IEEE TRANSACTIONS ON VERY LARGE SCALE INTEGRATION (VLSI) SYSTEMS, VOL. 20, NO. 1, JANUARY 2012 157
Transactions Briefs
Impact of Random Dopant Fluctuations on the Timing
Characteristics of Flip-Flops
Faiz-ul-Hassan, Wim Vanderbauwhede, and
Fernando Rodríguez-Salazar
Abstract—In this work, we have analyzed the effects of variability, due to
random dopant fluctuation (RDF), on the timing characteristics of flip-flops
for the future technology generations of 25, 18, and 13 nm, based on exten-
sive Monte Carlo simulations. The results show that RDF has a significant
impact on all of the timing parameters and that these parameters do not
follow a normal distribution; in particular, they are skewed and exhibit a
large tail. Moreover, the dispersion and skewness of the timing parameters
increase with technology scaling. The study of the exact shape of these dis-
tributions, especially in the tail section, is of fundamental importance in
the design and modeling of high-performance, reliable, and economically
feasible circuits. In this paper, the distribution tails are estimated based on
simulation data, with the aid of statistical nonparametric probability den-
sity functions, and it has been found that timing distributions can better be
represented by certain nonparametric distributions, in particular Pearson
and Johnson systems. The use of these representations during the statistical
static timing analysis will provide more accurate results as compared with
the normal approximation of distributions and will eventually reduce the
probability of yield loss.
Index Terms—Device variability, flip-flops (FFs), nonparametric estima-
tion, parameter fluctuations.
I. INTRODUCTION
Technology scaling facilitates the integration of large systems-
on-chip (SoCs) with complex functionalities. In the pursuit of de-
signing high-performance systems with large throughputs, the circuit
speed is continuously being increased, shrinking the clock period
and leaving fewer design margins. Such circuits normally incorporate
deeply pipelined architectures using flip-flops (FFs), in order to meet
throughput requirements. In sequential circuits, FFs are used to store
the evolving system state and the speed of computation is limited by
the speed at which this state can be made to change. In clock-based
digital circuits, FFs are considered to be essential to meet strict timing
constraints. Generally speaking, a modern digital integrated circuit
without FFs is simply unthinkable.
In high-performance circuits, the clock period is optimized to ob-
tain the best possible performance while meeting stringent timing con-
straints. The circuit performance can meet the design targets if there is
a strict control over the delay of the signals arriving at the clocked el-
ements (FFs). However, the signals are not completely isolated on the
chip and are subjected to various sources of delay variability, such as
process and environmental parameters and on-chip interconnect noise.
Besides these parameters, the performance of the circuits depends on
the inherent timing characteristics of the combinational logic and FFs.
Manuscript received February 05, 2010; revised June 06, 2010 and September
14, 2010; accepted September 28, 2010. Date of publication November 22,
2010; date of current version December 14, 2011.
F. Hassan and F. Rodríguez-Salazar are with the School of Engineering, Uni-
versity of Glasgow, G12 8LT Glasgow, U.K. (e-mail: F.Hassan@elec.gla.ac.uk;
F.Rodriguez@elec.gla.ac.uk).
W. Vanderbauwhede is with the School of Computing Science, University of
Glasgow, G12 8QQ Glasgow, U.K. (e-mail: wim@dcs.gla.ac.uk).
Digital Object Identifier 10.1109/TVLSI.2010.2088409
Due to technology scaling, the performance of the circuits is becoming
more sensitive to the variability in the devices so variability has be-
come one of the major problems in deep-submicrometer (DSM) tech-
nologies [1]–[3]. Statistical variability has become the most dominant
form of variability in 45-nm technology generation [4] and is expected
to further prevail in the future technology generations [5]. The design
margins, which are already reduced due to the increased clock fre-
quency, are greatly affected due to the variability which, in turn, limits
the achievable performance and yield.
In nano CMOS devices, random dopant fluctuation (RDF) is the
most dominant source of statistical variability in bulk MOSFETs. The
atomic-level intrinsic fluctuations in the number and location of dopant
atoms in the channel region produce variations in the electrical charac-
teristics of the devices (which are mainly due to difference in threshold
voltage) [6]. Therefore, different transistors on the chip may exhibit
large variations in the threshold voltage, resulting in delay variations of
the logic gates and circuits [7]. Further, it is increasingly difficult to pre-
cisely control dopant fluctuations during the manufacturing process and
this problem will aggravate in DSM technologies beyond 45 nm [8].
For the timing analysis of the sequential circuits, accurate character-
ization of the timing parameters of the FFs and latches is vitally impor-
tant [9]. This need becomes more crucial in the design of high-speed
circuits due to the strict design margins required [10]. The task be-
comes more challenging when statistical device variability effects are
also considered. While the maximum achievable performance and yield
of a circuit depends on the magnitude of the variability in the timing pa-
rameters, a better estimate of these parameters can only be made by the
transient analysis of the circuits through the SPICE simulation using
detailed device models. Statistical static timing analysis (SSTA) is a
powerful analysis tool used in the design of digital circuits with the
consideration of variability. During the SSTA of large digital circuits,
the probability density function (pdf) and cumulative density function
(cdf) of the timing parameters of different circuit elements are analyti-
cally processed to estimate the timing characteristics of the circuit pa-
rameters. The accuracy of the analysis depends on the characterization
data of the individual circuit elements, accurate representation of the
characterization data in the form of pdfs, and, finally, the correctness
of different analytical operations used during SSTA, like min, max, or
sum.
In this work, we have studied the impact of statistical variations due
to RDF on the timing characteristics of a standard CMOS FF. An accu-
rate characterization of the timing parameters of FFs constructed with
bulk MOSFETs of 25-, 18-, and 13-nm physical gate length has been
performed using Monte Carlo simulations. Due to the unavailability of
the characterization data from the manufacturing industry, predictive
device model cards have been used. The first four moments for all of
the timing parameters have been computed and can be used for estima-
tion of the probability distributions.
In current state-of-the-art chips, the device count has already ex-
ceeded one billion, mandating the estimation of the distributions more
precisely, especially in the tail regions. These distributions can then
be used to estimate the probability of occurrence of specific events
and, therefore, aid in the design of the system. Parametric analysis,
in which a known parametric distribution (e.g., normal) is fitted on
the experimental data, can be used to undertake this estimation. How-
ever, the limitation of this approach is that its accuracy depends on the
choice of a particular a priori density function [11]. For this reason,
1063-8210/$26.00 © 2010 IEEE