Deep Reinforcement Learning-Based Power Distribution Network Design Optimization for Multi-Chiplet System

This study employs a reinforcement learning (RL) technique, Dueling Double Deep Q Network (DDDQN), to optimize the power distribution network (PDN) in a high-density fan-out wafer-level package (HD-FOWLP) for a four-chiplet system, addressing the complexities of the PDN design with multiple power domains. A four-chiplet system is arranged in a ring topology, with each chiplet featuring seven distinct power islands, resulting in a total of 28 separate power islands within the power layer of the HD-FOWLP package's redistribution layer (RDL). The reinforcement learning algorithm used in this work addresses the difficulties of expansive design space and offers efficient optimization as well as shorter development time. A grid of 13 x 13 cells (in the format of a matrix) was used as the design environment to represent one quadrant of the package PDN; the design rules were translated into analytical expressions as the reward functions for RL. Training the RL model to generate a viable solution took 6 CPU hours. The proposed model demonstrates the ability to reduce development time and cost, easing engineers' workload and development difficulty.