, the Intel variant 6T_(1:2:3)
shows the fastest read delay, followed by many 8T designs and some 6T cells.
However, if we want to lower the soft error rate, despite of increasing the voltage
supply, the best SRAM design option is the 8T_(2:(2:2):2) cell, which unfortunately
occupies a very large area.
120%
110%
100%
90%
80% Read delay
70% Qcrit
SVNM
60% WTV
50%
40%
30%
20%
10%
0%
-10%
-20%
-30%
8T_2:(1:1):1
8T_1:(2:1):2
8T_1:(1:2):2
8T_2:(2:2):2
8T_1:(2:2):3
6T_1:2:1
6T_1:2:2
6T_2:2:2
6T_1:2:3
6T_2:1:1
6T_1:1:2
8T_1:(1:1):1
6T_2:2:1
8T_1:(1:1):2
8T_1:(2:2):1
8T_2:(2:2):1
8T_1:(2:2):2
8T_1:(3:2):2
8 FINS 10 FINS 12 FINS 14 FINS 16 FINS
Fig. 7.10 Comparison of read delay, SVNM, WTV and Qcrit for different design variants
normalized to the results of the 6T_(1:1:1) bit cell
Write delay
25% Leakage (avg)
Power (SW)
0%
-25%
-50%
-75%
-100%
-125%
-150%
6T_2:1:1
6T_1:1:2
6T_1:2:1
8T_1:(1:1):1
6T_2:2:1
6T_1:2:2
8T_2:(1:1):1
8T_1:(1:1):2
6T_2:2:2
6T_1:2:3
8T_1:(2:2):1
8T_1:(2:1):2
8T_1:(1:2):2
8T_2:(2:2):1
8T_1:(2:2):2
8T_2:(2:2):2
8T_1:(2:2):3
8T_1:(3:2):2
8 FINS 10 FINS 12 FINS 14 FINS 16 FINS
Fig. 7.11 Comparison of write delay, leakage, and switching power (SW) for different design
variants normalized to the results of the 6T_(1:1:1) bit cell
In Fig. 7.11 we plotted the additional design parameters. We see that all 8T
cells have much increased leakage power, while Intel’s 6T_(1:2:3) offers a 10 %
decreased write delay with an moderate increase of dynamic power consumption.
However, the leakage power increase is for this option also above 100 % of a
6T_(1:1:1) cell. An interesting alternative design option with 12 fins could be the
8T_(1:(2:2):1) cell, as it shows 25 % decreased write delay with no impact on
dynamic power. The leakage power increase of this variant could be controlled by
lowering VDD in standby or by active backward body biasing.
a 30 10 b 30 10
9 9
3σ (% of Median Value)
25 8 25 8
Read Delay [ps]

Read Delay [ps]
20 7 20 7
6 6
15 5 15 5
4 4
10 3 10 3
5 2 5 2
1 1
0 0 0 0
20 16 14 10 7 20 16 14 10 7
Technology Node [nm] Technology Node [nm]
6T_(1:1:1) cell 8T_(1:(1:1):1) cell
Fig. 7.12 Impact of VDD variation (˙10 %) in absolute values (box plots) and relative increase
compared to the median (dashed line) for a 6T_(1:1:1) high density (a) and a standard
8T_(1:(1:1):1) cell (b)
From a design perspective there is no clear winner for all evaluated design
variants. The best choice depends on the individual requirements and priorities,
which can be either high density, high robustness or high performance.
3.4 SRAM Scaling Trends
In this section we want to analyze different scaling related trends for SRAM cells
using PTM SPICE models from [18]. First of all we want to analyze the impact of
VDD variations. Afterwards, we quantify trends considering transistor technology
variations as described in Sect. 2.1.
To evaluate the different technology nodes we run Monte-Carlo (MC) simula-
tions with 1,000 samples to get an impression of the general trend.
We quantified the impact of supply voltage noise by assuming a maximal VDD
variation of ˙10 % for a 3 yield. Figure 7.12 shows the corresponding results of
the respective Monte Carlo simulations. We recognize, as expected, an increasing
influence of VDD on the read delay for the smaller nodes (up to 9.4 % for the 7 nm
node). Additionally, we observe a very similar trend for the 6T and 8T architectures.
To evaluate the scaling trends of the chosen resilience key parameters when
scaling down to the 7 nm node, we run MC simulations considering process
variations. Figures 7.13 and 7.14 show box plots and the percentage deviation for
a 3 yield with respect to median values for a 6T_(1:1:1) high density and an
8T_(1:(1:1):1) SRAM cell, respectively.
Although the read delay decreases with technology scaling (see Figs. 7.13a and
7.14a), which theoretically enables a higher working frequency, its relative 3
variation becomes as high as 50 % at the 7 nm node. This compromises its robustness
and cripples the possible increase in frequency.
a 35 60
b 800 35
700
Read delay [ps]
30 50 30
25 600 25
Qcrit [aC]
40 500
20 20
30 400
15 300 15
10 20 10
200
5 10 100 5
0 0 0 0
20 16 14 10 7 20 16 14 10 7
c 600 70 d 500 100
500 60
400 80
SVNM [mV]
WTV [mV]
400 50
40 300 60
300
30 200 40
200 20
100 100 20
10
0 0 0 0
20 16 14 10 7 20 16 14 10 7
Fig. 7.13 Scaling trends for a 6T_(1:1:1) high density SRAM cell: (a) read delay, (b) Qcrit , (c)
WTV and (d) SVNM
a b
35 60 800 35
700
Read delay [ps]
30 50 30
25 600 25
Qcrit [aC]
40 500
20 20
30 400
15 300 15
10 20 10
200
5 10 100 5
0 0 0 0
20 16 14 10 7 20 16 14 10 7
c 600 60 d 600 30
500 50 500 25
SVNM [mv]
WTV [ mV]
400 40 400 20
300 30 300 15
200 20 200 10
100 10 100 5
0 0 0 0
20 16 14 10 7 20 16 14 10 7
Fig. 7.14 Scaling trends for a 8T_(1:(1:1):1) SRAM cell: (a) read delay, (b) Qcrit , (c) WTV and
(d) SVNM
Fig. 7.15 (a) NAND and (b)

NOR cells with obtained
sizings (=fin numbers) in
parentheses
The critical charges Qcrit shows a similar trend, where its decrease with technol-
ogy scaling is accompanied by a noticeable increase in 3 variability (around 30 %
at 7 nm). Nevertheless, the 8T_(1:(1:1):1) design provides an approximately 10 %
higher Qcrit than the 6T_(1:1:1), and a slightly smaller 3 variability. Regarding
WTV, a quick examination of Fig. 7.7 shows that both designs are identical in terms
of writing procedure, which is exemplified in Figs. 7.13c and 7.14c. In both cases
their 3 variability increases from 40 % at 20 nm to nearly 58 % at 7 nm. Such
high variations have to be taken into account to guarantee enough voltage swing
during a write procedure to flip the cell’s content as expected. Perhaps the biggest
advantage of the 8T_(1:(1:1):1) design over the 6T_(1:1:1) version with respect to
the resilience key parameters presented here, is its much improved SVNM. Not only
reaches the 8T design an approximately 22 % higher SVNM than the 6T design, but
it is also much more robust in terms of 3 variability (28 % for 8T 7 nm compared
to 90 % for 6T 7 nm).
This overall analysis of the resilience key parameters presented here (read
delay, Qcrit , WTV, SVNM) shows clearly that the variability increases rapidly as
technology is scaled down, which has to be taken into account when designing
robust SRAM cells.
4 Trends in Logic
Before making predictions regarding the scaling trend of FinFET standard cells,
we first want to evaluate in detail how process variations influence the behavior of
FinFET-based logic.
4.1 Influence of Process Variations
To evaluate the influence of process variations we utilize the a 14 nm PTM model,

which has a nominal channel length of 18 nm [18]. As characteristic representatives
we choose a NAND and a NOR cell, which are shown in Fig. 7.15 together with
their obtained sizings using a discrete sizing algorithm from [15].
Fig. 7.16 Q-Q plot—delay and slope. (a) NAND—delay. (b) NOR—delay. (c) NAND—slope.
(d) NOR—slope
To evaluate these cells we run a Monte-Carlo simulation with 1,000 samples and
measure delay, slope, and leakage power for every input timing arc. Figures 7.16
and 7.17 show the corresponding Q-Q plots.
Obviously, delay is very well standard normal distributed as we already expect it
to be from bulk technologies. For the slope inverse (i.e., switching time) this holds
also for most timing arcs. Only for the timing arcs where the entire transistor stack
is involved (e.g., Br ! Zf for the NAND) this property does not hold. Figure 7.17
shows the lognormal Q-Q plot for leakage power. Leakage variability fits very well
a lognormal distribution for most samples. However, the tail seems to not fulfill this
assumption, hence, for accurate yield predictions more advanced distribution types
will probably be required here.
Additionally we conduct a worst case analysis as described by [4]. Based on
the gradient pointing to the worst-case point in the process parameter space, we
Fig. 7.17 Q-Q plot—leakage. (a) NAND. (b) NOR
Fig. 7.18 Delay and leakage variability—NAND. (a) Variability sources. (b) Global vs. local
compute the contribution of the different variability sources as well as the single
transistors compared to global variability (Figs. 7.18 and 7.19).
In both cell types, metal gate granularity is the major variability source for delay
and leakage. Besides, fin thickness also contributes considerably, especially for
leakage. The contribution of random dopant fluctuations is well suppressed due to
the low channel doping. Most variability is caused by local variations, which can
again be attributed to the dominance of metal gate granularity, followed by line
edge roughness effects.
Fig. 7.19 Delay and leakage variability—NOR. (a) Variability sources. (b) Global vs. local
Fig. 7.20 FO4 circuit
4.2 Scaling Trends
To evaluate scaling trends of nominal performances we utilize a fanout of four (FO4)

inverter circuit as depicted by Fig. 7.20. The driving inverter is represented by the
transistors M1 and M2, while the remaining transistors M3 to M10 represent its
load. We assume that all p-channel transistors (i.e., M1, M3, . . . ) are equally sized
and that all n-channel transistors (i.e., M2, M4, . . . ) are equally sized. The transistors
were sized so that the rise and fall delay as well as the rise and fall output slope of the
driving inverter (composed of M1 and M2) are approximately equal. Additionally
an approximately equal input and output slope was targeted.
We sized this circuit for different technology nodes using models from [18]
and [23] and extracted characteristic performances such as P/N sizing ratio, output
slope, delay and equivalent input capacitance of the driving inverter (Fig. 7.21).
While in planar CMOS it is required to size P- and N-transistors differently
to obtain similar behavior (i.e., equal rise and fall delay), in FinFET P- and N-
transistors behave more equally. This leads to an equal number of fins for both
Fig. 7.21 Scaling of minimal-sized inverter performances. (a) P/N ratio. (b) Delay. (c) Slope.
(d) Input capacitance
transistor types. The delay of the FO4 inverter initially scales with a very steep slope,
but after some generations this trend quickly reverts to an less steep scaling as we
know from planar technologies. Interestingly, slope scales very steeply in FinFET
technology. Input capacitance is in general much higher for FinFET compared to
planar technologies in the same technology node. Both, the scaling of slope and the
value of the capacitance, can be very well explained by the three-dimensional gate
controlling the channel.
To evaluate the scaling trend in FinFET technologies in the presence of variabil-
ity, we size a mid-performance inverter and evaluate its behavior under different
specifications, which are listed in Table 7.2. Specification 0 represents nominal
conditions without any variability. In each subsequent specification one additional
variation source is added: Process variations in spec 1, device temperature variation
in spec 2, and supply voltage variation in spec 3.
Table 7.2 Specifications for Spec-Nr. Yield VDD T (ı C)

performance evaluation
0 Nominal Nominal 27
1 (P) 3 Nominal 27
2 (PV) 3 ˙10 % 27
3 (PVT) 3 ˙10 % 0–60
Fig. 7.22 Scaling trends for timing performances of a medium-sized inverter. (a) Nominal delay.
(b) Delay degradation. (c) Nominal slope. (d) Slope degradation
Figure 7.22 shows the change of timing performances (delay and slope) over
the respective technology nodes, while Fig. 7.23 shows the change in power
performances (leakage power and switching energy).
Delay degradation due to process variations increases from 17 % up to 30 %
for a 3 -design. A detailed analysis reveals that from this increase about 5 % can
be accounted to an increase of sensitivity to process variation due to technology
scaling. The remaining larger increase results mostly from the increase of gate
work function variability due to metal gate granularity. This corroborates that metal
gate granularity is the major source of process variability in FinFET as metal grain
size does not scale. Another interesting insight is that the influence of temperature
increases over the studied technologies. While we observed in this study almost no
Fig. 7.23 Scaling trends for power performances of a medium-sized inverter. (a) Nominal leakage.
(b) Leakage degradation. (c) Nominal switching energy. (d) Switching energy degradation
influence of temperature at 24 nm channel length, the additional increase in delay at

14 nm channel length is about 10 percentage points.
Leakage power shows temperature and process variation as its major variability
contributors, while supply voltage mainly determines the variation of switching
energy.
5 Conclusion
We presented in this chapter an analysis of ongoing scaling related trends of

integrated circuits. For this purpose, we discussed major trends observed in process
technology, as well as in memory and logic designs. While newer technology styles,
such as FinFETs provide necessary advances to enable further technology scaling,
we can observe that there exists also a trend for reduced resilience with ongoing
scaling. Our hope is that with this roadmap, which concentrates on FinFET-based
designs, we are able to sensitize the research community to present and near
future problems associated with this technology. Additionally, we analyzed different

architectures for SRAM memories to demonstrate the different design choices
available and tradeoffs that have to be made.
Acknowledgements The authors want to gratefully acknowledge the contributions of Sani Nassif,
Nikhil Mehta, Helmut Graeb, and Javier Varela to this work. Veit Kleeberger was employed at the
Institute for Electronic Design Automation (TUM) during this work. This work was supported
in part by the German Research Foundation (DFG) as part of the priority program “Dependable
Embedded Systems” (SPP 1500—http://spp1500.itec.kit.edu).
References
1. Asenov, A.: Random dopant induced threshold voltage lowering and fluctuations in sub-0.1 m
mosfet’s: A 3-d atomistic simulation study. IEEE Trans. Electron Devices 45(12), 2505–2513
(1998)
2. Austin, T., Blaauw, D., Mudge, T., Flautner, K.: Making typical silicon matter with Razor.
IEEE Computer 37(3), 57–65 (2004)
3. Dadgour, H., Endo, K., De, V., Banerjee, K.: Modeling and analysis of grain-orientation effects
in emerging metal-gate devices and implications for SRAM reliability. In: IEEE International
Electron Devices Meeting (IEDM), pp. 1–4 (2008)
4. Graeb, H.: Analog Design Centering and Sizing. Springer Verlag (2007)
5. Grossar, E., Stucchi, M., Maex, K., Dehaene, W.: Read stability and write-ability analysis of
SRAM cells for nanometer technologies. IEEE Journal of Solid-State Circuits 41(11), 2577–
2588 (2006)
6. Ibe, E., Chung, S.S., Wen, S., Yamaguchi, H., Yahagi, Y., Kameyama, H., Yamamoto, S.,
Akioka, T.: Spreading diversity in multi-cell neutron-induced upsets with device scaling. In:
IEEE Custom Integrated Circuits Conference (CICC), pp. 437–444 (2006)
7. International technology roadmap for semiconductors (2013)
8. Jan, C.-H. et al.: A 22nm SoC Platform Technology Featuring 3-D Tri-Gate and High-k/Metal
Gate, Optimized for Ultra Low Power, High Performance and High Density SoC Applications.
In: IEEE International Electron Devices Meeting (IEDM), pp. 44–47 (2012)
9. Kleeberger, V.B., Barke, M., Werner, C., Schmitt-Landsiedel, D., Schlichtmann, U.: Compact
Model for NBTI Degradation and Recovery under Use-Profile Variations and ist Applica-
tion to Aging Analysis of Digital Integrated Circuits. Microelectronics Reliability 54(6–7),
pp. 1083–1089 (2014)
10. Kleeberger, V.B., Graeb, H., Schlichtmann, U.: Predicting Future Product Performance:
Modeling and Evaluation of Standard Cells in FinFET Technologies. In: ACM/IEEE Design
Automation Conference (DAC) (2013)
11. Lu, D.D., Lin, C.H., Niknejad, A.M., Hu, C.: Compact Modeling of Variation in FinFET
SRAM Cells. IEEE Design & Test of Computers 27(2), 44–50 (2010)
12. Messenger, G.C.: Collection of charge on junction nodes from ion tracks. IEEE Trans. Nuclear
Science, vol. 29, issue 6, pp. 2024–2031 (1982)
13. Nassif, S.R., Kleeberger, V.B., Schlichtmann, U.: Goldilocks failures: Not too soft, not too
hard. In: IEEE International Reliability Physics Symposium (IRPS) (2012)
14. Nassif, S.R., Mehta, N., Cao, Y.: A resilience roadmap. In: IEEE Conference on Design,
Automation & Test in Europe (DATE) (2010)
15. Pehl, M.: Discrete Sizing of Analog Integrated Circuits. Ph.D. thesis, Technische Universität
München (2012)
16. Predictive technology model (ptm). available at http://www.eas.asu.edu/~ptm
17. Saha, S.K.: Modeling process variability in scaled CMOS technology. IEEE Design & Test of
Computers 27(2), 8–16 (2010)
18. Sinha, S., Yeric, G., Chandra, V., Cline, B., Cao, Y.: Exploring sub-20nm FinFET design with
predictive technology models. In: ACM/IEEE Design Automation Conference (DAC) (2012)
19. Wang, W., Reddy, V., Krishnan, A.T., Vattikonda, R., Krishnan, S., Cao, Y.: Compact modeling
and simulation of circuit reliability for 65nm cmos technology. IEEE Trans. on Device and
Materials Reliability 7(4), 509–517 (2007)
20. Wann, C., Wong, R., Frank, D., Mann, R., Ko, S.B., Croce, P., Lea, D., Hoyniak, D., Lee, Y.M.,
Toomey, J., et al.: SRAM cell design for stability methodology. In: International Symposium
on VLSI Technology (VLSI-TSA-Tech), pp. 21–22 (2005)
21. Wirth, G., Vieira, M., Neto, E., Kastensmidt, F.: Generation and propagation of single event
transients in cmos circuits. In: Design and Diagnostics of Electronic Circuits and systems, 2006
IEEE, pp. 196–201 (2006)
22. Zhang, M., Mitra, S., Mak, T., Seifert, N., Wang, N.J., Shi, Q., Kim, K.S., Shanbhag, N.R.,
Patel, S.J.: Sequential element design with built-in soft error resilience. IEEE Transactions on
Very Large Scale Integration (VLSI) Systems 14(12), 1368–1378 (2006)
23. Zhao, W., Cao, Y.: New generation of predictive technology modeling for sub-45nm early
design exploration. IEEE Trans. Electron Devices 53(11), 2816–2823 (2006)
Chapter 8
Layout Aware Electromigration Analysis
of Power/Ground Networks
Di-an Li, Malgorzata Marek-Sadowska, and Sani R. Nassif
Abstract In this chapter, we briefly introduce physical foundations of electromi-

gration (EM) and present two classical EM-related theories. We discuss physical
parameters affecting EM wire lifetime and we introduce some background related
to the existing EM physical simulators. In our work, for EM physical simulation
we adopt the atomic concentration balance-based model. We discuss the simulation
setup and results. We present VEMA—a variation-aware electromigration (EM)
analysis tool for power grid wires. The tool considers process variations caused
by the chemical–mechanical polishing (CMP) and edge placement error (EPE).
It uses a compact model that features critical region extraction and variation
coefficient calculation. VEMA is a full-chip EM analysis tool; it extracts the
effective jL product values and performs a via-centric EM lifetime calculation on
ideally manufactured EM-mortal wires. It analyzes process variation effects on EM
reliability and reports variation tolerances of EM-sensitive power grid wires.
1 Introduction to Electromigration
1.1 Basic Concepts
Electromigration (EM) is the transport of material caused by the gradual movement

of ions in a conductor due to the momentum transfer between conducting electrons
and diffusing metal ions that make up the lattice of the interconnect material [1, 2].
The tensions caused by accumulating ions will gradually nucleate a void or an
D. Li
Qualcomm Inc, 5775 Morehouse Dr, San Diego, CA 92121, USA
e-mail: dianl@qti.qualcomm.com
M. Marek-Sadowska ()
University of California, Santa Barbara, CA 93106, USA
e-mail: mms@ece.ucsb.edu
S.R. Nassif
Radyalis, 7000 North Mopac Expressway, Suite 200, Austin, TX 78731, USA
e-mail: srn@radyalis.com

DOI 10.1007/978-1-4614-4078-9__8
146 D. Li et al.
a b
extrusion
void
Fig. 8.1 EM failures. (a) Void (b) extrusion
extrusion which will grow in size over time and eventually form an open or short
circuit as shown in Fig. 8.1.
In reality, void formation [3] occurs much faster than extrusion growth, which
is very unlikely to occur [4]. Therefore voids are considered the major EM-
caused failures. In this chapter, we only focus on the EM-induced void damage
of interconnects.
The ion movement is caused by several forces, the two main ones are: (1) The
direct electrostatic force Fe caused by the electric field, and (2) The force from the
momentum transfer with moving electrons Fp , which is in the opposite direction to
the electric field [5]. The resulting force Fres acting on an ion is given by (8.1).

Fres D Fe Fp D q Ze C Zp E D qZ j (8.1)
In (8.1), q is the electron charge, Ze is the nominal valence of the metal, Zp accounts
for the magnitude and direction of the momentum exchange between the conducting
electrons and the metal ions, Z* is the effective valence, E is the electrical field, j is
current density and is the metal resistivity.
1.2 Classical Theories

1.2.1 Black’s Equation
Black’s Equation is a mathematical model for the mean time to failure (MTTF) of a
wire due to electromigration [6]. The equation is given by (8.2).
Ea
M T TF D Aj n e . kT / (8.2)
In (8.2), A is an experimental constant, j is the current density, n is a model

parameter, Ea is the activation energy, k is Boltzmann constant, T is the absolute
temperature in K.
The model is abstract, not based on a specific physical model, but it flexibly
describes the failure rate dependence on the temperature, the electrical stress, and
the specific technology and materials. The value of Black’s equation is that it maps
8 Layout Aware Electromigration Analysis of Power/Ground Networks 147
Fig. 8.2 Back stress buildup

- - + - + + - + +++++
+
experimental data taken at elevated temperature and stress levels in short periods of
time to the expected component failure rates under actual operating conditions.
1.2.2 Blech Length Effect
For electromigration to occur, there is also a lower limit on the interconnect

length [2]. It is known as “Blech length” [7], and any wire that has a length
below this limit (typically on the order of 10–100 m) will not fail by electro-
migration. Here, the mechanical stress buildup causes a reversed migration process
which reduces or even compensates the effective material flow towards the anode
(Fig. 8.2).
Although known as “Blech length” effect, the actual criterion to determine if a
wire is EM-immortal or not, is the product of its current density j and length L.
Specifically, a conductor line is not susceptible to electromigration if the product
of the wire’s current density j and length L is smaller than a process technology-
dependent threshold value jLthreshold . Below, we briefly explain the phenomena
behind the Blech length effect. The electromigration induced drift velocity is
determined by (8.3).
Da jZ j ej
vd D (8.3)
kT
I.A. Blech observed that the metal ions stopped moving, when the wire length was
reduced to a certain length. Also, he observed that no ion drift could be detected
below a threshold current density.
These observations can be explained by considering the flux due to electromigra-
tion and the gradient of the chemical potential via a gradient of mechanical stress
[7–10] according to (8.4)

D v Cv @
Jv D jZ j ej (8.4)
kT @x
where ˝ is the atomic volume, and is the hydrostatic stress. This equation shows
that a gradient of mechanical stress acts as driving force against electromigration.
Thus, electromigration stops, when the opposing stress gradient, commonly referred
to as “back stress” [11], equals the electromigration driving force, so that Jv D 0.
This steady-state condition is the so-called “Blech Condition”, given by (8.5)
@ jZ j ej
D (8.5)
@x
148 D. Li et al.
Integrating (8.5) over the interconnect line length yields (8.6)
jZ j ej
.x/ D 0 C x (8.6)

where 0 is the stress at x D 0. This equation shows that the stress varies linearly
along the line, when the backflow flux equals the electromigration flux. Given that
the maximum stress a conductor line can withstand is th , the critical product for
electromigration failure can be stated as (8.7)
.th 0 /
.jL/c D (8.7)
jZ j e
This is the so-called “Blech Product”. The critical product provides a measure of
the interconnect resistance against electromigration failure and several experimental
works have reported that the critical product for modern copper interconnects is in
the range from 2,000 to 10,000 A/cm [12–15].
1.3 Affecting Factors

1.3.1 Current Density
Current density of a wire is a ratio of the total current flowing through that wire, I,
and its cross-section area, A (8.8).
I
j D : (8.8)
A
According to Black’s equation, MTTF of a wire is inversely proportional to its
current density. And their correlation depends on the current exponent n. There is
some controversy across EM literature regarding the exact value of n. A common
agreement is that the typical value of n is between 1 and 2. Interestingly, the most
important factor that determines the value of n is the current density itself. The
value of n is close to 1 when the current density is small, and it gradually increases
to two as the current density increases. For example, in [16], the authors report
n D 1 for j 0.1 MA/cm2 , n D 1.5 for 0.1 MA/cm2 < j < 1 MA/cm2 and n D 2 for
j 1 MA/cm2 ; in [17], the authors report n D 1.1 ˙ 0.2 for j 2.5 MA/cm2 and
n D 1.8 for j > 2.5 MA/cm2 . For n D 1, the time-average current density can be used
in (8.2) and the current waveform shape does not affect EM lifetime; else if n > 1,
transient current density should be considered and the shape of a current waveform
matters.
1.0
Ea=0.7eV
Normalized MTTF 0.8
0.6
0.4
0.2
0
300 310 320 330 340 350 360
Temperature (K)
Fig. 8.3 MTTF–temperature dependency
1.3.2 Temperature
According to Black’s equation, MTTF of a wire has an exponential dependency on

wire temperature. Figure 8.3 depicts the MTTF trend when temperature changes.
In Fig. 8.3, MTTFs are normalized to MTTF at 300 K. A sharp MTTF decrease
( 0.5X) can be observed when temperature rises from 300 to 310 K. Some papers
report huge temperature rise (hundreds of K) on wires [18]. We believe that these
results are exaggerated due to unrealistic assumption of extremely high current wire
densities. However, in practice, 10 K wire temperature rise above the substrate
temperature is common, which can still cause a non-negligible effect on wire EM
MTTF.
1.3.3 Wire Length
Wire length affects two parameters: jlimit and jeffect . According to Blech length effect,
there is a wire length limit that any wire shorter than that limit is considered EM-
immortal. From a different perspective, it can be also interpreted as given a wire
length, there is a current density limit that prevents the EM on the wire, as stated
in (8.9).
.th 0 /
jlim it D (8.9)
jZ j eL
Since the Blech length effect is caused by the mechanical stress buildup with a
reverse migration process, even if the current density on a wire is greater than jlimit ,
the reverse stress still exists. Therefore, the effective current density jeffect that is
150 D. Li et al.
a I1
I2
Ivia= I1+I2 = I3+I4
I3 I4
b c d
I1=1A I1=1.2A I1=1.9A
I3=1A I4=1A I3=0.8A I4=1.2A I3=0.1A I4=1.9A
I2=1A I2=0.8A I2=0.1A
Fig. 8.4 Via-array current distribution. (a) 3D orthogonal wire structure with 4-by-4 via-array. (b)
Even distribution (c) uneven distribution (d) extreme uneven distribution
used in Black’s equation is modified by j jlimit as in (8.10) [19].

Ea
M T TF D A.j jlimit /n e . kT / (8.10)
1.3.4 Interconnect Structure
The interconnect structure that affects EM mainly refers to the via-array design. For
a single via, all current is flowing through the via and its current density can be
easily obtained. However in a via-array, the current is not evenly distributed into
each via. References [20, 21] discuss this effect.
In Li and Guan [22], a 3D orthogonal wire structure with a 4-by-4 via-array
is built as shown in Fig. 8.4a. Several current configurations are applied to the
structure, and current densities through the horizontal cross-sectional plane of
multiple vias are recorded as shown in Fig. 8.4b–d. It can be observed that when
all wire currents are identical as in Fig. 8.4b, current distribution into the vias is also
uniform. Greater differences among the wire segment currents cause the current
split into the vias to be more non-uniform.
This effect can be explained by the differences among path resistances through
different vias. A SPICE simulation on a pure resistive mesh validates this observa-
Fig. 8.5 MTTF-activation energy dependency
tion. Based on the resistive mesh model, the current distribution into each via can
be determined and the EM lifetime of the via-array can be estimated. A detailed
discussion can be found in [22].
1.3.5 Activation Energy
Activation energy is the minimum energy required to cause thermal diffusion of

metal ions. It plays an important role due to its exponential effect on MTTF [23].
Figure 8.5 shows the MTTF trend when activation energy changes, while tempera-
ture is fixed at 300 K. In Fig. 8.5, MTTF is normalized to its value at Ea D 0.7 eV.
The exponential dependency can be observed.
Interconnect material is the dominant factor that determines the value of Ea .
Aluminum was the typical material used for interconnects before copper was first
introduced in 1997 [24]. Due to its generally higher activation energy, copper has a
significantly better electromigration resistance than aluminum does. In this chapter,
we focus our discussion on copper interconnect EM reliability.
1.3.6 Diffusion Paths
In a homogeneous crystalline structure, because of the uniform lattice structure of

the metal ions, there is hardly any momentum transfer between the conducting
electrons and the metal ions. However, this symmetry does not exist at the grain
boundaries and material interfaces, so here the momentum is transferred much more
vigorously [25]. Since in these regions, the metal ions are bonded more weakly
than in a regular crystal lattice, once the electron wind has reached certain strength,
152 D. Li et al.
anode
cathode Cu+
e- e- barrier
e- void anode layer
grain
boundary surface e- cathode
bulk
interface barrier Cu+
void
layer
Fig. 8.6 Dual-damascene copper EM diffusion paths
Table 8.1 Copper diffusion Diffusion path Ea (eV)

paths and the activation
energy Surface 0.50.7
Interface 0.81.25
Grain boundary 1.21.25
Bulk 2.1
atoms become separated from the grain boundaries/interfaces and are transported
in the direction of the current. Figure 8.6 shows the EM diffusion paths in a dual-
damascene copper interconnect structure, which is the main metallization process
in modern semiconductor fabrication [26].
In a dual-damascene structure, only a single metal deposition step is used to
simultaneously form the main metal line and the via beneath it. After the via and
trench recesses are etched, the via is filled in the same metal-deposition step that
fills the trench. Therefore, vias are always in the same copper (Cu) body as the
upper level metal, but are separated from the lower level metal by a barrier layer
which prevents copper diffusion into the dielectric as shown in Fig. 8.6. Typical
barrier layer materials are Ta, TaN, TiN and TiW. The diffusion processes caused by
electromigration can be divided into grain boundary diffusion, bulk diffusion and
surface diffusion. In general, surface diffusion is dominant in copper interconnects.
The diffusion path differences can be explained by their different activation energies
listed in Table 8.1.
2 Physical Simulation of EM
2.1 Background
One branch of the EM-related research is focused on its physical simulation.

Classical theories such as Black’s equation and Blech length effect provide macro-
level analysis tools for EM but are limited in explaining micro-level EM behavior.
The purpose of physical simulation is to better understand the physics behind EM
phenomena and more precisely predict interconnect EM behavior at microstructural

level. With such a complex system, typically a finite element method (FEM)
tool is used for computation and simulation. In [27], the author develops an EM
simulation method based on flux divergence caused by driving forces including
electromigration, thermo-migration, stress-migration and atomic concentration. In
[28], an alternative concept of EM modeling is proposed based on the system energy
reduction as the simulation criterion. In this section, we mainly apply the atomic
concentration model from [27] to simulate interconnect structures. As EM theory
becomes more complete, we expect the physical simulation to be closer to the real
situation.
2.2 Balance of Atom Concentration
The governing equation which describes the atom concentration evolution through-
out an interconnect segment, is the conventional mass balance (continuity) equation
in (8.11) [1].
@N .x; t /
Cr J D0 (8.11)
@t
In (8.11), N(x, t) is the atom concentration at the point with coordinates x D (x, y, z)
at the moment of time t, and J is the total atomic flux at this location. The total
atomic flux J is a combination of the fluxes caused by the different atom migration
forces. The major forces are induced by the electric current, and by the gradients of
temperature, mechanical stress and concentration: J D JE C Jth C JS C JC .
To define the fluxes mentioned above, we have (8.12).
8!
ˆ
ˆ J E D kTN
eZ jD0 e kT
Ea
ˆ
ˆ !

