Artificial Intelligence
Two-Phase Micro-Channel Copper Coldplates Sustain 1,500W per Accelerator Socket at 1.04 Facility PUE
Silicon Valley thermal engineers demonstrate pumpless dielectric vapor-chamber boiling across 72-GPU rack manifolds with 45-degree Celsius warm-water heat rejection.

Your listening Queue & catch up
Add a story or build a catch-up queue to start listening.
Thermal-mechanical engineers and datacenter infrastructure architects across Santa Clara, San Jose, and UC Berkeley published 2,000-hour endurance benchmarks on Saturday morning through the Open Compute Project Advanced Cooling Solutions (ACS) committee, validating a direct-to-chip Two-Phase Micro-Channel Coldplate architecture that cools 1,500-watt multi-reticle AI accelerators while holding junction-to-fluid thermal resistance below 0.014 °C/W. As next-generation GPU and custom ASIC packages push localized heat flux across HBM4 and logic chiplets beyond 180 watts per square centimeter, conventional single-phase water-glycol coldplates require high-pressure pumping loops and chilled facility water that degrade datacenter Power Usage Effectiveness (PUE).
Inside the new 40-micron skived-copper micro-channel coldplate, a non-conductive, zero-ozone-depletion hydrofluoroolefin (HFO) dielectric fluid enters as a subcooled liquid and undergoes nucleate boiling directly atop the processor lid at a uniform saturation temperature of 54 degrees Celsius. Because latent heat of vaporization absorbs 6x more thermal energy per gram than sensible liquid heating, mass flow rates drop by 78 percent and all eight accelerators on a server sled stay isothermal within 1.2 °C from inlet to outlet, eliminating thermal-throttling skew across synchronous tensor-parallel ranks.
At the rack Rear-Door Condenser Heat Exchanger, the vapor condenses back into liquid against a 45-degree Celsius warm facility water loop that rejects heat directly to outdoor dry-air coolers year-round in San Jose and Santa Clara climates without running mechanical compressor chillers or evaporating cooling-tower water. Across a 120-kilowatt rack testbed, parasitic cooling pump and fan power fell to 3.8 percent of IT load, achieving a measured facility PUE of 1.04 and zero water consumption (WUE = 0.0 L/kWh).
Open mechanical blind-mate quick-disconnect dripless coupling CAD specifications, pressure-drop correlations, and Redfish telemetry schemas have been published to the OCP Cooling Environments repository. Datacenter engineers can download the thermal test reports on the Open Compute Project Advanced Cooling Portal.
Sources
Comments (0)
No comments yet. Be the first to share a thought on this story.