Routability-driven Power/Ground Network Optimization Based on Machine Learning

Abstract

The dynamic IR drop of a power/ground (PG) network is a critical problem in modern circuit designs. Excessive IR drop slows down circuit performance and causes potential functional failures. Most industrial practices tend to over-design the PG network for the dynamic IR drop constraints, reducing routing resources and incurring routing congestion. Existing machine learning-based approaches target only dynamic IR drop prediction without considering the routability affected by the P/G network. This article develops a machine learning-based method to solve the dynamic IR drop and routing resources tradeoffs. Our model can predict the two targets accurately by adopting a multi-task learning scheme, achieving a 0.99 high correlation coefficient. We show that our trained model is generalizable by testing different placement results. Our algorithm also achieves significant speedups of up to 29× compared to the time-consuming dynamic IR drop simulation by a leading commercial tool. Experimental results show that our algorithm can save about 13% routing resources without worsening the dynamic IR drop peak value.