ˆ
ˆ
D E
0 kTa
ˆ
<! J th D NQ kT 2
e grad T
N D0 E a
J s D kT e kT gradH (8.12)
ˆ
ˆ !

ˆ
ˆ
Ea
J D D0 e kT gradN
ˆ
ˆ dNc
:̂ ! !
!

!
dt
C d i v J E C J th C J s C J c D 0
In (8.12), JE , Jth , JS and JC are atomic fluxes induced by electrical, thermal, stress
and atomic concentration forces, respectively. N is the atomic concentration, Q* is
the specific heat of transport of metal, and H is the local hydrostatic stress. EM
lifetime is considered to be proportional to atomic flux divergence (AFD D divJ), so
we have here MTTF / AFD.
154 D. Li et al.
Fig. 8.7 3D view of a

simulated wire structure
2.3 Simulation Setup and Result
The 3D interconnect structure used for simulation is a copper wire segment with
a via below the wire. The wire geometry is specified by its 5 m width, 100 m
length and 1.5 m height; the via is of 2.5 m size and is 1.5 m off the wire edge;
the initial temperature is set to 300 K; the activation energy is set to 0.7 eV for the
copper body and is set to 1.25 eV along the surface to mimic the behavior of the
interface between the copper body and the cap layer. The current is flowing from
the via bottom up to the wire, then to the other end of the wire. A 3D view is shown
in Fig. 8.7.
We build an EM physical simulator using ANSYS [29] and a CCC program.
ANSYS is a multi-physics FEM-based tool, which provides AFD distribution for
specified physical parameters including temperature, current density and stress. The
whole simulation time is divided into many small time steps 4t. During each time
step, ANSYS simulates the wire structure and CCC program calculates the AFD
using (8.12), then the AFD is fed back into ANSYS for simulation of the next time
step. This ANSYS/CCC loop continues until an EM failure state is reached (e.g.
10 % resistance increase).
Figure 8.8 shows the AFD contour map of the interconnect structure. The wire
end that connects to the via below has large AFD, while small AFD is found at the
other end. The result matches well with the theoretical analysis and experimental
observations from real EM tests in literature [30, 31].
3 Variation-Aware Compact MTTF Model
3.1 Variation Sources

3.1.1 CMP Dishing
Chemical Mechanical Polishing/Planarization (CMP) is a process of smoothing

surfaces with the combination of chemical and mechanical forces [32]. Copper and
Fig. 8.8 Physical simulation

result of a wire segment
the adjacent dielectric are removed from the wafer at different rates during CMP,
creating surface anomalies and a varying topography [33]. Many factors, including
pattern geometry (e.g., line density), affect the material removal rates. The CMP
process strives to achieve flat topography to improve yield. Dishing is a copper
surface anomaly caused by CMP. Dishing occurs when copper recedes below the
level of the adjacent dielectric. Although dummy metal fills are added in order to
achieve a flat topography, dishing still occurs (Fig. 8.9).
3.1.2 EPE
As technology continues to scale down, lithography no longer produces ideal

geometric shapes. The most typical error is the Edge Placement Error (EPE) [34].
EPEs outside the feature are considered positive errors (bumping) and EPEs inside
the feature are negative errors (necking) as shown in Fig. 8.10. Although techniques
such optical proximity correction (OPC) are applied to reduce EPE, these variations
still exist.
dishing
Fig. 8.9 CMP dished copper

wires
156 D. Li et al.
3.2 Compact Model

3.2.1 Observations
Figure 8.11 shows an example of AFD simulation result by ANSYS of a CMP

dished wire. A maximum AFD occurs at the cathode and a minimum AFD at the
anode end of the wire.
We make several observations based on the simulations with CMP/EPE varia-
tions:
(1) A bumping on a wire does not relax the overall EM stress, whereas necking
reduces the EM lifetime;
(2) On a wire, there exists a critical region where maximum AFD occurs. Necking
in this region affects EM significantly; necking elsewhere affects EM much less.
The position of a critical region depends on the wire length, current density, and
current direction;
(3) When CMP dishing affects the critical region, the resulting EM lifetime is
proportional to the square of the wire height;
(4) When necking affects the critical region, the resulting EM lifetime is a function
of the necking size and location.
These observations inspire us to develop a compact model that captures depen-
dencies between wire geometry imperfections and EM lifetime. We run simulations
with various configurations, and develop a reasonably accurate model within a
practical range of geometric variations.
necking
bumping
Fig. 8.10 EPE affected wires
Fig. 8.11 EM physical

simulation of a CMP dished
wire
distance (um) left bound j=40mA/mm2

right bound
30
25
20
15
10
10 20 30 40 50
wire length (um)
Fig. 8.12 Critical region boundaries
3.2.2 Critical Region
We define the critical region of a wire as a region where the AFD is greater than
10 % of AFDmax (the maximum value of AFD along the entire wire). The critical
region boundary position is found to be a piecewise function of wire length and
current density as shown in (8.13), where k1 D 0.85, k2 D 1.65e-12, CR D 0.85e6,
CL D 1.02e16 and CC D 0.0915 are constants, l is the wire length and j is the current
density.
(
0 0l < CL
BL D k1 j
k1 l CL
l CL
( j k1 j
(8.13)
k2 .j C CC / l 0 l < CR
k1 j 2
BR D
k1 l Cj R2 l CR
k1 j 2
A sample curve for current density 40 mA/mm2 is shown in Fig. 8.12. In this figure,
the square- and circle-marked lines represent distances from the cathode via to the
right and left boundaries of the critical region. The double arrowed lines mark the
lengths of critical regions. It can be observed that for a fixed current density, the
length of critical region does not change.
To quantify the effects of process variation we introduce the variation coefficient.
Definition 1
Variation coefficient Cvar is the ratio of an ideal wire EM lifetime tlife0 and EM
lifetime tlife of the corresponding wire with geometric variations.
158 D. Li et al.
Cvar
simulated
1.0
fitted h0
Cvar=kCMP.(h'/h0) h’
0.8
0.6
0.4
0.2
h'/h0
0.2 0.4 0.6 0.8 1.0
Fig. 8.13 Compact model curve for Cvar due to CMP
tlif e 0
Cvar D (8.14)
tlif e
Unlike for process variation induced global effects on EM lifetime, when consider-
ing local effects, we assume that the amount of charge passing through a wire does
not change.
3.2.3 Compact Model
There are many factors that affect the EM lifetime of a wire with CMP/EPE
variations. However, instead of investigating those detailed physical effects, we
try to extract the major components that are most important to EM lifetime with
variations and develop a simplified model that can be easily applied to full-chip
interconnect EM analysis.
The ratio between an imperfect and ideal wire height (h’ /h0 ) is the dominant
factor for Cvar of CMP dishing. This dependency is given by (8.15), and shown in
Fig. 8.13, for kCMP D 1 and m D 2.
m
h0
Cvar D kCMP (8.15)
h0
To determine Cvar of EPE necking, we first consider the case where a single
circular dent is present on a wire. The dent can be completely described by its depth
d, length l and location x, and Cvar is a function of these three parameters captured
by (8.16), where b and w are the length of the critical region and the width of a wire;
Fig. 8.14 Compact model x

curve for Cvar due EPE l
d
BL BR
Cvar
12 simulated
d=0.03um, l=1um
10 fitted
2
X(um)
10 20 30 40
kEPE is a fitted parameter equal to 7.2e18. Figure 8.14 shows a sample plot for Cvar
of EPE necking with d D 0.03 m and l D 1 m.
8
< 1x BL or x BR
d l
Cvar D f ; ;x D 1 C kEPE bl dw .x BL / BL < x .BL C BR / =2
w b :
1 kEPE bl dw .x BR / .BL C BR / =2 < x < BR
(8.16)
When multiple dents are present on a wire, their cumulative effect is found to be the
product of Cvar caused by each dent, given by (8.17).
Y
Cvar D Cvar;i (8.17)
i
This variation-aware EM lifetime model provides a mapping methodology from an

assumed variation to the resulting EM lifetime. A more useful way to apply this
model is to estimate the variation tolerance of EM-sensitive wires. For example,
we may want to know how much CMP dishing or how much EPE is allowed for a
given wire. These are later referred to as variation criticalities, and will be given to
designers.
160 D. Li et al.
Vdd Vdd
Vdd Vdd
C C D D
C C D D
A A B B
A A B B
Natural
boundary
Fig. 8.15 Circuit-level power grid model
4 Full-Chip EM Analysis
4.1 Power Grid Model
Power grid of a VLSI chip carries unidirectional currents of large magnitudes and it
is the most EM-vulnerable part of the on-chip interconnect network. In this chapter,
we focus on power grid wires only.
In Fig. 8.15, we show the circuit-level model of power grid used in our work.
We assume that power is delivered through a controlled collapse chip connection
(C4) package. Wires and vias are modeled as resistors, external voltage supplies are
modeled as ideal voltage sources, and switching transistors are modeled as current
sources.
There are many different power grid EM analysis flows depending on data format
types used to describe the power grid. In this chapter, we use the SPICE format. The
link between the SPICE netlist naming and numbering scheme for the circuit and
the original geometry of the power grid is described below.
• Node name:
n<net-index>_<x-location>_<y-location>
• Data associated with each layer starts from:
* layer: <name>,<net>_net: <net-index>
Each layer/net combination is associated with a unique net-index.
• Vias starts from:
* vias from: <net-index> to <net-index>
Vias are implemented as resistors or as zero voltage sources.
• Current source:
150
VSS Contact
125
100
M2 VDD M2 VSS
75
M1 VDD
50
M1 VSS
25
VDD Contact
0
0 25 50 75 100 125 150
Fig. 8.16 Small power grid sample
iB<block-number> <node> 0 <value> and iB<block-number> 0 <node>

<value>iB
Each current source is split into two components: from VDD to ideal ground and
from ideal ground to VSS.
Figure 8.16 shows a sample small power grid. Reference [35] provides more
details of this model.
4.2 Effective jL Product Extraction
Blech length effect can be used to quickly filter out a great number of EM-immortal
wires which do not require further consideration. The remaining EM-mortal wires
will be processed by a more time-consuming EM lifetime calculation procedures.
The studies of Blech length effect usually consider simple straight point-to-point
wires, for which the values of j and L can be easily determined. For complex
interconnect topologies the accumulated jL product along the longest possible path
is usually taken. However, in [36], the authors show that it is non-trivial to extract
jL product in a complex interconnect topology, and they develop an effective jL
extraction strategy based on EM physics, as stated in (8.18). In (8.18), jk and Lk are
the current density and length of a wire segment. The sum is taken over all possible
paths in the network, with (jL)eff being the maximum of these sums. The effective
162 D. Li et al.
Path A Path B Path C

L1 L2 L3 L4
Power e-
track
Via1 j1 Via2 j2 Via3 j3 Via4 j4 Via5
Fig. 8.17 Effective jL extraction
jL product matches well with experimental results in [37].

!
X
.jL/eff D max jk Lk (8.18)
k
The effective jL product extraction (8.18) can be applied to any complex intercon-
nect structure. But instead of computing all paths in a power grid, the effective jL
can be obtained as follows. For any power wire segment, find the longest consistent
current path that contains it and compute the sum of individual jL products of all
wire segments along the path. For example, in Fig. 8.17, paths A, B and C are
longest consistent current paths in the given power track, so we have (8.19).
jL1;eff D jL2;eff D jL3;eff D jLpathA D j1 L1 C j2 L2 C j3 L3

jL4;eff D jL5;eff D jLpathB D j4 L4 C j5 L5 (8.19)
jL6;eff D jL7;eff D jLpathC D j6 L6 C j7 L7
4.3 Analytical Lifetime Calculation
The Black’s equation given by (8.2) has its limitations, since it ignores some
physical factors that can substantially affect EM lifetime such as critical stress for
void nucleation, material effective modulus and diffusivity. Many physics-based EM
lifetime calculations are developed in literature. We adopt the method presented in
[36], which obtains EM lifetime based on the void nucleation and growth model. It
also provides EM lifetime calculation for complex interconnect topologies, which
can be applied to a power grid.
Here, we briefly explain how to calculate EM lifetime of a power grid wire
segment. Detailed discussion can be found in [36]. In Fig. 8.17, arrows denote
electron flow in a wire segment. First, each wire segment EM failure is associated
with its cathode via. So we have (8.20).
tL1 D tvia1
tL2 D tvia2 (8.20)
tL3 D tL4 D tvia4
Table 8.2 Physical parameters

Parameters Values
¢ void 40 MPa
Z* 1
B 28 GPa
1.18 1029 m3
¡ 2.1 cm (metal 5, 6), 3.9 cm (metal 2, 3, 4), 4.8 cm (metal 1)
Next, EM lifetime of a via depends on the metal atomic flux divergences contributed
by all the connected wire segments. Due to the flux divergence, a void is nucleated
when threshold stress is reached. The time for a void nucleation comprises the first
part of (8.21). After a void is nucleated, it keeps growing until it reaches a critical
length such that EM failure occurs. The time for void growth comprises the second
part of (8.21).
0 X p 12
r r Di
B void kT i C
tlif e B
D@ X C C Lc kT X 1 (8.21)
eZ 4 B A eZ
D i ji D i ji
i i
In (8.21), void is the critical stress required to nucleate a void, ji and Di are the
current density and the effective diffusivity of each wire segment i connected to
a via, B is the effective modulus of the materials surrounding the metal, Lc is the
critical void length, and T is the temperature. In reality, if the actual chip temperature
profile is available, it should be used for EM lifetime calculation. However, here
we assume temperature is uniformly fixed to 300 K. The effective diffusivity D is
computed using equation (8.22):
Ea w
D D D0 e kT (8.22)
h
In (8.22), w is the wire width and h is its height. When all Di values are the same,
which is typical in a power grid, the lifetime of a wire segment and thus the lifetime
of its cathode via is inversely proportional to the current density through the via.
The parameter values we used are listed in Table 8.2.
Wires are classified into three subgroups based on their analytical lifetime: EM-
safe, EM-sensitive and EM-weak. The classification criteria are listed below.
EM-safe: tlife > (1 C ˇ) treq ;
EM-weak: tlife < treq ;
EM-sensitive: treq tlife (1 C ˇ)treq ,
where treq is the required lifetime, assumed here to be 10 years. ˇ is the lifetime slack
threshold; its value is selected according to the maximum lifetime reduction found
164 D. Li et al.
by the variation-aware analysis. The wires with lifetime tlife > (1 C ˇ) treq are EM-
safe and should not fail even if they experience the worst process variation effects.
Wires with treq tlife (1 C ˇ)treq are considered EM-sensitive and their status is
determined by the variation-aware analysis using the compact model. A user can
either use a pre-estimated upper bound of ˇ, or run several iterations of our flow
to tune the value of ˇ. For example, if initially the value of ˇ is set to 10 %, and
later variation-aware analysis reports a 12 % maximum lifetime reduction, then the
user would adjust ˇ value to a number greater than 12 %. In practice, ˇ is within a
range of 10–15 %, thus the convergence of tuning ˇ is not considered a problem in
our work. All EM-sensitive wires will then be processed by the variation-aware EM
analysis.
4.4 Global Current Redistribution
CMP/EPE variations not only locally affect EM lifetime, but also cause global
power grid current redistribution from full-chip analysis point of view. However,
we need to know the exact CMP/EPE variations on the power grid to study those
redistribution effects.
For experimental purpose, to create the environment for studying global current
distribution changes due to resistance change, we apply Monte-Carlo simulation.
Monte-Carlo simulation samples are generated by first randomly throwing some
darts on the P/G network. Each such selected resistor, referred to as c-resistor, is a
variation center. Its own resistance and those of a few wires around it are changed.
Each c-resistor randomly picks a value according to a Gaussian distribution
N(, 2 ), where is mean and 2 is variance. This variation then propagates to
its neighbors with decreasing magnitude as the distance to the c-resistor increases.
The assumption is that wires in close layout proximity have similar imperfections
due to similar process conditions. Figure 8.18 gives an example.
In Fig. 8.18, the nominal value of each horizontal resistor is 5 and of vertical
resistor is 4. A is an EM-sensitive wire. B and C are c-resistors in the perturbed
network. The perturbed resistance values are shown in Fig. 8.18b. An EM-sensitive
wire segment may also become a c-resistor.
Currents of the EM-sensitive wires have to be computed in the perturbed
grid. Solving Kirchhoff’s Current Law (KCL) or Kirchhoff’s Voltage Law (KVL)
equations for the whole system each time a perturbed network is generated is
infeasible due to a huge computational cost. Here, knowing the node voltages in the
nominal network, we apply random walk method [38] to compute current flows in
its perturbed variant. Random walk is a technique that allows localized computation;
its principle is briefly described below using Fig. 8.19.
A random walk starts from any node x, and then travels to some adjacent
node according to a certain probability as in (8.23) which depends on electrical
conductance between the two nodes. In (8.23), Vx is the voltage of the node x,
degree(x) is the number of nodes adjacent to x, gi is the electrical conductance
between node i and x, Ix is the current load connected to x (Ix D 0 if no load is

connected to x). The travelling continues until a boundary node (a ground node or
an ideal voltage source node) is reached. Multiple travels are executed from a given
node, and then the average value is taken as its node voltage. So the voltages of
those nodes of interest (nodes of EM-sensitive wires) in a perturbed network can be
obtained without solving the whole system.
X
degree.x/
gi Ix
Vx D Vi (8.23)
iD1 X
degree.x/
X
degree.x/
gj gj
j D1 j D1
Lifetimes of all EM-sensitive wires are then recalculated for each perturbed
network. Two distributions are obtained: lifetime distribution of each EM-sensitive
wire and cumulative distribution function (CDF) of the full-chip EM reliability.
4.5 VEMA Flow
The variation-aware EM analysis tool VEMA includes four major steps: fast jL filter,
analytical EM lifetime calculation, global current redistribution and local variation
effect analysis. Figure 8.20 shows a flowchart for our variation-aware EM analysis
tool VEMA.
Fig. 8.18 P/G network. (a) Original (b) perturbed

166 D. Li et al.
Fig. 8.19 Random walk 2

from node x
g2
g1
1 3
x
g3
g4
Ix
5 Experimental Results
We implemented our tool in CCC and tested it with several publicly released IBM
P/G benchmarks [35]. Experiments were carried on GUN/LINUX server with Intel
Xeon E5440 2.83 GHz CPU and 16 GB memory. The specification required EM
lifetime (referred to as treq ) was set to 10 years and the temperature was assumed to
be 300 K for all experiments.
In Fig. 8.21, we compare jL filter results from the accumulated and effective
jL product extraction for metal three layer in IBM PG1 benchmark. An obvious
reduction in EM-mortal wires can be observed in effective jL product distribution.
This is because the accumulated jL extraction uses the longest possible path for an
entire power track, thus a large number of wires are unnecessarily classified as EM-
mortal. This imposes an onerous burden on further EM analysis and significantly
diminishes the efficiency of fast jL filter.
Moreover, there is a risk of missing EM-mortal wires using the accumulated
jL product. For example, in Fig. 8.22, two currents with the same magnitude but
opposite directions are flowing from two ends of a power track and meet at the
middle point. The accumulated jL extraction will classify this wire as EM-immortal,
because the accumulated jL product would be zero, whereas the effective jL product
extraction suggests computing jLs for the two consistent paths separately. The
quantitative results of jL filter are listed in Table 8.3.
When considering global effects of process variations, we assume that 10 % of
wires experience variations, with each c-resistor Ri following N(Ri , 0.001Ri 2 ) distri-
bution. We simulated the global effects of process variation on a randomly selected
EM-sensitive wire, named W2 , taken from IBM PG2 benchmark. Figure 8.23 shows
the lifetime histogram with a trend line for W2 . The solid vertical line denotes tsp and
the dotted line marks MTTF for W2 . It can be observed that with probability 0.84,
W2 is safe from EM failure taking into account global effects of process variations;
Psafe,w2 D 0.84.
Table 8.3 Experimental results
Mortal Worst MTTF Avg. prob. no Avg. CMP Avg. EPE
Benchmarks Total #wires A E Miss (year) #EM-sensitive failure (%) tolerance (%) tolerance (%)
IBM PG1 30027 25196 10032 623 4.07 46 32.7 15.74 8.50
IBM PG2 208325 111245 23124 77 4.94 63 89.7 16.24 15.41
IBM PG3 1401572 33872 33309 546 4.90 917 82.0 14.37 12.04
IBM PG4 1560645 198103 66226 0 5.53 554 74.6 12.53 11.06
8 Layout Aware Electromigration Analysis of Power/Ground Networks
167
168 D. Li et al.
Begin
Read in P/G
Network (res, via)
N
jL>jLBlech? EM-immortal
Y
Initialize β N
tlife>(1+β)*treq? EM-weak
Apply variation-aware analysis

Adjust β
N
β>Cvarmax-1?
Y
Output EM
reliability report
End
Fig. 8.20 VEMA flowchart
We determine an overall EM failure cumulative distribution function (CDF)

using (8.24) considering global effects from process variation. In (8.24) PCDF (q)
is the probability of no more than q EM-sensitive wire failures, p is the total number
of EM-sensitive wires, Psafe,i is the EM-safe probability of ith wire. EM-sensitive
wires are sorted according to the order of their increasing EM-safe probability.
Y
pq
PCDF .q/ D Psaf e;i (8.24)
iD0
Figure 8.24 shows the CDF curve for IBM PG2 benchmark. Point A on the CDF
curve corresponds to probability 0.95 of no more than eight EM-sensitive wire
failures, and the gray-shaded region beneath point A is highly sloped, elsewhere
it is gently sloped. This can be interpreted that strengthening the worst eight EM-
sensitive wires can boost EM-safe probability to 0.95; but beyond that due to gently
sloped region, the benefit of strengthening more EM-sensitive wires will be very
small.
Using the variation-aware compact EM lifetime model, we perform EM analysis
considering CMP/EPE variations. In Fig. 8.25, we show the EM lifetime distribution
for a randomly selected EM-sensitive wire W4 from IBM PG4 benchmark. For W4 ,
we assume 10 % EPE-caused area loss within the critical region. Local effects
caused by CMP and EPE are similar and manifest themselves as a shift of EM
lifetime distribution which reduces the ideal EM lifetime MTTF0 to MTTF’ and
causes a higher EM failure probability. The maximum achievable distance between
MTTF0 and treq indicates how critical EPE effect on MTTF is.
Fig. 8.21 EM-mortal map for IBM PG1. (a) Accumulated mortal map (b) effective mortal map
J J
Path A Path B
L L
Accumulated jL: jL+(-j)L=0
Modified jL: jLA=jL
jLB=jL
Fig. 8.22 jL extraction comparison

170 D. Li et al.
Fig. 8.23 EM lifetime

distribution for W2 with
global effect
Fig. 8.24 CDF of 1

Probability (PCDF)
EM-sensitive wires for

IBM PG2 0.8
A (8, 0.95)
0.6
0.4
0.2
0
0 10 20 30 40
No. of EM-sensitive wires (q)
Fig. 8.25 EM lifetime distribution shifted by EPE
Definition 2
CMP criticality is defined as a ratio of the original wire height over the maximum
wire height loss that a wire can tolerate while keeping tlife > treq . EPE criticality is
defined as the ratio of original total critical region area over the maximum area loss
that a wire can tolerate while maintaining its lifetime greater than the required limit,
tlife > treq .
Fig. 8.26 Variation criticality map for IBM PG4. (a) CMP criticality (b) EPE criticality
CMP _cri t i cali ty D h0

max.h0 h’/
(8.25)
EPE_cri t i cali ty D Ac
max.Aloss /
In Fig. 8.26, we report CMP/EPE criticality map for metal five layer in IBM
PG4 benchmark. For high criticality EM-sensitive wires process variation induced
printing imperfections should be minimized by techniques such as better OPC or
more careful insertion of dummy metal fill.
All these results are quantitatively summarized in Table 8.3. In Table 8.3, the
sub column labeled A under column Mortal provides the accumulated jL extraction
values, the sub column labeled E refers to effective jL extraction and the sub column
labeled Miss gives the number of EM-mortal wires missed by the accumulated jL
extraction. The column labeled Avg. prob. no fail is the average probability that
EM-sensitive wires will not fail.
Conclusions
As technology scales down, EM is becoming progressively a more serious
problem. It posts huge challenge on analysis tools and limits the application
of conventional EM theories such as Black’s equation and Blech length effect.
Moreover, the process variation brings in more uncertainty to EM analysis.
In this chapter, we introduced the basic background knowledge of EM, and
described the tool VEMA, which is a post-design EM analysis and reliability
evaluator. Fast jL filter and ideal EM lifetime calculation are performed to
quickly filter out EM-immortal wires and classify all EM-mortal wires into
three subgroups: EM-weak, EM-safe and EM-sensitive. Variation-aware EM
analysis is applied on EM-sensitive wires. The variation effects mainly refer
(continued)
172 D. Li et al.
to the changes of wire geometry caused by CMP and EPE that affect the EM
lifetime of individual wires. A compact variation-aware EM lifetime model
is developed. Using this model, variation tolerance of each EM-sensitive wire
can be fed back to designers. As technology is scaling down, process variation
on wires becomes increasingly more significant. Our work shows that its
effect on EM reliability is non-negligible, and a variation-aware EM analysis
such as VEMA may provide a more realistic assessment of the EM reliability
of power grids.
Acknowledgments Research support was provided in part by SRC through IBM Corporation and
in part by NSF grant CCF-1115663.
Many people have contributed at various stages to this book chapter. In particular we would
like to thank our collaborator from IBM, Dr. Baozhen Li for motivating this work and for support.
We also thank Xiang Qiu, Zhong Guan and Ali Abbasinasab from the University of California,
Santa Barbara for their helpful discussions.
There also have been many discussions with professors at the University of California, Santa
Barbara. In particular, we would like to thank Professor Tim Cheng and Professor Li-C Wang.
References
1. http://en.wikipedia.org/wiki/Electromigration.
2. J. Lienig, “Invited Talk: Introduction to Electromigration-Aware Physical Design,” Interna-
tional Symposium on Physical Design, 2006, pp. 39-46.
3. F. Wei, C. L. Gan, T. L. Tan, et.al, “Electromigration-induced extrusion failures in Cu/low-k
interconnects,” Journal of Applied Physics, 104, 023529 (2008).
4. R. S. Sorbello, “Theory of the Direct Force in Electromigration,” Phys. Rev. B, vol. 31, no. 2,
pp. 798-804, 1985.
5. J. R. Black, “Electromigration - A Brief Survey and Some Recent Results,” IEEE Trans. on
Electron Devices, Vol. ED-16 (No. 4), pp. 338-347, April 1969.
6. I. A. Blech, “Electromigration in thin aluminum films on titanium nitride,” J. Appl. Phys.,
vol. 47 (1976), pp. 1203–1208.
7. I. A. Blech and C. Herring, “Stress Generation by Electromigration,” Appl. Phys. Lett., vol. 29,
no. 3, pp. 131-133, 1976.
8. I. A. Blech and K. L. Tai, “Measurement of Stress Gradients Generated by Electromigration,”
Appl. Phys. Lett, vol. 30, no. 8, pp. 387-389, 1977.
9. C. Herring, “Diffusional Viscosity of a Polycrystalline Solid,” J. Appl. Phys., vol. 21,
pp. 437-445, 1950.
10. http://www.iue.tuwien.ac.at/phd/orio/node26.html.
11. E. T. Ogawa, A. J. Bierwag, K.-D. Lee, H. Matsuhashi, P. R. Justinson, and et al., “Direct
Observation of a Critical Length Effect in Dual-Damascene Cu/Oxide Interconnects,” Appl.
Phys. Lett., vol. 78, no. 18, pp. 2652-2645, 2001.
12. D. Ney, X. Federspiel, V. Girault, O. Thomas, and P. Gergaud, “Stress-Induced Electromi-
gration Backflow Effect in Copper Interconnects,” Trans. Dev. Mater. Reliab., vol. 6, no. 2,
pp. 175-180, 2006.
13. L. Doyen, E. Petitprez, P. Waltz, X. Federspiel, L. Arnaud, and Y. Wouters, “Extensive
Analysis of Resistance Evolution due to Electromigration Induced Degradation,” J. Appl.
Phys., vol. 104, p. 123521, 2008.
14. A. S. Oates and M. H. Lin, “Void Nucleation and Growth Contributions to the Critical Current
Density for Failure in Cu Vias,” Proc. Intl. Reliability Physics Symp., pp. 452-456, 2009.
15. J. W. McPherson and P. B. Ghate, “A methodology for the calculation of continuous dc electro-
migration equivalents from transient current waveforms,” in Proc. Symp. on Electromigration
of Metals, New Orleans, LA, pp. 64-74, Oct. 7-12, 1984.
16. C. K. Hu, R. Rosenberg, H. S. Rathore, et.al, “Scaling Effect on Electromigration in On-Chip
Cu Wiring,” International Conference on Interconnect Technology, 1999, pp. 267-269.
17. N. Srivastava, K. Banerjee, K. E. Goodson, “Scaling Analysis of Multilevel Interconnect
Temperatures for High-Performance ICs,” IEEE Trans. on Electron Devices, volume 52, issue
12, 2005, pp. 2710-2719.
18. B. Li, C. Christiansen, C. Burke and et al., “Short Line Electromigration Characteristics and
their Applications for Circuit Design,” International Reliability Physics Symposium, 2013,
pp 3.F.2.1-3.F.2.5.
19. M. Lin, N. Jou, W. Liang and K. C. Su, “Effect of Multiple Via Layout on Electromigration Per-
formance and Current Density Distribution in Copper Interconnect,” International Reliability
Physics Symposium, pp. 844-847, 2009.
20. N. Raghavan and C. M. Tan, “Statistical Modeling of Via Redundancy Effects on Interconnect
Reliability,” International Symposium on the Physical and Failure Analysis of Integrated
Circuits, pp. 1-5, 2008.
21. D. Li, Z. Guan, M. Marek-Sadowska and S. R. Nassif, “Multi-Via Electromigration Lifetime
Model,” International Conference on Simulation of Semiconductor Process and Devices,
pp. 308-311, 2012.
22. http://en.wikipedia.org/wiki/Activation_energy.
23. http://en.wikipedia.org/wiki/Copper_interconnect.
24. R. L. D. Orio, “Electromigration Modeling and Simulation,” doctoral dissertation, Institute for
Microelectronics, TU Wien, 2010.
25. D. Li, M. Marek-Sadowska and S. R. Nassif, “A Method for Improving Power Grid Resilience
to Electromigration-Caused Via Failures,” to appear in IEEE Trans. on VLSI Systems.
26. D. Dalleau, “3-D Time-depending Simulation of Void Formation in Metallization Structures”,
Doctoral thesis, University of Hannover, 2003.
27. W. Li, C. M. Tan and N. Raghavan, “Dynamic simulation of void nucleation during electro-
migration in narrow integrated circuit interconnects,” Journal of Applied Physics 105, 014305
(2009).
28. http://www.ansys.com.
29. M. Lin, N. Jou, W. Liang and K. C. Su. Effect of Multiple Via Layout on Electromigration
Performance and Current Density Distribution in Copper Interconnect. IEEE 47th Annual
International Reliability Physics Symposium, Montreal, 2009, pp. 844-847.
30. B. Li, J. Gill, C. J. Christiansen, et. al. Impact of Via-Line Contact on Cu Interconnect Electro-
migration Performance. IEEE 43rd Annual International Reliability Physics Symposium, San
Jose, 2005, pp. 24-30.
31. http://en.wikipedia.org/wiki/Chemical-mechanical_planarization.
32. F. M. Serry, D. Dawson, “Minimizing Dishing and Erosion in Copper CMP,” http://www.veeco.
com/pdfs/database_pdfs/minimizing_de_in_copper_cmp_45.pdf.
33. https://www.si2.org/openeda.si2.org/dfmcdictionary/index.php/Edge_Placement_Error.
34. S. R. Nassif, “Power grid analysis benchmarks,” Asia and South Pacific Design Automation
Conference, 2008, pp. 376-381.
35. S. P. Hau-Riege, “New Methodologies for Interconnect Reliability Assessments of Integrated
Circuits,” Doctoral thesis, Massachusetts Institute of Technology, 2000.
36. C. W. Chang, Z. –S. Choi, C. V. Thompson, et.al, “Electromiration resistance in a short three-
contact interconnect tree,” Journal of Applied Physics 99, 094505 (2006).
37. H. Qian, S. R. Nassif, and S. S. Sapatnekar, “Power Grid Analysis Using Random Walks,”
IEEE Trans. on CAD, vol. 24, no. 8, August 2005, pp. 1204-1224.
38. S. M. Alam, C. L. Gan, C. V. Thompson, et al, “Reliability computer-aided design tool for
full-chip electromigration analysis and comparison with different interconnect metallizations,”
Microelectronics Journal 38 (2007) 463-473.
Chapter 9
Power-Gating for Leakage Control and Beyond
Andrea Calimera, Alberto Macii, Enrico Macii, and Massimo Poncino
Abstract The need of reliable nanometric integrated circuits is driving the EDA
community to develop new automated design techniques in which power consump-
tion and variability are central objectives of the optimization flow.
Although several Design-for-Low-Power and Design-for-Variability options are
already available in modern EDA suites, the contrasting nature of the two metrics
makes their integration extremely challenging. Most of the approaches used to
compensate and/or mitigate circuit variability (e.g., Dynamic Voltage Scaling and
Adaptive Body Biasing) are, in fact, intrinsically power inefficient, as they exploit
the concept of redundancy, which is known to originate power overhead.
In this work, we introduce possible solutions for concurrent leakage minimiza-
tion and variability compensation. More specifically, we propose Power-Gating as
a mean for simultaneously controlling static power consumption and mitigating the
effects induced by two of the most insidious sources of variability, namely, Process
Variations (PV) due to uncertainties in the manufacturing and Transistor Aging due
to Negative Bias Temperature Instability (NBTI).
We show that power-gating, when implemented through the insertion of dedi-
cated switches (called sleep transistors), has a double effect: On one hand, when
sleep transistors are enhanced with tunable features, it acts as a natural supply-
voltage regulator, which implements a control knob for PV compensation; on the
other hand, during the idle periods, it makes the circuits immune to NBTI-induced
aging.
We describe optimization techniques for the integration of a new concept of
power-gating into modern sub-45 nm design flows, that is, Variation-Aware Power-
Gating. The experimental results we have obtained are extremely promising, since
they show 100 % timing yield under the presence of PV and circuit lifetime
extension of more than 5 in the presence of NBTI.
A. Calimera • A. Macii • E. Macii () • M. Poncino

