Experimental Study of Two-Phase Cooling to Enable Large-Scale System Computing Performance

High Performance Computing (HPC) performance demands, which usually waterfall into mainstream servers later in time, show increasing need for higher power and hence pose cooling challenges for succeeding generations of compute elements. New generation processors have higher core count, multiple dies, and higher power to achieve generational performance CAGR. Typical high performance compute platforms have dual-socket boards that consume approximately 150-300 watts per socket. Each socket is directly fan-cooled or single phase liquid cooled. The thermal challenge is to cost-effectively increase the thermal envelope to enable greater performance and a target of doubling the power dissipation per socket. In order to push toward large scale system performance via thousands of compute nodes, there is a need to bring several high power processor nodes together with a fabric and a switch chip. Placing them on a single cost-effective motherboard enables high bandwidth fabric signaling and dense packaging, but with a high thermal load. A solution to this is transporting the heat away with two-phase liquid because it enables cooling of the silicon effectively and to a much lower junction temperature. In this paper, conclusive results from a prototype are shown using high efficiency Pumped Liquid Multiphase Cooling technology. 2kW of heat is moved from ten 200W devices (in series) while maintaining low, matched junction temperatures. The thermal management solution is designed for 1U height, keeping low weight as design constraint. The results from pumped two phase system are compared with single phase glycol water system as well and demonstrated advantages of the two phase system.