SNU wants to harness the power of AI, but energy limits mean lights out for many labs

At one of Korea's most prestigious universities, faculty members secured GPUs, a commodity essential to harnessing the power of AI.

There's just one problem: They don't have the power to run them.

A Seoul National University (SNU) computer engineering professor spent hundreds of millions of won on GPU servers for AI research last year, but they've sat idle as the server room hit its power limit.

The same energy crunch hit a professor in SNU’s Department of Naval Architecture and Ocean Engineering this year.

“I had the equipment, but there wasn’t enough power in the server room to run it,” the ocean engineering professor said. “I waited for months before I could finally get it installed.”

The growing power demands of AI research are straining university laboratories, many of which lack the electrical capacity to accommodate additional computing equipment.

SNU's energy consumption has risen for five consecutive years, climbing 22.9 percent from 54,788 tonnes of oil equivalent in 2021 to 67,319 tonnes last year.

Electricity use jumped 27.4 percent over the same period, from 193,623 megawatt-hours to 246,686 megawatt-hours. But infrastructure development has failed to keep pace with the steady rise in consumption.

In 2022, SNU unveiled plans for a 240-megawatt data center at its Siheung Campus in Gyeonggi, where it would centrally manage its GPUs.

A dedicated data center would ease the strain on the campus power supply. But construction has yet to begin after Kakao withdrew from the partnership and consultations with residents dragged on.

Server rooms at SNU’s College of Engineering are already at maximum energy consumption as AI research drives up electricity demand.

“SNU as a whole isn't short on electricity, but each building can only receive a fixed amount through its transformers, so server rooms in some buildings can run short of power,” said Lee Gyu-sub, a professor at SNU Electric Power Research Institute. “Adding transformers and expanding the power grid requires major construction, so increasing the power supply to an existing building is virtually impossible.”

With power capacity already maxed out, labs seeking to install new servers are finding it nearly impossible to gain access.

“For someone new to get server space, someone else has to give theirs up, and no space has opened up for more than a year,” another professor at SNU’s College of Engineering said. “New professors can end up waiting until another professor retires.”

The slot shortage has frustrated newer faculty members, a professor at the college who is a semiconductor expert said.

“New professors are complaining, ‘There should be some power left for other labs. If others take up all the capacity first, where are we supposed to put our servers?’” the professor said.

Now, faculty members compete for a limited space. The college allocates openings on a first-come, first-served basis, and more than 20 labs are now wait-listed.

Unable to add more servers on campus due to limited power and spatial capacity, some SNU labs are turning to cloud services from companies such as Amazon and Microsoft. But that comes at a steep price.

“When we urgently need GPUs or have to run a large-scale experiment, we use external cloud services, which cost about 10 times more than running our own servers,” another SNU computer engineering professor said. “We can’t afford to rely on the external cloud all the time, so we use less powerful GPUs or fewer of them to stay within our power limits.”

Many SNU engineering labs are being squeezed on two fronts: too little power and too few GPUs. A sharp rise in GPU prices has left research budgets unable to stretch far enough to buy all the equipment they need.

“Prices for the latest GPUs have risen more than fivefold in a year,” said another professor at the engineering college. “Some labs have GPUs but not enough electricity to run them. Others can’t buy GPUs even if they have money.”

University bureaucracy can make matters worse. GPU prices can rise during the months it takes to get a research plan approved. It often leaves the original budget too small by the time equipment can be bid on.

“Even if we budget enough at the outset, GPU prices can rise during the three or four months it takes to get approval,” the professor said. “By then, the budget may no longer cover the purchase and the tender can fail.”

The shortages are also raising a bigger concern: a brain drain.

As power and GPUs become increasingly important to AI research, universities that lack them risk losing researchers to better-equipped companies and overseas universities.

“Many researchers are drawn to overseas universities and companies that have the infrastructure to support the work they want to do,” another professor said. “One professor who studied algorithms left for a U.S. university last year for that reason and has since joined OpenAI.”

Between 2021 and May of last year, 12 SNU engineering professors left for overseas schools, according to the university.

Students are also gravitating toward companies with better research infrastructure rather than staying at universities for graduate study.

“Students seem increasingly resigned to the idea that there is only so far they can take their research at Korean universities,” said the professor, who noted a growing inclination among colleagues to go overseas. Many master’s and doctoral students are also interested in internships at overseas companies, the professor added.

Yet the Korean school noted that the figure combines full-time and part-time faculty.

“Departures among full-time professors were significantly fewer,” SNU said. “Of the nine new professors who joined us this year, four came from overseas.”

Academics say the stakes extend beyond SNU. As AI infrastructure becomes increasingly critical to global technological competition, Korea needs a strategy to ensure universities have computing resources to compete.

“Better GPUs produce better research results,” the chip expert professor at the College of Engineering said. “Korea needs a national initiative to provide universities at the forefront of research with affordable GPUs and enough power to run them.”

Calls are also growing for faster investment in data centers and other infrastructure. Professors say the current model, in which individual labs and buildings buy and operate their own servers, has reached its limits. They argue that universities or the government should instead manage shared pools of GPUs.

“We need to move quickly to secure data center sites and the staff needed to operate them,” said the professor who earlier noted that many researchers want to move to overseas universities or companies. “We need to expand research infrastructure as soon as possible.”

Others argue that efficiency should come first. Before adding more power and equipment, universities could make better use of existing resources by sharing GPUs across labs.

“If we cut back on GPUs dedicated to individual labs and increase the number shared across the university, we could use the equipment more efficiently within the power we already have,” the professor said.

BY LEE GYU-RIM [lee.soojung1@joongang.co.kkr]