Panel: Coolant Temperatures for Durable Data Center Designs

Author: Chris Malone (Meta, Data Center Technology Strategy), Steve Wills (OCP Steering Committee), David Haley (Nvidia, Senior Thermal Engineering Manager), Shri Damaraju (Intel, Fellow System Design), Paul Artman (AMD, Fellow System Design), My Truong (Field CTO, Equinix), Kelsey Wildshire (Principal Infrastructure Architect, Microsoft) Collection: 2023 Global Summit - Cooling Environments

Summary

This panel presented and debated an industry-wide recommendation to standardize liquid coolant supply temperature at 30°C for direct-to-chip (cold plate) liquid cooling in AI data centers. The initiative, originated by Meta's Chris Malone at OCP two years prior, brought together seven organizations — Meta, Microsoft, AMD, Samsung, Intel, Nvidia, and Equinix — to align on a temperature setpoint that balances competing pressures: data center efficiency (which favors higher temperatures), chip thermal headroom (which requires lower temperatures as GPU power escalates past 1,000W+), air-cooling compatibility, and 15-year data center asset durability.

The fundamental problem driving the initiative is the collision of two opposing trends: GPU module power is rising sharply (toward 1,500–1,600W for current AMD and Nvidia high-end GPUs, and projected to continue rising), while the 30°C supply fluid temperature is approaching the thermal ceiling that higher-power chips will require. The panel's power trend charts showed GPU power as a green line increasing steeply from 2020 to 2032 while the red CPU power line rises more gradually; the right chart showed HBM (High Bandwidth Memory) stack counts growing from 4–8 stacks to 12 and then 16 stacks by the early 2030s, each generation increasing cooling difficulty.

Data centers are 15-year capital assets requiring 2–4 years to plan and build; once a coolant temperature design point is chosen, changing it requires costly retrofits. The panel quantified the efficiency penalty of going below 30°C: every 4°C reduction in coolant temperature increases chiller energy consumption by approximately 20%. The panel called 30°C the "equally unhappy" balance point — data center operators would prefer higher temperatures for efficiency, silicon suppliers need 30°C as a floor to support 1,500W+ GPU modules, and facility operators need a consistent standard to avoid over-provisioning.

The 30°C target applies specifically to the secondary fluid loop — the liquid entering the server (not the facility-level chilled water). Primary-to-secondary approach temperature deltas of 3–5°C are typical. ASHRAE's TC99 handbook update is aligning with this recommendation, defining 30°C as the lowest explicit set point (with 50°C on the high end). OCP's faster publication cadence is seen as complementary to ASHRAE's broader industry standardization role.