AI is pushing chips to run hotter than ever. Next-generation GPUs are expected to dissipate 2,000 watts each—roughly the heat of a space heater, concentrated on a chip smaller than your palm. Traditional air cooling can’t keep up, and the usual backup plan, compressor-based chillers, adds substantially to a data center’s energy bill. Evaporative cooling saves energy, but at the price of deepening a data center’s water footprint.
Researchers at UC Davis’s Western Cooling Efficiency Center (WCEC) have designed a modular data center for AI computing that packs 1.5 megawatts of computing power into the volume of a standard 40-foot shipping container—and keeps it cool without energy-intensive compressor-based systems. Better still, the sealed-loop design rejects heat directly to the air, consuming no water at all.
The system pulls heat directly off the chips using a custom copper cold plate built from thousands of microscopic pins—3D-printed using electrochemical additive manufacturing—and rejects it to the outside air through a novel WCEC-designed microchannel polymer heat exchanger (MPHX).
To prove the concept, WCEC built and tested lab scale MPHX prototypes, where measured heat transfer deviated from model predictions by just 5.7% on average in the relevant operating range. System-level modeling built on these validated components shows the design keeps GPUs at safe temperatures even on a 104 °F day, with cooling consuming just 2% of the computing power—and under 1% in best-case scenarios.
Read more in ASME’s Journal of Electronic Packaging.