Feasibility Evaluation of Immersion Cooling in Hyperscale Data Centers

Sammanfattning

The demand for higher power densities of the IT in data centers is increasing signifcantly, driven in part by the broader implementation of artifcial intelligence (AI). As nearly all the energy consumed by the IT is converted into heat, effcient thermal management systems are essential. When power densities exceed 20-30 kW per rack, traditional air cooling methods become insuffcient, necessitating alternative approaches such as liquid cooling. One such method is immersion cooling, which offers a promising solution for high power density applications. This thesis aims to evaluate the feasibility of immersion cooling in hyperscale data centers. A comparative analysis was conducted between immersion cooling and direct to chip cooling, alongside an evaluation of commercially available immersion fuids and their associated considerations. The Figure of Merit (FOM), which combines thermal properties of a fuid based on various heat transfer correlations, was used to assess fuid performance. Three different FOMs were calculated, corresponding to natural convection, forced convection and pressure drop, for a selection of commercially available immersion fuids. These fuids were evaluated based on both performance, environmental impact and safety. Additionally, a 960 kW data center using immersion cooling was also modeled in MATLAB Simulink to assess the potential for free cooling in Stockholm under ambient conditions. The simulation was conducted over a full annual cycle, with variations in both cooling water and immersion fuid temperature setpoints to assess system performance under different operating conditions. The literature comparison indicated that immersion cooling offers several advantages over alternative cooling methods, including improved temperature uniformity across servers and the potential for achieving lower Power Usage Effectiveness (PUE) values at high power density. However, it also presents certain drawbacks, such as increased whitespace area requirement within the data center and challenges related to material compatibility between server components and the immersion fuid. The fuid evaluation demonstrated that certain immersion fuids show low environmental impacts and can be operated without signifcant safety concerns. However, performance varied notably among the fuids, largely due to differences in viscosity. Among the evaluated fuids, SmartCoolant and DC20 were the best performing fuids, both in FOM and in the system simulations. The simulation results further indicated that a 960 kW immersion cooling system could operate using free cooling in Stockholm’s ambient conditions when coolant inlet temperatures were maintained at 45◦C and 50◦C. At a lower temperature of 40◦C, the performance was limited by the capacity of the dry cooler.