Data Centre Cooling for Vera Rubin (the new Nvidia data CPU+GPU)
- Mar 27
- 3 min read
Updated: Apr 30
NVIDIA’s Vera Rubin platform shifts data center cooling toward high‑temperature, fully liquid systems that can often run without traditional chillers, using warm‑water direct‑to‑chip loops and dry (or hybrid) coolers instead of big mechanical plants. And when we say warm water, we mean 45 °C flow, 65 °C return. As anyone in the HVAC game knows, that's wild - suddenly it becomes a whole bunch easier to cool (you can achieve 65 > 45 °C pretty easily with free cooling), and at the same you could have an entire district heating circuit straight off the other side of it. It looks likely to be a game changer in data centre energy efficiency.
Key thermal characteristics
The liquid loop is 100% responsible for removing IT heat; forced air is now mainly for secondary components and room mixing, not primary chip cooling.
Coolant supply is around 45 °C with return temperatures up to about 65 °C, so the system operates as “warm water” single‑phase direct liquid cooling (DLC).
To compensate for tighter temperature differentials at these higher setpoints, Rubin increases liquid flow rates to achieve almost double the effective rack‑level thermal performance vs Blackwell.
Rack and loop-level cooling
At rack level, Rubin servers use cold plates or embedded liquid modules on GPUs, CPUs and high‑power networking, plumbed into a closed liquid loop inside the rack (NVL72 and similar trays). Warm water enters at about 45 °C, passes through the cold plates, and exits around 55–65 °C depending on design delta‑T and flow. This loop connects to a CDU (cooling distribution unit) or manifold that interfaces with the facility’s secondary water loop at similar temperatures and pressures aligned with OCP MGX specifications (up to 5 Bar operating pressure).
Because the rack is fully liquid‑cooled, power densities per rack are expected to exceed 200 kW and push toward 600 kW in future Rubin deployments, far beyond what any practical air‑only system can support. This concentration of heat makes leak detection, materials compatibility, and redundant pumping within each loop critical design elements, but it also means much less heat is dumped into the white space air volume.
Facility-side cooling and “no chillers”
Vera Rubin’s big claim is that at a 45 °C supply temperature you can reject heat using ambient air in many climates, without mechanical chillers. In moderate regions (for example, North Virginia or New Jersey), analysts estimate that raising the water temperature window by ~10 °C can cut spend on heat‑rejection components by roughly one‑third, because dry coolers and simple fluid coolers can run in “free cooling” mode more of the year.
In practice, facility cooling for Rubin looks like this in cooler climates:
Warm‑water secondary loop from the racks to outdoor dry coolers or hybrid coolers.
Little or no compressor‑based chilling except for rare hot‑day conditions, redundancy, or mixed‑load halls.
Much lower dependence on evaporative cooling, which reduces site potable water use and alleviates local sustainability concerns.
However, in hotter markets (e.g. Mumbai, Melbourne design days), the ambient wet‑bulb and dry‑bulb temperatures can get close to or above the 45 °C inlet requirement once you include approach temperatures and pump/heat‑exchanger losses. In those locations, engineers expect that at least some mechanical chilling will still be required at peak conditions, meaning Rubin reduces chiller hours rather than eliminating chillers entirely.
Example table: cooling stack vs older AI racks
Aspect | Grace Blackwell era | Vera Rubin era |
Typical rack power | ~120–130 kW per high‑end rack | >200 kW, roadmap toward ~600 kW racks |
Primary chip cooling | Mix of air + DLC/immersion | Fully direct liquid to chip |
Supply water temp | ~35 °C–40 °C typical | 45 °C warm water standard |
Return water ceiling | ~60–65 °C | Up to ~65 °C expected |
Facility heat rejection | Chillers + evap + some dry | Dry/hybrid coolers dominant in many sites |
Chiller dependence | High | Reduced or zero in mild climates |
Impact on data center design
Rubin‑class loads force a shift from legacy 5–15 kW air‑cooled racks to purpose‑built “AI halls” or retrofitted spaces with liquid distribution, CDUs, and much higher rack densities. Existing sites will often add manifolded warm‑water loops, isolate Rubin pods with tailored heat‑rejection gear, and leave the traditional chilled‑water CRAH/CRAC systems to support the rest of the IT load.
Hyperscalers are already using digital‑twin tools (for example, Omniverse‑based DC design) to model chip‑to‑campus thermals, pump energy, and free‑cooling hours before deploying Rubin at scale. Analysts expect the combination of high‑temperature DLC and simplified plant to cut overall facility cooling power by on the order of 30–40% in favourable climates, significantly improving PUE for AI‑heavy sites.
