Key points are not available for this paper at this time.
The continuous increase in computational power of GPUs, essential for advancements in areas like artificial intelligence and data processing, is driving the adoption of liquid cooling in data centers. Skived copper cold plates featuring parallel straight channels are a mature technology, but they lack design freedom due to manufacturing limitations. As chips become increasingly complex in their design with the transition towards heterogeneous integration, these parallel straight channels are not able to address critical areas of concentrated high heat flux (hotspots) on a chip. A single hotspot exceeding the upper temperature limit can cause the full chip to throttle and hence limit performance. In addition, this would require a reduction in coolant inlet temperature in the data center, causing an increase in electricity and water consumption. Ideally, areas of the cold plate in contact with hotspots of the chip need smaller channels to increase convective heat transfer, whereas areas with low heat flux may benefit from larger channels to compensate for the increased pressure drop. However, manual optimization of such a cooling design is challenging due to the nonlinearity of the problem. In this paper, we explore the usage of topology optimization as a method to tailor microfluidic cooling design to the power distribution of a chip to address the hotspot temperatures in high-power chips, using a platform called Glacierware. We compare the hotspot-aware, topology-optimized microfluidic design to straight channels of various widths to benchmark its performance. Evaluations of this optimized design show a 13% lower temperature rise or a 55% lower pressure than the best-performing straight channels, indicating highly competitive performance in industrial settings where both pressure and flow rate are constrained.
Boutsikakis et al. (Sun,) studied this question.