- Nvidia’s Rubin-generation AI servers run coolant that’s often hotter than a hot tub, as it enters a closed circuit at up to 45°C and exits at around 55°C, with no performance penalty in tow.
- The principle used here is a temperature gradient with server chips still running hotter than the coolant, allowing heat transfer to occur seamlessly
- The approach enables Nvidia to save 2.6 million gallons per megawatts annually to near zero and can save a 50MW site over $4 million a year in costs
Nvidia’s Rubin generation is the first of its kind in several ways, but the one that perhaps stands out is the fact that it is 100% liquid-cooled, implementing it as a core design feature at the platform level, with every single chip and network component covered by a closed loop.
The reference design from Nvidia aims to allow AI factories that effectively consume zero water by implementing a closed circuit that does not utilize evaporative water cooling 99% of the time.
The approach goes a step further by potentially eliminating industrial refrigeration from the equation by using liquid cooling, where the temperature is as much as 45 degrees Celsius for coolant entering the system and approximately 55 degrees Celsius as it leaves.
A numbers game for a data center that crushes them on an industrial level
Nvidia’s approach is not the first of its kind in terms of thermal design or implementation; However, IBM beat it by 16 years by using 60-degree coolant to cool its prototype supercomputer at ETH Zurich.
The premise was the same: a much lower carbon footprint, significantly lower power consumption and a loop that fills once and leaves alone. What Nvidia adds is scale—its Vera Rubin-based DSX design extends the approach across an entire facility rather than a single machine.
“The NVIDIA DSX reference design for AI factories has zero water consumption – we’ve eliminated huge amounts of power consumption and virtually all water consumption,” said Ali Heydari, Nvidia’s director of cooling and data center infrastructure.
The coolant is a 3:1 water-to-propylene-glycol mixture, and it easily pulls heat away from chips that run much hotter than the coolant itself.
The savings aren’t just water-based: Industrial refrigeration plants reduce their energy costs by about 4% for every degree they raise their operating temperature, and Nvidia estimates that a 50-megawatt hyperscale plant could save over $4 million annually in cooling-related energy and water costs by adopting its design.
With cooling historically accounting for as much as 40% of a data center’s total power bill, Nvidia’s approach is ambitious, but it’s one that should quickly pay for itself—eliminating server fans entirely while allowing coolers to run only where needed and at significantly lower power than a loop that requires colder coolant.
It should be noted that DSX is a reference design: a set of best practices for building an AI factory, not a tally of what the industry has actually built, where many designs still consume very large amounts of water, even as AI data centers continue to pop up in drought-stricken regions of the US.
Nvidia may appear to be championing a “greener” AI data center for the future, at least in its own designs, but that’s only part of the story — and the part it leaves out is a little less flattering. Nvidia didn’t adopt hot water cooling because it’s elegant, or green, or smart, although it is all three. It adopted it because its accelerators became so power-dense that it was no longer physically viable to move air across them, and once you’re committed to liquid cooling, running it hot is simply the most efficient way to do it.
The sustainability gain is real, and it’s a consequence of a thermal problem that Nvidia created for itself by building the most power-tight chips in the industry. Hot water is not the trick. It is the bill that is due, paid in the most efficient currency available.
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews and opinions in your feeds.
