COMPARATIVE CFD ANALYSIS OF THE EFFECT OF AIR AND LIQUID COOLING ON THE FORM FACTOR OF GPU SERVER

Abstract

Artificial intelligence (AI) has becoming an important area which may significantly impact everyone’s daily life. This means a lot of high-performance chips such as high-performance CPU, GPU, FPGA (Field-programmable gate array), ASIC (application specific integrated circuit) devices may need. The thermal design power (TDP) of these chips are high, and now it is very common to see a processor TDP reaches as high as 300Watts[1]. When these high-power density processors are packaged together, the power density of the servers and compute boxes are significantly high. Traditional air-cooling is approaching its cooling capability limitation, especially when a cluster of racks is fully populated with computing-intensive nodes. The air cooling may still capable for high density thermal management, however there will be a significant amount of cost associate with it. Liquid cooling may provide a higher cooling efficiency, which can easily reject larger amount of heat generated from processors. This lowers the cooling PUE. In this work, the server used for study is FACEBOOK BIGBASIN which has 3OU chassis, consisting 8 Nvidia Volta V100 GPU, 4 PCIe cards, and 8 hot swappable fans. On the GPUs, there are 8 heatsinks on in the air-cooled model and 8 impingement type cold plates in the liquid cooled model. This work is proposed to identify, evaluate and develop effective liquid cooling by using cold plate which gives high performance compute with compared to air cooled solution. The GPU chips are assumed to be running at maximum Utilization. The cold plate is optimized for better thermal performance in 6sigma ET software.