Efficient Monte Carlo Based Methods for Variability Aware Analysis and Optimization of Digital Circuits.

Abstract

Process variability is of increasing concern in modern nanometer-scale CMOS. The suitability of Monte Carlo based algorithms for efficient analysis and optimization of digital circuits under variability is explored in this work. Random sampling based Monte Carlo techniques incur high cost of computation, due to the large sample size required to achieve target accuracy. This motivates the need for intelligent sample selection techniques to reduce the number of samples. As these techniques depend on information about the system under analysis, there is a need to tailor the techniques to fit the specific application context. We propose efficient smart sampling based techniques for timing and leakage power consumption analysis of digital circuits. For the case of timing analysis, we show that the proposed method requires 23.8X fewer samples on average to achieve comparable accuracy as a random sampling approach, for benchmark circuits studied. It is further illustrated that the parallelism available in such techniques can be exploited using parallel machines, especially Graphics Processing Units. Here, we show that SH-QMC implemented on a Multi GPU is twice as fast as a single STA on a CPU for benchmark circuits considered. Next we study the possibility of using such information from statistical analysis to optimize digital circuits under variability, for example to achieve minimum area on silicon though gate sizing while meeting a timing constraint. Though several techniques to optimize circuits have been proposed in literature, it is not clear how much gains are obtained in these approaches specifically through utilization of statistical information. Therefore, an effective lower bound computation technique is proposed to enable efficient comparison of statistical design optimization techniques. It is shown that even techniques which use only limited statistical information can achieve results to within 10% of the proposed lower bound. We conclude that future optimization research should shift focus from use of more statistical information to achieving more efficiency and parallelism to obtain speed ups.