Dip. di Automatica e Informatica, Politecnico di Torino, 10129 Torino, Italy
e-mail: enrico.macii@polito.it

DOI 10.1007/978-1-4614-4078-9__9
176 A. Calimera et al.
1 Introduction
MOS devices are approaching the size of atoms, which is a fundamental barrier for
the scaling process of the bulk technology. The research community is thus actively
pursuing alternative materials, fabrication processes, devices and architectures
to be adopted in mainstream circuit and system manufacturing. While several
new solutions have shown good potential (e.g., carbon nanotubes and graphene,
memristors, spin-based devices, ferromagnetic logic, atomic switches, NEMS), the
debate of which of them will prevail is still open; therefore, nanometer CMOS will
remain the dominant technology on the electronics market for a few years.
The 2009 ITRS Roadmap [17] reported that static power consumption and
variability are the most serious concerns for the design of nanometer ICs below
45 nm. Static power due to internal leakage mechanisms represents the main source
of power consumption in modern CMOS circuits [28], which have shown to be
power-hungry even when not switching. Variability, instead, refers to the marked
tendency of a manufactured circuit to show a deviation from its nominal behavior.
Main sources of variability include random and systematic process variations [4],
environmental condition variations [33] (e.g., temperature and Vdd fluctuations)
and aging effects (e.g., Negative Bias Temperature Instability and Hot Carrier
Injection) [1]. While static power translates to low energy efficiency, variability
originates lower reliability and lower fabrication yield; both factors make electronic
circuits less reliable.
Static power and variability are not new in the EDA field, and various options for
their management are already available. However, while in the past considering the
two as independent variables in the design space was accurate enough to obtain rea-
sonable results, advanced technology nodes are showing the need of holistic design
strategies that could provide concurrent static power optimization and variability
compensation. Unfortunately, this challenge is complicated by the fact that most of
the design solutions for compensating variability are intrinsically power inefficient:
Fault-tolerance approaches, such as “fail and correct”, or adaptive strategies, such as
“dynamic voltage regulation” or “forward body biasing”, are based on the concept
of redundancy, which is in contrast with low-power requirements.
In this work, we address this critical issue and we propose the use of
power-gating [2] as the enabling technology for achieving simultaneous leakage
optimization and variability compensation.
Power-gating (PG) is based on the insertion of a dedicated MOS switch, called
sleep transistor, between the gated block and the actual ground rail. This provides
the circuit with two power modes: A low-power mode, during which the sleep
transistor is turned-off and the leakage power is reduced by more than one order
of magnitude, and an active mode, during which the sleep transistor, which is
turned-on, guarantees a normal connection to the global power-rails. What is
interesting to note is that, in terms of variability, the sleep transistor shows important
and unique properties, that is: During the active periods, when enhanced with
tunable-width features, it can act as a natural supply-voltage regulator usable to
9 Power-Gating for Leakage Control and Beyond 177
implement adaptive control schemes for process-variation mitigation [14]; during

the low-power mode, its effect is to make the gated circuit immune by the aging
mechanisms (NBTI and HCI) [9].
Based on these observations, we claim that PG can represent a viable solution for
the implementation of low-power, aging-free reliable ICs. Obviously, the beneficial
effects that PG can provide must be weighted against the amount of overhead it
introduces. A direct Sleep Transistor Insertion (STI) methodology, in fact, may
originate excessive timing, area and power overhead, which can off-set the beneficial
effects that PG offers in terms of aging and variability compensation.
To overcome this drawback, we propose new optimization techniques based on
the concept of Variation-Aware Clustered PG, which consists of a methodology
for clustering and power-gating critical cells only, that is, to apply variation-aware
PG only to cells whose process variation-induced and/or aging-induced variations
have a direct impact on the overall performance of the circuit. To enable such a
strategy, we opted for an STI flow in which sleep transistors are inserted in layouts
with row granularity. This allows a finer control of variability/aging compensation,
while reducing the design overhead. Experimental results performed on a set of
benchmark circuits and mapped to an industrial 45 nm CMOS library prove the
effectiveness of the proposed solutions, as well as their integrability into industrial
design flows.
The remainder of the chapter is organized as follows. In Sect. 2, we discuss the
main challenges to be faced during nanometric IC design; we introduce models for
sub-threshold leakage power consumption, also describing the sources of variability
in scaled nanometric technologies. Section 3 addresses the key design issues related
to power-gating, with particular emphasis on automated solutions for physical sleep-
transistor insertion. In the last two sections, we provide detailed background and
models for process variation and NBTI effects and we present automated power-
gating strategies for process variation compensation (Sect. 4) and NBTI mitigation
(Sect. 5). Finally, section “Conclusions” gives some concluding remarks.
2 Design Issues for Nanometric CMOS Circuits
2.1 Sub-threshold Leakage Power Consumption
Among all the leakage current mechanisms induced by Short Channel Effects
(SCEs), the sub-threshold current Isubth has proven to be the major contributor
to the total static power consumption [28].
Isubth is defined as the drain-to-source current which flows when the transistor
operates in the weak inversion region, i.e., when the gate voltage Vg is below the
threshold voltage Vth . Under this condition, the channel shows a small, but non-
zero concentration of minority carriers that are diffused from the drain to the source
terminal whenever a potential greater than 0 is applied between drain and source,
i.e., Vds > 0.
A well-known model for the Isubth of a single nMOS transistor is given by the
following equation [28]:
W Vg Vth Vds
Isubth D Cox .m 1/v2T e mvT .1 e vT / (9.1)
L
with
Cd m
mD1C (9.2)
COX
where vT D K T =q is the thermal voltage, COX is the gate oxide capacitance; is

the carrier mobility; m is the sub-threshold swing coefficient, with Cd m representing
the capacitance of the depletion layer.
The magnitude of the sub-threshold current is a function of several parameters,
such as the operating temperature (Isubth increases as T increases), the supply
voltage (Isubth increases for larger Vg ), and the device size (Isubth increases as
the transistor gets larger and shorter). However, the parameter that affects most
(i.e., exponentially) Isubth is the threshold voltage Vth : Decreasing the Vth by
100 mV increases the leakage current by a factor of 10. That is why scaled CMOS
technologies (characterized by ever smaller Vth ) suffer from sensible sub-threshold
leakage.
2.2 Sources of Variability
With variability we commonly refer to the marked tendency of a manufactured

CMOS circuit to show a deviation from its nominal behavior. The sources of varia-
tion can be broadly classified according to their nature (statistical vs. deterministic),
their spatial reach (local vs. global), and their temporal rate of change (static vs.
dynamic). Figure 9.1 summarizes the typical variations arising in nano-scale CMOS
circuits and systems.
Under the label statistical it is possible to include all those variations that are
induced by stochastic events; they differ from deterministic variations, which can
be somehow predicted at design time. Global variations affect all the transistors
on the die, while local variations are limited to a few transistors in the immediate
vicinity of each other. Finally, the classification between static and dynamic depends
on the actual rate of change with time. Static variations, e.g., process variations,
remain effectively invariant over the entire lifetime of the manufactured chips,
while dynamic variations change over the lifetime of the chips. The changes can
manifest on a large time-sale (that is the case of slow-variations like aging effects:
NBTI, HCI and TDDB) or in a short time-scale (fast-variations like IR-drop, clock
jitter, coupling noise, temperature and Vdd variations). Although all these sources of
variability have deleterious effects on the reliability of CMOS digital circuits, two of
them have been recognized as particularly critical, thus worth specific consideration:
Process variations and aging.
Fig. 9.1 Types of variability in CMOS-based circuits and systems
Process variations (PV) [4, 5] are mostly due to random fluctuations of dopant
atoms, which result in the mismatching of the electrical characteristics of transistors
in the same die, and to systematic or non-systematic impreciseness of the manu-
facturing process (like lithography, etching, and chemical-mechanical polishing).
Process variations have a significant impact both on the power dissipation and
performance of a design: 20 variation in leakage power for a 1.5 variation in
delay between fast and slow dies has been reported in the literature. Given the
power/performance tradeoff, those figures translate into an increase of dies that
must be discarded because either too slow or too power consuming. Therefore,
as a relevant side effect, increased variability decreases yield, with important cost
implications.
Aging, instead, includes all those wear-out mechanisms that induce time-
dependent degradation of the operating characteristics of devices [1]. Two are the
main sources of aging in active devices: Bias Temperature Instability (BTI), and
Hot Carrier Interface (HCI) [26]. Both these physical/chemical effects result in the
generation of interface traps at the silicon/oxide interface and cause a drift of the
threshold voltage over time. These irreversible effects, and BTI in particular, have
traditionally been regarded as “reliability issues”, and have only recently received
some consideration in the CAD community as a factor affecting performance
of digital circuits. Traditional VLSI design, in fact, bypasses the analysis and
optimization of such dynamic networks by approximating the problem to the
optimization of uniform static networks with certain guard-band. This may induce
unacceptable design overheads.
3 Power-Gating for Leakage Power Reduction
3.1 Power-Gating Basics
Power-gating has proven to be a very effective approach to reduce standby leakage,

while keeping high speed in the active mode. It is based on the principle of adding
devices, called sleep transistors, in series with the pull-up and/or the pull-down of
logic gates, and turning them off when the circuit is idle, thereby decreasing the
leakage power component due to IDS sub-threshold currents. When a nMOS sleep
transistor is used on the pull-down path, a SLEEP signal controls its active/standby
mode (i.e., SLEEP = 1 during standby and SLEEP = 0 during active mode). In the
standby mode, the sleep transistor is off, thus disconnecting its insertion point, called
virtual ground, from the physical ground. In active mode the gated circuit operates
normally, but it incurs a delay degradation due to the series resistance of the sleep
transistor. Figure 9.2 shows a logic block with a nMOS sleep transistor connected.
The source terminals of the logic gates in the logic block are connected to
the virtual ground which is, in turn, connected to the drain terminal of the sleep
transistor.
Effective use of power-gating requires a proper sizing of the sleep transistor, since
that affects the performance of all the gates connected to it. While a small transistor
unacceptably slows down the circuit in active mode due to its high resistance, a large
one implies a significant overhead in area and a non-negligible energy (i.e., power
and delay) for ON$OFF transitions. One additional difficulty is that sleep transistor
sizing is determined by the maximum current injected by the circuit, which leads to
maximum drop across the sleep transistor drain-source path and, as a consequence,
causes worst-case delay degradation during active operation.
VDD
Logic Block
VGND
SLEEP nMOS
Fig. 9.2 A logic block with

nMOS sleep transistor GND
Several power-gating styles have been proposed, differing in the granularity

of the blocks to which sleep transistors are applied. Granularity may range from
individual cells (the fine-grained sleep transistor insertion approach [21]) to large
chip sub-units, (the block-level power-gating scheme), in which very large sleep
transistors are placed on the root of the power distribution networks of large chip
areas. While the fine-grained approach suffers from high area overhead and an
excessive buffering of the sleep signal due to very high capacitance load to be
driven by this signal, block-level power-gating has the disadvantage of having
long transition delays between sleep and active state, caused by the large RC time
constant of the sub-unit’s power distribution network.
Moreover, and this is the most important aspect, its coarse granularity reduces
the degrees of freedom available to the designer. If the decision of power-gating a
block is taken, then all gates in the block are gated. This may not be desirable if, for
instance, there is a critical subset of gates for which parametric variations (typically
due to fabrication process or aging mechanisms) induce speed degradation of the
whole circuit. In this case, a finer control of the sleep transistor insertion could
guarantee a perfect match between leakage savings and design overhead, and, as
we demonstrate afterwards, it also allows to exploit the beneficial effects of power-
gating in terms of variability only where necessary, i.e., only for those subsets of
gates that are more critical.
We refer to this finer strategy as the clustered sleep transistor insertion, a very
effective solution where multiple subsets of cells (i.e., the clusters) are connected
to dedicated sleep transistors distributed across the layout. It is worth mentioning
that clustered power-gating is not new in the low-power design domain, and many
clustering solutions have been proposed. In [20], the authors propose a solution
in which gates having mutually exclusive current discharge patterns are grouped
together; the resulting clusters allow optimal sleep transistor sizes. Instead, in [31],
the authors show an effective timing-driven clustering strategy that is able to
handle simultaneously timing and area constraints. Differently from previous works,
in Sects. 4 and 5 we propose variation-aware clustering methods in which gates
are grouped based on variation-induced timing criticality. Independently of the
clustering algorithm, physical design details as well as constraints posed by the
adoption of industrial EDA frameworks need to be considered while developing a
strategy for sleep transistor insertion.
Using layout rows as atomic clustering objects greatly simplifies the physical-
level management of virtual ground distribution, which is the major practical
obstacle in clustering approaches that work on a cell-by-cell basis. In fact, having
a mix of power gated and normal cells on the same row imposes drastic changes
in power routing within a single row. The ensuing disruption of routing regularity
makes it very difficult to control congestion and to ensure fast design convergence.
Furthermore, since all the sleep transistors are placed in dedicated rows the sleep
transistor placement is simplified and their overhead can be easily estimated and,
consequently, traded for leakage reductions.
Fig. 9.3 Power-gating design

flow
Placed Design
Clustering
Peak Current Est.
Sleep Transistor Sizing
Layout Modification
Sleep Transistor Insertion
Validation
To Router
In order to fully integrate clustering and sleep transistor insertion with a state-of-
the art physical design flow, placement and routing information have to be taken into
account within the core clustering algorithm that selects gates to be power-gated, as
well as in sleep transistor sizing and insertion.
Section 3.2 describes a common sleep transistor insertion flow suitable for the
variation-aware clustered strategies proposed in this work.
3.2 Clustered Row-Based Sleep Transistor Insertion

Methodology
State-of-the-art power gating methodologies follow a well known design flow,

depicted in Fig. 9.3.
The entry point is a standard-cell placed design. The first step is clustering,
in which cells are grouped together in order to be controlled by the same sleep
transistor. Next, maximum current estimation is performed for each cluster. This
information is essential to drive the selection of the appropriate sleep cells to
be connected to the various clusters (sleep transistor sizing). The layout is then
modified in order to accommodate the sleep transistor cells and the routing of the
sleep signals. Finally, the modified layout is validated before it is fed to the routing
tool.
This flow is fully compliant with industry standard back-end tools and it supports
various power gating strategies (i.e., different sleep transistor insertion types),
differing in the granularity of the blocks to which sleep transistors are applied, as
described in Sect. 3.2.3.
3.2.1 Peak Current Estimation
This section describes a methodology for estimating the maximum current drawn
by a cell cluster [30]. We define as Switching Window, T , of a gate under an
input pattern the interval between the arrival time and the output transition of the
gate. There is one switching window for each path through a particular gate. The
width of the switching window of a gate is equivalent to the propagation delay
of the gate for a rising or falling transition. For a given gate in the cluster, its
Maximum Switching Window (MSW) is the time interval encompassing all its
possible switching windows. If a gate has n switching windows, the set T1 ; : : : ; Tn
is called the Full Switching Window (FSW) of the gate.
The peak current estimation algorithm consists of two steps. The first one
computes the MSW for each gate in the cluster. The motivation behind extracting
the MSW and not all the switching times of a gate is that, for a complex circuit and
for a gate very deep in the logic, there is possibly a very large number of switching
time intervals, each one corresponding to a path being activated through a gate.
Moreover, it is very time consuming to extract all these windows for all the gates,
which form a cluster. Conversely, it is very fast to extract the MSW for each gate. In
S window T1 and
fact, it only requires the calculation of the first (earliest) switching
of the last (latest) one Tn ; the MSW is simply obtained as (T1 Tn ). We thus need
to extract only two switching time intervals for each gate in the cluster, which can
be accomplished very fast.
The second step consists of the construction of a current plot over time which
records the gates that switch at a particular time interval and, hence, the total
discharge current in that time interval. This plot is the superposition over time
of rectangles whose bases correspond to the MSWs of the individual gates of the
cluster (which will, in general, partially overlap), and whose heights correspond to
the currents drawn by the gates (e.g., derived by a technology library). From this
current plot, the time interval during which the maximum current discharge occurs
and the gates that contribute to this current discharge are obtained.
Figure 9.4 shows an example of current plot, in which five MSWs (from a to e)
are shown. The interval in which the current drawn is maximum corresponds to the
overlapping of the c, d , and e MSWs.
The current plot allows the identification of the time interval during which the
maximum current discharge could occur. To tighten this upper bound, we need to
extract the detailed FSW information only for those gates which contribute to this
maximum value, which are normally a very small percentage of the gates in the
cluster.
The algorithm enters then an iterative phase; it proceeds by finding, based on the
current plot, the subset of gates that contribute to this maximum current value. Gate
ordering on this subset of gates is executed first, so as to maximize the probability
of non-overlapping switching of these gates. After ordering the gates, gate-by-gate
FSW extraction is performed on the set of gates which contribute to the maximum
Fig. 9.4 Example of current plot
current. Once all the possible switching time intervals for this small subset of
gates are extracted, the current plot is updated and the process is repeated until
convergence is reached.
Convergence is guaranteed since, in the worst case, we have to extract the FSW
for all the gates in the cluster and compute their maximum currents. When full
FSW extraction is computationally too expensive, we may terminate the iteration
early, for example by using a time bound. Since the maximum current estimate
is monotonically non-increasing as the iteration proceeds (i.e., the upper bound is
progressively tightened), the use of a time bound provides us with a fine-grained
control of the accuracy vs. time trade-off.
3.2.2 Sleep Transistor Sizing
An effective use of power-gating requires a proper sizing of the sleep transistor.

In fact, while a small sleep transistor may unacceptably slow down the circuit in
the active mode due to its high resistance, a larger one implies a large area and a
significant energy cost to drive it [15]. Designers usually define an IR-drop threshold
(e.g., 10 % of VDD ) that is used as a constraint that must be met when sizing the sleep
transistor resistance.
The maximum sleep transistor channel resistance can be computed using
Eq. (9.3):
VDD ˛drop
Rst D (9.3)
Ion
where VDD ˛drop is the allowed voltage drop across the sleep transistor, expressed
as a percentage of the supply voltage, while Ion is the maximum discharge current
that power-gated cells inject into the sleep transistor during the active mode. The
estimation of the active current is not a trivial task. In fact, an erroneous estimation
of Ion translates to a sub-optimal sleep transistor sizing, which may result in area
and power increase or, even worse, in timing violations during the active mode [30].
Considering that the sleep transistor operates in the resistive region and knowing
Rst , its size can be properly evaluated using Eq. (9.4):

W 1
D (9.4)
L st Rst st COX .VDD Vthst VDD ˛drop /
where W =L is the ratio between width and length of the transistor channel, st is
the carrier mobility, COX is the oxide capacitance and Vthst is the threshold voltage.
The size of the sleep transistor is strongly influenced by its physical implemen-
tation, where carrier mobility, threshold voltage and gate length are key parameters.
Hence, for a proper sleep transistor sizing, indicating the type of MOS device is
mandatory: pMOS (i.e., header) or nMOS (i.e., footer). Although both devices can
be used as power-switches without distinction (i.e., maintaining the same leakage
reduction), nMOS transistors are usually more suitable. In fact, pMOS devices are
less leaky than nMOS ones, but they show a lower carrier mobility, which limits their
ON-current. As a result, in order to guarantee a certain current capability, pMOS
sleep transistors require more silicon area compared to nMOS.
In a typical power-gating approach, switch devices are high-threshold-voltage
transistors (i.e., HVT). Since a HVT transistor is less leaky, this choice helps
in reducing the total static consumption when the circuit is in stand-by mode.
Unfortunately, a larger threshold voltage reduces the current capability of the
transistor and, as demonstrated by Eq. (9.4), more area is required to achieve the
optimal resistance. On the contrary, with a low-threshold-voltage (i.e., LVT), the
silicon area taken by the switch device can be sensibly reduced, but its leakage
becomes significant.
Another design parameter that should be taken into account is the gate length.
Usually, the sleep transistors are designed with minimum channel length, but
for today’s nanometric technologies, this implies high leakage currents due to
Short Channel Effects (i.e., SCEs) and more power consumption. Increasing the
gate length allows to reduce the power consumption [15], but the current sinking
capability of the sleep transistor is drastically reduced and more area is required.
Moreover, since the majority of the fab processes are optimized for a single channel
length (typically, the minimum length), using devices with multiple lengths may
drastically increase process variation and sensibly reduce fabrication yield.
As discussed above, the sleep transistor has to be sized in order to limit the IR-
drop on the virtual rail when the maximum discharge current of the power-gated
cells is injected into the virtual ground. Under this IR-drop constraint, designers
can trade off channel length L and threshold voltage Vth (see Eq. (9.4)) to achieve
area and power optimization. In principle, maximum area efficiency (i.e., minimum
W ) is obtained using transistors with the largest current density capability, namely,
low-Vth transistors with minimum gate length. Obviously, this causes larger leakage
consumption due to an increase of the sub-threshold current. On the other hand,

in order to achieve minimum power overhead, high-Vth devices with larger gate
length are the best choice. This helps in reducing leakage power consumption at
the cost of a larger area. It is worth emphasizing that by increasing the equivalent
sleep transistor area, also the load capacitance that the driving circuit has to charge
and discharge increases, thus negatively impacting the dynamic power. Clearly,
depending on the system constraints, designers can play with the values of W
and Vth to achieve the required power/performance constraints. Most recent works
propose optimum sleep transistor synthesis in which, for a given IR-drop and area
constraints, the threshold voltage is selected to minimize the power [29]. To achieve
the optimum Vth , both body-bias and multi-Vth transistors can be exploited.
3.2.3 Power-Gating Strategies
As anticipated in Sect. 3.1, when power gating is applied to a generic logic block,
an important dimension of the problem concerns whether the gating is applied to:
(1) The entire logic block (block-level or full power-gating); (2) a subset of the cells,
typically, the non-timing critical ones (partial or clustered power-gating); (3) all
the cells individually (cell-level power-gating). The block-level approach has the
drawback of having an high reactivation period since it may have long transition
delay between sleep and active state. The cell-level strategy is characterized by high
area overhead and huge sleep signals buffering. The most promising power-gating
style is then the clustered one, in particular the row-based style, which offers a
reduced overhead with respect to the cell-level approach and limited active/sleep
transition times with respect to the block level strategy.
3.2.4 Layout Modification
The modifications to the layout required to accommodate sleep transistor insertion

depend on the chosen clustering granularity. Figure 9.5 shows an example of a
modified layout in the case of row-based clustering; the sleep transistor is placed
in a dedicated row indicated as ST row in the figure. Open channels indicate where
the ground lines of adjacent rows are split to accommodate for clustering rows which
share a common ground line.
In the case of block-level power-gating, the layout must be modified so that the
rows are extended on both sides of the layout to make space for the sleep transistor
cells. All side pins must be moved and an extra pin must be added to connect the
SLEEP signal with the outside. For row-based power-gating, the clustering phase
returns a set of rows to be disconnected from the ground and connected to Virtual
Ground. If those rows share the ground line with any non-power-gated row they
need to be spaced. If two rows sharing the ground line are both to be clustered, we
can simply connect the ground line to the Virtual Ground. Besides opening channels,
Fig. 9.5 Example of a complete layout after a row-based sleep transistor insertion
extra rows needed to host the sleep transistor cells are added. As in the block-level
case the pins have to be moved and the SLEEP pin added.
For any clustering strategy, the power grid and its connection to the standard cells
are created. The smallest region that can be power-gated is limited by the distance
from the Vgnd power rail. The maximum distance from the Vgnd is equal to the
distance between two power stripes of the same type divided by two. To avoid
having regions of different sizes, the rows are always cut in the same points and
those are determined by the position of the vertical stripes. Each cut, if needed,
is done at the maximum distance from the Vgnd lines. To avoid wrong substrate
polarizations the distance between two stripes of the same type has been kept equal
to the one without power gating, while the width of the stripes has not been changed
in order to avoid the IR drop increase.
In the remainder of this chapter, we will consider row-based sleep transistor
insertion as the reference power-gating strategy. However, the solutions we present
could be successfully adapted to other kinds of sleep transistor insertion approaches.
4 Clustered Tunable Power-Gating for PV Compensation
Parametric variations introduced by manufacturing represent the main cause of

performance variability and yield degradation in modern VLSI circuits.
As discussed in Sect. 2, process variations may range from few percent to several
orders of magnitude, following a deterministic scheme or a random distribution, also
depending on the spatial-scale they reach: Wafer-to-wafer (W2W), die-to-die (D2D)
and with-in-die (WID) variations. While the binning method has been successfully
adopted for tackling W2W and D2D variations, considering WID variations (which
are random, more localized and thus harder to manage) require dedicated counter-
measures that must be taken since the early phases of the design flow.
Several design methodologies have been proposed in the last years, from those
based on Design for Manufacturability approaches [13, 18, 25] (like litho-friendly
and Restricted Design Rules layout design), where variability of a given design is
either mitigated or amortized, to Adaptive strategies, which attempt to solve the
variability issues by sensing and correcting the desired parameters using various
knobs that affect them. These schemes are also called Monitor & Control (M&C)
strategies, to emphasize their analogy with closed-loop control systems.
M&C approaches have proven to be extremely efficient and they are usually
preferred to other strategies for their flexibility. Different embodiments of the
M&C paradigm have been proposed recently. The main differentiating factor is
represented by the strategy used to control the circuit. Most solutions use Dynamic
Voltage Scaling (DVS) [35], or Adaptive Body Biasing (ABB) [34] as control
knobs. Although both DVS and ABB are effective in adjusting circuit performance,
as a side-effect they have a dramatic impact on the power consumption of the
circuit; in fact, dynamic power is quadratically related to the supply voltage,
while sub-threshold leakage current shows an exponential relationship to the body
voltage. This makes their implementation energy inefficient, thus less effective for
leakage-dominated CMOS technologies.
In this section, we show how power-gating may represent a viable solution to
achieve concurrent power reduction and performance control for mitigating process
variations (and random WID variation in particular) and increase the timing yield.
We show that, when enhanced with tunable features, the sleep transistors can act
as natural supply-voltage regulators for the power-gated circuit, thus providing a
low-power, yet low-cost, control knob.
4.1 Modeling Process Variations and Timing Yield
WID variations are mostly due to systematic and random variations of several
physical device parameters, such as the concentration of doping atoms in the
substrate (Ni ), the effective channel length (Leff ) and width (Weff ) and the oxide
thickness (Tox ). The resulting effect is the shift of the electrical characteristics of
Fig. 9.6 Effects of parameter drifts due to process variations
the transistors, like the threshold voltage (Vth ) and the maximum current density
(Ids =W ), with a significant impact on the power dissipation and the performance of
a manufactured circuit [6].
Under these conditions, device parameters (and design metrics) have to be treated
as random variables, whose probability distribution function (PDF) depends on the
actual fabrication process. As pictorially summarized in Fig. 9.6, deviations from
the ideal case (i.e., where a given parameter X has a nominal, deterministic value)
are represented by a PDF with a given shape with average value avg.X / coincident
with the nominal value Xnom of the parameter; variability is related to the variance of
the PDF (e.g., the range of values between ˙3 ). Under the typical assumption that
values slightly exceeding the nominal value (Xmax in the figure) can be considered
as acceptable, we define the timing yield as the probability that X < Xmax , which
can be immediately obtained by the cumulative distribution function of X (area
below the PDF and delimited by Xmax ; striped region in the figure).
4.2 Controlling Performance with Tunable Sleep-Transistors
The active current Ion , drained by a circuit from the power supply during normal
operation, may represent a suitable metric for quantifying the performance degra-
dation induced by process-variation. In fact, the imperfections of the fabrication
process may affect several electrical parameters (such as channel dimensions and
threshold voltage), which in turn alter the actual current capability of the active
transistors, and thus, the intrinsic speed of the devices. Therefore, the larger the
speed degradation due to process variations, the smaller the resulting Ion .
During the active periods, the sleep transistor, that is turned-on and that operates
in the linear region (in Fig. 9.7, Ron represents the channel resistance in the linear
region), behaves as a current-to-voltage transducer that transforms the flowing
current Ion into a voltage drop (Vdrop in Fig. 9.7). The virtual-ground rail is now
Fig. 9.7 Voltage drop across

the sleep transistor during
active periods
Table 9.1 Quantitative dW C (ps)

effect of sleep transistor
dnom (ps) W 2W 4W 8W 16W
sizing on delays
130.2 152.7 147.1 140.8 138.7 134.9
at a potential higher than ground, and the circuit operates at a scaled supply voltage
(Vdd Vdrop ). By modulating Ron , and thus Vdrop , it is therefore possible to change
the operating point of the entire circuit: A lower Ron implies a faster circuit; a larger
Ron , on the contrary, will increase Vdrop thus making the circuit slower.
In summary, circuits made slower by PV can recover their speed by means of
Ron reduction, where Ron is proportional to the width of the sleep transistor.
The key aspect of this strategy is that the voltage regulation obtained through
Ron appears only temporarily while the circuit is switching. Once all the internal
switchings are completed, Ion falls to zero and so does Vdrop . In this sense, the
supply voltage Vdd is automatically adapted to the load, and there is no need of
an external voltage regulator. A quantitative evaluation of the above dependency is
given in Table 9.1, which reports delay values of a sample design (10 power-gated
chains in parallel, each consisting of 10 inverters) as a function of the total sleep
transistor width, in the presence of process variations.
Column dnom reports the nominal delay when the circuit is power gated with a
transistor of a given size W (in the example 20 “fingers” of 0.8 m each for a total
of 16 m) and ignoring the presence of variations. Columns dW C report worst case
delay values for different values of W . By worst case, we mean the extreme value
of the delay distribution obtained by randomly selecting 1,000 different instances of
the circuit with Monte-Carlo. The column dW C –W shows then the slowest circuit
instance (152:7 ps) when a transistor of size W is used for gating, i.e., the same
conditions for which dnom is measured. This instance is approximately 17 % slower
than the nominal one. If we then upsize the sleep transistor for such circuit instance,
we see that, as expected, we can progressively speed up the instance, asymptotically
reaching the nominal delay dnom .
4.3 Design Issues and Architectures

