Characterizing Power and Performance of GPU Memory Access

Abstract

Power is a major limiting factor for the future of HPC and the realization of exascale computing under a power budget. GPUs have now become a mainstream parallel computation device in HPC, and optimizing power usage on GPUs is critical to achieving future goals. GPU memory is seldom studied, especially for power usage. Nevertheless, memory accesses draw significant power and are critical to understanding and optimizing GPU power usage. In this work we investigate the power and performance characteristics of various GPU memory accesses. We take an empirical approach and experimentally examine and evaluate how GPU power and performance vary with data access patterns and software parameters including GPU thread block size. In addition, we take into account the advanced power saving technology dynamic voltage and frequency scaling (DVFS) on GPU processing units and global memory. We analyze power and performance and provide some suggestions for the optimal parameters for applications that heavily use specific memory operations.