Symbolic-functional representation inference for gate-level power estimation

Abstract

We propose SyfriPow, a method for estimating the vectorless average power consumption of gate-level circuits using sparse symbolic matrix inference. SyfriPow employs a probability-based approach, utilizing signal and transition probability to assess power consumption across various nodes. We present a symbolic representation probabilistic model for reasoning signal and transition probability through polynomial-based symbolic inference, incorporating a polynomial approximation strategy for spatial correlations and functional mapping for quick secondary computation. The model is sparsified with GPU accelerated polynomial sparse arithmetic engine, achieving sparse symbolic inference and sparse functional mapping which are implemented on Pytorch. Experiments demonstrate that SyfriPow achieves node-level probabilistic accuracy and significantly improves power accuracy compared to industrial software and academic algorithms, with an average power error of less than 4%. SFM efficiently analyzes large-scale circuits, processing up to 250k nodes in 100 s. SyfriPow can also function as an independent logic prediction engine, surpassing state-of-the-art algorithms in accuracy and speed.