4.3.1 Design Issues
As mentioned above, Ron modulation allows to compensate delay variations

with a simple mechanism and with an arbitrarily fine granularity. However, such
powerful control mechanism does not come for free, and designers must take
into consideration the amount of overhead introduced by a tunable power-gating
architecture.
Upsizing the sleep transistor may result in a significant area overhead, which
causes extra static and dynamic power consumption. In fact, the sleep transistor
itself leaks when turned off, and its leakage is proportional to its size. More-
over, driving a huge sleep transistor during power-mode transitions would imply
additional load capacitance and larger logic effort. Such overhead may nullify the
leakage power savings obtained by gating the circuit.
As an example, consider again the data of Table 9.1. If a 5 % delay increase can
be tolerated (i.e., dmax D 130:21:05 D 136:71 ps), then the sleep transistor must be
as large as 16W to achieve the required speedup; this will cause the circuit to have a
delay of 134:9 ps < dmax , but a power consumption due to the sleep transistor that
is 16 times larger than in the nominal case. Clearly, without a dedicated architecture
the use of power-gating as a M&C strategy may not be feasible.
4.3.2 Architectures
As described in Sect. 3, a key design variable in determining the area/power tradeoff

of a power-gated circuit is the granularity at which the sleep transistor is inserted.
This is true also when considering tunable power-gating architectures.
A first option consists of power-gating the whole circuit using a single tunable
sleep transistor (left configuration in Fig. 9.8). In this case, even if the transistor is
set to the proper width according to the detected delay value, its size must be the
largest of the range Wmax . Therefore, although effective in mitigating the process
variation effects, large area and power overheads can not be avoided.
Fig. 9.8 Tunable

power-gating (TPG) options:
Full-TPG (Left) and
clustered-TPG (Right)
A clustered sleep transistor insertion, on the other hand, may represent the
most appropriate architecture. In fact, tunability is required only for the cells that
determine the critical paths (CPs). On the contrary, all the other cells (i.e., cells for
which the delay increase due to process variations do not slow down the critical
paths) can be power gated with a regular, non-tunable, small-sized sleep transistor.
We call this approach Clustered-Tunable Power-Gating (right configuration in
Fig. 9.8). Under this scheme, two distinct sets of cells (i.e., clusters) use separate
sleep transistors: The critical cells identify the critical cluster C , which is gated by
a tunable transistor of size WC ; the rest of the cells form the non-critical cluster
NC , which is gated by a regular transistor of size WNC . Concentrating tunability
only where needed allows to reduce the total sleep transistor width (WNC C WC
Wmax ), and guarantees power savings while keeping the same delay compensation
capability.
4.3.3 Design Flow and Results
The design methodology for the automated implementation of clustered TPG does
not differ from that described in Sect. 3. Starting from a (row-based) placed design,
clustering, peak-current estimation, sleep transistor sizing and sleep transistors
insertion are the main phases of the flow. However, working with clustered tunable
power-gating architectures requires some changes and additions to the algorithms.
Design of the Tunable Sleep Transistor Cell Tunable sleep transistor cells are
not available in standard CMOS libraries. Then, it is important to support design
kits with new customized cells that contain parallel sleep transistors of different
sizes and driven by dedicated control signals. Figure 9.9 shows the schematic of
a possible architecture, as described in [14] and [32]. Each parallel sub-transistor
is driven by a NAND gate that receives an external configuration bit (b3, b2, b1,
b0, in Fig. 9.9), and whose value can enable (in case of 1-logic) or disable (in case
of 0-logic) the corresponding transistor. An additional sleep signal provided by an
external power-management unit is in charge of defining the operating mode of the
gated circuit.
Statistical Static Timing Analysis At design time, characterizing the effects
induced by process variations is key. In fact, this allows the identification of
the critical paths in presence of process variations. To do so, an option consists
of integrating probabilistic models into standard Static Timing Analysis (STA)
engines. For instance, it is possible to resort to Monte Carlo statistical sampling,
where device parameters are stochastic variables described by process-dependent
PDF. During each sample, the circuit timing is computed using traditional STA
tools. The output of a Monte Carlo analysis includes the path delay distributions.
From that, we can extract the list of statistical critical paths, i.e., the list of paths that
show a certain probability to have a slack smaller than a user defined threshold.
Fig. 9.9 Architecture of the tunable sleep transistor cell [32]
Clustering and Sleep Transistor Sizing The clustering phase is crucial, since it
defines which portion of the circuit has to be assigned to which cluster (critical C or
non-critical NC ). At this stage, it is important to identify the granularity at which
the sleep transistors are inserted into the layout. As described in Sect. 3, different
options are available, but row-based insertion can guarantee the finer granularity
at the minimum cost in terms of area and layout disruption. Hence, the clustering
problem comes down to the selection of which rows have to be considered as critical
or non-critical. An effective solution to this problem is to mark as critical those
rows that host at least one gate belonging to a statistical critical path [32]. Once
the clusters C and NC are formed, it is possible calculating their maximum active
currents, IC and INC ; respectively, and performing the sleep transistor sizing which
returns the actual width of the two transistors WC and WNC .
Sleep Transistor Insertion During this stage, the sleep transistor cells belonging
to a customized power-gating library are placed into the layout. If a row-based
approach is adopted, one can follow the strategy proposed in [31], where the
sleep transistors cells are placed in dedicated layout rows, called sleep rows. It is
worth mentioning that the sleep transistor of the critical cluster C is implemented
by means of modular tunable sleep transistor cells connected in parallel. All the
programmable sleep transistor cells are centered in their middle configuration,
namely, the configuration word is set in the middle of the dynamic range (1000
for the 4-bits cell shown in Fig. 9.9). This guarantees the maximal extension of the
compensation range. For the non-critical cluster one can use the same scheme, but
using non-tunable sleep transistor cells of fixed size [8].
Figure 9.10 offers a quantitative analysis of the timing yield and leakage
savings that a clustered tunable power-gating architecture (CLUSTERED-TPG)
may guarantee w.r.t. standard power-gating (FULL-PG) and block-level tunable
Fig. 9.10 Quantitative comparisons of various power gating options
power-gating (FULL-TPG). The results are obtained from the simulation of a subset
of the ITC’99 benchmarks mapped onto an industrial 45 nm technology. Only the
averages are reported.
The tunable approach (i.e., CLUSTERED-TPG and FULL-TPG) originates a
timing yield increase from 80.67 % (FULL-PG) to 100 %; in other terms, tunable
power-gating allows to fully recover any speed degradation induced by process
variation, thus making the number of circuits that must be discarded because they
incur timing violations fall to zero.
Concerning leakage power, we observe that a FULL-TPG architecture can
successfully tackle the problem of variability compensation at the price of a drastic
reduction of the achievable leakage savings; due to a larger sleep transistor (leakage
savings go from almost 100 % for the FULL-PG to a mere 24.36 % for FULL-TPG).
On the contrary, a clustered scheme (i.e., CLUSTERED-TPG) guarantees the same
timing yield as FULL-TPG, while maintaining reasonable leakage savings (72.1 %).
Clearly, applying a clustered architecture is what makes the tunable approach
applicable to real-life circuits.
5 Clustered Power-Gating for NBTI-Induced Aging

Minimization
Besides the non-deterministic variations described in Sect. 2, another, and possibly

more insidious, type of non-ideality of scaled devices concerns time-dependent
deviations in their operating characteristics [1].
There are two types of sources of such time-dependent variations: Bias Tempera-
ture Instability (BTI), and Hot Carrier Interface (HCI) [26]. Both phenomena cause
the generation of traps at the interface between the silicon and the oxide, resulting
into an increase over time of the threshold voltage of the transistors.
BTI affects both pMOS (Negative BTI—NBTI) and nMOS transistors (Positive
BTI—PBTI); in current technologies, the impact of PBTI is much lower than that of
NBTI, although its importance is expected to increase with the adoption of high-k
dielectrics in the gate-oxide interface [24]. Conversely, HCI effects are much more
significant in nMOS transistors, and are at least two orders of magnitudes larger
than for pMOS devices [12].
Between NBTI and HCI, NBTI is regarded as the most significant effect, because
the surface along which interface traps are created (i.e., the whole silicon-oxide
interface) is much larger than that of HCI (i.e., in the neighborhood of the drain
area) [1].
We can thus consider NBTI as the dominant source of aging of transistors
in current sub-45 nm bulk CMOS technologies. As we will show in the rest of
this section, the peculiar properties of NBTI and, in particular, its state-based
manifestation will allow to use power-gating as a powerful knob for aging control
while keeping the usual benefits in static power reduction.
5.1 Background and Models
NBTI effects occur when a pMOS is negatively biased (i.e., negative Vgs , or a
logic ‘0’ is applied—the stress state) and originate an increase of the threshold
voltage. When a zero-bias voltage is applied (i.e., a logic ‘1’), NBTI stress is
actually removed, resulting in a partial recovery (i.e., a decrease) of the threshold
voltage (the recovery state). While there is no full consensus on the exact quantum-
mechanical mechanisms that govern the NBTI effects, the reactivation-diffusion
model is accepted as accurate enough for pMOS NBTI aging [1]. A simplified
version of such a model is the following:
Ea 1
Stress State W Vth D ks e .t tst r / 4
kT (9.5)
r
t trec
Recovery State W Vth D kr .1 / (9.6)
t
where ks and kr are two parameters whose magnitude depends on few technological
parameters (such as channel strain and nitrogen concentration), k is the Boltzmann
constant, T is the device temperature, Ea is the activation energy, and tst r and trcv
denote stress and recovery times, respectively. t is the free variable and expresses
the temporal evolution of the threshold voltage drift.
Equations (9.5) and (9.6) show the qualitative behavior of NBTI stress and
1
recovery: Vth increases
p during stress with t 4 dependency, it decreases during
recovery with 1 1=t dependency, and it depends exponentially on temperature
during stress. The most relevant feature of the model, however, is that there are
two different behaviors based on the state of the device. This observation, coupled
with the experimental evidence that NBTI aging is frequency-independent [1, 22],
implies that it is the total stress time that matters rather than the actual dynamics of
stress-recovery.
A generic signal applied to a pMOS transistor can then be modeled as a periodic
waveform with equivalent amount of stress time, paving the way to a probabilistic
modeling of NBTI aging. This allows to lump the models of Eqs. (9.5) and (9.6) into
a single macromodel suitable to circuit-level simulation:
1
Vth D K ˇ t 4 (9.7)
where K includes all the technological and environmental (e.g.,temperature, supply

voltage) constants and ˇ denotes the stress probability of the signal connected to
the gate input of the pMOS transistors, that is, the probability of a logic ‘0’. This
macromodel is also more suitable for translating the increase of threshold voltage
into a delay degradation, which better corresponds to the intuition of “aging”.
Under the alpha-power law, we can write the delay of a logic gate as:
CL Vdd
dD (9.8)
.Vgs Vth /˛
where CL is the load capacitance, Vgs is the gate voltage and ˛ is a technology-
related exponent that can be approximated to 1 for sub-90 nm technology.
Because of the NBTI-induced threshold voltage increase, the “aged” delay d 0 >
d then becomes:
CL Vdd
d 0 .t / D (9.9)
.Vgs .Vth C Vth .t ///
which can be expressed as a penalty with respect to d as:
K .ˇ t /1=4
d 0 .t / D d .1 C / (9.10)
VGT K .ˇ t /1=4
where VGT D Vgs Vth;0 and Vth;0 is the threshold voltage at time 0.
Equation (9.10) allows translating circuit operations (in terms of signal probabil-
ities) to aging (i.e., delay increase over time) for the individual gates, similarly to
what it is done for estimating dynamic power based on switching probabilities.
5.2 Power-Gating and Aging Reduction
Equation (9.7) clearly shows that, for a specific manufactured device and for specific
operating conditions (temperature and Vdd ), there is one single knob that can be used
for mitigating the aging effects: The stress probability ˇ.
Ideally, one would like to make ˇ as small as possible. This would imply having
as many 1’s in the logic network as possible for the largest possible fraction of time.
Obviously, under normal conditions, this is infeasible, because: (1) To implement
meaningful functions, circuits require logic inversion, which by definition prevents
achieving arbitrarily small probability for a predetermined signal value (0 or 1).
(2) Circuit structure affects probability. On the other hand, information theory
suggests that a signal probability of 0 or 1 carries no information—entropy is 0, and
it is maximum for a 0.5 probability. Therefore, to implement a realistic function,
there must be a fair distribution of 0’s and 1’s.
Although the ideal objective is impossible to reach, technology-independent
and technology-dependent synthesis techniques can be used to minimize the
0-probability of internal signals. Kumar et al. [23] proposed multi-level synthesis
and technology mapping algorithms that adopt a modified, NBTI-oriented metric,
while Wu and Marculescu [19] used logic restructuring and pin reordering to exploit
the fact that not all transistors are identically important in determining the delay of
a gate in the pull-up network. These strategies can reduce the aging in the typical
state of the circuit (i.e., determined by the implementation and by the most common
input patterns) and achieve a sort of local optimum.
We propose power-gating as a knob for improving over these results, aiming
at a global optimum. We suggest leveraging a well-known weakness of power-
gating: The logic values of the nodes of a power-gated block are lost when the
block is disconnected from the power supply or the ground nets. This poses serious
problems when storing values in memory elements and when interfacing power-
gated to non power-gated regions. Solutions based on the usage of special types of
memory elements [3] or by proper design of the sleep transistor cell [16] do exist.
For the combinational portion of a circuit, the behavior of the internal nodes
depends on the type of switch (footer or header) implementing power-gating. If
a footer switch is used (and thus the block is disconnected from the ground), as
described in Sect. 3, an interesting behavior occurs. Nodes that are at logic value
‘1’ before opening the switch do keep their values, whereas nodes with value ‘0’
become floating. Both the virtual ground line and the ‘0’ nodes then get charged to
‘1’ by the leakage current of the pull-up network of the cells [27]. The speed of this
charge process depends on the design of the sleep transistor cell, and can be speed
up by using a proper pull-up boosting mechanism, as shown in [7].
For the sake of illustration, Fig. 9.11 plots the signals of a simple two-inverter
circuit connected to a footer switch after the latter has been turned off (0.0 on the
timescale) [10]. We observe that the signals a and c, originally at logic value ‘1’,
stay unchanged, while signal b, at ‘0’ when the sleep signal is activated, goes to ‘1’
quite abruptly, and it reaches about 85 % of Vdd in a few tens of nanoseconds.
a b c
Virtual
Sleep Ground
Fig. 9.11 Internal signals during standby intervals [10]
What it matters for our purposes is that all the nodes inside the gated block
region (more or less quickly) reach a logic value of ‘1’, that is, the gated block
recovers from aging during the standby intervals. This is clearly a non-logical state
that cannot be obtained by any mechanism that operates on the circuit function or
the input data, and is therefore orthogonal to such design approaches.
5.3 Design Issues and Architectures

5.3.1 Design Issues
From the analysis of the previous section, two key issues deserve a deeper
evaluation. First, it is evident that the benefit in terms of aging goes together with the
potential leakage reductions: If the gated block is put in a standby state sporadically,
the aging reduction will be marginal. Clearly, the amount of time spent in the
standby state is not a quantity that is controllable by the designer and it is determined
by the environment. Therefore, it is important to parameterize the aging in terms of
this quantity. Second, the implementation of power-gating does not come for free.
As discussed in Sect. 3, the addition of a footer switch in series with the pull-down
network increases the on-resistance of each cell and results into a (nominal, i.e., at
time zero) delay penalty. Since by tackling aging we are trying to compensate the
1.15E+01
1.13E+01 Psleep=0.7
1.10E+01
1.08E+01 Psleep=0.8
Delay
1.05E+01 Psleep=0.9
1.03E+01 No PG
1.00E+01
9.75E+00 β=0.5
9.50E+00
00
06
07
07
07
07
07
07
07
07
07
E+
E+
E+
E+
E+
E+
E+
E+
E+
E+
E+
00
50
75
75
75
75
75
75
75
75
75
0.
7.
1.
2.
3.
4.
5.
6.
7.
8.
9.
Time (s)
Fig. 9.12 Delay degradation as a function of sleep probability
increase of delay over time, it is fundamental considering the fact that in a power-
gated circuit we start from a larger nominal delay value.
Equation (9.10) can be adapted so as to incorporate the above described effects
by adding two more parameters: Psleep , the probability of the sleep signal (0 means
always active, 1 always in standby), and , the delay penalty the due to the addition
of the sleep transistor (i.e., ds D d .1 C / is the time-zero delay of the gated
block).
The new expression of the delay model of a gated block becomes:
K .ˇ .1 Psleep / t /1=4
ds0 .t / D ds .1 C / (9.11)
VGT K .ˇ .1 Psleep / t /1=4
Notice that the stress probability ˇ is now modulated by the complement of the
sleep probability .1 Psleep /; this is equivalent to an “effective” stress probability
(ˇ 0 D ˇ .1 Psleep /, with ˇ 0 t is now the effective stress time.
Figure 9.12 plots Eqs. (9.10) and (9.11); the curve for ds0 .t / is parameterized with
respect to Psleep , and it refers to a specific value of . Both curves for d.t / and ds0 .t /
refer to a value of ˇ D 0:5. The plot is qualitative just for the sake of illustration of
how the aging profiles change based on the various parameters.
The aging curves for power-gating start at a higher time-zero delay value (ds ), but
they exhibit an amount of recovery that is proportional to the percentage of standby
time Psleep . The diamonds in correspondence to the intersections between the non
power-gated curve and the power-gated ones show the breakeven points, that is, the
points in time after which the power-gated circuits start aging more slowly than
their non power-gated counterparts. Clearly, the intersection moves earlier in time
as Psleep increases, denoting better aging profiles.
Fig. 9.13 Power-gating

options: non-critical gates
only (left) and clustered
(right)
There are three parameters that affect the above analysis: Psleep ; ˇ; and . The
first two are not under the designer’s control and depend on the functional and
environmental characteristics of the circuit; for a given implementation, they assume
a well defined value. On the contrary, is a design variable; its magnitude is related
to the sleep transistor size [2–31]: A larger (smaller) implies a smaller (larger)
sleep transistor. The ideal case of D 0 would correspond to a transistor of infinite
size.
5.3.2 Architectures
The discussion above assumes the entire logic block being power-gated by a single
sleep transistor. But, the aging benefits of power-gating are proportional to ; a
large value of ds may result in a breakeven point which might be beyond the
typical lifetime horizon of the circuit. Therefore, it is essential to keep as small as
possible.
One way for achieving this is to selectively apply power-gating. A first option is
partial power-gating, in which power-gating is applied only to the non-critical gates
of a circuit (left schematic of Fig. 9.13). Assuming a coarse-grain implementation
of power-gating, gating only a portion of the circuit requires the interaction with
the physical placement of the cells [11, 31]. By not slowing down critical gates,
this solution has the same delay as the nominal one (ds
d ) and it requires a
sleep transistor width, W , smaller than that required for power-gating the entire
circuit. However, from the aging viewpoint, partial power-gating is almost identical
to the non power-gated case. In fact, keeping the critical gates un-gated means that
they will age as in the original circuit. Then, partial power-gating allows achieving
leakage reductions at zero delay overhead, but it does not yield any aging benefit.
To get concurrent leakage and aging benefits, we can adopt a clustered architec-
ture similar to the one of Sect. 4. The entire circuit is power-gated using two sleep
transistors (see the right part of Fig. 9.13): One for the critical gates (of size WC )
and another one for the non-critical gates (of size WNC WC ) [10]. The idea
is to decouple the sleep transistor problem by using a small transistor for the non-
critical gates (with the only purpose of saving leakage) and a larger one with smaller
performance penalty for critical cells. If WC W , with the whole circuit gated, we
can achieve leakage savings comparable to the case of full power-gating.
5.3.3 Design Flow and Results
As for the case of PV-aware power-gating, the methodology for an automated

implementation of an aging-aware clustered power-gating requires specialized
algorithms to be integrated with the standard flow shown in Sect. 3.
Library Characterization Cell libraries are usually not characterized for aging.
To this purpose, the designer has to fill look-up tables containing the delay
degradation of each cell, parameterized with respect to the static 0-probabilities of
the inputs, stress voltage (i.e., Vgs ), and temperature. This task could be split into two
phases by first characterizing the pMOS transistors (and modeling the corresponding
Vth variation) and then incorporating it into the cell as a negative voltage-source
on the gate-terminal of pMOS transistors.
Aging-Aware Timing Analysis This step consists of a probabilistic simulation that
uses the parameterized cell delay models and statistics of the input signals (static
probabilities) to determine the paths with a delay that is within a given percentage
of the nominal critical path d .
Clustering Entails the determination of the critical and non-critical clusters, i.e.,
two subset of the cells that maximize the leakage and aging benefits. The selection
of which cells go in which cluster strongly interacts with the type of sleep transistor
insertion strategy. In particular, the granularity of the gating unit coincides with
the granularity of the cluster assignment. For instance, if the row-based strategy of
Sect. 3 is adopted, clusters consist of rows, and therefore the assignment to clusters
is done on a row-by-row basis. In [10], the clustering problem is formulated as a
0–1 nonlinear program, where the unknowns are the membership of a gating unit
(in that work, layout rows) to one of the two clusters.
Sleep Transistor Sizing This phase is dictated by the timing ( ) constraints, and
involves, as for clustering, the estimation of the maximum current drawn by each
cluster, as discussed in Sect. 3.
The results of leakage power savings and lifetime extension achieved with the
application of the clustered power-gating methodology are quite promising, and
they are summarized in Fig. 9.14. Three power-gating schemes are compared: NO-
PG (no power-gating), PG (the entire circuit is power-gated) and CLUSTERED
(the two-cluster architecture in the right part of Fig. 9.13. Data are normalized with
respect to the NO-PG case (both leakage and lifetime are assumed to be 1).
The lifetime of a circuit is defined as the time at which the circuit degrades
its performance by 15 % beyond its nominal value (i.e., performance of the non
Fig. 9.14 Quantitative comparisons of various power-gating options
power-gated circuit). The results in the chart represent an average over a set of
standard benchmarks and some medium-sized industrial designs, and refer to a
45 nm industrial CMOS technology.
We observe that the PG scheme does not fully exploit the huge leakage reduction
(about 88 %) for lifetime extension; in fact, only a 57 % increase is achieved.
Conversely, the CLUSTERED architecture yields a 5 increase with respect to the
non-gated version, with a negligible loss in leakage reduction (about 0.5 % penalty
with respect to PG). These numbers demonstrate experimentally the effectiveness
of clustered power-gating as a tool for simultaneous leakage power minimization
and aging effects mitigation.
Conclusions
There is wide consensus within the electronics industry that the scaling
process of the CMOS technology will continue as far as the fundamental
barrier of the atom size will be reached. Therefore, CMOS-based circuits and
systems represent the future vehicle for digital applications of the next decade.
Unfortunately, nano-scale CMOS technologies show intrinsic mechanisms,
like internal leakage consumption and parametric variations, which make their
use extremely challenging.
In this work, we explored the possibility of exploiting low-power tech-
niques for concurrent leakage optimization and variability compensation,
and both static (i.e., due to process variation) and dynamic (i.e., due to
NBTI-induced aging mechanisms) variability have been considered. More
specifically, we have shown how power-gating, when implemented through
a clustered strategy, offers a suited performance control knob to compensate
process variations, as well as a natural solution for reducing NBTI effects.
(continued)
New design methodologies have been implemented to support variation-

aware clustered power-gating and experimental results, conducted on bench-
mark circuits mapped onto an industrial 45 nm technology, have highlighted
their effectiveness: 100 % of timing yield in the presence of process variations
have been achieved, with substantial (i.e., 5) lifetime extensions w.r.t. non
power-gated circuits.
References
1. M. Alam (2008) “Reliability- and process-variation aware design of integrated circuits,” in

Microelectronics Reliability, 48(8):1114–1122.
2. M. Anis, S. Areibi, M. Elmasry (2003) “Design and optimization of multi-threshold CMOS
(MTCMOS) circuits,” IEEE Transactions on Computer-Aided Design of Integrated Circuits
and Systems, 22(10):1324–1342.
3. P. Babighian, L. Benini, A. Macii, E. Macii (2006) “Enabling fine-grain leakage management
by voltage anchor insertion,” IEEE Design, Automation and Test in Europe (DATE’06),
pp. 868–873.
4. D. Boning, S. Nassif (2000) “Models of process variations in device and interconnect,” in
Design of High Performance Microprocessor Circuits, Wiley.
5. S. Borkar (2005) “Designing reliable systems from unreliable components: The challenges of
transistor variability and degradation,” IEEE Micro, 25(6):10–16.
6. S. Borkar, T. Karnik, S. Narendra, J. Tschanz, A. Keshavarzi, V. De (2003) “Parameter
variations and impact on circuits and microarchitecture,” ACM/IEEE Design Automation
Conference (DAC’03), pp. 338–342.
7. A. Calimera, A. Pullini, A. Sathanur, L. Benini, A. Macii, E. Macii, M. Poncino (2007) “Design
of a family of sleep transistor cells for a clustered power-gating flow in 65nm technology,”
ACM/IEEE Great Lakes Symposium on VLSI (GLSVLSI’07), pp.501–504.
8. A. Calimera, L. Benini, A. Macii, E. Macii, M. Poncino (2009) “Design of a flexible
reactivation cell for safe power-mode transition in power-gated circuits,” IEEE Transaction
on Circuits and Systems - Part I, Regular Papers, 56(9):1979–1993.
9. A. Calimera, E. Macii, M. Poncino (2009) “NBTI-aware power gating for concurrent leakage
and aging optimization,” International Symposium on Low Power Electronics and Design
(ISLPED’09), pp. 127–132.
10. A. Calimera, E. Macii, M. Poncino (2010) “NBTI-aware clustered power gating,” ACM
Transactions on Design Automation of Electronic Systems, 16(1):3.1–3.25.
11. L. Changbo, L. He (2004) “Distributed sleep transistor network for power reduction,” IEEE
Transactions on Very Large Scale Integration (VLSI) Systems, 12(9):937–946.
12. Z. Chen, K. Hess, J. Lee, J. Lyding, E. Rosenbaum, I. Kizilyalli, S. Chetlur, R. Huang (2000)
“On the mechanism for interface trap generation in mos transistors due to channel hot carrier
stressing,” IEEE Electron Device Letters, 21(1):24–26.
13. C. Chiang, J. Kawa (2007) Design for manufacturability and yield for nano-Scale CMOS,
Springer-Verlag.
14. H.-S. Deogun, D. Sylvester, R. Rao, K. Nowka (2005) “Adaptive MTCMOS for dynamic
leakage and frequency control using variable footer strength,” IEEE International SOC
Conference (SoCC’05), pp. 147–150.
15. D. Flynn, R. Aitken, A. Gibbons, K. Shi (2007) Low Power Methodology Manual, Springer-
Verlag.
16. S. Henzler, G. Georgakos, M. Eireiner, T. Nirschl, C. Pacha, J. Berthold, D. Schmitt-Landsiedel
(2006) “Dynamic state-retention flip-flop for fine-grained power gating with small design and
power overhead,” IEEE Journal of Solid-State Circuits, 41(7):1654–1661.
17. ITRS (2009) “Process integration, devices & structures,” in International Technology Roadmap
for Semiconductors, ITRS.
18. T. Jhaveri, V. Rovner, L. Liebmann, L. Pileggi, A. Strojwas, J. Hibbeler (2010) “Co-
optimization of circuits, layout and lithography for predictive technology scaling beyond
gratings,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems,
29(4):509–527.
19. W. Kai-Chiang, D. Marculescu (2009) “Joint logic restructuring and pin reordering against
NBTI-induced performance degradation,” IEEE Design, Automation and Test in Europe
(DATE’09), pp. 75–80.
20. J. Kao, S. Narendra, A. Chandrakasan (1998) “MTCMOS hierarchical sizing based on
mutual exclusive discharge patterns,” ACM/IEEE Design Automation Conference (DAC’98),
pp. 495–500.
21. V. Khandelwal, S. Srivastava (2007) “Leakage control through fine-grained placement and
sizing of sleep transistors,” IEEE Transactions on Computer-Aided Design of Integrated
Circuits and Systems, 26(7):1246–1255.
22. S. Kumar, C. Kim, S. Sapatnekar (2006) “An analytical model for negative bias temperature
instability,” IEEE/ACM International Conference on Computer-Aided Design (ICCAD’06),
pp. 493–496.
23. S. Kumar, C. Kim, S. Sapatnekar (2007) “NBTI-aware synthesis of digital circuits,” ACM/IEEE
Design Automation Conference (DAC’07), pp. 370–375.
24. S. Kumar, C. Kim, S. Sapatnekar (2009) “Adaptive techniques for overcoming performance
degradation due to aging in digital circuits,” IEEE Asia and South Pacific Design Automation
Conference (ASPDAC’09), pp. 284–289.
25. M. Lavin, F. L. Heng, G. Northrop (2004) “Backend CAD flows for restrictive design rules,”
IEEE/ACM International Conference on Computer-Aided Design (ICCAD’04), pp. 739–746.
26. S. Mahapatra, D. Saha, D. Varghese, P. Kumar (2006) “On the generation and recovery of
interface traps in mosfets subjected to NBTI, FN, and HCI stress,” IEEE Transactions on
Electron Devices, 53(7):1583–1592.
27. E. Pakbaznia, F. Fallah, M. Pedram (2008) “Charge recycling in power-gated CMOS cir-
cuits,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems,
27(10):1798–1811.
28. K. Roy, S. Mukhopadhyay, H. Mahmoodi-Meimand (2003) “Leakage current mechanisms
and leakage reduction techniques in deep-submicrometer CMOS circuits,” Proceedings of the
IEEE, 91(2):305–327.
29. A. Sathanur, A. Pullini, L. Benini, A. Macii, E. Macii, M. Poncino (2008) “Optimal sleep
transistor synthesis under timing and area constraints,” ACM/IEEE Great Lakes symposium on
VLSI (GLSVLSI’08), pp. 177–182.
30. A. Sathanur, L. Benini, A. Macii, E. Macii, M. Poncino (2011) “Fast computation of discharge
current upper bounds for clustered power-gating,” IEEE Transactions on Very Large Scale
Integration (VLSI) Systems, 19(1):146–151.
31. A. Sathanur, L. Benini, A. Macii, E. Macii, M. Poncino (2011) “Row-based power-gating:
A novel sleep transistor insertion methodology for leakage power optimization in nanome-
ter CMOS circuits,” IEEE Transactions on Very Large Scale Integration (VLSI) Systems,
19(3):469–482.
32. L. D. Silva, A. Calimera, A. Macii, E. Macii, M. Poncino (2011) “Power efficient variability
compensation through clustered tunable power-gating,” IEEE Journal of Emerging and
Selected Topics in Circuits and Systems, 1(3):242–253.
33. D. Sylvester, K. Agarwal, S. Shah (2008) “Variability in nanometer CMOS: Impact, analysis,
and minimization,” Integration - The VLSI Journal, 41(3):319–339.
34. J. Tschanz, J. Kao, S. Narendra, R. Nair, D. Antoniadis, A. Chandrakasan, V. De (2002)
“Adaptive body bias for reducing impacts of die-to-die and within-die parameter vari-
ations on microprocessor frequency and leakage,” IEEE Journal of Solid-State Circuits,
37(11):1396–1402.
35. J. Tschanz, S. Narendra, R. Nair, V. De (2003) “Effectiveness of adaptive supply voltage and
body bias for reducing impact of parameter variations in low power and high performance
microprocessors,” IEEE Journal of Solid-State Circuits, 38(5):826–829.
Chapter 10
Soft Error Rate and Fault Tolerance
Techniques for FPGAs
Fernanda Kastensmidt and Ricardo Reis
Abstract Different fault tolerance techniques can be applied to FPGAs according

to their type of configuration technology, architecture and target operating envi-
ronment. This chapter will present a set of fault mitigation techniques for SRAM,
FLASH and ANTIFUSE-based FPGAs and a test methodology to characterize those
FPGA under radiation. Results from neutron-induced faults will be presented and
compared.
1 Introduction
Integrated circuits operating in radiation environment are sensitive to transient faults

caused by the interaction of charge particles with the silicon [1]. At ground level,
the main source of radiation are neutron particles that interact with the material
provoking secondary particles such as alpha particles and muons that can ionize the
silicon provoking transient upset in circuits fabricated in nanometer technology [2].
The interaction of the charged particles with the transistor may provoke transient
and permanent effects. The effects that are caused by a single event interaction are
called Single Event Effects (SEE) and they can be transient as the Single Event
Upset (SEU) and Single Event Transient (SET) or permanent as single event latchup
(SEL), single event gate rupture (SEGR), or single event burnout (SEB) [3]. The
effect that is caused by accumulation of particle interaction is named Total Ionizing
Dose (TID) and it represents degradation in the performance of the transistors as it
modifies the threshold voltage and leakage current [4]. In this chapter, we focus on
transient effects caused by a single event as SEU and SET.
Field Programmable Gate Array (FPGA) components are very attractive for
aerospace applications [5], as well for many applications at ground level that require
a high level of reliability, as automotive, bank servers, processing farms, and others.
F. Kastensmidt • R. Reis ()

