New cooling design cuts NVIDIA GB200 chip temps by 140°C

Thermal management is a critical bottleneck for next-gen high-performance computing, especially in heterogeneous multi-chip packages like NVIDIA's GB200 Grace Blackwell Superchip. Researchers Michael Acquah and Zheng Liu have developed a computational framework that optimizes embedded cooling channel layouts by coupling steady-state heat conduction with a porous media coolant transport model. Their interdigitated cooling architecture, parameterized by channel count, width, and expansion over chip regions, allows systematic exploration of design space. Using surrogate-based optimization with mixed-integer quadratic programming, the framework minimizes a weighted objective of peak and average chip temperatures, while constraining channel placement to prioritize GPU hotspots.

Applied to a representative GB200 configuration (two GPUs and one CPU), the optimal design achieved a dramatic 140.45°C reduction in peak chip temperature and 35.87°C reduction in average chip temperature compared to baseline. This breakthrough directly addresses the thermal challenges of packing more compute into smaller footprints, enabling higher clock speeds and reliability in AI and HPC workloads. The method is adaptable to other multi-chip architectures, potentially accelerating the roadmap for next-gen supercomputing systems.