Improving energy efficiency of nvidia GPUs

Abstract

The project focuses on enhancing the energy efficiency of Nvidia GPUs, crucial for graphics, AI, and high-performance computing. The rising demand for GPUs has led to increased energy consumption, necessitating optimization for sustainability and cost-effectiveness. The project employs Dynamic Voltage and Frequency Scaling (DVFS) to dynamically adjust GPU frequency at runtime, aiming to find the optimal frequency for each task in real-time. The developed GPU profiling library, Annalist- Nvidia, utilizes NVIDIA Management Library (NVML) and CUDA Profiling Tools Interface (CUPTI) for online profiling with minimal overhead. Two DVFS policies, static and dynamic, are explored, showing substantial energy reductions, particularly in dynamic DVFS for the Stream benchmark. Future work includes extending DVFS policies, employing power-capping, and expanding the profiling library for broader language support and multi-GPU profiling.