Stochastic logical effort as a variation aware delay model to estimate timing yield☆
Introduction
Decreasing sizes of transistors result in manufacturing of digital integrated circuits (IC) to become much more difficult and prone to variations of parameters like transistor gate length, threshold voltage, etc. Performance (speed) variability due to the statistical parameter variations and environmental fluctuations has become more significant. IC designers need to estimate the timing yield and optimize their design accordingly until it reaches the desired yield before manufacturing. A survey [1] in 2011 by Solido Design Automation over 486 IC design professionals shows that two-thirds of designers and managers name variation-aware design as a top segment where technology advancement is needed.
In traditional VLSI design methodologies, designers prefer using Spice tool for detailed transistor level (TL) circuit simulations as a final verification before timing sign-off because of its accuracy. One would ideally like to perform a similar transistor-level, but statistical timing analysis for timing yield estimation. According to the above survey, more than half of the participant IC designers and managers want Spice simulators to be variation aware in the first place. Taiwan Semiconductor Manufacturing Company (TSMC) has already announced the insertion of transistor-level statistical timing analysis into its reference design flow in order to enhance timing accuracy [2]. There has been intense academic research on statistical timing analysis and timing yield estimation topics especially in the last decade [3], [4]. The researchers have to cope with hard problems like modeling inter- and intra-die variations with spatial correlations, accurate delay approximations without solving the actual non-linear and differential delay equations, propagation of the non-Gaussian random variables, etc. Previously negligible problems have become more and more important with the shrinking technology. Today, intra-die variations are at least as important as inter-die variations [5]. The combination of all these problems either increases the computational complexity of the solution or decreases the accuracy of the resultant timing yield estimate. Making too many assumptions to decrease complexity results in far-off estimates. The most accurate approach in statistical timing is the Monte Carlo (MC) method based on costly transistor level Spice simulations, which is the called golden method, however it is computationally too complex to be applicable.
The main contributions of this paper can be summarized as follows: a variation aware delay model, stochastic logical effort (SLE), is proposed to capture the tendency of a gate׳s delay with respect to the random device parameters. The characterization of a standard cell library to prepare it for SLE delay computation and efficient methods for the computation of SLE model parameters are presented. SLE is combined with a previously proposed importance sampling based timing yield estimation technique [6] to build a new estimator called ISLE. The theoretical error of the resultant ISLE estimator is derived in detail. The empirical tests over ISCAS’85 benchmark circuits considering both inter- and intra-die variations with spatial correlations show the accuracy of the SLE method. The results show that ISLE timing yield estimation is about 180 times faster on the average than the traditional methods with the same accuracy. ISLE is not meant to be a replacement of the less accurate but faster statistical static timing analysis methods, but instead a complementary method to be used as a final verification for the statistically critical paths in the circuit.
In Section 2, we present the variation aware gate delay model SLE. In Section 3, ISLE yield estimator is explained and a precise theoretical error analysis is given. Finally in Section 4, we present the experimental results and the computational cost.
Section snippets
Variation aware gate delay model: stochastic logical effort (SLE)
Section 2.1 provides an overview of the well-known logical effort approach. In Section 2.2, we introduce SLE for approximating circuit delay in the presence of statistical variations. Section 2.3 explains the extraction of the model parameters for the SLE gate delay models, Section 2.4 presents the characterization of a cell library and Section 2.5 explains the detection of the SLE parameter values for a given sample point from the random parameter space.
Preliminaries and previous work
Timing yield is the fraction of dies, which satisfy the timing requirements, in other words which have circuit delay smaller than a timing constraint .
In our previous work [6], we had introduced a novel timing yield estimation methodology based on importance sampling. In this paper, we will utilize the same methodology, but this time in conjunction with the SLE gate delay model. The timing yield estimation model is based on the transistor level Monte Carlo estimation of timing yield (Yield)
Experimental setup
Among all process parameters, the most significant are channel length and threshold voltage [10]. Therefore, in our experiments, we assumed two random device parameters: transistor gate length and threshold voltage , whose statistical variations are regularly reported by International Technology Roadmap for Semiconductors (ITRS) reports due to their impact on the performance of the integrated circuits. The random parameter variations are set according to the 2011 report of ITRS [11]:
- •
Conclusion
We have demonstrated in this paper that for a variance aware IC delay analysis stochastic logical effort can be used to approximately but efficiently capture the gate delay variations due to the parameter variations and that importance sampling in conjunction with stochastic logical effort (ISLE) can serve as a very accurate yet computationally viable timing yield estimation method as a final stage verification.
The proposed SLE formalism based ISLE timing yield estimation is applied to ISCAS’85
Alp Arslan Bayrakci received the B.S. degree in electrical and electronics engineering from Middle East Technical University, Ankara, Turkey, in 2004 and direct Ph.D. degree in computer engineering from Koc University, Istanbul, Turkey, in 2010. He has been with the Gebze Institute of Technology, Kocaeli, Turkey, since June 2011. His current research interests include statistical timing analysis, computer arithmetic, hardware trojans and computer-aided design methodologies.
References (17)
- et al.
Statistical static timing analysisa survey
Integr. VLSI J.
(2009) - et al.
Invited papervariability in nanometer cmos: impact, analysis, and minimization
Integr. VLSI J.
(2008) - 〈http://www.solidodesign.com/files/variation-aware-custom-design-survey-2011.pdf〉. Variation Aware Custom IC Design...
- Taiwan Semiconductor Manufacturing Company (TSMC) Reference Flow 9.0 Press Release. New TSMC Reference Flow 9.0...
- et al.
Statistical timing analysisfrom basic principles to state of the art
IEEE Trans. Comput. Aided Des. Integr. Circuits Syst.
(2008) - Ji-Seong Doh, Dae-Wook Kim, Sang-Hoon Lee, Jong-Bae Lee, Young-Kwan Park, Moon-Hyun Yoo, Jeong-Taek Kong, A unified...
- et al.
Fast monte carlo estimation of timing yield with importance sampling and transistor-level circuit simulation
IEEE Trans. Comput. Aided Des. Integr. Circuits Syst.
(2010) - et al.
Logical Effort: Designing Fast CMOS Circuits
(1999)
Cited by (6)
Statistical analysis of asynchronous pipelines in presence of process variation using formal models
2016, Integration, the VLSI JournalCitation Excerpt :In [46,47] the variation impacts on GALS NoC links have been explored; however the authors focused on synchronous and asynchronous interfaces, but asynchronous components in design were not covered. In [48,49] the timing delay has been modeled for yield improvement and voltage and temperature impacts analysis. From verification angles, reliability analysis of digital circuits especially transparent latch in presence of variation are discussed in [50,51].
Acceleration of Spatial Correlation Based Hardware Trojan Detection Using Shared Grids Ratio
2020, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)A Chip-Level Multi-parametric Yield Prediction Method Based on Copula Theory
2017, Tien Tzu Hsueh Pao/Acta Electronica SinicaHardware Trojan detection based on correlated path delays in defiance of variations with spatial correlations
2017, Proceedings of the 2017 Design, Automation and Test in Europe, DATE 2017An efficient prediction framework for multi-parametric yield analysis under parameter variations
2016, Frontiers of Information Technology and Electronic Engineering
Alp Arslan Bayrakci received the B.S. degree in electrical and electronics engineering from Middle East Technical University, Ankara, Turkey, in 2004 and direct Ph.D. degree in computer engineering from Koc University, Istanbul, Turkey, in 2010. He has been with the Gebze Institute of Technology, Kocaeli, Turkey, since June 2011. His current research interests include statistical timing analysis, computer arithmetic, hardware trojans and computer-aided design methodologies.