Meet the Researchers Tackling AI’s Heat Problem

Data centers have a heat problem. For Thomas Schutzius, associate professor of mechanical engineering at UC Berkeley, that problem presents an opportunity.

“We have this massive buildup that’s happening,” Schutzius tells me of the data center construction that began around the start of publicly accessible AI. “There’s [a] need to measure what’s happening.”

“Presumably, these [data centers] are going to get sold. And if they’re going to get sold, you probably want to know the value of your performance.” Schutzius goes on, “There’s an opportunity not just in the sale, but also on the operational side: if you know how efficiently things are running, you might be able to make it even more efficient.”

The hope to use engineering to design sustainable systems was what prompted Schutzius to pursue this kind of work in the first place. “What drew me to [thermal sciences was] energy. We have a lot of opportunities to make…sustainability impacts there,” he recalls.

As data centers become more important and more costly, Schutzius argues that there is a growing need to know exactly how they perform. That’s where Schutzius’s CDA-funded research project comes in: documenting the data outputted by these systems and using it to create a more sustainable future.

“Tracking things and having good data that an operator can go back and see…is done to some extent, but it’s not really done to the full potential,” Schutzius tells me. “If you have that data…you can make these things run a lot more efficient and sustainable, and we can put a value to that now.”

Schutzius’s research, which he conducts with PhD students Yash Lolla and Terry Ma, centers on a physical model of a data center that the team is using to test a different approach to cooling, documenting the data via digital twins and blockchain.

“We would like to build a [system] where we can track everything in terms of the power output, the flow rate of the water, the amount of heat being absorbed, and basically put it all in blockchain so that it can’t be changed,” Lolla tells me.

Cooling Differently

Instead of relying on conventional air cooling, the team’s data center model is hooked up to a liquid cooling system. Liquid cooling, according to Schutzius, is the best way to retain the massive amounts of energy already being pumped into data centers and expelled as heat.

“When you start to make [data centers] bigger and more powerful,” Schutzius tells me, “at some point you can’t cool with air…and you cross over into liquid cooling.”

“If you dump heat into [liquid] you don’t lose the value of the heat,” Schutzius says. “If you have this data…then we could know this for the future and make it more sustainable.”

“We’re able to use the energy more responsibly [with liquid cooling]. Not just: I take this, and I have to cool it, and then it’s useless to me—it can be recovered.”

One potential application of Schutzius’s work is building heating. “Most of the world’s energy consumption is basically buildings, especially building heat,” Schutzius says. Schutzius sees this use case as particularly promising for the distributed computing systems that will eventually make up the majority of AI infrastructure.

Data Centers & Digital Twins

Capturing the heat is only part of Schutzius’s project, however. The team also wants to understand exactly what is happening inside the system, in real time.

“We have the actual hardware [model] that we run, and then…you have this twin that you run alongside,” Schutzius says. “The twins generally give you more insight into the behavior [of the hardware], and then also future opportunities.”

“For a given temperature for a given flow rate for a given pressure, if we feel that this is not the optimal output, temperature, pressure, or flow rate, then there is no point in us doing that experiment. So, what we are doing with the digital twin is basically making informed decisions before we do the experiments,” Ma explains.

The final piece of Schutzius’s research is making the data history of these models verifiable via blockchain. This information, Schutzius tells me, is useful for buyers, sellers, and operators.

“Everyone benefits,” Schutzius says. “An operator benefits because you have records that don’t go away and cannot be manipulated. Someone who’s looking to acquire now also has…a full report on how this was done: Was it really hard? Was this well-maintained? And then, if one chose to report statistics about things like power, things like water, you have verifiable records of that.”

Modeling the Future

Schutzius sees the data center as a starting point rather than the endpoint of his work.

“Any sort of energy system would benefit from this type of framework,” Schutzius says. “There is the part of the data center, but I would say any energy system, solar, or battery energy storage systems benefit [from] having real data that you can’t modify.”

To Schutzius, the impactful work he’s doing would not be possible anywhere but Berkeley. “This place could only exist at Berkeley, specifically with the support of the Center for Digital Assets. There’s a more holistic view…bringing together people from different domains into mechanical engineering. That’s quite special,” Schutzius says. “Fresh approaches [are] totally encouraged by the fact that we have a center that’s financially supportive and pushes us to do that.”

As AI infrastructure expands, Schutzius expects increasingly sophisticated computing systems to become part of the physical world around us. “You’re going to do a ton of build out now from nothing,” Schutzius says of the data center boom. “This is an opportunity to do something differently than was done before.”

“Ultimately, we will be able to use AI to make AI better,” Schutzius says, “but I think a lot of that comes down to…good data that [is] organized, so we can…really take advantage of the opportunities that AI presents.”