PATH DELAY TUNING FOR PERFORMANCE GAIN IN THE FACE OF
RANDOM MANUFACTURING VARIATIONS
by
Kautalya Mishra
A thesis submitted to the Graduate Faculty of
Auburn University
in partial ful llment of the
requirements for the Degree of
Master of Science
Auburn, Alabama
May 9, 2011
Keywords: Process variability, Path delay improvement, Performance tuning
Copyright 2011 by Kautalya Mishra
Approved by
Adit Singh, Chair, James B. Davis Professor of Electrical and Computer Engineering
Vishwani Agrawal, James J. Danaher Professor of Electrical and Computer Engineering
Victor Nelson, Professor of Electrical and Computer Engineering
Abstract
Moore?s laws predictions of transistor densities doubling every two years in an integrated
circuit has held true for the past  fty years, and is predicted to hold true in the coming few
years. But as technologies shrink to smaller dimensions and scaling becomes more aggressive,
a number of factors are beginning to hinder the urge towards miniaturization.
Process variability, that introduces parametric variations in a device, is one such factor
that is today seriously limiting clock rates in large synchronous designs. These variations
are predicted to increase signi cantly with device scaling as silicon technology approaches
the end of roadmap.
Prior research work on process variability shows that factors contributing to process
variations a ect the threshold voltage (Vth) of transistors. Varying Vth corresponds to
varying path delays in a circuit, a few of which are signi cant outliers, with path delays
much larger than the nominal delays the circuits were designed to have.
These variations make it hard to make any pre-fabrication predictions on the slowest
paths in designs that would make room for a speeding methodology because of the random
way in which they are distributed in a chip. With circuits today including several hundred
million transistors, virtually every design will have a dozen of these outlier transistors that
would limit the maximum clock frequency at which the chip can be correctly operated.
This thesis studies a new design architecture that allows for tuning and speedup of
exceptionally slow paths in a chip to recover the lost performance and signi cantly increase
the average clock speed attainable by the manufactured parts.
ii
Acknowledgments
I?m very thankful and honored for having got the opportunity to work with Dr. Adit
Singh, whose deep insight, energy and conviction was a source of inspiration and motivation.
His words of advice and his work ethics have not just helped me work better academically,
but also perform beyond it in other areas of life. I?m deeply indebted to him for having given
me the opportunity to pursue research for him, and thankful for the lessons learned and the
knowledge gained.
I?d like to thank Dr. Vishwani Agrawal for being on my thesis defense committee. Over
the period of my Master?s pursuit I?ve learned a number of lessons on dedication and humility
while observing Dr. Agrawal both on and o classroom hours. His inquisitive nature and
ease of approachability have been instrumental in me overcoming a number of brick walls in
my understanding of various concepts.
I?d also like to thank Dr. Victor P. Nelson for also being on my thesis defense committee,
and for the knowledge imparted through his courses, that have been of utmost importance
as it satis ed a prerequisite requirement for entering the industry and starting my electrical
engineering career.
I?d like to thank Auburn University for providing me with the necessary facilities and
softwares that were required for my research work, and the people at Auburn who?ve been
a part of my memorable stay at Auburn.
Last but not the least, I?d like to thank my family for their love, support and words of
wisdom, and for letting me know that they are always there for me no matter what.
iii
Table of Contents
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ii
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix
List of Abbreviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Literature Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1 Historical Overview of Process Variability . . . . . . . . . . . . . . . . . . . 4
2.2 Sources of Variability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2.1 Random Dopant Fluctuation (RDF) . . . . . . . . . . . . . . . . . . 6
2.2.2 Line-Edge Roughness (LER) and Line-Width Roughness (LWR) . . . 7
2.2.3 Oxide thickness variation due to interface roughness . . . . . . . . . . 9
2.2.4 Polysilicon granularity (PSG) . . . . . . . . . . . . . . . . . . . . . . 10
2.2.5 High K dielectric morphology . . . . . . . . . . . . . . . . . . . . . . 10
2.2.6 Other sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Impact of Process Variability . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3 E ects of Outlier Presence on Path Delays . . . . . . . . . . . . . . . . . . . . . 13
3.1 Normal distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 VOL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.3 E ects of process variability on high performance applications . . . . . . . . 17
4 CMOS tunable gate architecture . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.1 CMOS tunable gates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.2 Tuning strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
iv
4.3 Sizing of tuning transistors . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.4 Faulty tuning transistors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.5 NAND gate simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5 SPICE Netlist Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.1 Mentor Graphics Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.1.1 ModelSim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
5.1.2 Leonardo Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.1.3 Design Architect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.1.4 IC Station . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.1.5 LVS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.1.6 PEX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
5.2 Tuning transistors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
6 Tuning implemented on very large circuits . . . . . . . . . . . . . . . . . . . . . 35
6.1 Building large circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
6.1.1 Di culties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
6.1.2 Basic algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
6.1.3 Standard circuit used to create larger circuits . . . . . . . . . . . . . 36
6.2 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
6.2.1 Primitive model  les . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
6.2.2 Single EXOR tree simulations . . . . . . . . . . . . . . . . . . . . . . 40
6.2.3 Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6.2.4 EXOR tree simulation: example . . . . . . . . . . . . . . . . . . . . . 44
6.2.5 Picking ?N? sub-circuits to create a larger circuits . . . . . . . . . . . 46
7 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
7.1 Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
7.2 Observation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
7.3 Bene ts obtained . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
v
7.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
8 Power Dissipated with tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
8.1 Static power dissipation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
8.2 Simulation di culty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
8.3 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
8.4 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
8.4.1 Bins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
8.4.2 SPICE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
8.4.3 Picking 10,000 sub-circuits to create a larger circuit . . . . . . . . . . 56
8.5 Simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
vi
List of Figures
1.1 Processor clock frequency versus Time . . . . . . . . . . . . . . . . . . . . . . . 2
2.1 Random Dopant Fluctuations in sub-micron technologies . . . . . . . . . . . . . 7
2.2 De nitions of LER and LWR . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.1 Threshold voltage variation with scaling . . . . . . . . . . . . . . . . . . . . . . 14
3.2 Standard inverter with an input and an output load . . . . . . . . . . . . . . . . 15
3.3 VOL versus Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
3.4 Process variability e ects on pre-optimized designs . . . . . . . . . . . . . . . . 17
4.1 Tunable CMOS architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.2 Ho man model of Sequential circuits with tunable gates . . . . . . . . . . . . . 21
4.3 Rising transition speed-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.4 Falling transition speed-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.5 Voltage division on tuned nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.6 Tunable NAND gate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
5.1 Tuning circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
5.2 Inverter as a tuning circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
vii
5.3 Inverted inverter as a tuning circuit . . . . . . . . . . . . . . . . . . . . . . . . . 33
6.1 EXOR tree circuit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
7.1 Simulation case : Vdd=1.5V sigma=0.075 . . . . . . . . . . . . . . . . . . . . . 48
7.2 Simulation case : Vdd=1.2V sigma=0.075 . . . . . . . . . . . . . . . . . . . . . 48
7.3 Simulation case : Vdd=1.5V sigma=0.150 . . . . . . . . . . . . . . . . . . . . . 49
7.4 Simulation case : Vdd=1.2V sigma=0.150 . . . . . . . . . . . . . . . . . . . . . 49
viii
List of Tables
3.1 Normal distribution - Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 Delay versus VOL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
4.1 Tunable NAND gate simulations . . . . . . . . . . . . . . . . . . . . . . . . . . 27
6.1 Number of circuits tuned . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
6.2 Tunable EXOR tree simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6.3 Performance gain for a Single EXOR tree . . . . . . . . . . . . . . . . . . . . . 45
7.1 Performance gain for a circuit with 10,000 EXOR trees . . . . . . . . . . . . . . 50
8.1 Bins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
8.2 Power dissipated with tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
ix
List of Abbreviations
SiO2 Silicon-dioxide
 Mean
 Standard deviation
.pm Primitive Model  le
VHSIC Very High Speed Integrated Circuits
CMOS Complementary Metal-Oxide Semiconductor
GND Ground
HfO2 Hafnium-dioxide
ITRS International Technology Roadmap for Semiconductors
LER Line Edge Roughness
LWR Line Width Roughness
MOSFET Metal-Oxide Semiconductor Field E ect Transistor
MOS Metal-Oxide Semiconductor
NMOS N-type MOS transistor
OPC Optical Proximity Correction
PMOS P-type MOS transistor
PSG Polysilicon granularity
x
PSM Phase-Shifting Mask
RDF Random Dopant Fluctuations
RET Resolution Enhancement Techniques
SDE Source-Drain Extension
Si Silicon
TOX Gate oxide thickness
Vdd Supply voltage
VHDL VHSIC Hardware Description Language
VOL Di erence between Vdd and Vth
Vth Threshold voltage
xi
Chapter 1
Introduction
The electronic industry has come a long way ever since the integrated circuit was in-
vented in 1958. Gordon E. Moore?s predictions, popularly known as Moore?s Law, on tran-
sistor densities doubling every two years in an integrated circuit have held true so far, and is
expected to hold true in the coming next decade at least. Increased complexity, faster clock
rates, increased storage capacity and the ability to put in a lot more in to the same area
have been the major driving forces favoring scaling.
But sustaining the scaling trend as forecast by Moore?s Law is now facing serious chal-
lenges [1] in the fundamental limitations imposed by the inherent physical properties of the
underlying materials, required tolerances down to a few atomic dimensions in the manu-
facturing processes, and statistical anomalies arising from the  nite number of atoms con-
stituting device structures in highly scaled technologies. Even though the scaling trend as
predicted by the International Technology Roadmap for Semiconductors (ITRS) is expected
to continue for another decade at least, performance gains measured in terms of processor
speeds is expected to saturate or even fall back a little. Figure 1.1 represents the processor
clock frequency variation with time.
A majority of the work done in this area, to compensate for the drop in processor
speeds, has been in the  eld of multiprocessing, where microprocessors are being designed
with multiple cores to improve performance through parallelism. Indeed impressive gains
have been observed in cases of dual and quad core processors, which have even inspired
thoughts of incorporating a dozen or even several hundred cores in future advanced designs.
But according to Amdahl?s law, and backed by extensive research, it has been shown that
additional processors only yield diminishing returns in most general purpose applications.
1
Figure 1.1: Processor clock frequency versus Time
Hence, in the long run, having multiple cores is not going to be a viable solution to improving
performance. It is imperative that focus be given to addressing challenges of scaling that
limit clock frequencies.
At the small dimensions devices are being fabricated today, manufacturing variations
that produce changes in device parameters are found to occur randomly and in signi cant
amounts. Because of their random nature, these variations are beyond the control of de-
signers for them to make any prefabrication changes to their designs to overcome the e ects
of these variations. Such random process variations occur signi cantly in nanoscale devices,
and are expected to be more signi cant in the 22nm technology and beyond.
