Direct to Silicon Microfluidic Cooling for Datacenters
作者
M. Baris Dogruoz,Sashikanth Majety,Venkata Chivukula,Cam Turner,Bharath Ramakrishnan,Husam A. Alissa,Srikant Bharadwaj,Camille Couturier,Malik Fahrni,Yann Meier,Remco van Erp,Samuel Higginbotham
出处
期刊: [Institute of Electrical and Electronics Engineers] 日期:2026-05-26卷期号:: 926-929被引量:1
标识
DOI:10.1109/ectc51846.2026.00156
摘要
The rapid growth of AI and HPC workloads has driven unprecedented thermal design power (TDP) requirements, challenging conventional cooling solutions. This work presents the design, system integration, thermal performance, and reliability of direct-to-silicon microfluidic cooling for NVIDIA GH200 GPUs, with applicability extending to Hopper and Blackwell architectures as well as future GPU variants. Our approach leverages precision etched microchannels at the back of the silicon (of GPU SoC die) to bring the coolant directly into the active die enabling localized heat removal at the die level. We also utilized novel hybrid design architectures for GPU and HBM cooling for reducing flow turbulence, improving heat transfer performance, and reducing the pressure drop while enhancing heat transfer, lowering thermal resistance. We detail the thermal and hydraulic co-design methodology for package integration and silicon etching. Experimental results demonstrate up to 60% reduction in junction-to-inlet temperature rise on GPU SoC die and ~35% reduction in junction-to-inlet temperature rise on the HBM die compared to traditional cold plates without compromising the requirements for datacenter availability; using industry standard ML benchmarks. Reliability testing spanning package-level mechanical stress, long-duration clogging and erosion studies, and node-level GPU operation confirms robust performance with no measurable silicon erosion or clogging related failures. Our findings establish microfluidic cooling as a viable path in advanced liquid cooling for datacenters to enable >1 kW per package thermal envelopes in future GPUs while meeting sustainability and energy-efficiency targets for future cloud infrastructure.