Numerical investigation on thermal performance of single-phase immersion cooling for GPU servers comparing with CPU servers

The growth of artificial intelligence computing has intensified the cooling energy demand of high-power GPU servers. Single-phase immersion cooling (SPIC) offers a promising route to improve data center energy efficiency, but its operating characteristics and design requirements for GPU servers remain insufficiently understood. This study develops a numerical model of a 1U eight-node GPU server with a total power of 5.6 kW under forced SPIC, accounting for the heat dissipation of GPU dies and high bandwidth memory stacks.

A coolant flow through fraction (FTF) is introduced to quantify coolant distribution efficiency. The effects of inlet-outlet coolant temperature difference, corresponding flow rate, inlet temperature, and power scenario are evaluated, together with a comparison against a CPU server. GPU heat sinks show approximately threefold higher thermal impedance than CPU heat sinks, while mixed convection and distinct thermal shadowing mechanisms govern their cooling requirements.

Three thermal response regimes are identified, with an inlet-outlet temperature difference of 4–10 °C providing stable chip temperatures and improved temperature uniformity. Increasing the inlet temperature raises chip temperatures but enhances heat removal, highlighting the need to balance thermal safety and cooling energy consumption. Both server types exhibit FTF values below 50%, revealing substantial bypass flow and considerable potential for flow path optimization.

Thermal resistance network analysis further shows that packages above 80 W/cm2 require a thermal interface material (TIM) thermal impedance below 0.03 °C·cm2/W and additional heat transfer enhancement. These findings provide guidance for energy efficient operating condition selection and thermal design of immersion cooled GPU servers.