Multi-Objective Optimization of an Embedded Jet Microchannel Heat Sink for Thermal Management of High-Heat-Flux Ai Chips With Multiple Hotspots

Abstract The rapid development of AI technology has placed higher demands on AI computing power, and the power of GPUs is increasing rapidly, which in turn imposes greater requirements on thermal management. Due to the non-uniform component arrangement, the heat distribution within the chip is also uneven, making temperature control of hotspots extremely important. This study focuses on the thermal performance of jet manifold microchannel (JMMC) heat sinks employing a high-thermal-conductivity diamond-copper composite material to mitigate hotspots on large-area, high-power AI chips containing six heat sources. Combined with CFD numerical simulation, regression analysis, and nonlinear programming, the present work establishes a high-precision mathematical model to quickly design the structural parameters of JMMC for high-power AI chips. This study analyzed the impacts of individual structural variables, including jet diameter, jet height, inclination angle, mass flux, and number of outlets, and identified the key factors affecting the maximum chip temperature, temperature uniformity, and pressure drop. Multi-objective algorithms combined with the Pareto front are used for structural optimization, achieving optimal structural performance. The optimized JMMC structure demonstrated significant performance improvements compared with the pre-optimized design. Under a peak heat flux of up to 1200 W/cm2 and a flow rate of 15.73 g/s, the maximum hotspot temperature was reduced by 10.98 °C, temperature uniformity was improved by 76.6%, and the thermal resistance of the JMMC was decreased by 35.47%.