Large complex designs today have several millions of transistors, and it is statistically
likely that every manufactured part will have dozens of these slow transistors that are per-
formance outliers, lying in the far ends of the distribution, beyond 4-6 standard deviations
away from the nominal value. A presence of such an outlier along a path can signi cantly
push the path delay up, and hence limit the frequency at which the device can be operated.
The focus of this thesis is to study a new tuning technique that would enable post-
manufacture speedup of exceptionally slow paths, and hence provide the ability to push
clock frequencies up to help get back much of the lost performance. The aim is to bring
2
down the outlier path delays as close to the nominal average case delay, without changing
the logic levels at which the gates switch, and ensuring the extra power dissipation is within
acceptable limits.
Also, variations in processor speeds make it imperative that every chip or core be indi-
vidually \speed binned" and operated at the fastest clock rate that it can reliably sustain,
to take advantage of this tuning.
Tuning, again, is done post manufacture and not during the design stage. Process
variability being random introduces parametric variations randomly, hence taking away any
scope of predicting where the outliers might be to make any changes during the design stage.
It is also assumed in this work that the digital circuitry is designed in CMOS logic.
This assumption is valid as CMOS logic continues to be the most power e cient logic, and
is hence expected to be the dominant logic to design digital circuits with.
The aim of this work has been to enable tuning by adding a minimum amount of extra
circuitry in each CMOS gate to speed up a particular transition at the output. Hence,
once an exceptionally slow transistor is identi ed, cells along the path are programmed to
speed up the slowest transition. This tuning, limited to only a few cells that are statistical
outliers, can help push the clock frequencies up by a signi cant amount. This also happens
with minimal impact on the total power dissipated in the chip, as will be discussed in the
following chapters.
The following sections of this thesis are arranged as follows. Literature survey is pre-
sented in Chapter 2, e ects of the presence of an outlier in Chapter 3, the CMOS tunable
gate architecture that is to be included in the design is introduced in Chapter 4, followed by
the steps taken to extract the SPICE netlist in Chapter 5, an outline of the method and the
algorithm used to show the e ectiveness of the design in Chapters 6, the simulation results
in Chapter 7, power dissipation simulation and results in Chapter 8, and  nally a conclusion
in Chapter 9.
3
Chapter 2
Literature Survey
This chapter presents a brief study on the present and past research work done on
process variability. The tuning architecture proposed in this work is novel, but the subject
of process variability as such has been dealt with and studied in the past. A historical
overview of process variability is  rst presented. That is followed by sections on the critical
sources of variability. A  nal section on the impact of process variability is also presented.
2.1 Historical Overview of Process Variability
The semiconductor industry has distinguished itself by its rapid pace of improvement,
which it owes largely to the industry?s ability to exponentially decrease the minimum feature
sizes used to fabricate integrated circuits [1]. But as the feature sizes get smaller, the
variability associated with fabrication becomes more signi cant. It may seem like process
variability is a new challenge, with the large amount of research literature today dedicated
to it. But in fact process variability has always been a critical aspect of semiconductor
fabrication [11].
Random variations in a semiconductor was  rst discussed in Shockley?s 1961 paper
titled \Problems related to p-n junctions in silicon," where the e ect of statistical spatial
 uctuations of donor and acceptor atoms with its resulting impact on the breakdown voltage
in the junction is considered [12]. After that a number of other works on the e ects of random
ion implantations have been done. The most prominent ones include works by Schemmert
and Zimmer on the sensitivity of threshold voltage (Vth) to implantation energy [13] , and
works by Yokoyama et al. [14] on studying Vth sensitivity using Monte Carlo approach.
4
As can be seen, process variation has been a subject of interest for a long time. It has,
though, become more signi cant in recent years as feature sizes get smaller and as we slowly
approach the end of the roadmap, which poses certain fundamental limitations on further
scaling.
2.2 Sources of Variability
Process variability may be categorized under two types in advanced CMOS technologies
- local or intra-die variations and global or inter-die variations [8]. Local variability is over
a short distance and normally within a die, while global variability is over larger distances
and normally seen between die-die and wafer-wafer.
Global variability causes a shift in the mean value of sensitive design parameters such
as channel length (L), channel width (W), layer thickness, resistivity, doping concentration,
and body e ect, while local variability introduces (1) systematic variability and (2) random
variability.
Systematic variability includes variability caused by optical proximity corrections (OPC),
phase-shifting mask (PSM), layout-induced strain, and well proximity e ect, and can be ad-
dressed by more controlled Resolution Enhancement Techniques (RET) and layout designs.
Random variability includes several sources of which the critical ones are Random
Dopant Fluctuations (RDF) [8, 11, 15, 16] , Line-edge and Line-width roughness (LER
and LWR) [8, 11, 17, 18, 19, 20] , variations in gate oxide thickness (TOX) due to interface
roughness [8, 11, 22] , polysilicon granularity [8, 11, 21, 22] , and high-K dielectrics with
metal gates [8, 11, 21, 22].
Of the above variability types, random variability is the most critical because its im-
pact on circuit performance is becoming increasingly signi cant for technology nodes below
45nm. The large outlier delays, observed in almost every device, that are almost four to
 ve times above the average case worst case delays arise out of random variability. Also,
addressing random variability requires innovative process and circuit design techniques and
5
device modeling. The focus of this work, hence, has been to address the issue of random
local variability.
A brief description of some important sources of random variability is presented below.
2.2.1 Random Dopant Fluctuation (RDF)
RDF has an increasingly signi cant e ect on the MOS Vth in the sub-micron technolo-
gies. In MOSFET?s, transistor channels are doped with dopant atoms to control the Vth. To
keep short channel e ects from degrading the Vth in technology nodes below 90nm, the dop-
ing concentrations are kept very high. For a MOSFET device with e ective channel length,
Leff, e ective width, Weff, channel doping concentration, NCH, and source-drain extension
(SDE) junction depth, xj, the total number of dopant atoms in the channel, according to
Saha, K. S. [8] , is given by:
Ntotal;chan = NCH:(Weff:Leff):xj (2.1)
According to Equation 2.1, as dimensions are scaled the number of doping atoms, given
by Ntotal;chan, comes down drastically even though the doping concentration goes up,
because the transistor dimensions Leff, Weff and xj reduce. Technology scaling involves
reducing transistor area by a half with every generation. Hence, Ntotal;chan would decrease
exponentially over technology generations; 45nm technology node has only around a 100
dopant atoms in its channel [11]. Any small change in the number of atoms would hence
imply a signi cant e ective change in the doping concentration, and hence a signi cant
di erence between the Vth of one transistor and another. This is what is referred to as
process variability. Its contribution towards device mismatch can be studied through the
following equation,
 Vtran =
 4p4:q3: 
Si: B
2
!
:TOX 
ox
:
0
@
4pN
q
Weff:Leff
1
A = 1p
2:
0
@ c2q
Weff:Leff
1
A (2.2)
6
Figure 2.1: Random Dopant Fluctuations in sub-micron technologies
where q,  Si,  ox, and  B are the electron charge, permittivity of the silicon, permittivity
of SiO2, and the built-in potential of S/D-to-channel PN junction of MOFETs, respectively.
C2 is a constant that depends on the NCH and TOX, and hence depends on Ntotal;chan.
Even though Ntotal;chan decreases with scaling, Weff and Leff reduce as well, and hence
e ectively the device mismatch becomes bigger with scaling [15].
A 3D numerical model with an adaptive local meshing scheme that allows prediction of
Vth for arbitrary dopant pro les (see Figure 2.1 ) was developed and simulation results for
45nm and 65nm were compared to the observed variations [16]. RDF was found to be 65%
of the total  Vt in 65nm silicon, and 60% of the total  Vt in 45nm silicon.
2.2.2 Line-Edge Roughness (LER) and Line-Width Roughness (LWR)
A second important factor responsible for process variability is one that arises out of
variations in gate patterning, leading to non-ideal (rough) line edges. LER is a result of sub-
wavelength lithography, which the semiconductor industry has been using for patterning
7
Figure 2.2: De nitions of LER and LWR
transistors since the 0.25 m technology node. For example, fabrication processes previously
used the wavelength of light,  =248nm, to pattern the minimum feature size, Critical Dimen-
sion CDmin=250nm and 180nm transistors. Wavelength  decreased to 193nm for 130nm
technology, and has remained the same since, even for 65nm transistors.
Until ultra-violet technology becomes available, sub-wavelength lithography will con-
tinue to be used for patterning, and will cause LER and LWR e ects in scaled MOSFETs
[8, 11].
Impacts of LER and LWR include increases in sub-threshold current and degradation
in the Vth characteristics [17, 18, 19, 20]. Diaz et al [17] found transistor performance
degradation to occur with an increase in the observed LER values. Experiments performed
by Kim et al. [18] showed that LER e ects began for gate lengths below 85nm. They also
observed a four-order increase in the standard deviation of the sub-threshold current for the
smallest gate lengths in their study. Fukutome et al. [19] in their experiments, observed
that roughness of extension edges induced by gate LER depended on the implanted dose,
halos (pockets), and various co-implantations. They showed an improvement of 4nm in
8
Vth roll-o with a decrease in the average LER, and con rmed that co-implants induced
a degradation of 5mV in the standard deviation of Vth. Another important observation
noted in experiments performed by Asenov et al. [20] was that LER and RDF e ects act
in a statistically independent manner, and that LER induced  uctuations have stronger
channel length dependence and are hence expected to supplant RDF as the dominant source
of variation.
The Vth-mismatch due to LER depends on the variability in Weff of the MOSFETs
and is given by [11] :
 VTH;LER/ 1qW
eff
< VTH;RDD (2.3)
2.2.3 Oxide thickness variation due to interface roughness
Conventional CMOS technologies with Silicon-dioxide (SiO2) gate dielectrics are sub-
jected to process variability introduced by the Si/SiO2 and SiO2/polysilicon-gate interface
roughness through TOX variation [8]. In advanced MOSFETs with TOX of about 1nm,
the interface roughness is found to be comparable to the TOX, and hence the variations
associated with it can be about 50%. Such variations in the TOX introduce variations in the
gate current Ig which produces a voltage drop in the polysilicon gate that changes the Vth
signi cantly. Studies done by Asenov et al. [22] show that intrinsic Vth  uctuations induced
by local TOX variations become comparable ( 30mV) to voltage  uctuations introduced
by RDF for conventional MOS devices with dimensions 30nm and below. It is also found
that process variations introduced by TOX variations due to interface roughness is found to
be statistically independent of process variations introduced by RDF [8, 11]. The e ective
contribution towards Vth variation through the two e ects is hence given by:
 VTH;total =
r 
 VTOXTH
 2
+ ( VRDDTH )2 (2.4)
9
2.2.4 Polysilicon granularity (PSG)
PSG enhances the gate dopant di usion along the grain boundaries which leads to non-
uniform polysilicon gate doping and potential localized penetration of the dopants through
the gate oxide into the channel region. The most signi cant source of  uctuation within the
polysilicon gate though, is the Fermi level pinning at the boundaries between grains due to
the high density of defect states. Local variations in the potential of up to 0.6V within the
gate would re ect on  uctuations in the surface potential ( s) within the MOSFET channel
leading to Vth mismatch between devices. In the sub 65nm MOSFETs these  uctuations
are comparable to those introduced by RDF [8, 11, 21, 22].
