Optimizing Chiplet Placement in Thermally Aware Heterogeneous 2.5D Systems Using Reinforcement Learning

The rapid advancements of heterogeneous 2.5D systems demand innovative solutions to optimize chiplet placement, ensuring a balance between thermal management, performance, and connectivity. This paper presents a reinforcement learning (RL) framework that dynamically arranges chiplets on a substrate to reduce thermal hotspots, improve heat dissipation, and enhance system performance. The RL framework models chiplet positions as the state space, with actions representing potential movements and rotations of chiplets, guided by a reward function that optimizes both temperature and interconnect length. The performance of the RL framework is validated on a multi-GPU system and a CPU-DRAM system, where the RL approach outperforms simulated annealing, demonstrating its potential to automate complex design tasks and accelerate innovations in semiconductor architecture by integrating AI-driven techniques.