Dynamic Modeling and Experimental Evaluation of a Controlled Two-Phase On-Chip Cooling System Designed for High Efficiency Datacenters

Abstract

An important challenge for the next generation of high performance datacenters is to guarantee sufficient, reliable, green and cheap cooling of the server blades. Since this seems unrealistic with air cooling technologies due to the expected ever higher chip heat fluxes (in excess of 100 W/cm2), a novel more effective cooling system is required. Switching to two-phase on-chip cooling systems would decrease the energy consumption, enhance the thermal performance and allow the reuse of the extracted heat. However, the design of such systems seems nontrivial because of the lack of simulation tools able to reproduce their behavior. In this context, the present thesis describes the one-dimensional dynamic modeling of an on-chip two-phase liquid pump cooling system for multiple micro-processors. The simulation code includes state-of-the art two-phase flow methods, two-dimensional conduction in the heat sink, heat flux dependent mass flow rate distribution between parallel micro-evaporators, and micro-condenser and in-line liquid accumulator models. Three different scales of simulations, namely the heat sink, multiple parallel heat sinks and entire system scales, are presented. An experimental test bench with a cooling capacity of about 300W is used to validate the code. Mass flow rate distribution, energy oriented performance maps, chip temperature profile, heat flux disturbances under controlled conditions and charge influence are studied experimentally. The experimental data are satisfactorily reproduced by the simulations with, for example, 266 steady-state mean chip temperatures, obtained for balanced and unbalanced heat fluxes conditions, predicted within ± 2K. The most significant discrepancy between simulations and experiments is found to be the pressure transient prediction. The validated code is then used to simulate a compact two-phase on-chip cooling system designed for a real datacenter blade server. At the heat sink level, start-up and hot-spots are studied using 2D temperature maps and time-strip analysis. Furthermore, the flow distribution between the parallel micro-evaporators of the studied layout is shown to be unique and thermal-hydraulic maps are generated. The system start-up and operation under random and periodic heat load disturbances with different control strategies are then analyzed. Finally, the thermo-economic integration of a datacenter with a district heating network is investigated.