Nvidia was awarded $5 million from the United States Department of Energy (DOE) to develop a data center cooling solution that combines two-phase direct-to-chip and immersion cooling techniques using low-global warming potential (GWP) refrigerants and per- and polyfluorinated alkyl substances (PFAs) that comply with environmental regulations.

Nvidia's funding is part of the DOE's Cooling Operations Optimized for Leaps in Energy, Reliability and Carbon Hyperefficiency for Information Processing Systems (COOLERCHIPS) program, which has also awarded funds to projects at HP, Flexnode, Intel Federal, universities and research centers. Nvidia's award of $5 million represents the most amount of funds given to a project through the DOE program.

Given "the recent relevance of generative AI [artificial intelligence] and its impact on the data center," Nvidia's research project will be one to watch, Dell'Oro Research Director Lucas Beran told SDxCentral. Nvidia GPUs boast one of the highest thermal design profiles (TDPs) across the industry, which represents how hot those chips' temperatures can get. Because the company is "at the forefront of those semiconductor TDPs, I'm definitely interested in what they're doing," Beran added.

Data centers produce a lot of heat and, traditionally, that problem is solved with air cooling – an array of fans circulates air and removes heat from processors. But that's far from the most efficient or sustainable way to cool down compute.

"With soaring demand for AI and [high-performance computing] applications driving exponential increases in computing power, this supply-demand mismatch means data centers have no choice but to become more efficient," Nvidia Senior Director of Data Center Engineering and Operations Jeremy Rodriguez told SDxCentral.

Nvidia's COOLERCHIPS research is unique for its deployment of two technologies never before combined: direct-to-chip (DTC) liquid cooling and immersion cooling. With DTC, a cold-plate made of a metal, like copper, makes direct contact with hot chips to transfer heat to a water-based fluid, which recirculates and removes heat from the data center. Immersion cooling constitutes a server bathing in a non-conductive fluid, which can transfer heat from the data center similarly to cold-plates.

By bringing "together all of the pieces of liquid cooling technology, ... that makes it unique," Beran noted.

Beran explained that GPUs and CPUs produce more heat than other data center components, and therefore require more targeted heat removal. Single-phase immersion cooling, however, "may struggle to do that with some of the higher TDP chips." To cool each component in the most logical way, Nvidia is "applying immersion cooling to some of the other components that still generate heat, but maybe not as much as the CPUs or GPUs, and then applying a two-phase direct liquid cooling solution to that CPU or GPU specifically."

"Using refrigerants is not common in today's data centers," Rodriguez pointed out. Nvidia's proposed solution, in contrast to air-cooled servers, "eliminates all fans while removing over 4kW of heat for each 1RU (1.75″ height, typical server height) server," he touted.

The project plans to use two different types of coolants: a "green refrigerant" with a GWP less than 10 — such as R1234yf for DTC — and PFAs-compliant dielectrics like Opticool-H for immersion fluid, Rodriguez said. Since it's also important to consider the impact of the liquids used in data center cooling, he noted Opticool-H biodegrades quicker and transfers heat more successfully than other dielectric coolants.

With the COOLERCHIPS project, a "wide net has been cast, and I think that's good," Beran said. "One thing that the data center industry knows is that we're going to use liquid. What they don't know is the form factor and the type of liquid that will need to be used and how that will change over time." This exploration of cooling technologies "is exactly what the industry needs to start to learn what works and what doesn't work," he said.