Instituto de Informática, Universidade Federal do Rio Grande do Sul (UFRGS), Porto Alegre, Brazil
e-mail: fglima@inf.ufrgs.br; reis@inf.ufrgs.br

DOI 10.1007/978-1-4614-4078-9__10
208 F. Kastensmidt and R. Reis
The high amount of resources available in programmable logic devices can be

applied to add flexibility to the on-board computer in satellites and to the automotive
industry, for example. As FPGAs can be configured in the field, design updates
can be performed until very late in the development process. In addition, new
applications and features can be configured after a satellite is launched, or updated
in hash environments. However, programmable devices have been found to be very
sensitive to radiation. It is fundamental to experimentally measure the soft error rate
of the available resources, as well as the output error rate of specific applications, to
evaluate their applicability in harsh environments.
The characterization of those programmable components is mandatory to sustain
its applicability under transient faults. The test methodology and characterization
of FPGAs under radiation will be presented. The fault mitigation methodology
to protect a design targeting those FPGAs will also be shown. Practical radiation
ground testing results under neutron for SRAM-based FPGAs and Flash-based
FPGAs will be presented and discussed.
2 FPGAs Under Soft Errors
Field-Programmable Gate Arrays (FPGAs) are configurable integrated circuit based

on a high logic density regular structure, which can be customizable by the end
user to realize different designs [6]. The FPGA architecture is based on an array of
logic blocks and interconnections customizable by programmable switches. Several
different programming technologies are used to implement the programmable
switches. There are three types of such programmable switch technologies currently
in use: SRAM, where the programmable switch is usually a pass transistor or
multiplexer controlled by the state of a SRAM bit (SRAM based FPGAs); Antifuse,
when an electrically programmable switch forms a low resistance path between
two metal layers (Antifuse based FPGAs); and EPROM, EEPROM or FLASH cell,
where the switch is a floating gate transistor that can be turned off by injecting
charge onto the floating gate.
Customizations based on SRAM are volatile. This means that SRAM-based
FPGAs can be reprogrammed as many times as necessary at the work site and
that they loose their contents information when the memories are not connected
to the power supply. The antifuse customizations are non-volatile, so they hold the
customizable content even when not connected to the power supply and they can be
programmed just once. Each FPGA has a particular architecture. Programmable
logic companies such as Xilinx, MicroSemi, Aeroflex (licensed for Quicklogic
FPGAs), Atmel and Honeywell (licensed for Atmel FPGAs) offer radiation tolerant
FPGA families. Each company uses different mitigation techniques to better take
into account the architecture characteristics.
10 Soft Error Rate and Fault Tolerance Techniques for FPGAs 209
2.1 Single Event Effects on SRAM-Based FPGAs
The SRAM-based FPGA is composed of an array of configurable logic blocks

(CLB), a complex routing architecture, an array of embedded memories (Block
RAM), an array of digital signal processing components (DSP) and a set of control
and management logic. The CLBs are composed of Look-up Table (LUT) that
implements the combinational logic, and flip-flops (DFF) that implements the
sequential elements. The routing architecture can be very complex and composed of
millions of pre-defined wires that can be configured by multiplexers and switches to
build the desirable routing.
The configuration of all CLBs, routing, Block RAMs, DSP blocks and IO blocks
is done by a set of configuration memory bits called bitstream. According to the size
of the FPGA device, the bitstream can contain millions of bits. The memory bits that
store the bitstream inside the FPGA is composed of SRAM memory cells, so they
are reprogrammable and volatile. When an SEE occurs in the configuration memory
bit of an SRAM-based FPGA, it can provoke a bit-flip. This bit-flip can change the
configuration of a routing connection or the configuration of a LUT or flip-flop in
the CLB. This can lead to a dramatic impact in the designed circuit, since an SEE
may change its functionality.
An SEE in the configuration memory bits of an SRAM-based FPGA has a
persistent effect and it can only be corrected when a new bitstream is loaded to the
FPGA. In the combinational logic, the effect of an SEE is related to a persistent fault
(zero or one) in one or more configuration bits of a LUT. Figure 10.1 exemplifies
an SEU occurrence in a LUT configuration bit and in a bit controlling a routing
connection. SEE in the routing architecture can connect or disconnect a wire in
the matrix. This is also a persistent effect and its effect can be a modification in
the mapped circuit, as a logic change or a short circuit in the combinational logic
implemented by the FPGA. It can take a great number of clock cycles before the
persistent error is detected and recovery actions are initiated, as the load of a faulty-
free bitstream. During this time, the error can propagate to the rest of the system.
Bit-flips can also occur in the flip-flop of the CLB used to implement the user’s
sequential logic. In this case, the bit-flip has a transient effect and the next load of
the flip-flop will correct it.
2.2 Single Event Effects on Flash-Based FPGAs
Well-known Flash-based FPGAs are from Microsemi, which presents the fami-
lies PROASIC3 and SmartFusion [7]. The reconfigurable array is composed of
VersaTiles and routing resources that are programmable by turning ON or OFF
switches implemented by floating gate (FG) transistors (NMOS transistor with a
stacked gate). The FG switch circuit is a set of two NMOS transistors: (1) a sense
transistor to program the floating gate and sense the current during the threshold
Fig. 10.1 Example of an SEU occurrence in a LUT and in the routing of an SRAM-based FPGA
Fig. 10.2 SET in the Switch In

Floating Gate
Flash-based FPGA
programmable switch
Sensing
Switching
Word
Switch Out
voltage measurement and (2) a switch transistor to turn ON or OFF a data-path in

the FPGA (Fig. 10.2). The two transistors share the same control gate and floating
gate. The threshold voltage is determined by the stored charge in the FG. Figure 10.3
illustrates VersaTiles used to implement some common logic gates. The VersaTiles
are connected through a four-level hierarchy of routing resources: ultra-fast local
resources; efficient long-line resources; high-speed, very-long-line resources; and
the high-performance VersaNet networks.
Each VersaTile can implement any 3-input logic functions, which is functionally
equivalent to a 3-inputs Lookup Table (3-LUT). But it is important to highlight
that the electrical implementation of the VersaTile is totally different than the
electrical implementation of a Lookup Table (LUT). Hence, the VersaTile may
Fig. 10.3 SET and SEU in the Flash-based FPGA VersaTile
have a different electrical behavior to variability effects with respect to a 3-inputs

LUT. The VersaTile can also implement a latch with clear and reset, or D flip-
flop with clear or reset, or enable D flip-flop with clear and reset by using the
logic gate transistors and feedback paths inside the VersaTile block. For each
configuration in the VersaTile block, the number of FG switches and transistors
in the critical path changes. Single Event Transient (SET) pulses can hit the drain
of the transistor at OFF state as presented in Fig. 10.2 provoking a transient pulse
in the configuration switches. Or it can hit the sensitive nodes of the transistors in
the VersaTile provoking SET or bit-flip according to the customization of the tile
(Fig. 10.3).
2.3 Single Event Effects on Antifuse-Based FPGAs
Antifuse-based FPGAs consists of a regular matrix composed of combinational

(C-cells) and sequential (R-cells) surrounding by regular routing channels. One of
well-known antifuse-based FPGAs is from Microsemi. All the customizations of the
routing and the C-cells and R-cells are done by an antifuse element (programmable
switch). Results from radiation ground testing have shown that programmable
switches either based on ONO (oxide–nitride–oxide) or MIM (metal–insulator–
metal) technology are tolerant to ionization and total dose effect [8]. Therefore,
the customizable routing is not sensitive to SEU, only combinational logic and the
flip-flops used to implement the design user sequential logic are sensitive to SEE.
Table 10.1 Summary of SEU and SET effects in FPGAs

SEU/SET in the logic of the
FPGA configuration basic block Routing connections Configurable switches
SRAM-based Persistent Persistent Persistent
Flash-based Transient No No
Antifuse-based Transient No No
Another well known antifuse-based FPGA is from Aeroflex and QuickLogic.

Its architecture is composed of a regular matrix of configurable logic cells used to
implement the combinational logic and flip-flops, surrounding by a regular routing
matrix. Programmable switches called ViaLink connector are used to do all the
customizations.
In order to summarize the SEU and SET effects in FPGAs, Table 10.1 shows
the susceptible parts of the architectures and classifies the effects as transient or
persistent, when it is needed reconfiguration to correct the fault.
3 Fault Tolerance Techniques for FPGAs
Different fault tolerance techniques can be applied to FPGAs according to their

type of configuration technology, architecture and target operating environment [6].
Techniques can be implemented by the user at hardware description language (HDL)
of the design before the design is synthesized into the FPGA. Or techniques can be
developed by the vendor, which provides a FPGA that is SEE robustness by layout.
Figure 10.4 illustrates these two options. Here we will focus on techniques that can
be applied by the user at the HDL design.
The main techniques are either based on spatial redundancy or temporal redun-
dancy [9]. Spatial redundancy is based on the replication of n times the original
module building n identical redundant modules, where outputs are merged into a
voter. Usually n is an odd number. The voter decides de correct output by choosing
the majority of the equal output values. The most common case of n-modular
redundancy (nMR) is when n is equal to 3, where it is called Triple Modular
Redundancy (TMR). In this case, a majority voter is used that is able to vote out
2 out of 3 values that are fault free. There is local TMR when only the flip-flops
are triplicated or global TMR where all the combinational and sequential logic is
triplicated. The TMR can be implemented in different ways by using large grain
TMR, or breaking into small blocks and adding extra voters. Each one can protect
SEU or SET, or both, as shown in Table 10.2.
When dealing with the routing, different techniques can be chosen to increase
or decrease fan-out, delay and set of connections, which may have a different
impact in the SEE sensitivity [10, 11]. Also embedded processors can use different
mitigations based on software redundancy, or processor redundancy like lock-step
and recomputation.
Fig. 10.4 Probability of a bit-flip as a function of the mean deposited charge by an incoming
particle for a 6T-SRAM cell before and after stress
Table 10.2 List of mitigation techniques that can be applied by the User in
Designs targeting FPGAs
Mitigation technique Abstraction level SET SEU
Local TMR HDL X
Global TMR HDL X X
Large grain TMR HDL X X
Voter insertion HDL X X
Reliability-Oriented place and Route Algorithm FPGA Flow X X
Temporal redundancy HDL X
Embedded processor redundancy HDL X X
Scrubbing System X
Time redundancy is based on capturing a value twice or three times in time to

vote out a transient fault. The values are shifted by a delay [9]. The idea is to be able
to capture 2 out of 3 upset free values to be able to mask the fault.
3.1 SRAM-Based FPGAs
In SRAM-based FPGAs, radiation-induced faults have a persistent effect so spatial

redundancy is needed to mask the upset combined with reconfiguration to correct
the fault in the configuration memory bits (bitstream). Figure 10.5 show the flow,
the design can be either protected by Global TMR (called XTMR by Xilinx) using
the XTMR tool or implemented by hand, and full or partial reconfiguration, called
scrubbing, must be applied to correct the upsets from time to time. Fault injection
can be used to evaluate the efficiency of the fault tolerance technique and ensure its
masking capability under single upsets.
The global TMR or XTMR consist on triplicating all the combinational and
sequential logic, input, outputs and clock trees as illustrated in Fig. 10.6. Note that
the majority voter can be applied after flip-flops in throughput logic, or after flip-flop
in a feed back path. In this second case, it is mandatory to have majority voters able
Fig. 10.5 Fault tolerance design flow for SRAM-based FPGAs
FPGA
TMR Output Voter
REDUNDANT REDUNDANT REDUNDANT

TMR flip-flop
TMR flip-flop
LOGIC (tr0) LOGIC (tr0) LOGIC (tr0)

INPUT REDUNDANT REDUNDANT REDUNDANT OUTPUT
REDUNDANT REDUNDANT REDUNDANT
package PIN
package PIN
REF
R0 R1 R2 MAJ
0 0 0 0
0 0 1 0
0 1 0 0
TMR flip-flop
0 1 1 1
1 0 0 1
1 0 1 0
tr0 MAJ 1 1 0 0
clk0 R0
R0 R1 R2 MAJ 1 1 1 0 3-state_0
O_voter
0 0 0 0 LUT: 00011000_00011000 R0
0 0 1 0
tr1 MAJ 0 1 0 0
clk1 R1
0 1 1 1 O_voter 3-state_1
1 0 0 0
1 0 1 1 R1
tr2 MAJ 1 1 0 1
clk2 R2
1 1 1 1 3-state_2
O_voter
LUT: 00010111_00010111 R2
CLB flip-flop
Fig. 10.6 XTMR technique example
to correct the bit-flip of the flip-flops to avoid accumulation of errors. The voter used
in the output is based on a minority voter, where it blocks the output if this one is
different to the other two.
140
Full device scrub time (ms)
120
100
80
60
40
20
0
Virtex Virtex E Virtex-II Virtex-II Virtex-4 Virtex-5 Virtex-6 Virtex-7 Spartan-3 Spartan-3 Spartan-6
(250nm) (180nm) (150nm) Pro (90nm) (65nm) (40nm) (28nm) E/A (90nm) (45nm)
(130nm) (90nm)
528 400 3200 528 640
Maximum interface bitrate (Mbps)
Fig. 10.7 Full configuration or scrubbing time of different SRAM-based FPGAs from Xilinx
FPGA
TMR Majority Voter
TMR Majority Voter

voters and refresh
REDUNDANT
TMR register with
tr0 tr0 tr0
Majorty Voter
LOGIC (tr0)
TMR Output
REDUNDANT
LOGIC (tr1) tr1 tr1 tr1
REDUNDANT
LOGIC (tr2) tr2 tr2 tr2
Fig. 10.8 Voting insertion in XTMR or Global TMR
Scrubbing or reconfiguration that can be full or partial can be performed by the

internal block called ICAP or by the SelectMap interface. Scrubbing is mandatory
to correct the upsets (bit-flips) in the configuration memory bits. Figure 10.7 shows
the time for scrubbing in different FPGA families. The time for performing full
scrubbing can be significant in large FPGAs, more than 50 ms. This means that
some applications will run many clock cycles between two corrections of scrubbing.
Consequently, spatial redundancy techniques are very important to ensure the fault
masking at the output between scrubbings.
Scrubbing can be costly also in terms of power. In average, the scrubbing rate
should be at least three times faster than the upset rate of the FPGA. In order
to reduce the scrubbing rate, it is necessary to ensure that the spatial redundancy
technique, such as XTMR for example, can tolerate more than a single fault.
Some improvements can be done as playing with different TMR granularities by
voting insertion, as shown in Fig. 10.8 [12].
Another option is to use Diversity Triple Modular Redundancy (DTMR) to
increase the tolerance to multiple and accumulated upsets in the configuration
memory bits. A DTMR system is designed through the association of three
Fig. 10.9 Diversity Triple Modular Redundancy (DTMR)
Fig. 10.10 nMR-based technique with SAV voter
diversified copies to a majority voter [13]. Figure 10.9 illustrates an example of

a DTMR system. Each copy can be implemented by using a different architecture,
or different processors, different logic granularities and others.
Another option to increase the masking capability to multiple upsets is to use
nMR-based technique. The nMR is composed of n functionally identical modules,
which receive the same m-bits input and deliver p-bits output to the Self-Adapted
voter (SAv), Fig. 10.10 [14]. The SAv receives n p bits from all modules and
Fig. 10.11 Local TMR to mitigate SEU in the flip-flops
generates the fault-free p-output, n error status flags (ESF), and a non-masked fault
signal (NMF). In this scheme, the system allows for the accumulation of defective
modules, while remaining at least two modules without fault. SAv is a majority
voter, considering as population fault-free modules.
3.2 FLASH-Based and Antifuse-Based FPGAs
The configurations of flash-based and antifuse-based FPGAs are not sensitive to

radiation. So, faults as SET and SEU can only occur in the combinational and
sequential logic and they have a transient effect. So, well-known techniques used
in Application Specific Integrated Circuits (ASICs) can be applied in those FPGAs.
In case SEU can be observed in the user’s flip-flops of the FPGA, a technique
based on local TMR, where only the flip-flops are triplicated and vote, can work
well, Fig. 10.11.
In case SET is an issue and also SEU in the configurable flash memory cells, a
technique based on global TMR, where all the combinational logic and flip-flops are
triplicated and vote, can work well, as is illustrated in Fig. 10.12. Or the designer
can choose on using a temporal filtering technique, where the SET will be filtered
by the added chosen delay in the clock trees or in the logic path, Fig. 10.13. In this
case, the idea is that each flip-flop will capture the data at a different moment, and
at least two flip-flops out or three will have the fault-free value.
4 Radiation Test Methodologies to Predict and Measure

SER in FPGAs
The test of FPGAs under radiation depends on a test plan developed for each type
of FPGA and design architecture. Here we will detail the radiation test for SRAM-
based FPGAs. There are two types of tests: the static test and the dynamic test.
TR0 TRV0
combinational logic D0 clk0 TR1 voter OK
TR2
E0
D1 TR1 TRV1
combinational logic clk1 X TR0 voter OK
TR2
E1
D2 TR2 TRV2
combinational logic clk2 TR0 voter OK
TR1
E2
Fig. 10.12 Global TMR as XTMR
Combinational clk Combinational X

clk
logic logic
OK
MAJ MAJ
clk+d clk+d
clk+2d clk+2d
Fig. 10.13 Temporal Redundancy or SET Filtering Technique with delay d applied in the clock
trees
The static test experiment consists of configuring the FPGA with a golden bitstream
containing the test-design and then constantly read back the FPGA configuration
memory with the Xilinx iMPACT tool through the JTAG interface. In the experiment
control computer, the golden bitstream is compared against the readback bitstream.
If differences are found, the FPGA is reconfigured with the golden bitstream and
the differences are stored in the computer. Faults are defined as any bit-flip in the
configuration memory detected by the readback procedure. In this case, it is possible
to calculate the upset rate in the configuration memory bits for that specific particle
flux, expressed by particles/s/cm2 .
The cross-sections (¢) is the sensitive area of a circuit, where one particle
may cause an upset event, and it is calculated using Eq. (10.1), where flu-
ence D flux time, expressed by particles/cm2 . Cross-section is expressed in cm2 .
#events
¢D (10.1)
fluence
There are two types of cross-section: static and dynamic cross-section. Static
cross-section, also known as device cross-section (¢ device) is defined as the ratio
between the number of upsets in the configuration memory bits of the SRAM-based
FPGA (events) and the fluence of hitting particles. Usually, it is used a normalized
cross-section or bit cross-section (¢ bit), where the cross-section is divided by
the total amount of configuration memory bits. Static cross-section quantifies the
sensitivity of the FPGA technology to a specific radiation source.
On the other hand, dynamic cross-section is defined as the ratio between the
number errors observed at the output at design configured into the SRAM-based
FPGA (events), divided by the fluence of hitting particles. Dynamic cross-section
quantifies the sensitivity of the implemented design application to any specific
radiation source. The rate at errors is defined as the soft error rate (SER). In this
case, the expected error rate is much lower than the static test. Based on the Xilinx
Reliability Report [15], in average it is necessary 20 upsets in the configuration
memory bits to provoke one error in the design output. This relation may of course
vary according to the logic density, mapping, routing and the chosen architecture
for the design. This number can triplicate when XTMR is used. And the relation
can increase even higher, 6 or more times, if DTMR or nMR is used.
Notice that SER is proportional to both device size sensitivity (cross-section) and
flux as shown in Eq. (10.2).
SER D flux ¢SEU (10.2)
When a charged particle (as neutron, protons or heavy ions) hits a device, part of
its charge is deposited in the device. This is known as linear energy transfer (LET)
and expresses the energy loss per unit length (dE/dx) of a particle and is a function of
the mass and energy of the particle as well as the target material density. The units of
LET are commonly expressed as MeV–cm2 /g. Radiation experiments with charged
particles commonly relate the relation between cross-section and LET, which also
depends on the incidence angle.
Fig. 10.14 illustrates a setup experiment under neutron at ISIS Facility in United
Kingdom, where we could measure the static and dynamic cross-section of SRAM-
based FPGAs and later on by knowing the sea level neutron flux, we can infer the
SER of the device and circuit application running at ground level.
For example, the device was exposed for 22 min in a neutron flux of 4.11 104
and 70 events were counted. The fluence can be calculated and consequently by
using Eq. (10.1), the cross-section obtained is 1.29 10-6 cm2 . And by using Eq.
(10.2), the SER is 5.3 10-2 errors per second. By knowing particles flux at see
level is 13 neutrons/cm2 /h, the expected error rate at sea level can be infer to
1.29 10-6 cm2 13 neutrons/cm2 /h D 1.67 10-5 errors per hour.
Fig. 10.14 Neutron-induced SEE in FPGAs Experiment Setup
Conclusions
Each type of FPGAs presents a different susceptibility to soft errors and
according to its upset rate and architecture, distinct fault tolerance techniques
may be chosen. For experiments under neutron-induced faults, it is possible
to calculate the static and dynamic cross-section of SRAM-based FPGAs and
to analyze its tolerance to multiple and accumulated upsets. The static cross-
section can be measured for each SRAM-based FPGA device and the dynamic
cross-section must be measured for each fault tolerant design implemented in
the FPGA. For a certain device with a specific static cross-section, there will
be many dynamic cross-sections according to the fault tolerance technique
chosen.
References
1. J. Barth, “Applying Computer Simulation Tools to Radiation Effects Problems”, in: IEEE
Nuclear Space Radiation Effects Conference Short Course, NSREC, 1997.
2. O. Flament, J. Baggio, C. D”hose, G. Gasiot, J.L. Leray, “14 MeV neutron-induced SEU in
SRAM devices,” In Nuclear Science, IEEE Transactions on, vol. 51, no. 5, pp. 2908–2911,
Oct. 2004.
3. M. Berg, “Fault tolerance implementation within SRAM based FPGA designs based upon the
increased level of single event upset susceptibility,” On-Line Testing Symposium, IOLTS 2006.
4. P. E. Dodd and L. W. Massengill, “Basic mechanisms and modeling of single-event upset in

digital microelectronics,” IEEE Trans. Nucl. Sci., vol. 50, no. 3, pp. 583–602, Jun. 2003.
5. T. R. Oldham, F. B. McLean, “Total Ionizing Dose Effects in MOS Oxides and Devices,” IEEE
Transactions on Nuclear Science, vol. 50, no. 3, 2003. pp. 483-498.
6. F. L. Kastensmidt, R. Reis, L. Carro, Fault-Tolerance Techniques for SRAM-Based FPGAs
(Frontiers in Electronic Testing), Springer, 2006.
7. Actel. ProASIC3, IGLOO and SmartFusion Flash Family FPGAs Datasheet. [Online]. Avail-
able: http://www.actel.com/documents/PA3_HB.pdf and http://www.actel.com/documents/
IGLOO_HB.pdf
8. Wang, J.J., RTAXS Single Event Effects Test Rep., Aug. 2004 [available on-line at http://www.
actel.com/documents/RTAXS_SEE_Report.pdf]
9. Anghel, L., Alexandrescu, D., Nicolaidis, M., Evaluation of a soft error tolerance technique
based on time and/or space redundancy, in the Proceedings of Symposium on Integrated
Circuits and Systems Design, SBCCI, 13, 2000. p. 237-242.
10. L. Sterpone, M. Sonza Reorda, M. Violante, RoRA: Reliability-oriented Place and Route for
SRAM-based FPGAs, PRIME05: IEEE Ph.D. Research In Micro-Electronics & Electronics,
2005, pp. 147-150
11. L. Sterpone, D. Boyang, D. Merodio Codinachs, V. Ferlet-Cavrois, Accurate Mitigation of
Single Event Effects on Flash-based FPGAs: A new Design Flow. RADECS 2013.
12. F. Kastensmidt, L. Sterpone, M. Sonza Reorda, L. Carro. On the Optimal Design of Triple
Modular Redundancy Logic for SRAM-Based FPGAs, DATE2005: IEEE Design, Automation
and Test in Europe, 2005, pp. 1290-1295
13. L. Tambara, F. Kastensmidt, J. Azambuja, E. Chielle, F. Almeida, G. Nazar, L. Carro, P. Rech,
C. Frost. Evaluating the Effectiveness of a Diversity TMR Scheme under Neutrons, RADECS
2013.
14. J. Tarrillo, P. Rech, F. Kastensmidt, C. Valderrama, C. Frost, Neutron Cross-section of
N-Modular Redundancy Technique in SRAM-based FPGAs. RADECS 2013.
15. Xilinx, Inc. (2013) “Device Reliability Report Third Quarter 2013” [Online]. Available: http://
www.xilinx.com/support/documentation/user_guides/ug116.pdf
Chapter 11
Low Power Robust FinFET-Based SRAM
Design in Scaled Technologies
Sumeet Kumar Gupta and Kaushik Roy
Abstract FinFETs have emerged as alternatives to conventional bulk MOSFETs

in scaled technologies due to superior gate control of the channel, lower short
channel effects and higher scalability. However, width quantization in FinFETs
constrains the design space of FinFET-based circuits, especially SRAMs in which
transistor sizing is critical for the circuit robustness. The adverse effects of width
quantization can be mitigated by appropriate device-circuit co-design of FinFET-
based memories. This chapter describes some of such techniques with an emphasis
on the device-circuit interactions associated with each methodology. The impact
of different technology options in FinFETs like gate-underlap, fin orientation, fin
height, gate workfunction and independent control of the gates on the stability,
power and performance of 6 T SRAMs is discussed.
1 Introduction
The past few decades have seen the evolution of semiconductor industry driven
by technology scaling. Miniaturization of bulk Metal Oxide Semiconductor Field
Effect Transistors (MOSFETs) along with scaling of power supply voltage (VDD )
has provided the benefits of higher performance, lower power and larger integration
density. However, for the past few technology generations, scaling of transis-
tors has been becoming increasingly difficult. Short channel effects (SCEs) like
drain-induced barrier lowering [1], device threshold voltage (VT ) roll-off [1] and
degradation in the ratio of ON current (ION ) to OFF current (IOFF ) have started
to manifest themselves to a larger extent in scaled MOSFETs. Another critical
challenge in scaled technologies is the increasing impact of process variations on
the device, circuit and system performance due to random dopant fluctuation, line
edge roughness etc. [2]. Such detrimental effects in scaled technologies make the
design and optimization of devices and circuits extremely challenging.
S.K. Gupta • K. Roy ()

School of Electrical and Computer Engineering, Purdue University,
West Lafayette, IN 47907, USA
e-mail: kaushik@ecn.purdue.edu

DOI 10.1007/978-1-4614-4078-9__11
224 S.K. Gupta and K. Roy
Fig. 11.1 Device structures with single gate viz. (a) bulk MOSFETs (b) partially depleted (PD)
SOI MOSFETs (c) fully depeleted (FD) SOI MOSFETs and multiple gates viz. (d) planar double
gate (DG) MOSFETs (e) FinFETs and (f) gate-all around MOSFETs. The arrows show the electric
field lines in the structures
1.1 Alternate Device Structures
In order to meet the challenges posed by transistor scaling, different device

structures have been explored [3–5] as alternates to the bulk MOSFETs. Partially
depleted silicon-on-insulator (PD SOI) MOSFETs [3] reduce the drain electric fields
penetrating to the source side due to the buried oxide (BOX), leading to mitigation
of short channel effects (Fig. 11.1b). However, the body in PD SOI MOSFETs is
floating which leads to indeterminacy in the threshold voltage (VT ) of the device
due to variable body potential [3]. The undesirable floating body effects in PD
SOI MOSFETs can be mitigated by reducing the thickness of the silicon body and
making it fully depleted (FD-SOI MOSFETs—Fig. 11.1c) [4]. Reduced thickness
of the devices also leads to lower short channel effects due to reduced penetration of
drain electric fields to the source side. Short-channel effects can be further reduced
in multi-gate structures (Fig. 11.1) like double-gate (DG) MOSFETs [5] and gate
all-around (GAA) MOSFETs [6] which exhibit superior control of the channel
electrostatics by the gate potential, compared to single-gate structures. The focus
of this chapter is DG MOSFETs.
Planar DG MOSFETs, shown in Fig. 11.1d suffer from a fabrication issue, known
as gate misalignment [7]. Alignment of the two gates of a DG MOSFET is required
in order to effectively control the short channel effects and obtain maximum ION in
the device. However, from the point of view of fabrication, gate alignment in planar
DG MOSFETs is extremely challenging. FinFETs (Fig. 11.1e) in which the gate is
formed by depositing the gate metal/poly-silicon around the fin-shaped body [8–9]
11 Low Power Robust FinFET-Based SRAM Design in Scaled Technologies 225
makes it possible to easily achieve gate alignment in DG MOSFETs. Henceforth,

all the discussion on DG MOSFETs will pertain to FinFETs.
1.2 Fundamentals of FinFETs
FinFETs exhibit reduced short channel effects (compared to bulk MOSFETs) viz.
lower drain induced barrier lowering (DIBL), lower sub-threshold swing (SS) and
higher ratio of ON current to OFF current (ION /IOFF ) due to ultra-thin silicon body.
Lower fin thickness (tSi ) limits the electric field penetration from drain to source and
enhances the scalability of FinFETs. However, tSi cannot be scaled indefinitely. As
tSi decreases, quantum confinement effects start to play a bigger role, increasing
the threshold voltage (VT ) of the transistors and leading to loss in the device
performance. Moreover, device variability increases at lower tSi .
Ultra-thin body (UTB) in FinFETs along with channel control by two gates make
it feasible to use undoped (intrinsic) body, while still achieving excellent device
electrostatics. Due to undoped body in FinFETs, the variations due to random
dopant fluctuations (a major source of random variations in bulk MOSFETs) is
eliminated. However, due to UTB, an additional source of variation viz. body
thickness variation is introduced in FinFETs.
Short channel effects (SCEs) in FinFETs can also be controlled by reducing gate
dielectric thickness (tOX ). Lower tOX results in superior gate control of the channel,
but leads to increase in the gate leakage (IG ). Use of high-k dielectrics expands the
device design space by allowing the scaling of effective oxide thickness (EOT) with
higher physical thickness of the dielectric, limiting the gate leakage current in scaled
FinFETs.
Another device parameter which plays an important role in determining the char-
acteristics of FinFETs is the gate-source/drain (S/D) overlap (LOVS/D ). Larger LOVS/D
leads to decrease in effective channel length (LCH ) for the same gate length (LG )
which worsens SCEs in the device. In case the electrostatic integrity of the device
cannot be maintained by scaling tOX and/or tSi , a negative overlap between gate
and S/D, called underlap (LUN ) may be introduced to increase the effective channel
length without increasing LG . This achieves improved electrostatics and higher ION -
IOFF ratio. Gate workfunction (˚ G ) is another important design parameter which
can be changed to obtain the desired VT of the transistor [10].
The width of FinFET is along the height of the fin (HFIN -Fig. 11.1e) and for
ensuring mechanical stability of the device, the aspect ratio HFIN :tSi is limited [11].
For a particular technology, HFIN is fixed and hence, the width of the device can
only be increased in quanta of HFIN . This effect is known as width quantization.
Restricted device sizing due to width quantization in FinFETs leads to constrained
design space of FinFET-based circuits.
The effect of width quantization becomes even more critical for circuits like
SRAMs and flip-flops in which transistor sizing has a significant effect on the
Fig. 11.2 Device-Circuit-System interactions in FinFET technology
circuit functionality [12]. Conflicting requirements for transistor sizes for stable read
and write operations in SRAMs further aggravates the challenge posed by width
quantization.
The adverse effects of width quantization on FinFET-based circuits can be
mitigated by employing device-circuit co-design techniques [13–24]. Figure 11.2
illustrates different technology options provided by FinFETs. The two gates of
FinFETs can be tied together or controlled independently to enhance the design
flexibility of FinFET-based circuits. In addition, device parameters like fin thickness,
gate-source/drain underlap, gate workfunction, fin orientation and fin height can be
optimized considering the circuit and system requirements. This chapter explores
such device-circuit interactions with a focus on FinFET-based SRAMs. Before we
present the discussions on the device-circuit co-design methodologies, let us review
the fundamentals of 6 T SRAMs.
1.3 Fundamentals of 6 T SRAMs
6 T SRAM cell (Fig. 11.3) comprises of 6 transistors: 2 access transistors (AXL

and AXR), 2 pull-down transistors (NL and NR) and 2 pull-up transistors (PL and
PR). Pull-down and pull transistors form the cross-coupled inverters which maintain
the state of the cell in the hold mode. Read and write operations in the cell are
performed through access transistors which are controlled by a word-line (WL), as
described subsequently. Cells in a row of the array share the same word-line. The
cell comprises of two bit-lines (BL and BLB) which read from or write into the cell
and are shared amongst the cells in the same column.
Fig. 11.3 Schematic of