2.2.5 High K dielectric morphology
Advanced CMOS technologies use high-K dielectrics like HfO2 with metal gates to pro-
vide a thicker physical TOX to reduce the amount of gate leakage and ensure a thin electrical
TOX required for the continuous scaling of MOSFETs. Its inclusion, however, introduces sig-
ni cant process variability due to the Si/high-K and high-K/MG interface roughness which
causes mobility degradation and TOX variation, and the phase separation that is created
between the crystallized grain and the amorphous SiOx matrix which causes  uctuations
in the channel potential under the gate. Also, the polycrystalline HfO2 with random grain
orientations causes dielectric constant to vary across the gate oxide. Hence, the surface
potential  s varies with L and causes variability in Vth. This Vth variability due to the
presence of high-K dielectrics increases with scaling, and is expected to be more than 30 mV
for 10nm MOSFETs with TOX of about 4nm [8, 11, 21, 22].
2.2.6 Other sources
Other sources of process variability, besides the critical sources mentioned above are
variations associated with patterning proximity e ects, variations associated with polish such
as shallow-trench isolation, gate and interconnect variation, variations associated with strain,
10
high stress capping layers, and embedded silicon-germanium, and variations associated with
implants and anneals.
2.3 Impact of Process Variability
Random variations introduced through RDF, LER/LWR, TOX variation, PSG, dielec-
tric variations, and other sources has a critical impact on the yield, performance and reli-
ability of the manufactured circuits. As a result parameters de ned for transistors at the
design stage vary signi cantly, post-manufacture, from their given nominal values. These
variations are expected to get only worse with scaling.
All of the sources mentioned above primarily degrade Vth. RDF and LER contributions
towards Vth variation are described in Equation 2.2 and Equation 2.3; of the two sources
LER is expected to become more critical as scaling continues. TOX variation due to interface
roughness has been shown to degrade threshold voltage as well, and for transistors with 1nm
TOX the Vth variation introduced is almost comparable to that introduced by RDF. PSG
and high-K dielectric e ects degrade the threshold voltage Vth by a ecting the surface
potential ( s).
Hence the focus of this work has been on studying threshold voltage variability in circuits
designed with CMOS logic. Only CMOS logic is assumed as it is the most power e cient
logic among all logic types, and is expected to continue to be the dominant logic type to
design with in future technologies as well.
Even though every process variability source a ects Vth in a statistically independent
manner, in this work values of Vth are drawn from a single normal distribution assuming the
e ects of all sources have been included. Hence for the purpose of simulation, every transistor
in every circuit simulated has Vth drawn from a nominal distribution whose mean and
standard deviation values are speci ed by the technology generation in which the transistors
are being designed.
11
In conclusion, random process variability is an increasing concern as dimensions shrink
and scaling continues. Di erent sources that contribute to variability have been studied and
their e ects analyzed. The next chapter studies the impact of variability on path delays.
12
Chapter 3
E ects of Outlier Presence on Path Delays
Extensive prior research suggests Deep Sub-Micron Technologies (DSM) are prone to
high process variability [1-24]. Distributions representing parametric values of transistors,
post fabrication, indicate a high standard deviation, implying a wide spread of the distribu-
tion, and a trend that the standard deviation would increase with scaling.
3.1 Normal distribution
Figure 3.1 compares two distributions of Vth for two di erent technology generations.
As can be seen, the spread of the distribution gets wider while the height of the distribution
gets shorter, as we scale from one technology generation to another.
Assuming a normal distribution for the spread, the results shown in Table 3.1 can be
arrived at [25] :
Range(a,b) Population within Approximately 1
( +a  ; +b  ) (a,b) in
(0,1) 34.13% 3
(1,2) 13.59% 7
(2,3) 2.14% 47
(3,4) 0.13% 770
(4,5) 0.003% 30000
(5,6) 0.000028% 3.5 million
(6,1) 0.0000001% 1 billion
Table 3.1: Normal distribution - Statistics
Table 3.1 represents the percentage of transistors that fall within the ranges speci ed
on the right side of the distribution. A similar table can be constructed indicating the
percentage of transistors that would fall in the left side of the distribution; being symmetrical
13
Figure 3.1: Threshold voltage variation with scaling
the percentages would be the same on both sides. Vth greater than nominal, i.e. falling on
the right side of the distribution, would correspond to a slow transistor, while Vth less than
nominal, i.e. falling in the left side of the distribution, would correspond to a fast transistor.
It is the slow transistors that are of interest, as their presence on a path pushes path delays
up by a factor which is greater than the speedup achieved by a symmetrically falling low
Vth transistor, and is hence in e ect responsible for the slow down.
Assume a nominal threshold voltage =0.5V, a standard deviation =50mV, and Vdd=1V,
for a particular deep sub-micron technology [26]. Here  +5* = 0.75V. By Table 3.1, 1 in
3.5 million transistors lie beyond 5 standard deviations, i.e. approximately 300 in a design
with a billion transistors.
Most designs today have billions of transistors in them, and hence, statistically there is
every chance to  nd at least a few hundreds of such outlier transistors in every fabricated
chip that lie far away in a distribution, at least 5 standard deviations away.
14
Figure 3.2: Standard inverter with an input and an output load
3.2 VOL
In the above example, Vth is greater than 0.75V beyond 5 standard normal deviations,
i.e. the di erence between Vdd and Vth, de ned as VOL, is less than 0.25V. As Vth gets
larger, it gets closer to Vdd, the supply voltage, and the delays increase exponentially.
To get a measure of this increase, simulations were done on a SPICE netlist of an
inverter (Figure 3.2) with an input load and an output load, by varying Vth of the pull-up
PMOS from 0.2V through the nominal value of 0.4V to 0.7V, for a supply voltage Vdd of 1V.
The simulations were done using 180nm technology  les, and the extraction of the SPICE
netlist through the various Mentor Graphics tools. More on the technology and tools used
is discussed later.
Vdd = 1V
VOL=Vdd-Vth Vth Delay Change from nominal
(V) (V) (ps) (%)
0.3 0.7 377 508
0.4 0.6 167 170
0.5 0.5 95 53
0.6 0.4 62 0
0.7 0.3 44 -29
0.8 0.2 33 -46
Table 3.2: Delay versus VOL
Table 3.2 gives a sense of how delay varies with Vth. As Vth increases, inverter delays
increase exponentially. For an increase of 0.2V above the nominal Vth, the increase in delay
15
Figure 3.3: VOL versus Delay
is close to three times the nominal delay value. Hence, for transistors that lie beyond  ve
standard normal deviations, the increase in delays will be more than four times the nominal.
The presence of such an outlier in a path can push up delays by a signi cant amount.
To illustrate using an oversimpli ed example, assume an 8 length inverter chain, with all
inverters having nominal delays ?X?. The total delay of the inverter chain would be ?8X?.
Now assume a ?4X? increase in the delay of an outlier inverter that is on such a chain. The
total delay is now pushed to ?12X?, implying a 50% increase in path delays, and hence a 33%
decrease in the clock frequency.
It is important to note that even though the distribution is symmetric, implying the
presence of an equal number of fast transistors along a path in the circuit, the total delays
on a path do not get averaged out. Figure 3.3, which is a plot between VOL and Delay,
illustrates that for an equal change in Vth, the speed up is a lot smaller than slow down.
Hence, there is a very small chance that a slow outlier node in a path would get compensated
by a fast switching node.
16
Figure 3.4: Process variability e ects on pre-optimized designs
3.3 E ects of process variability on high performance applications
The presence of a slow outlier on a path doesn?t necessarily imply that the path becomes
critical. There are chances of the path itself being small in comparison to other paths. The
large delays introduced by an outlier node in such a case might not make the path critical.
But most high performance applications today are speed optimized in the design stage
to have close to approximately equal path delays so as to save on power. In such high
performance designs where all paths are close to critical, the presence of an outlier node on
any path can push the critical path delay up by a signi cant amount.
Figure 3.4 illustrates the case. It represents a pipelined architecture with the rectangles
representing registers and arrow lengths representing path delays. The minimun clock period
of a design equals the worst case delay. Hence, in the presence of an outlier, even though
average case delays do not change by much, the clock period increases solely because of the
exponential increase in the delay of a few outlier paths.
In conclusion, statistically every fabricated design with close to a billion transistors will
have at least a few hundreds of outliers whose presence on any path can push clock delays
up by a signi cant amount and hence restrict the frequency at which the circuit can be
17
operated. The aim of post manufacture tuning is to bring down the clock period close to
the average case delays.
18
Chapter 4
CMOS tunable gate architecture
This chapter introduces the tunable CMOS gate architecture, which forms the basis of
this work. The tunable gate architecture design is  rst introduced, followed by the tuning
strategy, a study on the sizing of the tuning transistors, the occurrence of an interesting
case, and a  nal simulation result that shows the e ectiveness of the tuning methodology on
a single NAND gate.
4.1 CMOS tunable gates
Every CMOS gate has a pull-up network with PMOS transistors, and a pull-down
network with NMOS transistors. Under static conditions either the pull-up network is ?ON?
or the pull-down network is ?ON?. The output load capacitance of the gate charges to Vdd
when the pull-up network is ?ON?, while the capacitance discharges when the pull-down
network is ?ON?. It is this charging or discharging time that determines the speed at which
the gate switches.
The presence of a parallel path can speed up the charge time or the discharge time; it
is this principle that lays down the foundation of the tuning strategy. A parallel path for
charging and discharging the output load capacitance is provided to speed up an otherwise
slow transistor. To introduce this parallel path in the design, a parallel PMOS transistor with
tuning capability is introduced in to the pull-up network, and a parallel NMOS transistor
with tuning capability is introduced in to the pull-down network. The resulting architecture
resembles the design shown below in Figure 4.1.
To include the parallel path of charging or discharging, the corresponding tuning tran-
sistors have to be turned ?ON?. This is only done after diagnosis has been performed and
19
Figure 4.1: Tunable CMOS architecture
the slow outliers identi ed. The gate terminals of the tuning transistors are connected to
a switch which can be turned ?ON? or ?OFF?. The switches are ?OFF? when no tuning is
required, and ?ON? while tuning is done. Once turned ?ON?, logic of that particular gate
switches from a CMOS type to a pseudo-NMOS type.
The tuning transistors have to be sized appropriately to provide good speed up without
a ecting the output voltage logic levels. More is discussed about the sizing of the tuning
transistors in the subsequent sections of this chapter.
It is important to note that every node in the design comprising a pull-up and a pull-
down network has to have this tuning capability in it (Figure 4.2). This is because of the
inability to make any prefabrication predictions on the location of the outliers; outliers are
created randomly and can hence fall anywhere on the chip. The chances of a gate having
both a faulty PMOS and a faulty NMOS are very small; in its occurrence the fabricated chip
would be discarded as yield loss.
