University of Houston engineers find more efficient way to cool semiconductors for AI data centers

University of Houston engineers have discovered a way to cool semiconductors and release heat three times more efficiently  - a breakthrough they say could cut energy and water usage as new AI data centers spread across Texas. 

Hadi Ghasemi is among researchers at the University of Houston who are trying to tackle the energy "trilemma": the challenge to make energy affordable, reliable and sustainable. Growing public pushback against data centers has largely centered around affordability and sustainability, as the facilities are expected to drive major electricity demand in the coming years. The data centers in Texas are also driving a power plant boom.

"We can have a growth of the AI, but with a sustainable approach and less energy-intensive approach," said Ghasemi, a mechanical and aerospace engineering professor. "It's really important, especially for Houston, because we are the energy city of the world and we are supplying the energy for all of these data centers."

Higher ed: Texas A&M's semiconductor research fuels U.S.-made production ambitions

Ghasemi and his doctoral candidates, Amirmohammad Jahanbakhsh and Saber Badkoobeh Hezaveh, found a microscopic solution to the problem.

One common cooling method passes water through structures that sit atop semiconductors. They're about one centimeter across, and dozens of them fit on a single semiconductor wafer. Sometimes hundreds of thousands of the materials are needed for one data center alone.  

Many of the structures have vertical pillars that serve as a conduit to evaporate heat, though the pillars are in different shapes. Ghasemi said several liters of water run over the material per minute; they need to work quickly or else the semiconductors melt. (He said that a single wafer can generate 10,000 times the heat of the sun.)

Researchers have been using trial and error to find the best and most efficient designs for the materials. His team used AI to investigate: Of all the possible models, structures with branched or tree-like shapes cooled three times more efficiently. 

"We have so many examples of this structure in the nature, for example, in the lungs, in the rivers," Ghasemi said. 

The underlying problem is a paradox. Semiconductor chips - used in everyday electronic devices, military technologies and AI computing systems - have increased in efficiency and computing power as they become smaller. But that means they run more often, generating more heat. 

As a result, cooling systems have needed to increase in capacity, which has also increased their water and energy usage, Ghasemi said.

Ramanan Krishnamoorti, UH Vice President of Energy and Innovation, said that the U.S. is looking at a 30 to 50% increase in electrification over the next 10 years, and in Texas, it could be around 35%. Consumers can expect to face higher costs as a result of rising demand, meaning some will have to make hard choices about their electricity use, Krishnamoorti said. 

Aparajita Datta, an energy policy associate, said that the problem could affect some socioeconomic classes more.

"Across the university, not only are we thinking about newer and better technologies to help that whole process, we're also thinking about how do we bring the costs down," she said. 

Ghasemi began his research around 2020 and still continues, though he hopes that it will eventually be put into use at data centers. A $1 million grant from the U.S. Office of Naval Research is helping fund the project, which was published in the International Journal of Heat and Mass Transfer.

"Data centers will be everywhere in the future," Ghasemi said. "If we could make them more sustainable, it's definitely in the benefit of everyone in the society."

Subscribe

Read this story on houstonchronicle.com