MLTimer: Leakage Power Minimization in Digital Circuits Using Machine Learning and Adaptive Lazy Timing Analysis

摘要

The timing constrained discrete sizing technique (TC-DSP) is employed at all stages of the physical synthesis flow and has been studied extensively over the last 30 years. The ISPD gate sizing contests introduced industry standard benchmarks and library which motivated a lot of researchin this area. However most of the solutions employed were either sensitivity driven or based on analytical methods that required incremental timing analysis after every iteration with both consuming a significant amount of time to perform the optimization. The key observations reported inthis paper are (i) there exists a good correlation between the slack distribution among gates in a given iteration and the order of gate replacements in subsequent iterations; and, (ii) across the benchmark circuits there exists significant overlap in the number of sub-circuits that have similarstructures. This paper exploits the above observations to propose MLTimer, an iterative algorithm that uses adaptive lazy timing analysis in conjunction with a Support Vector Machine (SVM) engine for solving the TC-DSP quickly and efficiently. We observe that for large benchmark circuits (≥200,000)our proposed solution reduces the leakage power by 3% and the running time by over 50% when compared to the best reported heuristic in the literature. This significant decrease in running time is very useful to the industry for achieving timing and power closures of large designs within agiven deadline.