Universities suffer lack of advanced chips for AI research

South Korean universities, at the forefront of artificial intelligence (AI) research, are grappling with a significant hurdle- the inability to conduct in-depth research on generative AI models due to the lack of access to the latest technology. This issue is not limited to smaller institutions; even major universities such as Seoul National University and KAIST are struggling to secure the necessary budget for AI chips despite being part of government-supported projects. The situation is further complicated by insufficient power supply to operate these chips within university facilities.

"To develop generative AI models such as OpenAI's 'Sora,' hundreds of the latest NVIDIA Graphics Processing Units (GPUs) are required," Professor A from KAIST said Thursday. "Even if we win government projects, it's impossible to purchase NVIDIA GPUs in multiple quantities, so we are conducting research using outdated gaming GPUs."

The financial implications of this issue are significant. To create a video-generative AI service like 'Sora,' universities need 'AI accelerators’ such as GPUs capable of performing multiple calculations simultaneously. These GPUs, the majority of which are supplied by NVIDIA, are highly valued globally. The latest GPU by NVIDIA, the H100, is known to cost around 50 million won per unit. Given these costs, It's no wonder that universities are struggling to keep up with the latest technology.

Even if chips are secured, the power issue has to be addressed. "Although we want to operate additional GPUs, the school administration replied that it's impossible to provide additional power supply," Professor Kim Geon-hee from Seoul National University's Department of Computer Science said. "Professors have to search for buildings with spare power capacity themselves."

Even with secured budgets, equipment purchases require deliberation. If research equipment exceeds 100 million won in national research and development (R&D) projects, it must undergo review by the National Research Facilities and Equipment Center (NFEC), which typically takes over three months for results. "Typically, at least eight NVIDIA GPUs are needed for AI research. The approximate cost of purchasing eight GPUs and servers alone is 500 million won. Even if we secure the budget, going through the review process results in further delays in actual research," said Kim Jong-won, dean of the Gwangju Institute of Science and Technology (GIST) AI Graduate School. "Universities need alternatives such as establishing research centers to collectively share AI chips."


전남혁 기자 forward@donga.com