Reinforcement Learning-based Power Management Architecture for Optimal DVFS in SoCs

This paper presents a Power/Performance-optimal Markov Decision Processes (MDP) based DVFS (Dynamic Voltage and Frequency Scaling) controller. MDP framework is employed for DVFS switching decision-making under conditions of uncertainly. The approach yields an optimal Power/Performance DVFS technique based on provided reward/performance structure, presenting a dynamically adjusted DVFS mechanism in respect to the formulated model.