The rise of artificial intelligence is driven by massive data center expansion. Data centers are projected to account for 9 to 17 percent of total U.S. electricity usage by the end of 2020. Currently, about a third of data center power is spent cooling the chips that run AI models.
That’s the process Ferveret is working to streamline. The startup, founded by Reza Azizian, a former MIT nuclear engineering postdoctoral fellow, and Matteo Bucci, MIT’s Esther and Harold E. Edgerton Associate Professor, takes a nuclear reactor approach to cooling chips without water and with significantly less power.
The company’s cooling system bathes computer servers in a special liquid that absorbs heat much more efficiently than air from fans. What makes this solution different from other liquid cooling systems is the air bubbles. Ferveret’s Adaptive Phase Cooling (APC) solution creates very small air bubbles on the surface of the server that slough off more frequently to accelerate the heat transfer process.
Ferveret is already testing solutions with companies such as data center development and operator company CleanSpark, AI accelerator company FuriosaAI, and Switch, one of the largest data center operators in the United States.
In a recent study conducted in collaboration with the Samueli School of Computer Science at the University of California, Los Angeles, Ferveret found that its APC solution improved computing power efficiency by 15% compared to state-of-the-art liquid cooling solutions. By combining these savings with Ferveret’s power control system to optimize operating conditions, data centers can capture 35% more tokens (small text or data) from AI models with the same amount of power, the company says.
“Our goal is to make data centers as sustainable as possible and ensure that they can use every watt of power to generate the most useful output: tokens,” Azizian says. “Our system enables more powerful chips to operate, significantly reduces data center energy waste, and achieves all this while consuming zero water.”
From nuclear reactors to AI
Azizian was a postdoctoral fellow at MIT in 2013 when he met Bucci, then a research scientist. They worked on heat transfer in nuclear reactors before Azizian entered the industry, where they shifted their focus to chip cooling. Azizian first worked on Microsoft’s HoloLens augmented reality headset before joining Nvidia, which makes graphics processing units that companies use to train and run modern AI models. Meanwhile, Bucci continued his research at MIT, becoming an assistant professor in 2016.
When Azizian entered his first data center in 2017, he was struck by the huge, noisy fans that filled the building with cooling.
“I thought, ‘Oh my god, this is not the way to cool a facility,'” Azizian recalled, noting that air cooling can still account for 40% of the power going to a data center. “It wasn’t an efficient method, but it didn’t hurt performance so no one cared that the cooling technology was 50 years old.”
Azizian began talking with Butch about applying his knowledge of optimizing heat transfer in nuclear reactors to data centers. Scientists have spent decades finding better ways to move heat within nuclear reactors.
“Heat transfer determines how much energy can be extracted from the core, which directly translates into revenue,” Azizian explains.
The founders started Ferveret in 2021. A lot has changed since Azizian entered our first data center. The explosion in artificial intelligence has placed an emphasis on squeezing as much computing power as possible from limited power sources, and chip companies are cramming more and more components into their chips.
This has led data center operators to use liquids to cool chips. A technique called immersion cooling, in which the chip is submerged in liquid, is often used. The most effective form of immersion cooling involves boiling the liquid.
“Liquids are better heat transfer media than air, so if you stick your hand in room-temperature water, it still feels cold,” explains Bucci. “When the liquid boils, the heat removal becomes even more effective because the phase change requires a lot of energy. That energy is removed from the chip. This allows us to transfer a large amount of heat while minimizing the temperature difference between the chip and the liquid.”
Unfortunately, once the liquid boils, the system is complicated by the need for operators to capture and reliquefy air bubbles while controlling pressure, temperature, and liquid inventory.
Ferbele’s system is an adaptation of a nuclear reactor process called subcooled boiling. It uses low-boiling liquids and does not use the toxic PFAS “forever chemicals” that other approaches rely on. At the surface of the chip, Felverette’s liquid produces smaller air bubbles than other immersion cooling approaches. These bubbles detach more frequently and rapidly recondense in the surrounding liquid, accelerating the bubble rewetting cycle at the chip surface and enhancing heat transfer.
Ferveret offers APC systems in small boxes, each housing one server. The founders say the modular system makes the system easier to deploy and simplifies maintenance.
“Thanks to physics, we can now form elements that were not possible in the past,” Azizian says. “Most immersion cooling solutions are big tanks that people submerge their servers in. We have a small, modular rackmount solution that can adapt to current infrastructure, making it easier for people to adopt our technology.”
Ferveret also provides control software that adjusts the power delivered to each server in real time to further improve efficiency.
“We offer full-stack systems that include cooling boxes, racks, cooling distribution units, and sensors that measure temperature and pressure,” says Bucci. “Our software monitors these sensors and optimizes the operating conditions within each box to ensure that energy consumption within the system is minimized.”
AI with fewer resources
Ferveret not only helps data centers operate more efficiently, but also improves sustainability by making it easier to operate data centers in remote locations with more renewable energy.
“The sun shines even in places where there’s not a lot of water, so the advantage of not using water is that you can build data centers in places where there’s solar energy but nothing to cool the data center,” says Bucci. “This technology will help us bring data centers to regions where we wouldn’t normally have the resources to deploy data centers, such as Africa, the Middle East, and of course parts of the Americas. This is a huge liberation.”
Ferveret is in talks with major cloud computing companies known as hyperscalers and is currently participating in Nvidia’s Inception program for startups. The company plans to announce expanded partnerships later this year. From there, the founders plan to quickly scale the technology to help the AI industry continue to grow without further burdening the planet.
“The computing industry faces huge challenges with access to electricity, and in many regions, access to water,” Azizian said. “As the industry grows, those limits will become even more restrictive. The main goal for these data center operators will be to get more tokens out of their power. We have shown that we can do that.”
