Operating Temperatures to Support Performance, Energy Efficiency, and Heat Recovery

Author: Mark Steinke, Wade Vinson, and Bahareh Eslami, NVIDIA Collection: 2025 OCP Global Summit - Data Center Facility

Summary

NVIDIA engineers Mark Steinke and Wade Vinson present an analysis of how facility water supply (FWS) temperature selection affects cooling energy consumption, cooling power demand, and -- critically -- how much compute capacity can be deployed at gigawatt-scale AI factory sites. The study models 1 GW IT heat loads across three U.S. cities (Atlanta, Abilene, Phoenix) using ASHRAE hourly weather bin data, evaluating six heat rejection technologies: dry coolers, evaporative coolers, dry coolers with trim chillers, air free cooled chillers, water-cooled chillers, and air-cooled chillers. The analysis sweeps FWS set points from 40 degrees C down to 25 degrees C, charting the resulting energy (GWh) and power (MW) profiles for both dry cooler and trim chiller components at each temperature and geography.

The central finding is the concept of "chiller power sloshing" -- the idea that mechanical cooling power budgeted for worst-case conditions sits stranded for the vast majority of the year. In the Atlanta example at FWS 25 degrees C, the worst-case cooling power demand is 184 MW, yet approximately 100 MW of that capacity goes unused for 62% of the year (5,422 hours). NVIDIA proposes that this stranded mechanical power should be allocated to additional IT compute during off-peak cooling hours, effectively converting otherwise wasted power headroom into additional AI inference tokens. The servers themselves can throttle back performance by a trivial amount during the few hottest hours per year, making the tradeoff highly favorable. NVIDIA currently specifies a 45 degrees C TCS (thermal control setpoint) for its Blackwell/Rubin GPU products, which enables dry-cooler-only operation in many climates, but recommends that operators design sites for lower temperatures (e.g., 30 degrees C FWS) to future-proof against potential needs from next-generation silicon.