Power control in semiconductor chips: integration of reinforcement learning and dynamic modelling

As semiconductor technology enters the nanoscale era, power control has become a major challenge impeding chip performance. Traditional heuristic methods fail to handle complex dynamic workloads, while pure reinforcement learning lacks stability and safety. This paper proposes a hybrid intelligent control framework integrating reinforcement learning with dynamic physical modelling. Offline-trained power and performance models offer prior guidance and safety constraints for online decisions. Experiments using Google production cluster data show the framework achieves an average power of 101.3 W - 21.3% lower than on-demand strategies - and a tail latency of 34.1 ms with only 1.2% violation rate. The energy efficiency reaches 15.6 instructions per joule, outperforming existing methods. This study provides an effective solution for energy-aware chip-level power management and introduces new ideas for intelligent cyber-physical systems.