As Moore’s law slows down, computing systems must prioritize higher energy ef-ficiency to sustain performance scaling. CPUs and GPUs stand out as the primary workhorses of computing resources, underscoring the critical need for these resources to achieve high energy efficiency. However, the implementation of runtime power management on CPUs and GPUs presents significant challenges, given the high varia-tions and complexities stemming from diverse workloads and hardware configurations. These factors often diminish the efficacy of offline optimization and reactive-based power management methods.

In this thesis, we present a prediction-based runtime power management solution designed for multicore CPUs or GPUs. Our proposed approach utilizes dynamic volt-age and frequency scaling (DVFS) a...[ Read more ]

As Moore’s law slows down, computing systems must prioritize higher energy ef-ficiency to sustain performance scaling. CPUs and GPUs stand out as the primary workhorses of computing resources, underscoring the critical need for these resources to achieve high energy efficiency. However, the implementation of runtime power management on CPUs and GPUs presents significant challenges, given the high varia-tions and complexities stemming from diverse workloads and hardware configurations. These factors often diminish the efficacy of offline optimization and reactive-based power management methods.

In this thesis, we present a prediction-based runtime power management solution designed for multicore CPUs or GPUs. Our proposed approach utilizes dynamic volt-age and frequency scaling (DVFS) as the primary management technique, with the goal of minimizing energy consumption while meeting user-defined performance require-ments. Furthermore, we integrate a variety of valuable methodologies, including rein-forcement learning, transfer learning, and federated learning, to enhance the predictive capabilities and adaptability to varying workloads and environments, thereby improving overall energy efficiency outcomes. The proposed approach is customized for multicore CPUs and GPUs based on their characteristics and is evaluated separately. Experiments conducted on practical applications demonstrate that our method achieves maximum energy savings under user-defined performance constraints when compared to state-of-the-art designs.