*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Title: Machine Learning in Physical Design
Committee:
Dr. Lim, Advisor
Dr. Swaminathan, Chair
Dr. Mukhopdhyay
Abstract: The objective of the proposed research is to devise learning-based algorithms which can address the challenges that current state-of-the-art physical design (PD) implementation flows for 2D and 3D ICs are facing. Due to the non-generalizable heuristics adopted by most PD tools, chip designers are struggling to obtain desired power, performance, and area (PPA) targets in advanced technology nodes. In this research, we focus on devising Machine Learning (ML) algorithms that help mitigate this issue. The proposal focuses on the use cases of ML in the following three topics: 1. Design space exploration and optimization, 2. Netlist encoding for generalization across designs and technologies, and 3. Unsupervised loss functions for direct PPA optimization. Given that PD is a sequential process that is consisted of many design stages, in the preliminary research, we tackle each topic by devising novel ML algorithms at a dedicated PD stage. At the clock tree synthesis (CTS) stage, we propose a generative adversarial framework to perform CTS outcomes prediction and optimization. At the placement stage, we devise a methodology to perform placement optimization on commercial CPU designs by minimizing an unsupervised loss function through graph neural networks (GNNs). Finally, to improve the PD flow of monolithic 3D (M3D) ICs, we present a tier partitioning methodology using GNNs and the weighted k-means clustering algorithm. We believe the presented works shall demonstrate the great potentials of leveraging ML algorithms to revisit classic PD problems.