There is also a possibility of the occurrence of a rare and an interesting case; it is possible
that either the pull-up network or the pull-down network has a high resistance slow outlier
and the corresponding tuning transistor, that would have to be turned ?ON? to bring down
20
Figure 4.2: Ho man model of Sequential circuits with tunable gates
the gate delay, be a low resistance low Vth outlier. In such a case tuning would not be
possible because of the possible degradation of the voltage logic. This occurrence again is
very rare and is considered a yield loss. More is discussed on this case later.
There are drawbacks to adding tuning transistors. One of the most signi cant being
the additional parasitic capacitances that are added that push path delays up. But it will
be seen in subsequent sections of this work that on tuning the slow outliers, the worst case
path delays of the fabricated chip can be pulled well below the worst case path delays of a
fabricated circuit that has no tuning capability in it.
Also, turning ?ON? a particular tuning transistor to speed up the slow transition would
slow down the complementary transition. But it will be seen that the slowdown is much less
than the speedup.
There are other important e ects of including tuning transistors, a few being additional
power dissipation and increase in the amount of area required. It will be seen that these
21
e ects are minimal and within a ordable limits of adding tuning circuitry. More is discussed
in the subsequent chapters.
Such a tuning architecture has been studied before by Ashouei et al. [5] but only as
a defect tolerance methodology. In the presence of a fault, say in the pull-up network, the
entire pull-up network is disconnected from the power rail and is replaced by a properly sized
single always ?ON? pull-up transistor, hence converting the CMOS gate to a pseudo NMOS
structure. A similar technique is adopted to replace a faulty pull-down network with an
always ?ON? pull-down transistor. This method is similar to the one proposed in this work,
with the di erence that no additional circuitry is required to disconnect a network. Such
a defect tolerant methodology, however, has not been applied for performance tuning. As
will be seen, the potential exists to combine both defect tolerance and performance tuning
in aggressive CMOS technologies.
4.2 Tuning strategy
As discussed in previous chapters, statistically there is every chance that a large design
will have at least a few hundreds of outlier transistors that lie far away in the distribution,
whose single presence on a path pushes path delays up by a signi cant amount. The slow
transistor could either be in the pull-up network or the pull-down network; chances of both
the pull-up and the pull-down network having a slow transistor are very small, in which case
the chip would count as faulty and would fall under yield loss.
The presence of a slow transistor in the pull-up network would push up the rising delay
as it would take longer for the output capacitance to charge through it. To counter the slow
transistor, the tuning PMOS connected in parallel is turned ?ON? to provide a separate path
for charging the output capacitance, as shown in Figure 4.3. The red dot in the pull-up
network indicates the presence of a slow PMOS transistor.
In the same way, the presence of a slow transistor in the pull-down network would push
up the falling delays as it would take longer for the output capacitance to discharge through
22
Figure 4.3: Rising transition speed-up
Figure 4.4: Falling transition speed-up
23
it. Their presence is countered by turning ?ON? the NMOS tuning transistor which would
provide an additional path for discharge, as shown in Figure 4.4.
It is assumed here that an algorithm exists to diagnose the slow transistors in the design.
Once diagnosed, the appropriate tuning transistors are turned ?ON? to bring down the gate
delays.
4.3 Sizing of tuning transistors
Sizing of tuning transistors is an important factor determining the e ectiveness of the
tuning strategy. The tuning transistors should be sized large to provide good speed up, but
at the same time be sized small enough to ensure that logic levels are maintained.
For instance assume the pull-up network to be slow. To speed up rise time the tunable
PMOS transistor is turned ?ON?. Now, when the pull down network is activated, the circuit
would behave like a voltage division network (Figure 4.5), with the output voltage being
supply voltage times the fraction of pull down resistance to the total resistance. To ensure
that this output voltage stays within bounds of logic ?0?, the resistance of the tuning PMOS
should be high, or the W:L ratio of the transistor should be small.
A similar voltage division circuit is set up when an outlier transistor exists in the pull-
down network and the NMOS tuning transistor is turned ?ON? while the pull-up network is
activated.
Hence, a trade o between a large ratio and a small ratio is required for the strategy to
work correctly. For the purpose of simulations in this work, a 4:1 resistance ratio between the
tuning transistor and the corresponding complementary network is maintained. This ensures
a LOW voltage of Vdd/5, and a HIGH voltage of 4Vdd/5, which are within acceptable ranges
of logic ?0? and logic ?1?, respectively.
24
Figure 4.5: Voltage division on tuned nodes
4.4 Faulty tuning transistors
It is important to note that the tuning circuitry comprising the tuning transistors would
also be exposed to process variability, as it is fabricated along with the other functional
transistors of the chip. Hence, it could itself be a victim of process variability. A faulty
tuning transistor would be of relevance only if it is required during the tuning process,
otherwise it is of no importance.
There is a di erence though in how a defective tuning transistor is perceived. As men-
tioned earlier, process variability a ects Vth distribution. In other words the Vth of the
transistor could be high, corresponding to a slow, high resistance, low leakage transistor, or
the Vth could be low, corresponding to a fast, low resistance, high leakage transistor.
For a transistor in the pull-up network or the pull-down network, a high Vth would
imply a faulty slow transistor as it takes longer to switch; a low Vth transistor would be
fast to switch and hence not be an outlier candidate. But for a tuning transistor, a high
Vth would correspond to a high resistance transistor, and even though slow, the e ective
25
Figure 4.6: Tunable NAND gate
resistance of the slow outlier and the tuning transistor would be lower and hence some speed
up will be achieved.
On the other hand, a low Vth tuning transistor, even though faster to switch, could
degrade the output voltage logic so much that the output doesn?t switch anymore. Such
an occurrence would make it impossible to tune the outlier transistor to avoid voltage logic
degradation. Hence, for the case of tuning transistors, it is the low Vth transistor that is
considered defective and not the high Vth transistor. This occurrence is again not common,
and in the chance that it is to occur, the fabricated chip would be discarded as a yield loss.
4.5 NAND gate simulations
To test the e ectiveness of the tuning strategy, simulations were performed on a single
NAND gate with input and output loads. Mentor Graphics tools were made use of to
generate the NAND gate SPICE circuit with input and output loads. A gate level structural
netlist was  rst created as a verilog  le. This verilog  le was fed in as an input to Design
26
Architect to create design view points. The design view points created in Design Architect
are used in IC Station to generate the circuit layout. Once the layout was generated, it was
veri ed for over ows and shorts, after which a layout versus schematic check was done using
the LVS tool. Capacitances were  nally extracted through Calibre PEX.
Once the  nal SPICE netlist was generated, tuning transistors connected to the NAND
gate were added. This addition had to be done manually as the standard cell libraries do not
have tunable gates in them. The width over length ratios for the tuning transistors were  xed
so as to have a 4:1 resistance ratio between the tuning transistor and the complementary
network of the gate. Additional capacitance was added to the gate outputs to include the
e ects of the tuning transistors. More is discussed on the SPICE netlist extraction in the
following chapter.
The threshold voltages of the transistors were assumed to be nominal  rst, and the rise
and fall time delays were noted. Another set of simulations were performed on the same
NAND gate, but with one pull-up PMOS transistor having a threshold voltage 0.1V above
nominal; rise and fall time delays were noted before and after tuning. The same simulations
were repeated for Vth of PMOS being 0.2V above nominal, and 0.3V above nominal. Figure
4.6 shows the tunable NAND gate structure, and TABLE 4.1 provides the results of the
above simulations.
180nm Technology
Vth Rising Delay Falling Delay Delay change
(PMOS) Untuned Tuned Untuned Tuned (%)
0.4 113 89 130 142 -25.67
0.5 174 125 129 140 19.54
0.6 305 188 128 139 38.36
0.7 703 296 127 138 57.89
Table 4.1: Tunable NAND gate simulations
The transistor W:L ratios in the NAND gate are the standard ratios de ned by the
cell libraries. The W:L ratios of the tunable gates were  xed manually to ensure a 4:1 ratio
27
between the tunable PMOS and pull-down, and tunable NMOS and pull-up resistances. As
can be seen the performance gain for Vth = 0.4V is negative, as the complementary delays
are pushed up, but su cient gains are obtained as Vth of PMOS increases; for Vth = 0.7V
the reduction in gate delay is close to 58%.
In conclusion, the methodology seems to be very e ective in pulling back gate delays
by a signi cant amount. Its e ectiveness on a large circuit with long chains having multiple
gates needs to be tested though.
The steps taken in generating the SPICE netlist for the purpose of simulations are
discussed in the next chapter.
28
Chapter 5
SPICE Netlist Extraction
Extracting the SPICE netlist for the purpose of simulations is a very critical part of
the work done here. A description of the tools used and the manual additions implemented
are discussed in this chapter. Mentor Graphics tools consisting of ModelSim, Leonardo
Spectrum, Design Architect, IC Station, LVS and PEX, have been used for extracting the
transistor level SPICE netlist.
The circuits studied in this work include basic inverter chains, NAND gate chains and
EXORtree circuits. Any larger circuit constructed, is constructed out of these simple circuits.
The reason for studying these simple circuits is explained later.
The chapter starts with a brief description of the tools in the order in which they were
used, followed by a description of the method used to include the tuning transistors.
5.1 Mentor Graphics Tools
5.1.1 ModelSim
The behavioral model of the circuit is  rst written either in VHDL or Verilog and
then simulated in ModelSim. This VHDL/Verilog description forms the seed from which
ultimately the transistor netilst is extracted.
Process variability, in the context of this work, was studied on basic inverter chains,
NAND gate chains and EXORtree circuits. As the structural description of these circuits is
known, a gate level structural description was written in Verilog. The behavioral description
is not provided to ensure that no optimization is done on the circuit.
29
5.1.2 Leonardo Spectrum
Once the behavioral/structural description of the circuit is created in VHDL/Verilog,
Leonardo Spectrum is used to generate the structural description of the circuit in Verilog
format with the basic details of the gate delays speci ed through the .sdf  le. It is to
be ensured that no optimization is done while synthesizing an input structural netlist to
generate a Verilog gate level structural description; optimization if done could change the
gates used and a few inter-connections without of course changing the output functionality.
5.1.3 Design Architect
The next step in the extraction process is to load the Verilog  le generated by Leonardo
Spectrum into Design Architect to generate a schematic of the circuit. While generating the
schematic it will be seen that standard cells with  xed width and length ratios, for a given
drive, are used to construct the circuit. It is not possible to change the width or length
information of the standard cells. This inability is the reason why tuning transistors are
to be included manually at the end of the extraction process. More on this is discussed
later. Design Architect also generates design view points which are used by IC Station in
extracting the layout information.
5.1.4 IC Station
The next step is to open the schematic in IC Station to generate the layout. While
generating the layout all over ows are to be connected, all shorts removed, and input output
pins added. This would ensure that the design veri cations in IC Station are carried out
correctly.
30
5.1.5 LVS
The LVS tool is used next in performing the layout versus schematic check. This is done
by comparing transistor dimensions and the interconnects of the transistors of the layout,
with the transistor dimensions and interconnects of the transistors of the schematic.
