Interpretable machine learning in VLSI physical design

Abstract

Today's popularisation of portable devices largely depends on the progress in integrated circuits. Modern Very Large Scale Integration technology (VLSI) allows billions of transistors to be packed into the same chip. In the past years, digital design in VLSI has been developed compared to analogue design. The traditional method is hard to model the performance change in analogue or mixed-signal components caused by physical design. In the early 2000s, rapid advances in machine learning and computing power made analogue design automation possible. Despite their outstanding performance, the transparency issue has become significant. This paper introduces the history of VLSI physical design, which includes placement and routing in the early stages. The change that machine learning (ML) has made is mentioned in the third section. Analysis of the potential problem has been proposed, followed by a brief category of some well-known work in interpretable Machine Learning, which could be the primary direction for VLSI automation to be further popularised in the future.