Learning to Place Chiplets: A Multi-Objective Reinforcement Learning Approach

By Richard Chang ¹, Partha Pratim Kundu ¹, Jun Liu ², Dingjie Lu ², Sezin Ata Kircali ¹, Yubo Hou ¹, Jie Wang ¹, Gen Liang Lim ¹, Sridhar Narayanaswamy ², Rotaru Mihai Dragos ³, Rahul Dutta ³ and Ashish James ^1
1 Institute for Infocomm Research, Singapore
² Institute of High Performance Computing, Singapore
³ Institute of Microelectronics, Singapore

Abstract

As heterogeneous systems scale, traditional rule-based and stochastic methods used for chip placement face limitations in convergence, scalability, and thermal management. To address these challenges, we propose a reinforcement learning (RL) framework with multiobjective optimization, employing a customized reward shaping method to minimize interconnect wirelength and improve thermal distribution. We applied our approach to two generic use cases in 2.5D advanced packaging - a 4-chiplet RDL system and a multi-GPU system. Our results outperformed the state-of-the-art methods like Bayesian optimization (BO) by up to 40% in wirelength reduction with 4°C thermal improvements.

Keywords

Simulated annealing, Bayesian optimization, Multi-objective reinforcement learning, Advanced packaging

To read the full article, click here