5.1.6 PEX
The last step in the SPICE extraction of the functional transistors is carried out through
PEX. Capacitances are extracted by the tools, that makes use of lookup tables in the tech-
nology  le package. A  nal transistor level SPICE netlist with these capacitances extracted
is generated. It is this SPICE netlist that is used in HSPICE for simulation.
5.2 Tuning transistors
So far, tuning transistors have not been included in the design. There are di culties
associated with including tuning transistors in the design. The most important being that
standard cells do not come with the tuning capability. A number of di erent techniques were
studied to implement the automated addition of tuning transistors.
As can be seen from Figure 5.1 above, tuning transistors together closely resemble the
architecture of an inverter, with the di erence that the gate terminals are connected to
separate voltage sources instead of being connected together to an input voltage source, and
input and output nodes are connected together. With this in mind the following architectures
were implemented.
Figure 5.2 shows the close resemblance of an inverter connected to a gate output with
the actual tuning architecture. It was proposed that every gate be connected to an additional
inverter and the SPICE netlist extraction be done. Once the SPICE netlist was extracted,
one would manually alter the connections between the inverter and the gate to transform
the inverter architecture to a tuning architecture.
31
Figure 5.1: Tuning circuit
Figure 5.2: Inverter as a tuning circuit
32
Figure 5.3: Inverted inverter as a tuning circuit
The problem associated with this architecture is that additional capacitance is added
at the gate output nodes from the gate terminals of the tuning transistors. It is only the
di usion capacitance of the tuning transistors that is to be added. To overcome this problem
the architecture shown in Figure 5.3 was proposed.
Figure 5.3 better resembles the tuning circuitry because of the direct addition of the
di usion capacitance of the transistors in the inverters with the node output capacitance. It
was proposed that the inverter input node and its capacitance be removed once the SPICE
netlist extraction was complete.
But this architecture could also not be implemented. As it was later found, while gener-
ating the schematic, inverter transistor width and length values could not be altered; only the
standard cell values of width and lengths would be interpreted by IC Station in generating
the transistor layout. It is important that the transistor dimensions of the tuning circuitry
be altered to have a 4:1 resistance ratio between it and the corresponding complementary
network in the node, for good speed up to be achieved without disturbing the voltage logic.
As a result it was decided that all tuning transistors be included manually in the SPICE
netlist after the functional netlist was extracted from the various design tools described
33
earlier. The transistor dimensions were also manually  xed to the required ratio. Additional
parasitic capacitance that would have to be included to compensate for the tuning transistors,
is also added manually to every node.
The value of this additional capacitance at every node is determined by extracting the
di usion capacitance of transistors in an inverter. The di usion capacitance here happens to
be the gate output capacitance of the inverter. But this capacitance extracted corresponds to
the standard cell dimensions of the inverter whose L=180nm and W=450nm for the NMOS
and W=990nm for the PMOS. To scale the capacitance, the ratio by which the widths of the
transistors scale, is calculated. Width for the PMOS was scaled to 270nm and for NMOS
was scaled to 180nm. Hence on an average the dimensions were scaled by a factor of 3. The
inverter output capacitance, which happens to be the di usion capacitance, is then divided
by 3, and this value added to every node capacitance.
This strategy of adding tuning transistors was possible in the case of this work as simple
circuits entirely constructed using a single gate, such as NAND gate or a NOT gate, were
studied. For complex circuits with a mix of several gates, the above strategy would fail
as it would be very time consuming to manually  x transistor dimensions of every tuning
transistor connected to di erent types of gates; these dimensions would vary from one gate
type to another. Future studies involves incorporating tunable gates into the standard cell
libraries to make this automation possible.
Once the SPICE netlist extraction was complete, the SPICE  le was simulated using
HSPICE. The next chapter studies the algorithm implemented to study process variability
on very large circuits.
34
Chapter 6
Tuning implemented on very large circuits
The preceding sections indicate the potential for improving performances of circuits
through tuning. However, to truly identify the e ectiveness of such a strategy, it is necessary
to create an environment where the circuit has several million transistors. This chapter
presents the work in creating such an environment through clever innovative techniques,
and then simulating every circuit before and after tuning to calculate worst case delays and
measuring performance improvement.
6.1 Building large circuits
Simulating large circuits is necessary because it is in such large circuits that transis-
tors are exposed to process variability. In a small circuit, because of the small number of
transistors involved, statistical outliers will not be found to occur in most cases. Chips fab-
ricated today have a few hundred outlier transistors in them because of the close to billion
transistors that are there in the chip.
6.1.1 Di culties
There are a number of problems associated with simulating large circuits. First, every
node in the circuit has to have the tuning capability, which means adding a PMOS and an
NMOS transistor manually to every node. This has to be done manually because the library
 les that are used in generating circuit layouts and node capacitances, do not have gates
with tuning capabilities in them. Hence doing the addition manually in a circuit that has
several thousands of gates at least, would be very tedious.
35
Secondly, every circuit has to be simulated for di erent nominal threshold voltages and
standard deviations, and di erent supply voltages, each be repeated at least a 1000 times to
get an average case out.
It is also necessary to simulate a circuit with di erent sizes to get a sense of an approx-
imate size beyond which there would be su cient performance bene t to make this tuning
technique applicable. A single HSPICE simulation for a large standard circuit takes several
hours even on a fast computer. Such simulations are hence not very practical.
6.1.2 Basic algorithm
To get around this problem of being unable to simulate large circuits, an innovative
algorithm is adopted. First a standard small circuit with less than a 100 transistors is
simulated, with Vth for every transistor in the circuit dawn from a normal distribution. The
same circuit is again simulated, but with di erent transistor threshold voltages, again drawn
randomly from a normal distribution. This process of drawing Vth and simulating is done
20,000 times, and for each simulation the delays of the circuit are calculated.
Now, to construct a larger circuit from these smaller sub-circuits, a random ?N? number
of sub-circuits are drawn from the collection of 20,000 circuits, and the largest delays of the
?N? sub-circuits selected determine the delay of the larger circuit. ?N? represents the number
of small sub-circuits selected at a time to create the larger circuit. ?N? is allowed to vary
from 10 to 10,000. The random pick of ?N? sub-circuits is done 1000 times each to get an
average case for every size ?N?. In this way large circuits can be constructed from smaller
ones, and their delays calculated without having to simulate them.
6.1.3 Standard circuit used to create larger circuits
The standard small circuit picked in this work is an EXOR tree with 8 inputs and 1
output. The tree has 7 EXOR gates, each constructed with 4 NAND gates. Every NAND
36
gate has the tuning capability added to it. There are 8 paths to the output, and each path
has 6 active NAND gates on it.
The advantage of using an EXOR tree is that it is very testable, which would make the
diagnosis easier, and a single bit change in the input is re ected at the output. This property
is made use of in activating a single path individually, by sending out pulses through every
input, one input at a time, and every pulse sent with enough spacing to allow the output to
settle even in the presence of a slow outlier, before being pulsed at the next input to activate
the next path.
Figure 6.1A is a  gure of an EXOR gate constructed using NAND gates. The NAND
gates here are all tunable. Figure 6.1B is a  gure of an 8 input EXOR tree. Figure 6.1C
represents a large circuit constructed out of ?N? EXOR trees, the number ?N? varying from
10 to 10,000.
6.2 Simulation
This section details the logic used while programming in MATLAB and PERL to create
the environment described in the preceding section. The EXOR tree circuit was  rst con-
structed and synthesized as a SPICE  le. This  le was labeled ?without tuning circuitry?.
Another EXOR tree SPICE  le labeled ?with tuning circuitry? was created with tuning tran-
sistors connected to every NAND gate output. Two  les were created so that a comparison
could be drawn between the delays of the original circuit and the delays of the tuned circuit.
To start simulations on HSPICE, however, the primitive model (?.pm?)  les that would
be needed to run with the SPICE netlists need to be created. The standard ?.pm?  le that
is made available in the package cannot be used as it details the parameters of a single
transistor. It is, however, used as a seed to construct the ?.pm?  les required for the EXOR
tree simulations. The following subsection details the method used to generate primitive
model  les that incorporate variability.
37
Figure 6.1: EXOR tree circuit
38
6.2.1 Primitive model  les
As mentioned, the ?.pm?  les are included in the SPICE netlists before being simulated.
To ensure that every transistor has Vth drawn from a normal distribution, and is hence
di erent from every other transistor in the circuit, the ?.pm?  le for an EXOR tree circuit
would have instances of the standard single transistor ?.pm?  le for every transistor in the
circuit, with Vth of every instance di erent from the other.
In the 8 input EXOR tree circuit with tuning circuitry included, there are 84 PMOS
transistors and 84 NMOS transistors. Hence, there will be 84 di erent descriptions of PMOS
and 84 di erent descriptions of NMOS in a single ?.pm?  le to be included with the circuit
SPICE netlist; these descriptions only vary in the value of Vth.
It was also mentioned earlier, in the section on the basic algorithm implemented to
create large circuits, that 20,000 copies of the EXOR tree are to be simulated. All of these
20,000 copies will also have transistors that vary in Vth from every other transistor?s Vth in
any of the 20,000 copies. Hence, a MATLAB program to generate 20,000 * 84 ( = 1,680,000)
values of PMOS-Vth and 20,000 * 84 ( = 1,680,000) values of NMOS-Vth was written that
would extract the 1,680,000 Vth values from the standard normal distribution for a speci c
 (mean) and  (standard deviation).
These values are then assigned one by one to a ?.pm?  le. Every time 84 transistor
instances have been written, a new ?.pm?  le is created to have the next 84 set of instances.
In this way 20,000 di erent ?.pm?  les, with Vth for every transistor drawn from a nominal
distribution was created, and stored in a speci c folder, each  le numbered sequentially from
1 to 20,000. The entire process of creating the 20,000 ?.pm?  les was repeated for di erent  
values.
This process of generating ?.pm?  les was also done for di erent values of Vdd. The
reason being that, while drawing Vth from a distribution, there is a possibility of the value
being greater than Vdd or less than zero. This would not be correct and needs to be avoided.
In other words bounds are set for the Vth pick, depending on the value of Vdd.
39
The Vdd values chosen for simulation are 1.2V and 1.5V. For Vdd=1.2V, the lower
bound was  xed at 0.1V and the upper bound at 1.0V. While for Vdd=1.5V, the lower
bound was at 0.1V and the upper bound at 1.3V. Hence for every combination of Vdd and
 , ?.pm?  les were generated. All in all there were 4 sets of  les, each with 20,000 ?.pm?  les
in them, for the cases Vdd=1.2V  =0.075V, Vdd=1.2V  =0.150V, Vdd=1.5V  =0.075V,
and Vdd=1.5V  =0.150V.
It is to be noted that the circuit labeled ?without tuning circuitry? would also use the
same ?.pm?  les in spite of having fewer transistors; this circuit would have 56 PMOS tran-
sistors and 56 NMOS transistors. To ensure that transistors belonging to a gate in the two
circuits labeled ?without tuning circuitry? and ?with tuning circuitry? are the same, the tun-
ing transistors are only labeled from 57 to 84. This is important because if an outlier is to
exist, it is to fall at the same location in the two circuits. Only then can a comparison be
made between the delays of the two circuits.
