Ferveret, a startup from the Massachusetts Institute of Technology (MIT), is using cooling technology deployed in nuclear reactors to solve challenges faced by data centers today. The waterless cooling technology can also help establish data centers in areas where water is scarce while abundant solar energy powers the chips and cooling infrastructure. 

The recent surge in artificial intelligence (AI) applications has been made possible by high-performance computing chips from companies like NVIDIA. When AI needs to be deployed at scale, companies need to house many chips capable of processing information quickly enough. To save costs, they hire computing power from private data centers that buy and deploy the high-performance chips and rent them out over the long term. 

As more and more companies deploy AI in their workflows, the demand for data centers has shot up. With new data centers, energy demand is also growing. Estimates suggest that by the end of the decade, up to 17 percent of the US electricity usage will be consumed by data centers. The power consumption is not just for high-performance chips but also for cooling them. 

Liquid cooling chips

When data centers were first set up, chips were cooled using air. As their computing capacities increased, they began to generate more heat, which needed better cooling approaches. 

Liquids are better than air at removing heat from a chip. Their performance is even better when the heat is used to bring about a phase change. At the liquid’s boiling point, it can extract more heat from the chip without increasing the temperature difference between the liquid and the chip. 

The newest data centers use an immersion cooling approach in which the entire server rack is immersed in the cooling liquid to maximize heat transfer. However, with AI adding more computational power to the server pack, heat generation has increased, necessitating newer cooling solutions. 

Nuclear reactor-inspired tech

MIT spin-off Ferveret is looking to deploy cooling technology used in nuclear reactors and adapt it for data center use. Called subcooled boiling, it uses a liquid with a low boiling point, which produces smaller bubbles than immersion cooling approaches. 

While immersion cooling has to contend with issues such as capturing the bubbles and reliquefying them under constrained pressure and temperature conditions, in Ferveret’s approach, the bubbles recondense quickly into the surrounding liquid. 

In a recent study, Ferveret found that its cooling solution resulted in up to 15 percent improvement in computational power efficiency compared to state-of-the-art immersive cooling techniques. This, combined with Fereveret’s proprietary power control system, allows data centers to extract 35 percent more tokens from the AI models without any increase in power usage. 

Ferveret’s Adaptive Phase Cooling (APC) system is deployed in small boxes, each capable of housing one server. This modular approach allows its deploy and maintain its system much more easily. 

“The physics enable us to get to form factors that weren’t possible in the past,” said Reza Azizian, co-founder at Ferveret, in a press release. “Most immersion cooling solutions are large tanks that people submerge the servers in. We have a smaller, modular rack-mounted solution that makes it adaptable to the current infrastructure, so it’s easier for people to deploy our technology.”

The waterless cooling approach is also beneficial in areas where water is scarce, but sunlight is plentiful. Setting up solar power plants in these areas is simple, but with low water, data centers could not be cooled down. 

“This technology can help deploy data centers in regions where normally you wouldn’t have the resources to do so, including Africa, the Middle East, and of course parts of America. It’s a huge unlock,” added Matteo Bucci, associate professor at the Department of Nuclear Science and Engineering and also Ferveret co-founder.