Optimisation of liquid jet impingement coldplate for an overclocked CPU with multiple discrete cores

• Jet array impingement thermal-hydraulics sensitive to several geometric design parameters, with cross-dependencies between them. • Automated optimisation workflow developed in simulation computational fluid dynamics-based environment. • Maximum heat transfer requires focussed array of jets concentrated around regions of high heat flux. • Higher population of lower velocity jets favoured to mitigate pressure drop and reduce hydraulic power. • Balance between heat transfer and pressure achieved with high population of closely spaced, large diameter jets sustaining high convective heat transfer intensity over large area. • Approximately 30% of the total core-to-coolant thermal budget is associated with the heat transfer across the CPU silicon die, hence improvements to the heat exchanger have a diminished overall impact on the net core-to-coolant performance. This study details the development of an automated, multi-objective optimisation framework for liquid jet impingement cooling of commercial Central Processing Units (CPUs). Integrating Computer Aided Design (CAD), Computational Fluid Dynamics (CFD), and geometric optimisation software tools, optimised jet array configurations were tailored for cooling of a commercial CPU with multiple, heat-generating, high heat flux cores. Experiments were performed on a live test CPU under a highly overclocked compute workload and the measurements used to validate the simulation model both pre and post optimisation. A comparative analysis of both single and multi-objective optimisation design strategies illustrated trade-offs depending on the choice of objective functions, which were chosen to be the average core junction temperature and the hydraulic pumping power in this study. When maximised cooling of the cores was considered, the optimum jet design aggressively targeted the concentrated heat source region of the CPU cores, achieving a core junction temperature 10 K below the throttling limit, even under highly overclocked workload conditions. Conversely, the multi-objective design that equally weighted cooling performance and hydraulic power penalty produced a design with a significantly larger population of lower-velocity jets, leveraging a larger effective surface area of the heat spreader to compensate for the lower jet impingement convective heat transfer coefficients. Despite the higher core junction temperature, it still remained below the throttling temperature whilst achieving a fourfold reduction in pumping power. Analysis of the CPU thermal stack identified significant temperature gradients across the silicon die owing to the very high localised heat fluxes at the discrete cores. This contributed approximately 30% of the total core-to-coolant thermal budget, which is significant. This highlights the importance of applying realistic thermal boundary conditions when designing coldplate heat exchangers for practical CPU packages.