Аннотация
Concerning Deep Learning (esp. CNNs) workloads on integrated-GPU based heterogeneous programming models and mobile/embedded platforms, we propose a data-driven offline (batch) RL-oriented dynamic power management (DPM) agent/governor design that uses an offline RL method to learn a simple, effective and robust policy from previously-collected static datasets without further interaction with the environment, concerning intrinsically expensive training costs of generic offpolicy online RL or/and imitation learning (IL). Our preliminary results show improvements up to 20.3% in terms of energy-delay product (EDP) and a reduced inference time of 2.2% on Tiny- YOLOv3/v4, compared to the default separate CPU and GPU governors.