Investigation of Direct-to-Chip Cooling for High Power GPUs

This technical report investigates a direct-to-chip two-phase cooling solution designed for high power GPUs, specifically focusing on the Nvidia H100 architecture. The study addresses the limitations of conventional air-cooling methods, which struggle to manage the increasing power densities of modern CPUs and GPUs. A testing facility was developed to characterize the performance of this cooling solution, utilizing a thermal test vehicle that mimics the Nvidia H100 GPU. Systematic experiments were conducted to measure key parameters such as pressure drops, die temperature, pumping power, and thermal resistance under varying heat loads and flow rates. The results indicate that the thermal resistance of the cold plate assembly is between 0.02-0.03W/K, and the system performs well under high power tests, demonstrating its scalability and efficiency. The findings suggest that pumped direct-to-chip cooling is a viable and effective solution for managing the thermal demands of next-generation data center technologies.