6.2.2 Single EXOR tree simulations
The SPICE netlist  les are now ready for simulation. It is to be noted that even though
there are 20,000 di erent ?.pm?  les for a speci c  , there is only one copy of the SPICE
netlist  le; one for ?without tuning circuitry? and one for ?with tuning circuitry?. Only the
?.pm?  le that the SPICE  le points to is modi ed every time.
A MATLAB program was written to open a SPICE netlist  le, modify the ?.pm?  le
to be pointed to, run the SPICE  le on HSPICE, and extract the path delay information
to be stored in a  le; HSPICE was invoked from the MATLAB program. The path delay
information was stored in a  le labeled with the same number the ?.pm?  le was labeled with.
This  le had information on the delays of every path and also the worst case delay, found by
calculating the largest delay in the simulation, along with the worst case path number and
transition type. The worst case delays were also stored in a matrix in MATLAB to speed
up its use by another MATLAB program.
40
The SPICE  le had commands in it to extract every rising and falling transition delay
information. It also has the capability to extract the power dissipated, but this command is
made use of only later while simulating for power.
In this way 20,000 di erent delay  les for the 20,000 copies were created and stored in
a speci c folder, along with the worst delays also being stored in a matrix in MATLAB (the
matrix size was 20000 x 1). This process of simulating was repeated for di erent Vdd and  
values.
All in all there were 4 cases :
CASE 1 Vdd = 1.2V  = 0.075V
CASE 2 Vdd = 1.2V  = 0.150V
CASE 3 Vdd = 1.5V  = 0.075V
CASE 4 Vdd = 1.5V  = 0.150V
For all of the above cases, there were 2 basic simulations done on each of the 20,000
copies - one on ?without tuning circuitry? and the other on ?with tuning circuitry?. In the
process two di erent worst case delay matrices were also constructed. It is to be noted that
no tuning has been performed yet.
6.2.3 Tuning
Once all simulations were done for both ?without tuning circuitry? and ?with tuning circuitry?,
and the worst case delays calculated and stored in a matrix, tuning was implemented. Only
those circuits which have outlier transistors in them need to be tuned.
To pick outliers, a delay cut-o was established, and all worst case delays larger than
the cut-o were declared outliers. This cut-o was selected so as to have between a 100 and
150 outlier cases out of the 20,000 circuits simulated for every case.
Once the cut-o was established, a search is performed through a MATLAB program
to extract those circuit numbers that have their worst case delays greater than the cut-o .
41
The delay  le for an outlier circuit is then opened and the value of the worst case delay,
along with the path number of the largest delay and the transition type is extracted.
With this information in hand, the corresponding ?.pm?  le is opened, and a search is
performed on the largest Vth values in the  le. This search was done through a PERL script
that would extract every line one-by-one from the ?.pm?  le and search for the patern ?Vth0?
in the  le. The PERL script is invoked from MATLAB, and hence forms a part of the same
MATLAB code.
This search of Vth is restricted only to the transistors that fall on the critical path, and
only those that are involved in the slowest transition. This is done to speed up the search.
Details on which transistor lies on which path, for a particular transition, was predetermined
and stored in the MATLAB code.
Since a comparison needs to be done on a Vth belonging to a PMOS and a Vth belonging
to an NMOS, the average Vth of the NMOS transistors in a particular gate is compared with
the Vth of a PMOS transistor. This is done because, in NAND gates, the NMOS transistors
are connected in series, while the PMOS transistors are connected in parallel. Hence for a
falling transition to occur, the discharge would happen through the series connection of the
NMOS transistors; slowest charging would occur only through a single PMOS transistor.
Hence, Vth values of a PMOS transistor and average Vth values of the two NMOS
transistors in a particular gate were calculated and extracted by the MATLAB program, for
every gate in the critical path. Gates on this critical path are then sorted in descending
order of Vth.
To tune the circuit, the tuning transistor connected to the gate with the highest Vth
value, is to be turned ?ON?. If the the outlier transistor happens to be a PMOS transistor,
the PMOS tuning transistor of that gate is turned ?ON?. On the other hand, if the outlier
transistor happens to be an NMOS transistor, the NMOS tuning transistor of that gate is
turned ?ON?.
42
The tuning transistors in the SPICE  le are labeled with the gate number the tuning
transistor belongs to. This makes the search for the tuning transistors connected to the gate
under study easier.
Once the tuning transistor number is established, the voltage source connected to the
gate terminal of the tuning transistor is determined. Here again, the voltage source is labeled
conveniently to make its search possible. For an NMOS tuning transistor, the voltage source
is initially connected to the GND (logic ?0?). To enable tuning, the value of the voltage is
switched to Vdd (logic ?1?). In the case of a PMOS tuning transistor, the voltage source
which was initially connected to Vdd (logic ?1?), is switched to GND (logic ?0?) to enable
tuning.
Once the tuning voltage source values are modi ed in the SPICE  le, the MATLAB
program invokes HSPICE to simulate the SPICE  le. The delays again are calculated, and
the worst case delay is noted. If the worst case delay falls below the cut-o , the circuit is said
to be tuned, and the next outlier circuit is picked for tuning. If not, the second gate in the
sorted list of gates on the critical path is tuned. Again, HSPICE is invoked to simulate the
circuit, and path delays calculated and extracted by the program. If worst case delay stays
above the cut-o again, the next gate in the sorted list is tuned. This process of picking
a gate for tuning is carried on until tuning is achieved. If all single gate tuning cases are
exhausted, the program would check for multiple gate tuning to see if tuning multiple gates
on the critical path brings the worst case delays down.
In most cases, speedup was achieved by tuning the  rst gate in the sorted list of gates
on the critical path. For the case Vdd = 1.2V  = 0.150V, out of the 120 cases tuned, 112
cases were tuned by the tuning the  rst gate in the sorted list. For 7 circuits, tuning was
achieved by tuning 2 gates in the critical path. 1 circuit could not be tuned; this is because
of the occurrence of the special case. Table 6.1 gives the number of gates tuned for each of
the cases studied.
43
CASE Number of
circuits tuned
Vdd = 1.2V  = 0:075V 136
Vdd = 1.2V  = 0:150V 120
Vdd = 1.5V  = 0:075V 144
Vdd = 1.5V  = 0:150V 113
Table 6.1: Number of circuits tuned
Once tuning is done, a  nal third matrix with tuned delay values is constructed and
stored in MATLAB. Together the three matrices are used in measuring performance bene t.
The motive behind storing values in a matrix is to provide faster accessibility of these data
to other MATLAB programs.
It is to be noted that once an outlier path in a circuit is tuned, there is no possibility of
another path becoming an outlier as a result of tuning. It is true that the complementary
transition is slowed down, but this slow down would never be so large to become an outlier.
Only outlier cases, with path delays signi cantly larger than average case delays, are being
tuned.
6.2.4 EXOR tree simulation: example
Table 6.2 represents one EXOR tree simulation for the case Vdd = 1.2 and  =0.150V.
The table shows Rising Input 4 to have the largest delay (7.93ns) before tuning. The in-
crease in the delays after adding tuning transistors is attributed to the parasitic capacitances
introduced by the additional transistors. This critical path delay is brought down consider-
ably on tuning to 1.786ns - a 77.5% reduction in the path delay, which is very signi cant.
The example demonstrated here is that of an extreme outlier. On an average the bene ts
on single EXOR tree outlier circuits obtained were less. Table 6.3 gives the average single
EXOR tree circuit outlier bene ts.
It is to be noted that the entire process of  nding outlier circuits, extracting details of
the slowest path delays, sorting gates along a path in descending order of Vth, modifying
44
Input Transition Without any tuning circuitry Untuned delays Tuned delays
(ns) (ns) (ns)
INPUT 1 RISING 1.301 1.373 1.377
FALLING 1.222 1.295 1.295
INPUT 2 RISING 1.030 1.092 1.089
FALLING 1.295 1.371 1.371
INPUT 3 RISING 1.499 1.161 1.079
FALLING 0.9493 1.005 1.030
INPUT 4 RISING 7.930 8.323 1.716
FALLING 0.9136 0.9644 0.9996
INPUT 5 RISING 0.8768 0.9277 0.9302
FALLING 0.8978 0.9489 0.9482
INPUT 6 RISING 0.8381 0.878 0.8881
FALLING 1.689 1.788 1.786
INPUT 7 RISING 0.9795 1.034 1.036
FALLING 1.148 1.216 1.215
INPUT 8 RISING 1.010 1.068 1.067
FALLING 1.310 1.384 1.382
7.930 8.323 1.786
Table 6.2: Tunable EXOR tree simulation
CASE PATH DELAY IMPROVEMENT
Vdd=1.2V  = 0:150V 70%
Vdd=1.2V  = 0:075V 15%
Vdd=1.5V  = 0:150V 27%
Vdd=1.5V  = 0:075V 10%
Table 6.3: Performance gain for a Single EXOR tree
the value of the voltage sources connected to tuning transistors, and invoking HSPICE to
simulate the SPICE netlist, repeated over and over again until speedup is achieved, was done
entirely by a single MATLAB program.
Another MATLAB program was written to pick a subset of ?N? EXOR trees to create
a larger circuit. The ?N? EXOR trees constitute the sub-circuits put together to form the
larger circuit. More is discussed on this in the following subsection.
45
6.2.5 Picking ?N? sub-circuits to create a larger circuits
Once 20,000 simulations were done for every case, on the two circuits labeled ?with-
out tuning circuitry? and ?with tuning circuitry?, and then tuning done on the outlier circuits
to bring about speedup, larger circuits were constructed by randomly picking an ?N? number
of sub-circuits, and then measuring the largest worst case delay of the ?N? worst case delays.
This largest worst case delay would become the worst case delay of the larger circuit.
Picking up ?N? circuits here refers to drawing ?N? worst case delays from the three worst
case delay matrices that were constructed earlier for every case. This is done by generating
?N? random numbers in MATLAB, and then extracting the matrix values of that index. Once
the three values are extracted, bene t obtained on tuning is calculated.
This process of picking a speci c ?N? sub-circuits is repeated a 1000 times to get an
average case; doing only a few picks may include only special cases and not cover the average
case.
The above process of picking ?N? sub-circuits 1000 times was repeated for di erent values
of ?N? ranging from 10 to 10,000.
The next chapter presents the results of simulation done on large circuits, with plots
and explanations on the observed trends of curves in the plots.
46
Chapter 7
Simulation Results
This chapter presents the results of circuit tuning on large circuits of varying sizes. For
every size ?N? the pick and delay evaluation was done 1000 times and the average delay for
the 1000 cases was calculated.
7.1 Plots
The following plots, Figure7.1, Figure7.2, Figure7.3, Figure7.4, represent variations of
path delay versus circuit size for the cases Vdd=1.5V sigma=0.075, Vdd=1.2V sigma=0.075,
Vdd=1.5V sigma=0.150, and Vdd=1.2V sigma=0.150 respectively. These path delays are
the average path delays of the randomly generated 1000 circuits.
