Machine Learning Application for Early Power Analysis Accuracy Improvement: A Case Study for Cells Switching Power

Abstract

Power Analysis (PA) is an important task performed repeatedly throughout the flow to ensure the design closure within the power budget. One of the challenges of the state-of-art PA tools is the low accuracy at high abstraction levels such as gate and RTL. This accuracy gap can be reduced by feeding-back physical information such as Standard Parasitic Extraction File (SPEF) to higher abstraction level to have an estimate of the RC components of the nets. Unfortunately, with current design methodologies, the SPEF file is only available at a very late stage of the circuit development after passing the physical design stage. In this paper, we introduce a machine learning application that estimates accurately the switching power of the cells without needing the SPEF file (SPEF less PA flow). Three ML models (Multi-linear Regression, Random Forest, and Decision Tree) were trained and tested on different industrial designs at 7nm technology. They are trained using different cells' properties available, SPEF, and SPEF-less power numbers to accurately predict the Switching power and eliminate the need for the SPEF file. With this new ML approach, we were able to reduce the SPEF less flow average cell switching power error from 34% to 8%.