FinFET-based 6 T SRAM
Suppose the cell stores ‘0’ i.e. VQ D 0 and VQB D VDD . (Here, VQ and VQB are
voltages at nodes Q and QB). During the read operation, bit-lines BL and BLB are
pre-charged to VDD and word-line WL is asserted. VQ rises to some positive voltage
VREAD determined by the relative strengths of AXL and NL. At the same time BL
starts discharging while BLB remains at VDD . The difference in the voltages of BL
and BLB is amplified by the sense amplifier to obtain the value stored in the cell.
Since the storage node voltages are disturbed during the read, read stability of the
cell is less than the hold stability. If VREAD becomes greater than the trip point of
the inverter formed by AXR, PR and NR during read, the contents of the cell can
flip, causing a read failure [25]. During the write operation, BL is charged to VDD
and BLB is discharged to GND. On asserting WL, VQB is discharged to VWRITE ,
determined by the relative strengths of PR and AXR. Successful write occurs if
VWRITE is less than the trip point of the inverter formed by PL, AXL and NL.
With the basic understanding of FinFETs and cell operation of 6 T SRAMs, let
us discuss the design techniques for FinFET-based SRAMs. We describe the co-
optimization of fin ratio and tSi in Sect. 2, joint optimization of HFIN -tSi -tOX -VDD -VT
in Sect. 3, analysis of fin orientation in Sect. 4, spacer thickness optimization in
Sect. 5 and independent gate FinFETs in Sect. 6.
2 Co-optimization of Fin Ratio and Fin Thickness
Width quantization in FinFETs restricts the ratio of the transistor widths in 6 T

SRAM to integer values and results in constrained design space. As a result, the
conflict between read stability, write-ability, hold stability, performance and power
of 6 T SRAM is aggravated. The design space of FinFET-based 6 T SRAMs can be
expanded by co-optimizing the number of fins with the device design parameters.
In this section, we will explore the joint effect of the fin ratio and fin thickness (tSi )
on the cell stability, performance and leakage of FinFET-based 6 T SRAMs.
Let us start by explaining the effect of tSi on the device characteristics. Higher
tSi results in aggravated short channel effects leading to increase in IOFF and lower
ION -IOFF ratio. As in [13], let us perform the comparison of devices with different
Table 11.1 Comparison of FinFETs with different fin thickness (tSi ) at iso-
IOFF [13]. (IOFF D 25 nA/m)
tSi (nm) ˆG (eV) IDSAT (mA/m) IDLIN (mA/m) DIBL (mV/V)
7 4.61 2.18 0.88 51
9 4.73 1.95 0.92 115
11 4.80 1.56 0.8 194
tSi under iso-IOFF conditions, which is achieved by appropriately adjusting the gate
work-function (˚ G ). At iso-IOFF , lower short channel effects (SCE) in devices
with a reduced tSi tend to increase drain current (ID ). However, lower tSi results
in mobility degradation [26] and increase in source/drain (S/D) resistance, which
tends to decrease ID . At high VDS (saturation region), the effect of lower SCE
has a dominant effect on ID due to which saturation current (IDSAT ) increases with
decreasing tSi (Table 11.1 [13]). However, at low VDS (linear region), the impact of
SCE, mobility degradation and S/D resistance on ID are comparable. Due to this,
maximum value of linear current (IDLIN ) is obtained for tSi D 9 nm in Table 11.1.
With this discussion in mind, let us describe the effect of tSi on the cell stability,
performance and leakage of 6 T SRAMs.
During read, NL (Fig. 11.3) operates in the linear region while AXL is in
saturation. Due to increase in IDSAT at lower tSi and comparable IDLIN for different
tSi at iso-IOFF , the relative strength of AXL with respect to NL increases as tSi is
reduced. This results in higher VREAD which tends to lower the read static noise
margin (SNM). However, lower tSi also leads to higher ION -IOFF ratio, which tends
to increase the read SNM. Depending on the relative impact of the two effects
on read stability, read SNM might increase or decrease with tSi at iso-IOFF . The
access time of SRAM reduces with decreasing tSi due to increase in the strength
of the access transistor. Hold SNM increases for lower tSi due to higher ION -IOFF
ratio of the transistors. Write-ability of the cell reduces with decreasing tSi due to
reduction in the relative strength of AXR with respect to PR. Moreover, higher ION -
IOFF ratio at reduced tSi makes it difficult to flip the contents of the cell and lowers
the write-ability. Since the comparison is performed under iso-IOFF conditions, the
cell leakage is similar for different tSi .
Let us, now, discuss the impact of fin ratio on SRAM characteristics. We define
NAC , NPD and NPU to be the number of fins in the access, pull-down and pull-up
FinFETs, respectively. Increasing NPD :NAC results in lower VREAD and higher read
stability. Improvement in read stability can also be achieved by increasing NPU :NPD
which leads to higher logic threshold voltage (VM ) of the inverter formed by PR and
NR (see Fig. 11.3). Write-ability of the cell can be increased by increasing NAC :NPU ,
which results in lower VWRITE or by increasing NPU :NPD which results in higher
VM of the inverter formed by PL and NL. Hold stability improves with increasing
NPU :NPD . Note, the conflicting requirements for the fin ratios for achieving high
read stability, write-ability and hold stability. As an example, if NAC is decreased
to increase NPD :NAC and improve read stability, then NAC :NPU decreases, resulting
Fig. 11.4 Variability of gate

delay versus number of
fins [25]
Fig. 11.5 Layouts of FinFET-based 6 T SRAM showing increase in area with increasing number
of fins: (a) layout with using one/two fins in the transistors (b) layout for a high performance cell
in degradation in write-ability. Decreasing NPU results in improved write-ability;

however, this leads to lower NPU :NPD and reduced read and hold stabilities. Access
time decreases with (a) higher NAC due to increase in the strength of AXL and
(b) higher NPD due to lower VREAD which increases the gate-to-source voltage (VGS )
and drain to source voltage (VDS ) of AXL, increasing its strength. Larger number
of fins results in higher cell leakage. However, at the same time, more fins result
in lower parameter variations [27] as the variation effects get averaged out for
larger number of fins. This is illustrated in Fig. 11.4 which shows a decrease in
the variability of gate delay as the number of fins increases [27].
Increase in the number of fins may also lead to increase in the cell area.
Considering the spacer lithography technology for the fabrication of FinFETs [28],
FinFETs with 2i-1 and 2i fins can be fabricated in the same lithographic pitch (Here,
i is a natural number). For minimum area of the layout, all transistors with one fin
each may be used. To increase the cell stability, number of fins in the pull down
transistors can be increased to 2, as shown in Fig. 11.5. For high performance
cells, number of fins in the access transistors and the pull down transistors may
be increased at the cost of area (Fig. 11.5).
Considering the trade-offs discussed above, the design space for FinFET-based
6 T SRAM comprising of the fin ratio and fin thickness can be explored. As an
example, the joint impact of fin ratio and tSi on the read stability and write-ability
Fig. 11.6 (a) Read SNM and (b) Write margin as a function of fin ratio and fin thickness [13]
is shown in Fig. 11.6 [13] under iso-IOFF conditions. Increase in read SNM with
decreasing tSi can be observed and is attributed to improved ION -IOFF ratio for
low tSi , as explained before. However, higher ION -IOFF ratio for low tSi results in
degradation in write margin (WM). Read SNM and WM for different fin ratios
are also shown in Fig. 11.6. Fin ratio NP U W NAC W NPD D 1 W 1 W 2 provides
the optimal balance between read stability, write-ability and cell area. The plots
in Fig. 11.6 can be used for the co-optimization of tSi and fin ratio to achieve the
optimal cell stability of FinFET-based SRAMs.
3 Joint Optimization of Fin Height, Fin Thickness, Oxide

Thickness, Supply Voltage and Threshold Voltage
Due to quasi-planar structure of FinFETs, the device current can be increased by

increasing the fin height (HFIN ) with a minimal increase in the device footprint. This
is in contrast with the planar MOSFETs, in which the increase in the device footprint
is similar to that in the device width. The minimal increase in the device footprint
with increasing HFIN in FinFETs is due to the fact that in order to maintain the
mechanical stability, the aspect ratio HFIN :tSi has to be below a critical value. Hence,
increase in HFIN is accompanied by a proportional increase in tSi and consequently,

a small increase in the device footprint. Increase in tSi leads to aggravation of
short channel effects in FinFETs. In order to maintain the electrostatic integrity
of the device, gate dielectric thickness (tOX ) can be reduced. However, tOX scaling
is limited by the gate current. As a result, tSi and hence, HFIN cannot be increased
beyond a certain value. In order to increase the device strength further, one has
to increase the number of fins, which leads to width quantization in FinFETs, as
explained earlier. However, for small increase in the device current, the optimization
of HFIN can be a potentially useful technique, which we discuss in this section.
Increase in the device strength with the increasing HFIN allows (a) increase in
threshold voltage (VT ) to achieve exponential reduction in sub-threshold leakage
and (b) VDD scaling to achieve lower leakage and dynamic power with a similar read
access time. Considering these options, two kinds of co-design techniques have been
explored in [14]: (a) joint optimization of HFIN -tSi -tOX -VT (or device-VT ) design
space and (b) joint optimization of HFIN -tSi -tOX -VT -VDD (or device-VT -VDD ) design
space. The desired VT is obtained by adjusting the gate workfunction appropriately.
In the device-VT co-optimization, HFIN is increased along with increase in tSi
and reduction in tOX , keeping VDD constant. The gate workfunction is adjusted
to increase VT and achieve similar ION per fin. On the other hand, device-
VT -VDD co-optimization involves increasing HFIN and tSi , reducing tOX , scaling
VDD and increasing VT to achieve iso-ION per fin. VDD scaling results in lower
short channel effects, which allows higher tOX to be used in device-VT -VDD co-
optimization compared to device-VT co-optimization, leading to lower gate leakage
(Fig. 11.7a). Exponential reduction in sub-threshold current due to increase in VT
and consequently, higher ION -IOFF ratio for larger HFIN in device-VT co-optimization
leads to increase in read (Fig. 11.7b) and hold stabilities. For device-VT -VDD co-
optimization, higher ION -IOFF ratio for larger HFIN tends to increase read and hold
stabilities. However, operation at lower VDD for devices with larger HFIN tends to
decrease the read and hold stabilities. As a result, an optimal point in Fig. 11.7
is observed for design point D8. The optimal design point can be chosen by
considering the impact of the co-optimization techniques on cell leakage and SNM.
In Fig. 11.7a, the design point D9 (see Table 11.2) yields minimum leakage along
with a moderate improvement in read SNM for both device-VT and device-VT -
VDD co-optimizations. The read access time remains similar across different design
points (Fig. 11.7c) since the devices are designed under iso-ION . Increasing HFIN
leads to increase in capacitance at the storage nodes which leads to increased
immunity to soft errors due to increase in the cell charge (Fig. 11.7d). However,
increase storage node capacitance results in higher write time (Fig. 11.7c).
In summary, joint optimization of HFIN , tSi , tOX, VT and VDD leads to increase
in static noise margin, lower cell leakage and improved immunity to soft errors at
iso-read access speed at the cost of increase in write access time.
Fig. 11.7 (a) Cell leakage (b) read SNM (c) read and write access time and (d) cell charge for
different design points for Device-VT and Device-VT -VDD co-optimizations [14]
Table 11.2 Device Design name D7 D8 D9 D10 D11

parameters for design points
at iso-ION per fin [14] TSi (nm) 7 8 9 10 11
HFIN (nm) 35 40 45 50 55
HFIN -VT Co-optimization
VDD (V) 1:0 1:0 1:0 1:0 1:0
TOX (nm) 2:4 2:1 1:8 1:5 1:2
NMOS ˆG (eV) 4:6 4:72 4:82 4:89 4:95
PMOS ˆG (eV) 4:8 4:68 4:58 4:51 4:45
HFIN -VT- VDD Co-optimization
VDD (V) 1:0 0:95 0:9 0:85 0:8
tOX (nm) 2:4 2:15 1:9 1:65 1:4
NMOS ˆG (eV) 4:6 4:67 4:71 4:73 4:74
PMOS ˆG (eV) 4:8 4:73 4:69 4:67 4:66
Fig. 11.8 Fins aligned along

different crystal orientations
using fin rotation [15]
Fig. 11.9 Electron (NMOS)

and hole (PMOS) mobilities
along different crystal
orientations [15]
4 Fin Rotation and Orientation
Quasi-planar structure of FinFETs enables the alignment of transistors along a

particular crystal orientation by fin rotation (Fig. 11.8). Due to the difference in the
carrier effective masses, FinFETs with different fin orientations exhibit difference
in mobilities and scattering rates [15]. Figure 11.9 shows the mobility for different
crystal orientations for electrons and holes [15]. NMOS shows a higher mobility
along (100) direction while for PMOS, the highest mobility is along (110) direction.
Hence, the drive strength of FinFETs can be optimized by aligning them along
appropriate crystal orientations. In this section, we investigate the impact of fin
rotation and orientation on the characteristics of 6 T SRAMs.
Figure 11.10a shows read SNM for different fin orientations of the access (AX),
pull-down (PD) and pull-up (PU) transistors. Maximum read SNM is obtained
when AX is along (110) direction and PD is along (100) direction due to reduced
Fig. 11.10 (a) Read SNM (b) write margin (c) hold SNM and (d) access time for different fin
orientations [15]
strength of AX relative to PD resulting in lower VREAD . The optimal orientation

of PU for maximum read stability is (110) which increases the strength of PU
leading to higher VM of the cross-coupled inverters. However, from the point of
view of write margin, (100) orientation of PU is the most superior (Fig. 11.10b).
This is because of reduced strength of PU relative to AX leading to lower VWRITE .
For AX, (100) orientation yields maximum write-ability due to increase in the
strength of access transistor. For hold stability, stronger PU ((110) orientation) as
well as PD ((100) orientation) leads to maximum benefits (Fig. 11.10c) due to
increase in the coupling between the two inverters. For access time, strong AX
and PD (both (100) orientations) yield the minimum access time (Fig. 11.10d). Fin
rotation and orientation enhances the design flexibility of FinFET based SRAMs
by mitigating the effect of width quantization and allowing finer modulation of the
transistor strengths. Compared with all (110) transistors (conventional orientation
for (100) wafer), multi-oriented FinFETs in a 6 T SRAM show benefits in terms
of increased read stability with a negligible effect on the write margin. However,
when compared with all (100) FinFETs (conventional orientation for (110) wafer),
optimization of fin orientation shows less improvements in the cell stability. Thus,
the effectiveness of fin orientation optimization depends on the wafer orientation
and yields maximum benefits for (100) wafers. However, on a (100) wafer, rotation
of fins along (111) direction is not feasible, which restricts the combinations of fin
orientations that can be used in a 6 T SRAM. Fin orientation also leads to area
penalty and is expected to be more susceptible to process variations.
Fig. 11.11 (a) Conventional FinFET (b) Symmetric Spacer Extension (SSE) FinFET (c) LgDev
FinFET with larger gate length and (d) Asymmetric Drain Spacer Extension (ADSE) FinFET
5 Spacer Thickness Optimization
FinFETs exhibit significant reduction in the junction capacitance compared to bulk

MOSFETs due to lower source/drain (S/D) junction cross-section. As a result,
overlap capacitance becomes the dominant component of the parasitic capacitance
in FinFETs. In order to reduce the parasitic capacitance, a negative gate-S/D overlap
(called underlap) can be introduced in FinFETs by optimizing the spacer thickness.
In addition to the capacitance, gate underlap has a strong impact on the device
electrostatics, and in turn, on the device characteristics like ION , IOFF , sub-threshold
swing, gate leakage and drain-induced barrier lowering (DIBL). In this section, we
discuss the optimization of spacer thickness and evaluate the impact of introducing a
gate underlap in FinFETs on 6 T SRAM characteristics. We perform the analysis for
FinFET structures with both symmetric and asymmetric gate underlap (Fig. 11.11).
5.1 Symmetric Spacer Optimization
Figure 11.11b shows the FinFET structure with symmetric gate underlap (Sym-
metric Spacer Extension (SSE) FinFET) [16]. Gate sidewall spacer thickness is
increased by LSP on both sides of the gate to introduce an equal gate underlap
on the source and drain sides. Gate underlap results in (a) increase in the effective
channel length of the device and (b) change in the device electrostatics due to weaker
gate control in the underlapped portion of the channel. Due to these two effects, ION
degrades with increasing LSP . However, at the same time, FinFETs with higher
LSP exhibit lower capacitance due to the reduction in the gate overlap. ION and
Fig. 11.12 Dependence of

ON current, capacitance and
intrinsic device delay
(CD V/ION ) on LSP [16]
Fig. 11.13 Sub-threshold

and edge tunneling leakage as
a function of LSP [16]
capacitance versus LSP is shown in Fig. 11.12. It can be observed that for small
LSP , ION degradation is small and is due to increase in the effective channel length.
However, for large LSP , sharp decrease in ION is observed and is attributed to
a significant loss of gate control over the channel, especially on the source side.
This effect is explained in a greater detail in the next sub-section. Figure 11.12
also shows a decrease in gate and drain capacitance (CG and CD ) with increasing
LSP . The net effect of decreasing ION and capacitance on intrinsic device delay
(CD V/ION ) can be observed in Fig. 11.12 which shows a minimum in CD V/ION for
an optimal LSP . Thus, symmetric spacer optimization can lead to lower intrinsic
device delay. Moreover, increase in the effective channel length leads to lower
short channel effects viz. lower DIBL, lower sub-threshold swing and reduced sub-
threshold leakage with increasing LSP (Fig. 11.13). Reduced gate overlap also
results in lower edge tunneling current (IET ), as shown in Fig. 11.13.
Let us now explain the impact of symmetric underlap in FinFETs on SRAM
stability, leakage and performance. Improvement in short channel effects of the
transistors with increasing LSP results in higher read and hold SNM, leading to
reduction in the data retention voltage (DRV). At the same time, lower intrinsic
Table 11.3 Comparison of SSE FinFET SRAM with conventional (Conv.) and LgDev FinFET
SRAMs [16]
Metric Conv. (LSP D 0) SSE (LSP D 12 nm) LgDev (LSP D 12 nm)
Read SNM (mV) 129 132 137
Write Time (ps) : Nominal 25.4 25.7 23.8
:Worst Case 54.2 42.3 52.9
Worst Case Data 220 160 150
Retention Voltage (mV)
Leakage Power (nW): PSUB 13.05 2.52 2.10
(T D 27 ı C) : PGATE 34.05 13.78 60.9
:PTOT 47.1 16.3 63
delay of the optimized devices (as explained earlier) and lower storage node
capacitance result in lower write time. Moreover, lower sub-threshold leakage leads
to significant reduction in the cell leakage. This comes at the cost of (a) increase
in cell area due to increase in the device footprint and (b) higher access time due
to lower strength of the access transistors. Table 11.3 shows the comparison of
conventional FinFETs and SSE FinFETs [16]. Comparison is also performed with
a device which has the same spacer thickness as conventional FinFETs but with
channel length equal to that of SSE FinFET (see LgDev in Fig. 11.11c). Note, LgDev
has a larger gate length compared to SSE FinFET. Larger read SNM and lower DRV
is observed in LgDev based SRAM, compared to SSE FinFET SRAM because ION -
IOFF ratio in LgDev is higher due to superior gate control. However, since the overlap
capacitance is higher in LgDev, the worst case write time is larger compared to SSE
FinFET based SRAM. Moreover, large gate length results in higher gate current in
the ON state due to larger cross-sectional area of the gate. This leads to higher cell
leakage in LgDev FinFET SRAM.
5.2 Asymmetric Drain Spacer Extension FinFETs
In the previous sub-section, we discussed the pros and cons of SSE FinFETs. One of
the drawbacks of introducing gate underlap on the source side is that carrier injection
from the source to the channel is significantly affected by the underlap, which results
in a large degradation in ION for large LSP . Instead of introducing underlap on both
sides, if the underlap is introduced only on the drain side, the degradation in ION
can be mitigated. Figure 11.11d shows the FinFET structure with asymmetric gate
underlap (Asymmetric Drain Spacer Extension (ADSE) FinFET) [17]. Asymmetry
in the structure of ADSE FinFETs results in unequal ID for VDS > 0 and VDS < 0.
This asymmetry is used to mitigate the read-write conflict in 6 T SRAMs. In order
to explain the benefits of ADSE FinFET based SRAM, let us discuss the impact of
asymmetric gate underlap on the device characteristics.
Fig. 11.14 Conduction band profile of ADSE FinFETs (a) at different VGS (b) for different LSP
(c) at different positive VDS and (d) at different negative VDS [17]. The thick line on the device
symbol indicates the underlapped terminal
As in SSE FinFETs, gate underlap in ADSE FinFETs leads to increase in the

channel length and change in the device electrostatics. Due to weaker gate control
in the underlapped part of the channel, an extra potential barrier is introduced on
the drain side at high VGS (Fig. 11.14a). The height of the extra potential barrier
increases with increasing LSP (Fig. 11.14b). Figure 11.14c shows that the positive
drain bias (VDS ) pulls down the extra potential barrier (Fig. 11.14c). On the other
hand, negative VDS has a small effect on the extra potential barrier since positive
bias is applied on the non-underlapped terminal (Fig. 11.14d).
Increase in the channel length and extra potential barrier in the underlapped part
of the channel leads to reduction in ION (Fig. 11.15), as in SSE FinFETs. A knee
point (LKP ) can be observed in Fig. 11.15. For LSP < LKP , decrease in ID of ADSE
FinFETs is gradual and matches with that of LgDev. This implies that slow decrease
in ID is due to increase in channel length. For LSP > LKP , sharp decrease in ID
is due to increase in extra potential barrier. For VDS > 0, an increase in LKP with
increasing VDS can be observed (see Fig. 11.15). This is due to the fact that positive
VDS lowers the extra potential barrier (Fig. 11.14c), which results in the onset of
sharp decrease in ID at a higher LSP . For VDS < 0, LKP is relatively insensitive
to VDS (Fig. 11.15) because of weak dependence of the extra potential barrier on
Fig. 11.15 Drain current versus LSP of ADSE and LgDev FinFETs for (a) VDS > 0 and
(b) VDS < 0 [17]
Table 11.4 Comparison of Improvement Improvement over SSE

ADSE FinFET SRAM with Parameter over conventional (iso-device-footprint)
conventional and SSE
ION /IOFF 16X 2.3X
FinFET SRAMs
DIBL 51 % 20 %
SS 12.6 % 3.5 %
CGD 29 % 27 %
negative VDS (Fig. 11.14d). The difference in the dependence of extra potential
barrier and hence, LKP , on the polarity of drain bias creates an asymmetry in current
for positive and negative VDS .
In addition, ADSE FinFETs show improved short channel effects due to increase
in the channel length. Reduction in IET is also observed with increasing LSP for
positive VDS [17]. The comparison of ADSE FinFETs with conventional FinFETs
and iso-device footprint SSE FinFETs is summarized in Table 11.4. ADSE FinFETs
show improvement in short channel characteristics and lower gate-drain capacitance
compared to iso-area SSE FinFETs. In addition to superior device characteristics,
ADSE FinFETs exhibit asymmetry in ID for VDS > 0 and VDS < 0. This unique
feature of ADSE FinFETs is exploited to mitigate the read-write conflict in 6 T
SRAMs, which we discuss next.
Figure 11.16 shows the ADSE FinFET based 6 T SRAM. The underlapped
terminal is indicated by the thick lines. During the read operation, VDS across
AXR is negative, which results in lower ID due to extra potential and hence, lower
strength of AXR relative to NR. During the write operation, VDS across AXL is
positive which reduces the effect of the extra potential barrier on ID and increases the
strength of AXL relative to PL. This leads to the mitigation of the read-write conflict
(Fig. 11.17). Moreover, improved short channel effects in ADSE FinFETs results in
lower cell leakage (ILEAK ) and hold stability. At the same time, lower storage node
capacitance results in lower write time. However, at large LSP , degradation in ION
Fig. 11.16 Schematic of ADSE FinFET based 6 T SRAM showing current direction during
(a) read and (b) write [17]
becomes dominant over the reduction in capacitance which leads to increase in write
time. Weak access transistors during read leads to increased access time. ADSE
FinFETs also have larger cell area due to increased device footprint. Figure 11.17
also shows comparison between ADSE and LgDev FinFET SRAMs. ADSE FinFET
based SRAM exhibits superior read and write stability compared to LgDev based
SRAM due to asymmetry in ADSE FinFETs which mitigates the read-write conflict.
Moreover, ADSE FinFET SRAM shows lower cell leakage. IGATE (gate leakage
current) decreases with LSP for ADSE FinFET based SRAM. This is because of
reduced edge tunneling current and negligible change in direct tunneling current
(since gate length LG does not change with LSP ). On the other hand, LgDev based
SRAM shows an increase in IGATE due to increase in LG which results in larger
direct tunneling current. Also, larger storage node capacitance in LgDev leads to
larger write time. However, LgDev FinFET SRAMs show lower access time than
ADSE FinFET SRAMs due to higher strength of the access transistors. Moreover,
superior short channel control in LgDev results in larger hold stability.
6 SRAMs based on Asymmetrically Doped (AD) FinFETs
In the previous section, we discussed that transistors with asymmetric gate underlap
leads to the mitigation of the read-write conflict in SRAMs. In this section, we
discuss another technique of introducing asymmetry in the transistor by unequally
doping the source and drain of the transistors, as shown in Fig. 11.18. Similar to
the ADSE FinFETs, AD FinFETs exhibit unequal currents for positive and negative
VDS [23]. The current flowing from the terminal with lower doping (i.e. when the
terminal with larger doping acts as the source) is higher than the current flowing
in the other direction (i.e. when the terminal with lower doping acts as the source).
Fig. 11.17 (a) Read SNM (b) write margin (c) hold SNM (d) access time (e) write time and (f) cell
leakage versus LSP for ADSE FinFET and LgDev FinFET SRAMs [17]
The asymmetry in current increases as the doping concentration of the drain terminal
is reduced, with the source doping kept fixed. In order to achieve the mitigation of
the read-write conflict, the terminal with lower doping is connected to the storage
nodes (Fig. 11.18). During the read operation, the strength of the access transistor
reduces, increasing the read stability. During the write operation, higher drive of the
Fig. 11.18 (a) Asymmetrically Doped (AD) FinFETs (b) AD FinFET based SRAM and (c)
comparison of the metrics of AD FinFET SRAM with standard FinFET SRAM (Std.) [23]
access transistors leads to increase in the write-ability. In addition lower cell leakage
is achieved at the cost of increase in the access time. Unlike ADSE FinFET SRAM,
there is no area penalty associated with this technique.
7 Independent Gate FinFETs
Two gates in FinFETs offer an option to modulate the transistor strength by

independently controlling the front and back gates (FG and BG, respectively). Drain
current in independent gate (IG) FinFETs is strongly dependent on the coupling
between FG and BG. Due to ultra-thin body, FinFETs exhibit strong FG-BG
coupling which increases as tSi is reduced in scaled technologies. As an example,
Fig. 11.19 shows the ID -VGS characteristics of tied-gate (TG) FinFETs and IG-
FinFETs (with back gate voltage (VBG ) D 0) for LG D 32 nm and 10.8 nm with
corresponding tSi D 9 nm and 4.5 nm, respectively. The electric fields emanating
from BG reduce the inversion charge in IG FinFET formed due to FG. This leads
to the ratio of ID of IG and TG FinFETs less than 0.5. FinFETs with tSi D 4.5 nm
exhibits stronger FG-BG coupling compared to the FinFET with tSi D 9 nm due
to lower body thickness and increase in the quantum mechanical effects. As a
result, the ratio of the currents of IG and TG FinFETs is smaller for FinFET with
tSi D 4.5 nm.
Fig. 11.19 ID -VG

characteristics of FinFETs
with tSi D 9 nm and 4.5 nm
showing smaller ratio of ON
current of IG with respect to
TG for tSi D 4.5 nm
Fig. 11.20 (a) Schematic and (b) thin-cell layout of BMIG FinFET SRAM [20]
Independent control of two gates in IG FinFETs can be exploited to enhance the

design flexibility of FinFET based SRAMs and mitigate the read-write conflict. In
this section, we discuss some of the design techniques for SRAM using IG FinFETs
and compare each with standard tied gate (TG) FinFET SRAM.
7.1 Bi-mode Independent Gate FinFET SRAMs
Figure 11.20 shows the IG FinFET SRAM proposed in [20], in which the two gates
of the access transistors are controlled by two word-lines. The strength of the access
transistor can be modulated between two values in the ON state by independently
asserting/de-asserting the two wordlines. (Henceforth, the IG FinFET SRAM in
Fig. 11.20 is called bi-mode independent gate (BMIG) FinFET SRAM). The two
ON state modes for BMIG FinFETs are, henceforth, called high strength (HS)
and low strength (LS) modes. During the read operation, only RWL is asserted
(LS mode). Due to weak access transistor, read stability increases. This allows
one fin to be used in the pull down transistors (instead of two, as in standard
FinFET—see Sect. 2), thus reducing cell leakage. During the write operation,
Fig. 11.21 Two BMIG SRAM cell showing pseudo-read problem during write
both WL and RWL are asserted (HS mode), increasing the strength of the access
transistor. Hence, the conflicting requirements for the access transistor for high read
stability and write-ability are mitigated. However, this technique incurs a cell area
penalty (Fig. 11.20b), compared to TG FinFET SRAM due to the back gate contact
of the access transistors. Moreover, weak access transistor results in significant
degradation in the access time and therefore, limits the application of BMIG FinFET
SRAM, as proposed in [20], to low power low throughput applications. Further,
decreasing the number of fins in the pull-down transistors to 1 while using the access
transistor in the HS mode during write worsens the pseudo-read issue [29, 30] in
BMIG FinFET SRAM array, as illustrated in Fig. 11.21. Suppose write operation
needs to be performed on cell A for which the bitlines are charged to the appropriate
voltages. For another cell B in the same row, belonging to a different word, the
bitlines are precharged to VDD so that no new value is written on that cell. On
asserting the wordlines, the node of cell B storing ‘0’ gets charged to some positive
voltage VPSEUDO-READ depending on the resistive divider action of AXL and NL,
similar to the read operation. In the presence of process variations, VPSEUDOREAD
may become larger than the trip point of the inverter formed by AXR, PR and NR
causing the contents of cell B to flip. This is known as the pseudo-read or half-select
problem. In BMIG FinFETs, the relative strength of access transistors with respect
to pull-down transistor is higher than TG FinFET SRAM since a single fin is used in
pull-down FinFET. This worsens the pseudo-read problem in BMIG FinFET SRAM
compared to TG FinFET SRAMs. Mitigation of the pseudo-read problem can be
achieved by architectural techniques like write-back scheme [31] in which the whole
row is written simultaneously. Table 11.5 summarizes the differences between TG
and BMIG FinFET SRAMs [20].
In addition to low throughput applications, BMIG FinFETs are also suitable for
dynamic voltage frequency (DVF) scalable SRAMs [22]. During high workload
conditions (high voltage - VDDH operation), both the wordlines are asserted (HS
mode) during read and write to achieve fast operations. At low workload conditions,
Table 11.5 Comparison of Metric TG BMIG

