Thermal Performance Characteristic of Single-phase Immersion Cooling on Servers with Optimized Heatsink Design

The rapid development of artificial intelligence (AI), high performance computing (HPC), communication technology and cloud service has been requesting high performance of servers powered by CPU and GPU in data centers (DCs). In order to meet the commercial and industrial demand, the thermal design power (TDP) of both CPU and GPU has been rising from generation to generation due to constant increase on the core count, cache size and heat flux density. Intel XEON processor has reached 350W and Nvidia’s H100 GPU even up to 700W on TDP. This exposes huge challenges on system thermal design with traditional air-cooling solution under current boundary condition of heatsink keep out zone (KOZ), air flow and off-the-shelf heatsink technologies. Therefore, liquid cooling has more attention to address the thermal management challenges from increasing TDP of chips and rack power density in DCs. Among these novel liquid cooling solutions, the singlephase immersion cooling(1-PIC) has emerged due to the special advantages over air-cooling solution in terms of enhanced cooling capability, improved components operation reliability and optimal Power usage effectiveness (PUE).Tencent Cloud (Tencent) has huge amount deployment of servers in DCs globally to respond the demand of all walks of life and consumes huge electricity energy each year. In order to reduce carbon footprint and accomplish green and sustainable DC, one 1-PIC solution prototype with fluorochemical fluid is built to investigate the cooling performance and optimization for application in future. This paper presents both computational fluid dynamics (CFD) analysis and experimental investigation to study the thermal performance characteristic of 1-PIC on servers developed by Tencent on Intel Eagle Stream (EGS) platform. Thermal performance is characterized based on system level to represent the real operation scenarios. Meanwhile, this paper also describes the design of prototype tank, rework on servers for immersion, thermal interface material (TIM) selection and impact on thermal cooling performance from different heatsink design, fluid flowrate and fluid inlet temperature. It is demonstrated that 1-PIC solution could deliver better cooling capability and support higher TDP of CPU than traditional aircooling solution, which could be a promising cooling methodology for high-power server’s adoption in DCs.