TDPNavigator-Placer: Thermal- and Wirelength-Aware Chiplet Placement in 2.5D Systems Through Multi-Agent Reinforcement Learning

Researchers have unveiled TDPNavigator-Placer, a novel multi-agent reinforcement learning framework that optimizes chiplet placement in advanced 2.5D integrated circuits. It uniquely tackles the conflicting objectives of minimizing wirelength for speed and managing thermal design power (TDP) for heat. The system uses specialized agents for each goal, outperforming current state-of-the-art methods and delivering a significantly improved Pareto front for more balanced, practical chip designs.