BMIG FinFET SRAM [20]
with TG FinFET SRAM Read SNM (mV) 210 240
Access time (Norm.) 1 2.2
Write time (Norm.) 1 1
Fig. 11.22 VDD versus the

inverse of cell access time for
BMIG FinFET SRAM
showing HS and LS modes of
operation and switching
between the two modes. HS
mode is switched to LS mode
as transition is made from
FB3 to FB2 and vice versa.
The plot corresponds to the
worst case process corner for
read stability [22]
the cell is operated at low voltage (VDDL ). Although the frequency requirements
are low, VDD scaling may not be possible in conventional TG FinFET SRAM due
to increase in read and write failures. However, in BMIG FinFET SRAM, only
one wordline is asserted (LS Mode) during read at VDDL which increases the read
stability and allows low voltage operation. During write at VDDL , both the wordlines
are asserted (HS mode) to increase the strength of the access transistors and achieve
low write failures. Since both the wordlines are asserted during read at VDDH , two
fins need to be used in the pull-down FinFETs to achieve high read stability. This
makes the operation of BMIG FinFET SRAM at VDDH identical to TG FinFET
SRAM. Also, the cell leakage of BMIG FinFET SRAM is the same as TG FinFET
SRAM. Moreover, the pseudo-read issue in BMIG FinFET SRAM with a single
fin in the pull-down FinFET discussed above is mitigated in DVF scalable BMIG
FinFET SRAM since the write operation of TG and BMIG FinFET SRAMs are
identical.
Figure 11.22 shows VDD versus the inverse of access time for BMIG FinFET
at worst case process corner for read stability, illustrating the switching between
LS and HS modes during the read operation. In the frequency band FB4, access
FinFET is operated in the HS mode since the frequency requirements are high. In
FB3, VDD scaling is precluded by read failures in the HS mode. The cell is still
operated in the HS mode since LS mode of operation needs higher VDD to meet the
frequency requirements in FB3. As transition is made from FB3 to FB2, the access
transistor is switched to the LS mode, which allows further VDD scaling. Finally,
VDD scaling is limited by read failures in the LS mode in FB1. Note, in TG FinFET
SRAM, VDD scaling is not possible at low frequencies due to read failures; however
due to bi-modal operation during read, BMIG FinFET SRAM exhibits significantly
lower minimum supply voltage (VMIN ) at low frequencies. The write operation of
BMIG FinFET is identical to TG FinFET SRAM, hence write-ability of the cell is
not affected. Thus, read-write conflict is mitigated in BMIG FinFET SRAM at low
frequencies, at the cost of larger cell area due to back gate contacts of the access
transistors [22].
7.2 Tri-Mode Independent Gate FinFET SRAMs
In Sect. 7.1, we discussed the benefits of BMIG FinFETs for DVF scalable
SRAMs. The bi-modal operation of BMIG FinFET SRAMs results in significant
reduction in VMIN at low frequencies, compared to TG FinFET SRAM. However, no
improvement in write-ability is achieved. In this sub-section, we discuss tri-mode
independent gate (TMIG) FinFET SRAM which achieves improvement in write-
ability along with increase in read stability at low frequencies and similar access
time for high frequency operation [22].
This technique exploits the fact that spacer lithography technology used for
fabrication of FinFETs [26] offers a single-fin and two-fin transistors in the same
lithographic pitch. Thus, depending on the contact size and the fin pitch, the
difference between the device footprint of FinFETs with one and two fins may be
minimal. In TG and BMIG FinFET SRAMs, one fin is used in the access transistor,
which requires etching of one of the fins after the two fins are fabricated using
spacer lithography technique. However, in TMIG FinFET SRAM, the second fin
is not etched away (Fig. 11.23). Instead, the gate metal is etched from the top
of Fin2 (Fig. 11.23) to allow independent control of the two gates of Fin2. One
of the gates of Fin2 is connected to the two gates of Fin1 (FG) while the other
gate of Fin2 is controlled by BG. This structure allows three-way modulation
of the strength of the access transistor, as shown in Fig. 11.24. Let us define
VFG and VBG to be the voltages applied at the FG and BG of TMIG FinFET.
Therefore, (VFG ,VBG ) D (VDD ,VDD ) leads to the highest strength of TMIG FinFET
(high strength (HS) mode) followed by (VFG ,VBG ) D (VDD ,0) (medium strength
(MS) mode) and (VFG ,VBG ) D (0,VDD ) (low strength (LS) mode).
Figure 11.25 shows the schematic and layout of TMIG FinFET based 6 T SRAM.
An increase in area in TMIG FinFET SRAM compared to TG FinFET SRAM is
Fig. 11.23 TMIG (tri-mode

independent gate) FinFET
[22]
Fig. 11.24 Four

configurations of TMIG
FinFET [22]
Fig. 11.25 (a) Schematic and (b) thin-cell layout of TMIG FinFET SRAM
observed due to the back gate contacts of the access transistors, similar to BMIG
FinFET SRAM. Also note, an increase in x-dimension of TMIG FinFET SRAM
cell allows three fins to be used in the pull down transistor without further increase
in the area. The two wordlines modulate the strength of the access transistor between
three values and achieve reduction in cell VMIN along with the mitigation of read-
write conflict, as explained next.
During write operation, TMIG FinFETs are operated in the HS mode by asserting
both the wordlines. This leads to increase in the write-ability of the cell. During the
read operation at high workload conditions, the access transistors are operated in
the MS mode by asserting only WL_WRH to achieve similar access time and read
stability as TG FinFET SRAM. Read operation during low workload conditions is
performed by operating TMIG FinFETs in the LS mode to increase read stability
and allow low voltage operation. The switching between MS mode and LS modes
of operation during read in TMIG FinFET SRAM is similar to that between HS and
LS modes in BMIG FinFET SRAM, as described in Sect. 7.1 (see Fig. 11.22). Due
to larger number of fins the TMIG FinFET SRAM, cell leakage increases. However,
note, MS mode in TMIG FinFET SRAM has a larger strength compared to a single
fin access transistor in TG FinFET SRAM. Therefore, VT of TMIG FinFET SRAM
can be increased to achieve reduction in cell leakage while still achieving similar cell
access time at high VDD compared to TG FinFET SRAM. Hold stability in TMIG
FinFET SRAM is slightly degraded compared to TG and BMIG FinFET SRAM due
to increase in fin ratio of pull-down and pull-up FinFETs (NPD :NPU ).
Figure 11.26 shows the iso-leakage comparisons of VDD versus frequency curves
for TMIG and TG FinFET SRAMs at the process corners that are worst from the
point of view of read stability and write-ability. It can be observed that dynamic
switching between MS and LS modes of operation during read achieves significant
reduction in VMIN at low frequencies while still maintaining similar performance
at high VDD . The plot in Fig. 11.26 is obtained by considering the four frequency
bands similar to Fig. 11.22 in case of BMIG FinFET SRAM. However, unlike
BMIG FinFET SRAM, TMIG FinFET SRAM achieves significant improvement in
write-ability and reduction in VMIN for the process corner that is the worst for write
operation (Fig. 11.26). Hence, tri-modal operation of TMIG FinFET SRAM allows
significant reduction in active VMIN of the cell across different process corners at the
cost of slight degradation in hold VMIN [22].
Fig. 11.26 Comparison of VDD versus 1/cell access time of TG and TMIG FinFET SRAMs for
process corners that are the worst for (a) read stability and (b) write-ability [22]
Fig. 11.27 Schematics of (a) BMIG [19] and (b) TMIG FinFET SRAMs with pass gate feedback
(PGFB) FinFET 6 T SRAM [24] and (c) comparison with TG FinFET SRAM
7.3 Pass Gate Feedback in FinFET SRAMs
Pass-gate feedback is an interesting technique to mitigate the read-write conflict in

SRAMs by providing the feedback to the back gate of the access transistor from
the respective storage nodes. 6-T SRAM cell with BMIG FinFET employing pass
gate feedback (PGFB) [19] is shown in Fig. 11.27a. The back gate of the access
transistor is connected to the storage node. During the read operation, the access
transistor corresponding to the node storing zero (AXL) is rendered weak, thus
increasing read stability. The access transistor corresponding to the node storing
‘1’ (AXR) is strong because both of its gates are ON. This further contributes to
increase in the read stability. However, the access time increases significantly due
to weak access transistor AXL. Moreover, during the write operation, the node
QB discharges and decreases the strength of AXR. Depending on the coupling
between the front gate and the back gate, decrease in the drive of AXR may lead
to lower write-ability. This is especially critical in scaled technologies, in which
quantum confinement effects lead to increase in the coupling between the front
and back gates, as discussed before. However, pass-gate feedback is an interesting
technique to mitigate the design conflicts in SRAMs. Coupled with write-assist
techniques, pass-gate feedback can offer a useful design option for SRAMs in scaled
technologies. It may be noted that since the back gate of the transistor is connected
to the storage node, there is no area penalty with respect to TG FinFET SRAM.
It has been argued by the authors in [19] that write-ability of the cell can be
further improved by connecting the back gate of PR and PL to a write word-line
(WWL) which is charged to VDD during write. This weakens the pull-up transistors,
making it easier to flip the contents of the cell. During the read operation, WWL
is de-asserted which increases the strength of the pull-up transistors and raises
the logic threshold voltage of the inverter formed by NR, PR and AXR, thereby
increasing the read stability of the cell. However, the extra back gate contact can
lead to significant increase in the area of the bit-cell. Other techniques like gate-
workfunction engineering can be used to increase the write-ability. However, this
leads to lower hold stability and increase in the cell leakage [24].
Pass-gate feedback can also be employed in SRAMs using TMIG FinFETs as the
access transistors, as shown in Fig. 11.27b. During the read operation, the strength
of AXL (with BG connected to Q D ‘0’) is comparable to a single-fin FinFET [24].
In other words, AXL in TMIG FinFET SRAM is slightly stronger than AXL in TG
FinFET SRAM (since single-fin FinFETs are typically used as access transistors in
TG FinFET SRAM, as discussed previously). This tends to mildly decrease the read
stability of the cell. On the other hand, the strength of AXR (with BG connected to
QB D ‘1’) is twice of AXR in a TG FinFET SRAM. This increases the trip point
of the inverter formed by PR. NR and AXR and therefore, tends to increase the
read stability. The overall effect is that read stability of TMIG FinFET SRAM is
comparable to that of TG FinFET SRAM.
During the write operation, the strength of AXR (with BG connected to
QB D ‘1’) in TMIG FinFET SRAM is higher than that in TG FinFET SRAM. As a
result, large improvement in write-ability is achieved. As the voltage at node QB
discharges during write, the strength of AXR decreases, but is still sufficient to
provide high write-ability.
Hence, TMIG FinFET SRAM with pass-gate feedback achieves significant
improvement in the write-ability at comparable read-stability. In addition, co-
optimization of the gate workfunction leads to lower cell leakage along with
comparable hold stability at the cost of a mild increase in the access time. The
cell-level comparison of BMIG and TMIG FinFET SRAMs employing pass-gate
feedback is shown in Fig. 11.27c.
Connecting the back gate of the access transistor to the storage node may
have issues with the leakage and stability of the unaccessed cells during read
and write operations [19]. During the read operation, cell leakage and leakage
through AXL may increase significantly for the bit-cell belonging to an unaccessed
row. Considering the worst case, suppose the accessed bit-cell stores ‘0’ and all
the unaccessed bit-cells in the same column store ‘1’ (value opposite to that of
the accessed bit-cell), as shown in Fig. 11.28a. During the read operation, BL
discharges to some voltage (VDD - V). Hence, VGS for the back gate of AXL of all
the unaccessed cells becomes positive ( V), leading to an increase in cell leakage
and bitline leakage through AXL. Increase in bitline leakage delays the development
of voltage differential V between BL and BLB, resulting in increased access time.
During the write operation, the bit-cells belonging to the unaccessed rows may be
disturbed if they store the same value as the accessed node (see Fig. 11.28b—note
both the accessed and unaccessed cells store ‘0’ i.e. VQ D 0). Since the back gate
of AXR is turned ON, the storage node VQB of the unaccessed cells is disturbed
Fig. 11.28 Array-level implications of using pass gate feedback in SRAMs during (a) read and
(b) write. The access transistors shown in the figure are BMIG FinFETs or Fin2 of TMIG FinFETs
(Fin1 of TMIG FinFETs not shown)
due to current flowing through AXR, thus introducing an additional source of

failure under process variations. These issues are not just with 6 T PGFB SRAM
[19] but with other cells that uses PGFB configuration like 4 T SRAM cells in
[32]. It is interesting to note that increase in the coupling between front gate and
back gate may lead to the mitigation of these issues, as shown in the analysis in
[24]. Moreover, in TMIG FinFET SRAMs, these issues are significantly reduced
compared to BMIG FinFET SRAMs. Nevertheless, one needs to be aware of these
issues while designing SRAMs with pass-gate feedback.
8 Summary
FinFETs have emerged as promising alternatives to bulk MOSFETs in scaled tech-

nologies because of lower short channel effects and superior scalability. However,
quasi-planar structure of FinFET results in width quantization which reduces the
design flexibility of FinFET based circuits. This effect has an aggravated impact
on SRAMs, in which transistor sizing is critical for cell stability. Therefore,
device-circuit co-design techniques are required which mitigate the effect of
width quantization in FinFET based SRAMs. In this chapter, we discussed such

techniques and presented the implications of each on SRAM design. We showed
that techniques like fin ratio-tSi co-optimization, joint optimization of HFIN -tSi -tOX -
VDD -VT , fin orientation and spacer thickness optimization alleviate the effect of
width quantization and expand the design space of FinFET-based SRAMs. We also
described asymmetric drain spacer extension (ADSE) FinFETs and asymmetrically
doped (AD) FinFETs in which the asymmetry in the drain current for positive
and negative VDS was exploited to mitigate the read-write conflict in 6 T SRAMs.
Finally, we described various IG FinFET based SRAMs and discussed their benefits
in mitigating the design conflicts and the trade-offs associated with each technique.
References
1. Yuan Taur and Tak. H Ning, “Fundamentals of Modern VLSI Devices”, Cambridge University
Press, pp: 140–144.
2. Shekhar Borkar, Tanay Karnik, Siva Narendra, Jim Schantz, Ali Keshavarzi and Vivek De,
“Parameter variations and impact on circuits and microarchitecture”, Proc. IEEE Design
Automation Conference, 2003, pp: 338–342.
3. G.G. Shahidi, T.H. Ning, T.I. Chappell, J. H. Comfort, B. A. Chappell, R. Franch, C. J.
Anderson, P. W. Cook, S. E. Schuster, M.G. Rosenfield, M.R. Polcari, R.H. Dennard and B.
Davari, “SOI for a 1-volt CMOS technology and application to a 512 Kb SRAM with 3.5 ns
access time”, Int. Electron Device Meeting, Dec 1993, pp: 813–816.
4. K. K. Young, “Short-Channel Effect in Fully Depleted SOI MOSFETs”, IEEE Trans. Electron
Devices, vol. 36, no. 2, Feb. 1989, pp: 399–402.
5. M. Vinet, T. Poiroux, J. Widiez, J. Lolivier, B. Previtali, C. Vizioz, B. Guillaumot, Y. Le Tiec,
P. Besson, B. Biasse, F. Allain, M. Casse, D. Lafond, J.-M. Hartmann, Y. Morand, J. Chiaroni
and S. Deleonibus, “Bonded planar double-metal-gate NMOS transistors down to 10 nm”,
IEEE Electron Device Letters, vol. 26, no. 5, May 2005, pp: 317–319.
6. J.P. Colinge, M.H. Gao, A. Romano-Rodriguez, H. Maes, C. Claeys, “Silicon-on-insulator
‘gate-all-around device’”,Int. Electron Device Meeting, Dec. 1990, pp: 595–598.
7. J. Widiez, F. Dauge, M. Vinet, T. Poiroux, B. Previtali, M. Mouis, S. Deleonibus, “Experi-
mental gate misalignment analysis on double gate SOI MOSFETs”, Proc. Int. SOI Conf., Oct.
2004, pp:185–186.
8. D. Hisamoto, W. –C. Lee, J. Kedzierski, H. Takeuchi, K. Asano, C. Kuo, E. Anderson, T-J
King, J. and C. Hu;, “FINFET- a self-aligned double-gate MOSFET scalable to 20 nm”, IEEE
Trans. Electron Devices, vol. 47, no. 12, pp: 2320–2325, Dec 2000.
9. B.S. Doyle, S. Datta, M. Doczy, S. Hareland, B. Jin, J. Kavalieros, T. Linton, A. Murthy, R.
Rios, R. Chau, “High performance fully-depleted tri-gate CMOS transistors,” IEEE Electron
Device Letters, vol. 24, no. 4, April, 2003, pp: 263–265.
10. P. Ranade, H. Takeuchi, T. –J King and C. Hu, “Work Function Engineering of Molybdenum
Gate Electrodes by Nitrogen Implantation,” Electrochemical and Solid State Letters, vol. 4, no.
11, 2001, pp: G85-G87.
11. B. Yu, L. Chang, S. Ahmed, W. Haihong, S. Bell, Y. Chih-Yuh, C. Tabery, H. Chau, Q. Xiang,
T. –J. King, J. Bokor, C. Hu, M. –R. Lin and D. Kyser, “FinFET scaling to 10 nm gate length”,
Int. Electron Device Meeting, Dec 2002, pp: 251–254.
12. T. Ludwig, I. Aller, V. Gernhofer, J. Keinert, E. Nowak, R. V. Joshi, A. Mueller and
S. Tomaschko, “FinFET Technology for Future Microprocessors“, IEEE SOI Conference,
October 2003, pp. 33–34.
13. D. Lekshmanan, A. Bansal and K. Roy, ”FinFET SRAM: Optimizing Silicon Fin Thickness
and Fin Ratio to Improve Stability at iso Area,“ IEEE Custom Integrated Circuits Conference,
Sept. 2007, pp. 623–626.
14. H. Ananthan and K. Roy, ”Technology and circuit design considerations in quasi-planar
double-gate SRAM,“IEEE Transactions on Electron Devices, vol. 52, no. 2, Feb. 2006, pp.
242–250.
15. S. Gangwal, S. Mukhopadhyay and K. Roy, ”Optimization of Surface Orientation for High-
Performance, Low-Power and Robust FinFET SRAM,“ IEEE Custom Integrated Circuits
Conference, Sept 2006, pp. 433–436.
16. A. Bansal, S. Mukhopadhyay and K. Roy, ”Device-Optimization Technique for Robust and
Low-Power FinFET SRAM Design in NanoScale Era,“ IEEE Transactions on Electron
Devices, vol. 54, no. 6, June 2007, pp. 1409–1419.
17. A. Goel, S. K. Gupta and K. Roy, “Asymmetric Drain Spacer Extension (ADSE) FinFETs
for Low Power and Robust SRAMs”, IEEE Trans. Electron Devices, vol. 58, no. 2,Feb 2011,
pp:296–308.
18. K. Endo, S. –I. O’uchi, Y. Ishikawa, Y. Liu, T. Matsukawa, K. Sakamoto, J. Tsukada, K.
Ishii, H. Yamauchi, E. Suzuki, M. Masahara,” Enhancing SRAM cell performance by using
independent double-gate FinFET”, Int. Electron Device Meeting, Dec. 2008.
19. A. Carlson, Z. Guo, S. Balasubramanian, R. Zlatanovici, T. J. King Liu, B, Nikolic, “SRAM
Read/Write Margin Enhancements using FinFETs”, IEEE Trans. VLSI, vol. 18, no. 6, June
2010, pp:887–900.
20. Z. Liu, S.A. Tawfik, V. Kursun, “An independent-gate FinFET SRAM cell for high data
stability and enhanced integration density”, IEEE Int. SOI Conf., 2007, pp: 63–66.
21. S. A. Tawfik, Z. Liu, and V. Kursun, “Independent-Gate and Tied-Gate FinFET SRAM
Circuits: Design Guidelines for Reduced Area and Enhanced Stability”, Int. Conf. Microelec-
tronics, 2007, pp: 171–174.
22. S. K. Gupta, S. P. Park and K. Roy, “Tri-mode Independent Gate (TMIG) FinFETs for dynamic
voltage/frequency scalable 6 T SRAMs,” to appear in IEEE Trans. Electron Devices.
23. F. Moradi, S. K. Gupta, G. Panagopoulos, D. T. Wisland, H. Mahmoodi and K. Roy,
”Asymmetrically Doped FinFETs for Low-Power Robust SRAMs,“ IEEE TED, vol.58, no.12,
pp.4241,4249, Dec. 2011.
24. S. K. Gupta, J.P. Kulkarni, and K. Roy, ”Tri-Mode Independent Gate FinFET-Based SRAM
With Pass-Gate Feedback: Technology–Circuit Co-Design for Enhanced Cell Stability,“ IEEE
TED, vol.60, no.11, pp.3696,3704, Nov. 2013
25. Jan M. Rabaey, Anantha Chandrakasan and Borivoje Nikolic, “Digital Integrated Circuits, A
Design Perspective”, Prentice Hall Electronics and VLSI series, pp:657–661.
26. D. Esseni, A. Abramo, L. Selmi, E. Sangiorgi, “Physically based modeling of low field electron
mobility in ultrathin single- and double-gate SOI n-MOSFETs”, IEEE Trans. Electron Devices,
vol. 50, no. 12, Dec. 2003, pp:2445–2455.
27. D. Lekshmanan, A. Bansal, K. Roy, “Body Thickness Optimization and Sensitivity Analysis
for High Performance FinFETs,” Device Research Conf., June 2007, pp: 91–92.
28. Y. –K.Choi, T. –J. King, C. Hu, “A Spacer Patterning Technology for Nanoscale CMOS”, IEEE
Trans. Electron Devices, vol. 49, no. 3, March 2002, pp:436–441.
29. F. Bauer. G. Georgakos, D. Schmitt-Landseidel, “A Design Space Comparison of 6 T and 8 T
SRAM Core-Cells”, Lecture Notes in Computer Science, vol. 5349, 2009, pp: 116–125.
30. M. –L. Fan, Y. –S Wu, V P –H Hu, C –Y Hsieh, P. Su, C. –T. Chuang, “Comparison of 4T and
6T FinFET SRAM Cell for Subthreshold Operation Considering Variability – A Model-Based
Approach”, IEEE Trans. Electron Devices, vol. 58, no. 3, March 2011, pp:609–616.
31. K. C. Chun, P. Jain, J. H. Lee and C. H. Kim, “A 3 T Gain Cell Embedded DRAM Utilizing
Preferential Boosting for High Density and Low Power On-Die Caches”, IEEE, Journal Solid
State Circuits, vol. 46, no. 6. June 2011, pp: 1495–1505.
32. M –L Fan et al, “Comparison of 4T and 6T FinFET SRAM Cell for Subthreshold Operation
Considering Variability – A Model-Based Approach”, IEEE Trans. Electron Devices, vol. 58,
no. 3, March 2011, pp: 609–616.
Chapter 12
Variability-Aware Clock Design
Matthew R. Guthaus and Gustavo Wilke
Abstract High-performance clock network design has been a challenge for many
years due to the drastically increasing effect of process variability. In addition,
tight power budgets have lowered supply voltage levels which make designs more
sensitive to noise. Together, variability and noise present a colossal challenge to
clock designers in order to meet timing, yield, and power simultaneously. This
chapter discusses the different strategies that designers use to ameliorate variability
and noise problems in clock network design.
1 Introduction
The clock is the most important signal in any synchronous design, because it
controls all synchronous communication. A failure in the clock signal can lead to
errors in one or more sequential elements. Strict constraints on clock timing must
be used to guarantee the correct behavior of synchronous digital blocks. The clock
period must be defined in such a way that data will always be ready and stable before
clock edge arrives at the clock pin of the sequential elements which are also called
clock sinks.
Figure 12.1 shows the timing parameters that must be considered to safely
determine clock frequency and ensure correct operation. Assume that TCK .n/0 is
clock arrival time at flip-flop A, at clock cycle n and TCK .n C 1/00 is clock arrival
time at flip-flop B during clock cycle .n C 1/. Data propagation time through flip-
flop A is represented by TPFFA . The minimum (maximum) delay associated with
the combinational logic is represented by TC;MI N (TC;MAX ). Flip-flop B setup and
hold times are represented respectively by T SFFB and THFFB . Given that TCLOCK
M.R. Guthaus ()

University of California Santa Cruz, 1156 High St. MS SOE3,
Santa Cruz, CA 95064, USA
e-mail: mrg@ucsc.edu
G. Wilke
UFRGS, Av. Bento Gonçalves, 9500 - Campus do Vale - Bloco IV, Porto Alegre,
RS 91509-900, Brazil
e-mail: wilke@inf.ufrgs.br

DOI 10.1007/978-1-4614-4078-9__12
256 M.R. Guthaus and G. Wilke
Tpffa Tc TSffb
Data In Data Out

D Q Combinational Logic D Q
A B
Tck(n)’ Tck(n)’’
Clock
Fig. 12.1 Clock period definition
is the clock period, Eqs. (12.1) and (12.2) represent the timing constraints that must
be satisfied to guarantee the correct data storage for the case illustrated by Fig. 12.1.
TCLOCK TPFFA C TC;MAX C T SFFB C .T ck.n/0 TCK .n C 1/00 / (12.1)

0 00
THFFB TPFFA C TC;MI N C .TCK .n/ TCK .n/ / (12.2)
Equation (12.1) represents a lower bound to the clock period. The clock period
has to be chosen such that Eq. (12.1) is going to be respected for any path connecting
any two flip-flops in the design. Equation (12.2) limits the minimum delay TC;MI N
allowed for any combinational path. Both equations have to be simultaneously
respected to guarantee the correct circuit behavior.
The main challenge faced by clock designers is how to design robust clock
distribution networks respecting Eqs. (12.1) and (12.2) within the allowed power
budget. Traditional worst-case analysis is known to be robust but is usually too
pessimistic and leads to over-designed circuits that waste power. In the last decade
statistical techniques have begun to aid designers to safely reduce the amount of
pessimism in IC design. Statistical design, however, is taking longer than expected
to be applied to clock design, because the importance of clock timing on the circuit
robustness makes designers conservative.
An alternative to traditional clock tree based architectures began to appear
in microprocessor designs in the late 1990s. To decrease the effect of process
and environmental variability, microprocessor designers adopted clock meshes and
clock spines in many designs [3, 27, 35, 40]. Due to the redundant structure of
meshes and spines, fast and slow paths are able to compensate each other and
reduce the clock arrival time variability [18]. The main disadvantage presented by
these architectures, however, is the large power penalty caused by the extra wiring
resources and short circuit currents.
The remainder of this chapter discusses the impact of variation sources on the
clock design and presents different techniques to make the clock network less
sensitive to these variations. In the remainder of this section, we will present some
12 Variability-Aware Clock Design 257
important definitions and discuss the parameters that make a clock network robust.
In Sect. 2, the different sources of variability and their impact on clock networks are
discussed. Sections 3 and 4 present design strategies to create clock networks that
are less sensitive to variability.
1.1 Definitions
The clock network is designed so that the time it takes for a clock edge to travel from
the clock root to any sink is the same. This delay is defined as the clock arrival time
(At ). Due to design mismatch, the clock arrival time at the clock sinks is often not
exactly the same. The difference between the clock arrival time at different sinks is
defined as clock skew. Given clock sinks si ; sj 2 S and At .si / which represents the
clock arrival time at the clock sink si , the clock skew (S k) for a set of clock sinks S
can be defined in a more formal manner as follows:
S k.S / D max.i;j / jAt .si / At .sj /j: (12.3)
In addition to design mismatch clock skew, environmental variations such as

temperature, crosstalk, and power supply noise are also responsible for degrading
circuit performance along with the clock skew. Many of these variations can occur
over time resulting in uncertainty in the clock arrival time at each clock sink which
is called clock jitter.
When defining the clock period, clock skew and clock jitter must be budgeted.
The sum of both values represents an upper bound for the difference in the clock
arrival time at any two flip-flops. In Eq. (12.1), the effect of clock skew and clock
jitter is illustrated by the TCK .n/0 TCK .n C 1/00 term. For sake of simplicity, all
future references to clock skew will refer to the combined uncertainty due to both
clock skew and clock jitter.
1.2 Robustness
Clock robustness is key for the correct operation of any circuit. Unlike the other
signal lines, a fault in the clock signal can easily propagate to hundreds of sequential
elements generating errors in multiple logic bits. These logic errors can occur due
to two reasons: delay violations and glitches. Both types of faults can in turn create
logic errors.
Clock delay violations are caused by the non-robust budgeting of clock margins.
When the clock skew violates the budgeted margins due to process or environmental
variations, Eqs. (12.1) and (12.2) are also violated. This situation leads to the storage
of invalid values in some sequential elements. If the clock arrival time at the desti-
nation flip-flop is delayed or the clock edge at the transmitter flip-flop is expedited,
the data hold time constraints can be violated by the fastest paths in combinational
logic (Eq. (12.2)). When the inverse situation happens, i.e., transmitter clock is late
or receiver clock is early, the time for data to propagate through combinational logic
is shortened leading to setup time violations (Eq. (12.1)).
Clock glitches are the another potential cause of logic errors. It can be cause by
two main issues, radiation [28] and poor synchronization of clock gating control
signals on latch-free clock gaters. A clock glitch is a spurious transition in the
clock signal that can lead to invalid values stored in affected sequential elements.
Although invalid data can be stored in memory elements due to a clock glitch, it
is possible that the failure is corrected when the next clock edge arrives. In case
the invalid data is not corrected by the following clock edge (e.g., write enable
signal was turned off) the error may propagate to multiple sequential elements in
the following clock cycles. Even when the invalid data is quickly corrected it can
still cause errors. During the time that a flip-flop stores an invalid value a data glitch
is generated. The data glitch will cause errors if it has enough time to propagate to
a memory element.
A robust clock network has to minimize the probability that a delay fault will
occur and minimize the chances of a clock glitch reaching several clock sinks. In
the next sections, techniques to design robust clock networks are described.
2 Variability
Understanding the different variability sources and properly modeling them is a very
important task during the clock design. When process and environmental effects
were not the main cause of clock skew, the analysis of clock network performance
could be done using a corner-based approach. Corner-based analysis bounds all
the parameters in the design by minimum and maximum values. Corner analysis is
robust if the worst case corner can be identified, but it also became overly pessimistic
as the impact of variability increases.
Statistical analysis resurged in the early 2000s as an alternative to reduce
pessimism under scenarios where the impact of variation sources is large. The use
of statistical methods do not guarantee by itself an accurate estimation of clock
skew values. The way that each source of variation is modeled strongly affects the
accuracy in the clock skew estimation. The correct modeling of each variability
source is fundamental for the proper estimation of the delay margins. In this section,
we discuss the characteristics of the different types of variation and how they should
be modeled.
2.1 Types of Variation
Variability in clock networks can be due to manufacturing process variation (P),

on-chip supply voltage variation (V), on-chip temperature (T), and cross-talk
between the clock and signal wires (X). These are often referred to as PVT variations
(ignoring the X). PVTX variations, however, are a significant challenge in the
synthesis of clock networks, because they can add skew and jitter which reduces
timing margins.
Each form of variation can be modeled as either a statistical distribution or
as a bounded corner value. The distribution and, more importantly, the correla-
tion of parameters must be considered in order to model a source of variation
[29]. Variation that is observed from die-to-die (D2D) is called inter-die variation
while variation that occurs within a die (WID) is called intra-die variation. Often
the WID variation exhibits significant spatial correlation so that nearby locations
are correlated but distant locations are not. Many sources of variation are systematic
which can be predicted in terms of design or process details. Systematic variation
is distinctly different from correlated variation because it is not a random variable
when the cause of variation is considered [15].
2.1.1 Process
Process variation can be either front-end (device) variation or back-end (intercon-

nect) variation. Some front-end variation such as Vth are uncorrelated and do not
exhibit spatial correlation or systematic behavior [31]. Other device parameters
such as Leff are mostly systematic [15] with high spatial correlation except for
a small portion due to line-edge roughness (LER) [26, 37]. Interconnect variation is
primarily due to systematic chemical mechanical polishing (CMP) variation which
affects the metal height and therefore interconnect parasitics with high correlation
[38]. However, line-edge roughness (LER) can also affect interconnect as a highly
uncorrelated behavior.
2.1.2 Voltage
On-chip power supply variation is one of the most significant sources of variation in
a clock network [39,41]. Due to the high power consumption during peak switching,
large L ddti voltage drops can occur in the power grid. Chip-level clock distributions
include decoupling capacitors with the large clock drivers to reduce supply-
related variation. The decoupling capacitors low-pass filter the supply variation
and remove sharp current peaks. The remaining supply variation has high spatial
correlation [41].
2.1.3 Temperature
On-chip temperature fluctuations are also very significant for both device and
interconnect variations. Temperature directly affects the mobility of silicon and the
resistance of metal wires [2, 22]. The effect on interconnect capacitance, however,
is relatively small. The spatial correlation of temperature is highly dependent on the
package and heat sink model. For high-performance designs with large heat sinks,
there can be temperature gradients of 30–40 ı C on-chip. Whereas, for GPUs (with
high activity in all regions) or low-performance designs without active heat removal,
the ambient chip temperature will be higher yet gradients will be lower [41]. Despite
the high spatial correlation, significant temperature-induced skew can be present
between the initial (cool) chip and a steady-state thermal map due to mismatch
in clock sensitivities to temperature. The temperature map can also dynamically
change depending on activity patterns.
2.1.4 Crosstalk
The clock network is the largest cross-talk aggressor due to its large capacitance and
span. However, signal nets also affect the jitter of individual clock branches in trees.
Adjacent layers have different routing orientations and thicker oxides, so within-
layer coupling is the most significant problem. Clock wires are usually shielded to
reduce these crosstalk effects [39].
2.2 Impact of Variation Models
The law of large numbers suggests that independent WID variation can be reduced
simply by increasing the number of random variables. For example, the mean and
standard deviation of a sum of independent Gaussian distributions is
X sX
D i i nd D i2 (12.4)
8i 8i
which implies that increasing ji j will also decrease the total ind . As a limit, this will
approach zero with infinite Gaussians. As an example, consider a single transistor
which is known to have Vth variation due to random dopant fluctuation. Pelgrom’s
model [31] demonstrates this with the relation
1
Vth / p (12.5)
WL
that implies that larger transistors will have less variation. Clock buffers in general
are large devices ranging from 10 to 64 minimum size for tree-type distributions
to huge 29 cm drivers for clock grids [4]. The previous models correctly imply that
Vth is in fact not very significant for clock buffers due to the averaging over very
large device sizes.
The problem with the previous observation is that the law of large numbers only
holds true if the variations are uncorrelated. In the case of full correlation, the mean
and standard deviation become
X X
D i corr D i (12.6)
8i 8i
which will not decrease the total corr