Curves in red correspond to delay versus circuit size variations for circuits without any
tuning capability, while curves in blue represent the same for circuits with tuning capability
in them but without any tuning done, and curves in green represent the same variations for
tuned circuits.
7.2 Observation
It is observed that the blue curves are always above the red curves. This behavior is
attributed to the addition of parasitic capacitance to the load capacitance of every gate as
a result of adding tuning transistors. With increased capacitance it takes longer for every
load capacitance to charge and discharge.
As circuit size gets larger, the chance of the presence of a larger extreme outlier in the
circuit gets higher. This behavior explains the exponential increase in the delays of circuits
with increasing circuit size, indicated here by the exponential rise of the red and blue curves.
47
Figure 7.1: Simulation case : Vdd=1.5V sigma=0.075
Figure 7.2: Simulation case : Vdd=1.2V sigma=0.075
48
Figure 7.3: Simulation case : Vdd=1.5V sigma=0.150
Figure 7.4: Simulation case : Vdd=1.2V sigma=0.150
49
The aim of tuning is to get the path delays to fall well below the outlier path delay
value. As can be seen little or no bene t is obtained in small circuits, as the chance that an
outlier will exist is less. This explains the initial behavior of the green curve when it tends
to lie above the red curve. But beyond a certain circuit size we observe good returns, as the
green curve seems to saturate while the red and the blue curves keep increasing.
Delays of circuits for a  =0.150V distribution are greater than a  =0.075V distribution.
This is attributed to the fact that larger  distributions have larger spreads, and as the spread
gets wider the chances of picking a slow outlier is greater.
It is also observed that while the green curve appears to be saturating, a slight increase
towards the end of the plot may be observed. This is attributed to the fact that a few EXOR
trees out of the 20,000 random copies could not be tuned. The reason being the presence of
a high resistance slow outlier in a network and a low resistance fast transistor in the tuning
circuitry of the same network; an interesting case that was studied earlier. If tuning is done
on such a node the voltage logic of the node would be disturbed, and the gate would very
often not switch at all. However this special case is very rare and only 1 out of 20,000 cases
were found to exhibit this behavior in the Vdd=1V and  =0.150V case.
7.3 Bene ts obtained
Table 7.1 shows the average case path delay improvements obtained for a circuit size of
10,000 sub-circuits under the di erent simulation environments.
CASE PATH DELAY REDUCTION
Vdd = 1.2V  = 0:150V 59.45%
Vdd = 1.2V  = 0:075V 15.82%
Vdd = 1.5V  = 0:150V 39.52%
Vdd = 1.5V  = 0:075V 6.80%
Table 7.1: Performance gain for a circuit with 10,000 EXOR trees
50
7.4 Conclusion
Clearly the potential for performance improvement exists. Even though simulations were
done in 180nm technology node, the distributions used for the purpose of simulation mimic
current and future technology spreads. Hence similar or better bene ts can be expected with
smaller technology nodes.
The e ects of the implemented design on power dissipation are studied in the following
chapter.
51
Chapter 8
Power Dissipated with tuning
The tuning technique implemented provides performance bene t measured in terms of
gain in clock frequency. Its impact on the extra power dissipated is studied in this chapter.
8.1 Static power dissipation
In a circuit, once an outlier is identi ed, the required tuning transistor of that node is
turned ?ON? and remains ?ON? throughout. Hence when the complementary network in the
same node is activated to an ?ON? state, there exists a direct path for current to  ow from
Vdd to GND; in other words, static power dissipation exists.
To measure the amount of power dissipated, simulations were done on a circuit size
of 10,000 sub-circuits for the case Vdd=1.2V and  =0.150V, as this case mimics the case
impacted by process variability the most.
8.2 Simulation di culty
While simulating for power it is important to simulate the circuit at a certain clock
frequency, as dynamic power is a function of frequency. Most chips fabricated are run at a
clock period that provides a 10% bu er to the critical delay in the chip, thus operating at a
period 1.1 times the worst case delay.
With respect to the work done here, a circuit of size 10,000 would have to be selected
randomly out of the 20,000 sub-circuits, and then simulated again at a clock period that is
1.1 times the worst case delay of that circuit to estimate power. This process of selecting
10,000 random sub-circuits out of 20,000,  nding the worst case delay of the pick, then
simulating it with a clock period that is 1.1 times the worst case delay, and repeating the
52
process a 1000 times to get an average case out, would e ectively require another 10 million
(10000*1000) simulations of the EXOR tree sub-circuits, which is very tedious.
8.3 Algorithm
To get around the above problem, another innovative algorithm is adopted. First 1000
circuits of size 10,000 sub-circuits are drawn randomly from the 20,000 sub-circuits, and
their worst case delays noted.
These 1000 circuits are then speed binned depending on the value of the worst case
delay of the pick. For circuits without any tuning capability, 5 bins were created - those
with worst case delays less than 9ns, those with worst case delays between 9ns and 10ns,
those with worst case delays between 10ns and 11ns, those with worst case delays between
11ns and 12ns, and those with worst case delays between 12ns and 13ns. Every circuit in
a bin can be operated with a clock period which is 1.1 times the largerst bin delay. Hence,
the bin labeled less than 9ns would have circuits simulated with a clock period of 9.9ns.
Similarly circuits in a bin labeled between 11ns and 12ns would be simulated with a clock
period of 13.2ns. In the same way bins were created for circuits with tuning capability added
but no tuning done.
Every one of the 20,000 circuits is now simulated again for the di erent bin clock fre-
quencies. In other words, for circuits without any tuning capability added, the 20,000 EXOR
trees would be simulated at clock periods 9.9ns, 11ns, 12.1ns, 13.2ns, and 14.3ns. Average
power dissipated is evaluated for every simulation.
To create a circuit with 10,000 sub-circuits and evaluating the power dissipated, 10,000
random sub-circuits are drawn from the 20,000 sub-circuits. The worst case delay of this
circuit would determine the bin that the circuit would fall in. Once the bin is identi ed, total
power dissipated for every sub-circuit is obtained from the pre-simulated power dissipated
values for that clock period, and then added up together to give the total power dissipated
in the circuit.
53
Tuned circuits were also simulated for di erent cases. In the  rst case only 20 out of
20,000 circuits were tuned, and in another case 85 out of 20,000 circuits were tuned. These
circuits were operated both at a clock period which was equal to the clock period at which
they were operating at before any tuning was done, and at 1.1 times the worst case delay of
the circuit after tuning was done.
In this way the number simulations has been cut down from 10 million to less than a
hundred thousand.
8.4 Simulation
To carry out the above algorithm, the SPICE  les generated to measure delays, are
again simulated with references to the ?.pm?  les, this time, however, to also include power
measurement. The same set of ?.pm?  les used to measure delays are again used here. Power
measurement is done by including a SPICE command that measures the instantaneous power
dissipated in a device over a time period. The total instantaneous power is then averaged
over the clock period to get the average power dissipated. This power dissipation result
includes both the dynamic and the leakage power.
8.4.1 Bins
As mentioned earlier, bins were created for simulating for power to reduce the number
of simulations required. Table 8.1 shows the bins created for the two circuits. As can be
seen bins for ?without tuning circuitry? start and end with a smaller clock period than bins
for ?with tuning circuitry?. This is due to the fact that circuits with tuning circuitry in them
would have larger gate delays, because of the presence of parasitic capacitances that are
included through the addition of tuning transistors.
An important point to note is that no clock is actually applied to the circuit input. It is
the width of the input pulse that represents the clock period. This is due to the fact that a
54
without tuning circuitry with tuning circuitry
8-9ns 9-10ns
9-10ns 10-11ns
10-11ns 11-12ns
11-12ns 12-13ns
12-13ns 13-14ns
Table 8.1: Bins
clock if applied must have a period which is greater than the worst case pulse width required
by the circuit to function correctly.
8.4.2 SPICE
Once the bin periods are established, a MATLAB program is written to simulate every
one of the 20,000 sub-circuits at 1.1 times the bin frequency. The values of the power
dissipated in every circuit are extracted and stored in a matrix.
The next set of simulations are done on tuned circuits. The challenge behind tuning is
that, an outlier circuit  rst needs to be tuned to bring about speedup, and then a search
needs to be done to decide which bin the tuned circuit would fall in. Once the bin is chosen,
the tuned circuit is again simulated, but this time at a clock period which is 1.1 times the
bin period.
In the beginning, simulation is done for delay measurement. Hence, the period at which
simulation is done, is kept large enough to accommodate the outlier delays. When simulating
 nally for power, the values of the pulse widths in the SPICE netlist had to be modi ed
to the bin period, before being simulated again to measure power. This modi cation of
pulse widths was done though a PERL script, invoked from the MATLAB program of power
simulations.
While simulating with a circuit size of 10,000, it was found that all circuits once tuned
had a worst delay in the range of 6-7ns. Hence, for tuned circuits, only 1 bin, with a period
of 7.7ns, was created.
55
Once tuned circuits are simulated, the power dissipation values are stored in a third
matrix. These matrices, created for power, are then made use of, along with the three delay
matrices created earlier, in generating larger circuits.
Power dissipated with tuning, but no modi cation done to the pulse widths, was also
calculated and stored for the purpose of comparisons, which will be studied in the proceeding
sections.
Tuning was also done for two cases di ering in the number of circuits tuned. In one case
only 20 out of the 20,000 circuits were tuned, and in another 85 out of the 20,000 circuits
were tuned.
8.4.3 Picking 10,000 sub-circuits to create a larger circuit
To generate a larger circuit of size 10,000 sub-circuits, a random set of 10,000 numbers
was generated by a MATLAB program. With this as an index, the delay of the circuits were
found from the three delay matrices, and the bins were then chosen, to decide which bin
to pick the power dissipation values from. Once the power dissipation values for the 10,000
circuits were picked, they were added to give the total power dissipated in the large circuit
with 10,000 EXOR tree circuits. This process of picking a random set of 10,000 numbers
and calculating power is repeated 1000 times to get an average case.
8.5 Simulation results
CASE Without any tuning circuitry Untuned Tuned Change
20 gates Tuned 0.0656 W 0.0652 W 0.0654 W 0.24%
85 gates Tuned 0.0657 W 0.0654 W 0.0657 W 0.41%
20 gates Tuned and Speeded 0.0655 W 0.0653 W 0.0902 W 38.2%
85 gates Tuned and Speeded 0.0656 W 0.0652 W 0.107 W 64.1%
Table 8.2: Power dissipated with tuning
56
Table 8.2 shows that the percentage increase in the power dissipated with tuning alone
and no speedup done is less than 1%. In the cases above more gates are being tuned than
prescribed earlier - 1 in 3.5 million that lie beyond 5 sigma should be tuned. But in our case
we are forced to tune a little more because of the smaller circuit size involved. In spite of
tuning more, the excess power dissipation observed is still only less than 1%, which would
be a lot lower in the 1 in 3.5 million case. On tuning 85 gates instead of 20, the increase in
power was also very substantial.
