Abstract
Along with advancements in microelectronics packaging, the power density of processor units has steadily increased over time. Data center servers equipped for high performance computing (HPC) often use multiple central processing units (CPUs) and graphical processing units (GPUs), thereby resulting in an increased power density, exceeding 1 kW per U. Many data center organizations are evaluating single phase immersion technology as a potential energy and resource saving cooling option.
In this work immersion cooling was studied at a power level of 2.7kW/U with a 5U-height immersion cooling tank. Heat generated by a simulated GPU server was transferred to the secondary loop coolant, and then exchanged with the primary loop facility coolant through the heat exchanger. The chiller supply and return temperature and flow rate was controlled for the primary loop.
The simulated GPU server chassis was designed to provide thermal power equivalent to a high power density server. Eight simulated power heaters, of which each unit was the size of a GPU chipset, was assembled in the comparable location to a real IT equipment on a 4U server chassis. Power for the GPU simulated chassis was able to support up to 2700 W maximum. Three investigations for this immersion cooling system evaluation were performed through comprehensive testing. The first is to identify the key decision making factor(s) for evaluating the thermal performance of 4 hydrocarbon-based dielectric coolants, including power parametric analysis, transient analysis, power cycling test, and fluid temperature profiling. The second is to develop an optimization strategy for the immersion system thermal performance. The third is to verify the capability of an 1U heat sink to support high density processor units over 300 W per GPU in an immersion cooling solution.
A Message-Passing Hardware/Software Cosimulation Environment for Reconfigurable Computing Systems