as ji j increases. As an example, assume that
every clock buffer has an uncorrelated power supply variation of Vdd D 1:0 ˙ 0:15.
According to the independence assumption, it is therefore better to implement many
instances of a smaller device as opposed to fewer instances of the larger device to
reduce the total variation while having the same total drive strength. By simply
doing this, one can achieve a relative reduction in variability of
qP
2
i nd 8i i
D P : (12.7)
corr 8i i
If a simple long-channel transistor model is used, a single 8 library buffer will

have 12 % variation in delay due to supply variations while eight 1 library buffers
in parallel will have a variation of 4.2 %. In reality, however, all of these buffers will
be adjacently placed and will have the same power supply rail so this is an artifact
of using the wrong correlation model.
Similarly, assume a wire model with an uncorrelated 10 % wire variation which
affects both resistance and capacitance. It is better to segment a 1 mm wire into 100
segments as opposed to 2 segments, because this will reduce the delay variation by
7. However, if this variation is due to CMP, it will have high spatial correlation at
distances less than 0.5–1 mm, so the above independent assumption is not valid and
leads to an incorrect observation. Conversely, if the variation is due to uncorrelated
LER, then segmenting the wires into large 500 m intervals is not correct and will
overestimate the variation.
It is also often not accurate to assume that a variation is a percentage of a nominal
value. As an example, consider an uncorrelated 10 % wire variation due to line-edge
roughness (LER). A wide wire will have less LER delay variation compared to a
minimum width wire because LER is an absolute number and not a percentage of
the nominal resistance or capacitance.
One overlooked approach for reducing variation is to improve the correlation
rather than minimize variation. This is especially true in clock distributions since
we are concerned with the relative difference in delays for skew. p If you consider two
insertion delays with D 0:1, the skew variation is 3 D 3 0:12 C 0:12 D 0:42
if they are independent whereas it is reduced to zero if they are perfectly correlated.
Correlation can be improved by using the same gate types/sizes, using the same
metal layers for D2D variation, placing items closer together when spatial variation
is present, etc.
2.3 Common Industry Models
There are several approaches used in industry to model variation. The oldest
approach is to increase the tbudget with a timing margin to “guard band” the setup
and hold times without specifically analyzing the variation. Most frequently, timing
sign-off will use worst-case corners that analyze the worst local skew when portions
of a clock distribution are “fast” and others are “slow”. The number of corners
in such approaches were traditionally only for PVT variations, but additional
corners have been added for individual metal layers, Negative Bias Temperature
Instability (NBTI), and other sources of variation. This is to prevent the process (P)
term from becoming overly pessimistic when all sources are assumed worst-case
simultaneously.
Despite an exponentially increasing number of corners [44], multi-corner timing
analysis is the current industry standard practice. Incremental algorithms are used
to perform common path pessimism removal (CPPR) [20, 49] due to shared clock
buffers that cannot be simultaneously fast and slow in a single corner. Most CPPR
algorithms are iterative and worst-case exponential time complexity. Other methods
have used data from multiple corners and normalized delay spreads to quantify the
effects of variation on clocks [33].
Statistical static timing analysis (SSTA) solves the previous problems [1, 7, 45],
but has a number of daunting challenges of its own such as obtaining accurate
variation models, efficiently considering non-Gaussian and non-linear sources of
variation, and gaining designer acceptance. There are several surveys and books that
cover the recent advances in SSTA [5,13,36] and are applicable to clock researchers.
2.4 Requirements for Variation Models
We now summarize the requirements that are needed to accurately assess the
robustness of a clock distribution with respect to variation.
Both device and interconnect models must be considered. Ignoring device
variation implies that one can over-buffer a design making interconnect variation
less significant. Similarly, without interconnect variation, one can under-buffer
the design which makes the delay of the devices less significant. The relative
importance of device and interconnect variability also depends on whether the
synthesis algorithm is considering local, regional, or global distributions. Local
distributions are affected more significantly by device variability whereas global
distributions are affected more significantly by interconnect variability.
The physical manifestation of variability must be considered in the model.
Blindly selecting variation as a percentage of delay or a parameter is not correct.
Most variations (Vth , LER, etc.) are not a percentage. The correct model requires
careful consideration of the specific source of variation.
Assuming that only WID variation matters is not correct. D2D variation also
matters because parameter mismatch (between layers, implants, etc.) can lead to
different insertion delay sensitivities and therefore skew.
Ignoring correlation for WID variation is not correct. If WID variation is
assumed to be completely uncorrelated the law of large numbers can always be
applied, but this may not be appropriate. It is necessary to consider the right
correlation model for each source of variability.
Both skew and jitter must be simultaneously considered. Otherwise, insertion
delay and jitter sensitivity can be increased to reduce skew or skew can be increased
to reduce jitter. In high-performance designs, skew and jitter are on the same order
of magnitude.
Variability must be analyzed statistically. Simple corner-based approaches are
too pessimistic and do not allow for the cancellation of correlated variation or the
reduction of variation due to the law of large numbers (when appropriate) [44].
Multi-corner methods do not scale since the number of corners will become too large
[44]. Therefore, statistical methods should be used with the appropriate variation
models.
2.5 Analysis Methodologies
When variability impact is considered, the clock network cannot be analyzed in a

single spice run. Statistical methods have to be applied for an accurate evaluation
of the variability effect on the clock performance. Many of the methods developed
for statistical timing analysis can be adapted to evaluate the clock network. Monte
Carlo analysis however is still the most common analysis methodology due to its
simplicity and accuracy.
The main limitation of Monte Carlo simulation is the number of samples required
for an accurate analysis. A single flat spice run on a full clock network can take
many hours. To enable the characterization of large clock trees the analysis can be
done level-by-level. This methodology however cannot be applied to clock meshes
and clock spines. Whenever clock buffers are shorted they have be analyzed as a
whole. In a clock mesh or a clock spine the simulation of the clock tree driving the
mesh/spine can still be separated from the stages in which the buffers are shorted
but every shorted set of buffers has to be handled in a single simulation. Large clock
meshes may be prohibitively large to be fit in a single spice run, so several strategies
have been proposed [9, 48] to enable the analysis of these large meshes.
Frequency domain analysis [46,50] can achieve speed ups in the order of 1,000
but these techniques require that the buffer model be linearized which introduces
error. Model order reduction techniques are more accurate but achieve speed ups of
at most 10 [48]. The method of [9] is the most flexible as it can be applied using
any electrical simulation, but it can only reduce the clock mesh characterization time
if the different runs are parallelized.
3 Clock Trees
The simplest style of clock distribution network is an H-tree. The H-tree, as shown in
Fig. 12.2, uses a recursive “H” routing pattern and inserts buffers at regular intervals
in the hierarchy. H-trees are advantageous because they are simple and fairly power
efficient. They are also symmetric and therefore resistant to inter-die sources of
variation. However, H-trees can still be plagued by intra-die sources of variation as
shown in [19]. Besides their robustness, H-trees are only typically useful for the
top levels of clock distribution since the sink loads and locations are never uniform.
Loads can be balanced with dummy gates, but this can increase the clock tree power
consumption considerably. In addition, individual sinks may have differing clock
edge requirements according to a useful skew budget.
In response to the non-uniformity of clock sinks, balanced clock trees offer
a low-power solution with equal delay to clock sinks as illustrated by Fig. 12.3.
The diagonal routes are later embedded into traditional rectilinear layers. Several
researchers have presented exactly zero skew [6, 8, 11, 43], bounded skew [10, 21]
and heuristic methods [23, 24] to construct these trees. Most balancing techniques
rely on simplified delay models such as Elmore delay [12], which are known to be
only approximations of more complex delays [32,34]. However, they are still useful
in generating initial solutions that can be refined with more accurate models later.
The main disadvantage of balanced trees is that they often assume process
parameters are nominally equal for different interconnect layers and do not consider
Fig. 12.2 H-Tree clock

distribution
Fig. 12.3 Balanced clock

tree distribution
Fig. 12.4 Non symmetrical

equal wirelength routing
example metal 5
metal 6
net A
net B
Fig. 12.5 Symmetrical

routing example
metal 5
metal 6
net A
net B
the impact of process variation on skew. The different variability sources described
in Sect. 2 can introduce skew in the clock tree even when it has zero-skew under
the nominal case. The effect of variations sources is very significant in today’s
technologies and its effect has to be account on every design.
One way to make a clock network more robust is by equalizing the way it affects
the clock network. Die-to-die variability effects on the clock network, for instance,
can be completely eliminated by equalizing the sensitivity of each parameter on the
delay on each path. Consider the two nets shown in Fig. 12.4, net A and B that
are routed using metal 5 (vertical) and metal 6 (horizontal) wires. Assuming that
the delay of nets A and B must be equalized, it is important that both nets have
exactly the same wire length. The routing shown in Fig. 12.4 guarantees nominal
zero skew between them, but if during fabrication metal 5 is faster while metal 6 is
slower, net B would become faster than net A. This situation can be avoided if both
nets are routed using the same amount of routing in each layer in the same order as
shown in Fig. 12.5. By doing so, if one metal layer gets faster while the other gets
slower both nets are affected in the same fashion. Sensitivity matching algorithms
have attempted to do this by sizing buffers and wires to equalize the sensitivities of
different process parameters using statistical timing analysis [17].
On the other hand, symmetrical clock trees are very efficient to reduce the
effect of die-to-die variations. Custom symmetrical clock trees can be built using
symmetrical routing algorithms such as [16, 30]. In general symmetrical clock trees
are more robust to die-to-die variations at the expense of higher power consumption
due to the longer total wirelength.
Fig. 12.6 Time available for Tc

signal to propagate between
flip-flops is unaltered if both
clock edges are modified by FFA FFB
the same amount
CK_FFA CK_FFB
Original Available Tc
New Available Tc
CK_FFA d
CK_FFB d
Original Available Tc = New Available Tc
Unlike die-to-die variation, within-die variation is not easily addressed. The

effect of within-die variation on skew can be minimized by increasing the correla-
tion between the clock signals. As illustrated in Fig. 12.6, if the variability affecting
the clock skew is correlated, a delay in the launching clock edge will coincide with
a delay in the capturing clock edge. This results in an identical clock period length.
This concept was used in [17] for a modified version of the DME algorithm in
which the correlation between clock sinks that share timing paths is increased which
reduces the variability of those paths.
Besides the techniques described above, simple design principles can be followed
to make clock networks more robust. Power supply noise is one of the main
contributors to clock skew, therefore inserting decoupling caps in the power supply
close to the clock buffers helps to decrease clock skew due to power supply noise.
Another strategy to minimize clock skew variation is to reduce the number of buffers
in the clock network as buffers are typically more sensitive to variation than wires.
Aggressively reducing the number of buffer stages is not always good, however, as
it can increase slew rates. Excessively large slew rates make the clock network more
sensitive to threshold voltage variations as buffer delays are more dependent on the
threshold voltages when the input signal slew is high. It should be noticed that those
techniques are general and can be applied to non-tree clock network topologies.
In many microprocessor designs the clock network uses active compensation
through de-skew buffers to reduce the effect of process variation on clock skew.
Active deskewing uses programmable delay buffers and a control feedback. The
control feedback is responsible for detecting and minimizing the clock skew.
De-skew buffers can dynamically monitor the clock network delays [14, 25] or
be configured only once after fabrication [42]. Active deskewing requires extra
circuitry to monitor the clock skew and adjust the delay at the variable delay buffer.
Static deskewing is much simpler as delays at the variable delay buffer are set only
once and no skew monitoring is required. Due to the simpler and more robust design,
static deskewing is more commonly found than dynamic deskewing in industry.
4 Clock Meshes
A clock mesh is a grid of orthogonal wires to which clock buffers are connected.
The grid structure allows mesh buffer outputs to interact to each other. Whenever a
buffer switches before the others, it is slowed down by the late switching buffers.
This behavior reduces the overall difference between early and late arrival time. The
capability to compensate for any variability source makes clock meshes popular in
high-performance designs. Meshes and spines are widely used in the design of the
clock distribution for microprocessors [3, 27, 35, 40].
The buffers that drive clock meshes are typically driven by a clock tree. Non-
homogeneous clock sink distribution and obstructions may make the clock mesh
irregular requiring the clock tree to be custom designed for it. Clock sinks are
usually not directly connected to the clock mesh. Local buffered clock trees are
designed to cluster clock sinks and decouple the actual sink capacitance from the
clock mesh. The total clock mesh power consumption is reduced by decoupling
clock sink capacitance.
The high performance offered by clock meshes comes at the cost of high
power consumption. The redundant routing that shorts all mesh buffers represents
extra capacitance that must be charged/discharged every clock cycle. Clock gating
techniques are not as effective on clock meshes as they are on clock trees. Clock
gaters must be placed in the local cock buffers after the clock mesh. The mesh must
be divided into smaller sub-meshes in order to gate the mesh itself.
Short circuit current is another important contributor to total mesh power. As
usual, internal short circuit current flows between PMOS and NMOS devices when
the clock signal is switching. However, the clock mesh also has significant short
circuit currents flowing between different mesh buffers. Moreover, during the time
that mesh buffers are compensating for variation, short circuit currents flow between
buffers as illustrated in Fig. 12.7.
The contribution of short circuit current to total power consumption is dependent
on two design parameters: the clock skew at the input of the mesh buffers and
the total load driven by the clock mesh. When the input clock signal is poorly
synchronized at the input of the mesh buffers, short circuit current can have a long
duration which increases the power consumption. The contribution of short circuit
current to the clock mesh power consumption as a function of the input clock skew
is shown in Fig. 12.8. Figure 12.8 shows power consumption on the y-axis while the
x-axis shows the maximum clock skew at the input of mesh buffers normalized to a
1 GHz clock period. A roughly linear increase in the power consumption is observed
when the input skew increases.
The other main contributor to the short circuit power consumption are the mesh
buffer sizes. Larger mesh buffers produce larger short circuit currents. Mesh buffer
sizes are determined by the load connected to the clock mesh, the higher the load
the larger mesh buffers are required to be so timing constraints can be met. The load
driven by the clock mesh can be reduced by adding more level of buffers between
T1 T2
VDD
OFF
CLK A
ON
GND
T1 < t < T2
VDD
ON
CLK B
OFF
GND
Fig. 12.7 Short-circuit current caused by different arrival time at mesh buffers
Fig. 12.8 Short-circuit current caused by different arrival time at mesh buffers [47]
the clock mesh and the clock sinks. By decoupling clock sink capacitances from the
clock mesh the contribution of the clock mesh power to the total power of the clock
network is minimized.
Besides the clear performance advantage offered by clock meshes, they also
facilitate fast clock design turn-around. Due to the tight skew requirements, clock
trees must be redesigned every time changes are made at the clock sink level. Clock
meshes, on the other hand, are less sensitive since every sink is driven by a large
set of nearby buffers. Therefore, the effect of changes at the clock sinks is shared
among many buffers. If many random changes are performed the overall effect tends
to average out. This property allows the design of the clock network to be kept
unchanged when minor changes happen at the clock sink level.
The clock mesh conveniently shields the top-level clock tree from any changes
in the clock sink distribution. However, the clock mesh requires that mesh buffers
be spread over the entire region covered by the clock mesh. Any large macro within
this region must accommodate space for mesh buffers. Leaving buffer-less regions
in the clock mesh drastically reduces variability tolerance and increases clock slew.
Clock spines are frequently used when macros cannot accommodate mesh buffers.
Clock spines are essentially one dimensional clock meshes with buffers placed
along the line so that the outputs are shorted and the inputs are driven by a clock
tree. Figure 12.9 illustrates a clock spine. Although the clock spine also allows
faster switching buffers to compensate for late switching ones, its performance is
significantly inferior than a clock mesh. The two dimensional structure of clock
meshes reduces the equivalent resistance between the mesh buffers and enables
more effective compensation for variability. Compensation between the spine
buffers is almost exclusively adjacent and leads to high skew between distant points.
Clock Root
Clock Sinks
Fig. 12.9 Clock spine example

In addition, high skew compared to clock meshes is observable between different

spines. Despite this, clock spines are popular in high performance designs due to the
floorplanning benefits and can be found in microprocessors such as in [27] and [40].
Conclusion
Clock robustness is a critical part of chip design as any timing fault can
propagate and produce multiple logic errors. A robust clock network must
properly account for the effect of the process and environmental variation
to ensure that timing margins are not violated. The sensitivity of the clock
network becomes even more evident as technology scales. Shrinking devices
amplify the effect of process variability due to fabrication challenges and
random dopant fluctuation in transistor threshold implants. Reduced supply
voltages are critical to reduce power consumption but also lead to reduced
noise margins which make designs more susceptible to power supply noise,
crosstalk and radiation effects. As discussed in the previous chapter, it is clear
that clock networks must be designed with robustness in mind. Variation-
tolerant clock trees, clock meshes and clock spines will continue to gain
importance in high performance designs.
References
1. Agarwal, A., Zolotov, V., Blaauw, D.T.: Statistical timing analysis using bounds and selective
enumeration. IEEE Transactions on Computer-Aided Design 22(9), 1243–1260 (2003)
2. Ajami, A., Banerjee, K., Pedram, M.: Modeling and analysis of nonuniform substrate
temperature effects on global ULSI interconnects. IEEE Transactions on Computer-Aided
Design of Integrated Circuits and Systems 24(6), 849–861 (2005)
3. Bailey, D., Benschneider, B.: Clocking design and analysis for a 600-MHz Alpha micropro-
cessor. Journal of Solid-State Circuits (JSSC) 33(11), 1627–1633 (1998)
4. Benschneider, B.J., Black, A.J., Bowhill, W.J., Britton, S.M., Dever, D.E., Donchin, D.R.,
Dupcak, R.J., Fromm, R.M., Gowan, M.K., Gronowski, P.E., Kantrowitz, M., Lamere, M.E.,
Mehta, S., Meyer, J.E., Mueller, R.O., Olesin, A., Preston, R.P., Priore, D.A., Santhanam,
S.: A 300-MHz 64-b quad-issue CMOS RISC microprocessor. Journal of Solid-State Circuits
(JSSC) 30(11), 1203–1214 (1995)
5. Blaauw, D., Chopra, K., Srivastava, A., Scheffer, L.: Statistical timing analysis: From basic
principles to state of the art. IEEE Transactions on Computer-Aided Design of Integrated
Circuits and Systems 27(4), 589–607 (2008)
6. Boese, K., Kahng, A.: Zero-skew clock routing trees with minimum wirelength. In: ASIC
Conf., pp. 1.1.1–1.1.5 (1992)
7. Chang, H., Sapatnekar, S.: Statistical timing analysis considering spatial correlations using
a single PERT-like traversal. In: International Conference on Computer-Aided Design,
pp. 621–625 (2003)
8. Chao, T.H., Hsu, Y.C., Ho, J.: Zero skew clock net routing. In: Design Automation Conference
(DAC), pp. 518–523 (1992)
9. Chen, H., Yeh, C., Wilke, G., Reddy, S., Nguyen, H., Walker, W., Murgai, R.: A sliding
window scheme for accurate clock mesh analysis. In: IEEE/ACM International Conference
on Computer-Aided Design, ICCAD, pp. 939–946 (2005)
10. Cong, J., Koh, C.K.: Minimum-cost bounded-skew clock routing. In: International Symposium
on Circuits and Systems (ISCAS), pp. 215–218 (1995)
11. Edahiro, M.: Minimum path-length equi-distant routing. In: Asia-Pacific Conf. on Circuits and
Systems, pp. 41–46 (1992)
12. Elmore, W.C.: The transient response of damped linear networks. Journal of Applied Physics
19, 55–63 (1948)
13. Forzan, C., Pandini, D.: Statistical static timing analysis: A survey. Integration, the VLSI
Journal 42, 409–435 (2009)
14. Geannopoulos, G., Dai, X.: An adaptive digital deskewing circuit for distribution networks.
International Solid-State Circuits Conference (1998)
15. Gupta, P., Heng, F.L.: Toward a systematic-variation aware timing methodology. In: Design
Automation Conference, pp. 321–326 (2004)
16. Guthaus, M., Sylvester, D., Brown, R.: Clock tree synthesis with data-path sensitivity
matching. In: Design Automation Conference, 2008. ASPDAC 2008. Asia and South Pacific,
pp. 498–503 (2008)
17. Guthaus, M.R., Sylvester, D., Brown, R.B.: Process-induced skew reduction in deterministic
zero-skew clock trees. In: Asia and South Pacific Design Automation Conference (ASPDAC),
pp. 84–89 (2006)
18. Guthaus, M.R., Wilke, G., Reis, R.: Revisiting automated physical synthesis of high-
performance clock networks. ACM Trans. Des. Autom. Electron. Syst. 18(2), 31:1–31:27
(2013)
19. Hashimoto, M., Yamamoto, T., Onodera, H.: Statistical analysis of clock skew variation
in H-tree structure. In: International Symposium on Quality Electronic Design (ISQED),
pp. 402–407 (2005)
20. Hathaway, D., Alvarez, J.P., Belkbale, K.P.: Network timing analysis between signals travers-
ing a common circuit path. United States Patent 5,636,372 (1997)
21. Huang, D.J.H., Kahng, A.B., Tsao, C.W.A.: On the bounded-skew clock and Steiner routing
problems. In: Design Automation Conference (DAC), pp. 508–513 (1995)
22. Im, S., Srivastava, N., Banerjee, K., Goodson, K.E.: Scaling analysis of multilevel interconnect
temperatures for high performance ICs. IEEE Transactions on Electron Devices 52(12),
2710–2719 (2005)
23. Jackson, M.A.B., Srinivasan, A., Kuh, E.S.: Clock routing for high performance ICs. In: Design
Automation Conference (DAC), pp. 573–579 (1990)
24. Kahng, A.B., Cong, J., Robins, G.: High-performance clock routing based on recursive
geometric matching. In: Design Automation Conference (DAC), pp. 322–327 (1991)
25. Kapoor, A., Jayakumar, N., Khatri, S.P.: A novel clock distribution and dynamic de-skewing
methodology. In: International Conference on Computer-Aided Design (ICCAD), pp. 626–631
(2004)
26. Kim, S.D., Wada, H., Woo, J.C.S.: TCAD-based statistical analysis and modeling of gate
line-edge roughness effect on nanoscale MOS transistor performance and scaling. IEEE
Transactions on Semiconductor Manufacturing 17(2), 192–200 (2004)
27. Kurd, N.A., Barktullah, J.S., Dizon, R.O., Fletcher, T.D., Madland, P.D.: A multigigahertz
clocking scheme for the pentium 4 microprocessor. IEEE Journal of Solid-State Circuits
36(11), 1647–1653 (2001)
28. Mallajosyula, A., Zarkesh-Ha, P.: A robust single event upset hardened clock distribution net-
work. In: Integrated Reliability Workshop Final Report, 2008. IRW 2008. IEEE International,
pp. 121–124 (2008). DOI 10.1109/IRWS.2008.4796101
29. Nassif, S.R.: Modeling and forecasting of manufacturing variations (embedded tutorial). In:
Asia and South Pacific Design Automation Conference (ASP-DAC), pp. 145–150 (2001)
30. Ozdal, M.M., Hentschke, R.F.: Exact route matching algorithms for analog and mixed signal
integrated circuits. In: Proceedings of the 2009 International Conference on Computer-Aided
Design, ICCAD ’09, pp. 231–238. ACM, New York, NY, USA (2009)
31. Pelgrom, M., Duinmaijer, A., Welbers, A.: Matching properties of MOS transistors. Journal of
Solid State Circuits (JSSC) 24(5), 1433–1439 (1989)
32. Pillage, L.T., Rohrer, R.: Asymptotic waveform evaluation for timing analysis. IEEE Transac-
tions on Computer-Aided Design 9, 352–366 (1990)
33. Rajaram, A., Damodaran, R., Rajagopal, A.: Practical clock tree robustness signoff metrics.
International Symposium on Quality Electronic Design (ISQED) pp. 676–679 (2008)
34. Ratzlaff, C.L., Gopal, N., Pillage, L.T.: RICE: Rapid interconnect circuit evaluator. In:
Proceedings of the 28th conference on ACM/IEEE design automation, pp. 555–560 (1991)
35. Restle, P., Carter, C., Eckhardt, J., Krauter, B., McCredie, B., Jenkins, K., Weger, A., Mule, A.:
The clock distribution of the POWER4 microprocessor. In: International Solid-State Circuits
Conference (ISSCC), pp. 144–145 (2002)
36. Sapatnekar, S.: Timing. Springer (2004)
37. Steinhogl, W., Schindler, G., Steinlesberger, G., Traving, M., Engelhardt, M.: Impact of
line edge roughness on the resistivity of nanometer-scale interconnects. Microelectronic
Engineering 76, 126–130 (2004)
38. Stine, B.E., et al., D.S.B.: The physical and electrical effects of metal-fill patterning practices
for oxide chemical-mechanical polishing processes. IEEE Transactions on Electron Devices
45(3), 665–679 (1998)
39. Sze, C.: Personal communication (2010). IBM Austin Research Lab, Austin, TX
40. Tam, S., Rusu, S., Desai, U.N., Kim, R., Zhang, J., Young, I.: Clock generation and distribution
for the first ia-64 microprocessor. IEEE Journal of Solid-State Circuits 35(11), 1545–1552
(2000)
41. Teodorescu, T.: Personal communication (2010). ATI Radeon 5870 Clock Designer, Sunny-
vale, CA
42. Tsai, J.L., Zhang, L., Chen, C.C.P.: Statistical timing analysis driven post-silicon tunable clock-
tree synthesis. In: International Conference on Computer-Aided Design (ICCAD), pp. 575–581
(2005)
43. Tsay, R.S.: Exact zero skew. In: International Conference on Computer-Aided Design
(ICCAD), pp. 336–339 (1991)
44. Visweswariah, C.: Death, taxes and failing chips. In: Design Automation Conference,
pp. 343–347 (2003)
45. Visweswariah, C., Ravindran, K., Kalafala, K., Walker, S.G., Narayan, S.: First-order
incremental block-based statistical timing analysis. In: Design Automation Conference,
pp. 331–336 (2004)
46. Wang, R., Koh, C.K.: A frequency-domain technique for statistical timing analysis of
clock meshes. In: IEEE/ACM international conference on Computer-aided design, ICCAD,
pp. 334–339. Piscataway, IEEE Press, San Jose, CA (2007)
47. Wilke, G., Fonseca, R., Mezzomo, C., Reis, R.: A novel scheme to reduce short-circuit power
in mesh-based clock architectures. In: Symposium on Integrated Circuits and System Design
(SBCCI), pp. 117–122 (2008)
48. Ye, X., Li, P., Zhao, M., Panda, R., Hu, J.: Analysis of large clock meshes via harmonic-
weighted model order reduction and port sliding. In: IEEE/ACM international conference on
Computer-aided design, ICCAD, pp. 627–631. Piscataway, IEEE Press, San Jose, CA (2007)
49. Zejda, J., Frain, P.: General framework for removal of clock network pessimism. In: Interna-
tional Conference on Computer-Aided Design (ICCAD), pp. 632–639 (2002)
50. Zhang, L., Yu, W., Zhu, H., Zhang, W., Cheng, C.K.: Clock skew analysis via vector fitting
in frequency domain. In: International Symposium on Quality Electronic Design, ISQED, 9.,
pp. 476–479. Los Alamitos, IEEE Computer Society, San Jose, CA (2008). DOI 10.1109/
ISQED.2008.4479780

Circuit Design For Reliability

Uploaded by

Copyright:

Available Formats

Circuit Design For Reliability

Uploaded by

Copyright:

Available Formats

Ricardo Reis · Yu Cao

Circuit Design for Reliability

ISBN 978-1-4614-4077-2 ISBN 978-1-4614-4078-9 (eBook)

© Springer Science+Business Media New York 2015

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

Ricardo Reis, Yu Cao, and Gilson Wirth

The scaling of CMOS technology to the nanometer regime inevitability increases

© Springer ScienceCBusiness Media New York 2015 1

B. Kaczer, T. Grasser, J. Franco, M. Toledano-Luque, J. Roussel, M. Cho,

B. Kaczer () • J. Franco • M. Toledano-Luque • J. Roussel • M. Cho

© Springer ScienceCBusiness Media New York 2015 5

of characteristic time scale, especially in pFETs—is then argued to point to a dis-

2 Brief Overview of BTI

Fig. 2.1 Shift in pFET 25

Fig. 2.3 Total degradation 100

ΔVth = R(trelax) + P (mV)

In an arbitrary FET of an arbitrary digital circuit, the average probability of a

5 Similarity Between BTI Relaxation and Low-Frequency

Long, log(t)-like behavior of Vth without a characteristic time scale is typically

6 Semi-quantitative Model for BTI Relaxation

In order to visualize this common property it is beneficial to consider an equivalent

Fig. 2.4 (a) A characteristic a 70 101

-dΔVth /dt (Vs-1)

C = 100 ... 10N C = 100 ... 10N

Fig. 2.6 The plateau in DF 7

7 Properties of Individual Defects

Fig. 2.7 Characteristic Vth

Threshold voltage shift x10-3 (V)

0.001 0.01 0.1 1 10 100

Fig. 2.8 Two-dimensional

8 Modeling Properties of Individual Defects

A model of the above-described properties of individual gate oxide defects can be

Fig. 2.11 Histogram of

then illustrates this concept on a simulated example of a deeply scaled pFET

9 BTI Distribution in Deeply-Scaled FETs

As CMOS devices scale toward atomic dimensions, device parameters become

11 Improving PBTI with Rare-Earth Incorporation

PBTI was considered a minor problem in technologies based on SiO2 . It arose

12 Improving NBTI in High-Mobility SiGe pFETs

1. J. H. Stathis and S. Zafar, Microelectronics Reliab. 46, 270 (2006).

22. V. Huard, C. Parthasarathy, C. Guerin, T. Valentin, E. Pion, M. Mammasse, N. Planes and

Gilson Wirth and Roberto da Silva

Abstract Charge trapping phenomena is known to be a major reliability concern

We start by discussing MOSFET low-frequency (LF) noise, which is known to

© Springer ScienceCBusiness Media New York 2015 21

Fig. 3.1 Traps within a few

2 Charge Trapping Events as a Source of Noise

In nanometer scale MOSFETs the alternate capture and emission of carriers at

3 Power Spectrum of the RTN Noise due to a Single Trap

Fig. 3.6 Energy band Gate voltage applied to the device

Gate Oxide Substrate

In order to derive the low frequency noise spectrum of cyclo-stationary RTN, we

A .s/ D< x .t/  x .t C s/ >average D P .x .t/ D 1/  P11 .s/ (3.2)

where T is the periodR of the cyclo-stationary excitation, and symbol <•> is an

4 Approximation for Excitation Frequencies Higher than the

D nT < £e .t/ > C < £c1.t/ >

For small values of T a Taylor expansion may be employed, leading to

Fig. 3.7 Noise Power

which is evaluated as being

“eq D <1=£e .t/ >=<1=£c .t/ > (3.10)

Figure 3.7 depicts the behavior of noise power as a function of “eq .

5 Average Power Spectrum of the RTN Noise due

6 Square Wave Excitation

 ˛e Eon =kB T C .1 ˛/ e Eoff =kB T

Assuming statistical independence of the random variables, the average noise of a

g .Et / D aEt2 a .Ec Ev / Et C k (3.18)

where a is a fitting parameter. The integral of g(Et ) from Ev to Ec is normalized to

Figure 3.10 depicts a U-shape trap density.

7 Variability in the Power Spectrum of the RTN Noise due

Fig. 3.10 U-shaped trap 0.5

9 The Charge Trapping Component of Bias

A .s/ D< x .t/ x .t C s/ >average D P .x .t/ D 1/ P11 .s/ (3.2)

˛e Eon =kB T C .1 ˛/ e Eoff =kB T