DVFS and Timing Optimization on GPU for Data Center Computation

Data center computing requires efficient GPU support, both in terms of functionality and power consumption. GPU performance efficiency can be reduced due to high power usage and reduced GPU work stability. So it requires an analysis of computational performance and power efficiency to improve performance and reduce power usage. Core voltage, core frequency, and memory timings are parameters that affect the efficiency of computing performance, power efficiency, and stability. Increasing computational efficiency and GPU power with the effect of modifying parameters can be done through the Basic Input-Output System (BIOS). This study analyzes the efficiency of computational performance by optimizing memory timings and analyzing power efficiency and stability by modifying the DVFS algorithm. Tests are carried out using computational benchmarks commonly used in data centers including the tessellation algorithm, rendering, image processing, pi calculation, image stitching, deep learning, molecular simulation, and N-body. The efficiency of computing performance and GPU power efficiency can be increased by optimizing memory timings and changing the voltage and frequency values on DVFS. Increased performance efficiency ranged from 33.3% to 66.7% and power efficiency increased from 19.9% to 32.6%. Modification of the DVFS voltage state can increase voltage stability and GPU core frequency stability.