Power Management for Chiplet-Based Multicore Systems Using Deep Reinforcement Learning

The trade-off between performance and energy to achieve high energy efficiency has become a critical design issue for computing systems. For emerging chiplet-based multicore systems, the explosive growth of system complexity exacerbates the design challenge to improve the energy efficiency of both processors and power delivery systems (PDSs). Previous works that co-manage processors and PDSs based on reinforcement learning can adapt to dynamic workload variations. However, they face poor scalability and PDS efficiency degradation issues. To tackle the above problems, we propose a deep Q-network (DQN)-based online control scheme for power delivery and consumption co-management of chiplet-based multicore systems. When evaluated on realistic applications, our proposed approach with a centralized DQN agent can achieve on average 4.6% and 33.2% greater energy-delay-product (EDP) reduction over the state-of-art modular Q-learning (MQL) approach and heuristic-based approach, respectively.