It is also observed that the power dissipation increases to 38% and 64% on operating
the tuned circuits at their maximum clock frequencies. This is expected because the circuits
are now being operated at approximately 38% and 64% higher clock frequency.
Also, the power dissipated in circuits without any tuning capability included is greater
than in circuits with tuning capability included, in spite of the larger leakage power pro-
duced, because these circuits are operated at a higher clock frequency. Circuits with tuning
capability included but no tuning done would have larger circuit delays than circuits with-
out any tuning circuitry added, because of the additional capacitance added by the tuning
transistors. Hence operating them at a lower clock frequency would generate less power.
The above results are in tune with the expected results. On an average, keeping the
frequency and Vdd supply voltage the same, switching from NMOS logic to CMOS logic
reduced the power dissipation by approximately 2 orders of magnitude. Hence, tuning gates
should also, on average, consume power that would go up by 2 orders of magnitude, because
e ectively the logic changes from CMOS to a pseudo NMOS type. Hence, assuming an
increase in per gate consumption of power by 2 orders of magnitude, the increase in power
consumed in a circuit with 280,000 gates, and having 25 gates that are tuned would cause
an increase in the power consumed by 0.89%, which is very close to the obtained results.
The next chapter presents the  nal conclusions and observations on the work done.
57
Chapter 9
Conclusion
In this work a new design methodology that targets slow down brought about by process
variability, and allows for post fabrication speedup is studied.
Random process variability is bound to occur and a ect every chip fabricated as scaling
goes beyond 22nm. Statistically every fabricated chip will have dozens of outlier transistors
that would limit the maximum clock frequency at which the chip can be operated. Often
considered as a defect, there exists a need to address this problem so as to reap the bene ts
of scaling.
The design methodology incorporates the idea of adding redundant transistors to the
circuit which can be turned ?ON? in case an outlier is detected. Tuning is limited to outlier
gates, which in a circuit size of a 100 million would equal a few hundred gates. This tuning
would be bene cial in pulling back worst case path delays close to the average case delay, at
the expense of a very small increase in the power dissipated. Simulation results indicate that
worst case path delays were brought down by 50%, while power dissipation only increased
by a meager 0.44%.
While reviewing the proposed tuning architecture, there were a few concerns about the
area overhead, diagnosis di culty, and implementing tuning itself.
It is true that there is an area overhead, but this area overhead is compensated by the
savings obtained on Silicon itself, through the fault tolerance capability of this architecture.
The implemented architecture is fault tolerant as turning ?ON? a tuning transistor could
provide a path for charging or discharging the gate output capacitance of an otherwise
faulty node. Hence, even though additional area is used up by the tuning circuitry, tuning
when implemented could increase the yield, and buy back a portion of the used up Silicon.
58
Diagnosis of outliers is di cult indeed, but a scheme that enables it is being developed
and studied. The following is a brief description of the work being done. The scheme  rst
involves running path delay tests on the circuit to detect the slow paths. Vectors pairs
that would detect slow paths are determined by running ATPG. When tested on a chip, an
incorrect output logic measured within a particular time frame for the vector pair applied
would implicate a set of gates that could be outlier candidates. Di erent sets of outlier
candidates are determined for di erent vector pair inputs and the corresponding outputs
where the fault was detected. Starting with this set of implicated gates, an algorithm is
implemented to  lter out the correct outliers from the set. This diagnosis scheme is being
worked on, and when ready would make it possible to exercise the tuning strategy.
Tuning implementation is perceived to be implemented by overlaying a programmable
memory like mesh on the chip, with the gate terminals of the tuning transistor connected to
the memory array. The memory array is programmable to imply that a tuning transistor can
be turned ?ON? by programming the memory appropriately. To make this implementation
a ordable, the memory mesh could be fabricated in thin- lm technology, where amorphous
Si is used in the fabrication process. Such a technology would have a very low drive, but for
the sake of tuning, would su ce. Hence, it is valid to assume that such a tuning architecture
has scopes of applicability in future technology nodes.
The work described in the thesis is only the beginning. The cases studied were naive in
the sense that a simple circuit with the same gates and equal path delays is assumed, and
adding tuning transistors and the additional loading capacitance had to be done manually
and with small approximations. Also, simulations were done in 180nm technology, which
might seem to indicate irrelevance of the obtained results, but the distributions worked with
are pertinent to the advanced technologies, and hence close to applicable.
Future work involves using 45nm technology for the purpose of simulations, adding
tunable gates to libraries so that its extraction can be automated, and working with standard
circuits to make the case more authentic and applicable. Also a de nite diagnosis strategy
59
that would enable the detection of outlier gates in a chip, and a tuning methodology that
would enable implementing the tuning scheme, needs to be studied.
60
Bibliography
[1] http://www.itrs.net/links/2009ITRS/Home2009.html, The International Techonol-
ogy Roadmap for Semiconductors, 2009.
[2] S. Nassif, K. Bowman, \Design for Variability in DSM Technologies," ISQED 2000.
[3] S. Borkar, \Design Challenges of Technology Scaling," IEEE Micro, Vol. 19, Issue 4,
pp. 23-29, August 1999.
[4] M. Ashouei, M. Nisar, A. Chatterjee, A. D. Singh, A. Diril, \Probabilistic Self-
Adaptation of Nanoscale CMOS Circuits: Yield Maximization under Increased Intra-Die
Variations," International Conference on VLSI Design, Bangalore, India, pp. 711-716,
January 2007.
[5] M. Ashouei, A. D. Singh, A. Chatterjee, \A Defect-Tolerant Architecture for End-of-
Roadmap CMOS," European Test Symposium, Freiburg, Germany, May 2007.
[6] A. D. Singh, \Scan Based Testing of Dual/Multi Core Processors for Small Delay De-
fects", in Proc. International Test Conference, 2008.
[7] A. D. Singh, \A Self-Timed Structural Test Methodology for Timing Anomalies due to
Defects and Process Variations", in Proc. International Test Conference, 2005.
[8] K. S. Saha, \Modeling Process Variability in Scaled CMOS Technology," IEEE Design
and Test of Computers, Vol. 27, Issue 2, pp. 8-16, March/April 2010.
[9] B. Cheng, D. Dideban, N. Moezi, C. Millar, G. Roy, X. Wang, S. Roy, A. Asenov,
\Statistical-Variability Compact-Modeling Strategies for BSIM4 and PSP," IEEE De-
sign and Test of Computers, Vol. 27, Issue 2, pp. 26-35, March/ April 2010.
[10] V. Wang, K. Agarwal, S. R. Nassif, K. J. Nowka, D. Markovic , \A Design Model for
Random Process Variability" in proc. ISQED pp. 734-737, 2008.
[11] C. Kenyon, A. Kornfeld, K. Kuhn, M. Liu, A. Maheshwari, W. Shih, S. Sivakumar, G.
Taylor, P. VanDerVoorn, K. Zawadzki, \Managing Process Variation in Intel?s 45nm
CMOS Technology." Intel Technology Journal , Vol. 12, Issue 2, pp. 92-110, June 2008.
[12] W. Shockley, \Problems related to P-N Junctions in Silicon." Solid-State Electronics,
Vol. 2, pp. 35-67, January 1961.
61
[13] W. Schemmert, G. Zimmer, \Threshold-Voltage Sensitivity of Ion-Implanted MOS
Transistors due to Process Variations." Electronics letters, Vol. 10, Issue 9, pp. 151-
152, May 1974.
[14] K. Yokoyama, A. Yoshii, S. Horiguchi, \Threshold-sensitivity Minimization of Short-
Channel MOSFET?s by Computer Simulation." IEEE Journal of Solid-State Circuits,
Vol. 15, Issue 4, pp. 574-579, August 1980.
[15] P. A. Stolk, F. P. Widdershoven, D. B. M. Klaassen, \Modeling Statistical Dopant
Fluctuations in MOS Transistors," IEEE Transactions on Electron Devices, Vol. 45,
Issue 9, pp. 1960-1971, September 1998.
[16] K. Kuhn, J. Kelin , \Reducing Variation in Advanced Logic Technologies: Approaches
to Process and Design for Manufacturability of Nanoscale CMOS." IEEE International
Electron Devices Meeting, IEDM Technical Digest, pp. 471-474, December 2007.
[17] C. H. Diaz, H. J. Tao, Y. C. Ku, A. Yen, K. Young, \An Experimentally Validated
Analytical Model for Gate Line Edge Roughness (LER) E ects on Technology Scaling."
IEEE Electron Device Letters, Volume 22, Issue 6, pp. 287-289, June 2001.
[18] H. W. Kim, J. Y. Lee, J. Shin, S. G. Woo, H. K. Cho, J. T. Moon, \Experimental Inves-
tigation of the E ect of LWR on Sub-100-nm Device Performance." IEEE Transactions
on Electron Devices, Vol. 51, Issue 12, pp. 1984-1988, December 2004.
[19] H. Fukutome, Y. Momiyama, T. Kubo, Y. Tagawa, T. Aoyama, H. Arimoto, \Direct
Evaluation of Gate Line Edge Roughness Impact on Extension Pro les." IEEE Trans-
actions on Electron Devices, Vol. 53, Issue 11, pp. 2755-2763, November 2006.
[20] A. Asenov, A. R. Brown, J. H. Davies, S. Kaya, G. Slavcheva, \Intrinsic Parameter
Fluctuations in Decananometer MOSFETs introduced by Gate Line Edge Roughness."
IEEE Transactions on Electron Devices, Vol. 50, Issue 5, pp. 1254-1260, May 2003.
[21] 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, Issue 4/5,
pp. 433-449, July 2006.
[22] A. Asenov, S. Kaya, J. H. Davies, \Intrinsic Vth Fluctuations in Decananometer MOS-
FETs Due to Local TOX Variations," IEEE Transactions on Electron Devices, Vol. 49,
Issue 1, pp. 112-119, January 2002.
[23] S. Changhwan, S. Xin, K. L. Tsu-Jae, \Study of Random Dopant Fluctuation (RDF)
E ects for the Trigate Bulk MOSFET," IEEE Transactions on Electron Devices, Vol.
56, Issue 7, July 2009.
[24] A. T. Putra, A. Nishida, S. Kamohara, T. Hiramoto, \Random Threshold Voltage Vari-
ability Induced by Gate-Edge Fluctuations in Nanoscale Metal-Oxide-Semiconductor
Field-E ect-Transistors," Applied Physics Express 2, Vol. 024501, January 2009.
62
[25] http://en.wikipedia.org/wiki/Normal_distribution, Normal distribution.
[26] http://ptm.asu.edu/modelcard/HP/32nm_HP.pm, 32nm High Performance PTM
models.
[27] A. D. Singh, K. Mishra, A. Faraz, A. Chatterjee, \Path Delay Tuning for Performance
Gain in the face of Random Manufacturing Variations," in proc. International Confer-
ence on VLSI Design, Bangalore, India, January 2011.
63