[Reading Science] 'The Excavator of the AI Era' Visiting the GIST Supercomputing Center

NVIDIA GPU Supercomputer Secured in Advance
Academia and Companies Line Up for AI Research
Center Director Kim Jong-won: "Expanding Shared Supercomputer Centers Is Essential for International AI Competitiveness"

The Supercomputing Center at Gwangju Institute of Science and Technology (GIST) is considered a sacred place not only for the Jeollanam-do region but also for AI researchers across South Korea. Before the National AI Data Center, located directly across the street, was established, domestic academia had to visit this center to conduct AI research using NVIDIA’s GPU-based supercomputers. Although the national supercomputer No. 5, ‘Nurion,’ has become the top supercomputer in the public sector in Korea, it is not a GPU-based supercomputer required for the AI era. Therefore, GIST has been like “a drop of rain in a drought” for domestic AI researchers.

Kim Jong-won, Dean of the GIST AI Graduate School (left), is explaining the 'DreamAI' supercomputer to reporters on the 22nd. DreamAI boasts the best performance among supercomputers owned by domestic universities. Photo by GIST

On the 22nd, guided by Kim Jong-won, Dean of the GIST AI Graduate School and Director of the Supercomputing Center, I visited the supercomputer data center located on the first floor of the graduate school building. At the entrance, promotional materials boasting the performance of the center’s computers were displayed. ‘Dream AI.’ In Korean, it means “a dreaming AI.” “Ranked 178th in the TOP500 as of November 2022.” The TOP500 ranking is compiled twice a year at the German Supercomputing Conference, ranking supercomputers worldwide. Dream AI’s ranking this year is 275th. When it first appeared, it ranked 6th domestically, and although two years have passed, it still holds the highest rank among computers owned by domestic academia.

Supercomputers are essential for scientific research. Although Dream AI’s ranking may seem low, it has GPUs, giving it an advantage in supporting AI computations. Dream AI houses a total of 320 NVIDIA A100 GPUs. Because of this, researchers who need GPU-based computations still come here. A research official lamented, “Even if you want to rent Dream AI, you have to wait a long time.” Considering the need for learning, inference, and acceleration functions through GPUs in the AI era, Dream AI is highly useful. All supercomputers secured by companies such as Naver, Kakao, Samsung Electronics, and KT include GPUs. Academia, public research institutes, and companies line up to conduct research through Dream AI. This is why Dream AI is a shared supercomputer.

An information guide on the TOP500 ranking of GIST's supercomputer 'DreamAI'. Photo by Baek Jongmin

Dream AI is arranged in a separate space with computers and storage units facing each other across a corridor. Standing in the computer room, cold air blows up from the floor and hits the face. This air cools the heat emitted by the computers and circulates outside through pipes for management. Director Kim said he pays special attention to temperature and power management. Momentary power outages due to power issues sometimes occur. In such cases, an uninterruptible power supply (UPS) operates.

Director Kim explained, “We also have a monitoring system to safely manage the computers.” The noise from the cooling fans is so loud that it is hard to hear Director Kim’s voice. Measuring the noise with a smartwatch showed 82?85 dB. This level is comparable to subway noise near railroad tracks. Working for long hours in such an environment poses a risk of hearing loss. Besides Dream AI, the center also houses servers and computers entrusted by companies.

The center also has NVIDIA-manufactured computers such as ‘DGXA100’ and ‘DGX-1V.’ It is a point of pride for the center to own DG computers equipped with eight A100 GPUs each. Director Kim said, “There are not many places that own systems directly manufactured by NVIDIA.” In 2019, he also signed a Memorandum of Understanding (MOU) for cooperation between NVIDIA and GIST. As a result, GIST has extended its collaboration with NVIDIA to education. On the day after the reporter’s visit to the Supercomputing Center, on the 23rd, NVIDIA experts delivered the latest deep learning model technologies and methodologies using Dream AI’s GPUs to GIST students as well as external participants.

An NVIDIA Korea official is giving a lecture on AI research using GPUs at GIST on the 23rd. Photo by GIST

◇ “Shovels and hoes won’t do... Excavator-level support is needed” = Director Kim repeatedly emphasized the ‘excavator’ theory for AI development. He argued that equipment like the shared supercomputer Dream AI must be widely distributed to solidify the domestic AI foundation. This is why he was obsessed early on with establishing a GPU-based supercomputing center.

Director Kim expressed concern, saying, “Most domestic AI research and development and services are carried out through small-scale computing infrastructure. This does not align with the global trend where countries competitively introduce ultra-large supercomputers to strengthen AI research.” He explained that instead of each lab installing and managing small-scale computers for research, securing high-performance computers for shared use is necessary to catch up. Preferring private cars means you cannot match the transportation capacity of buses. He added, “With Meta’s open-sourced LLaMA 3.1 having 405 billion parameters, fine-tuning it is impossible with small-scale equipment.”

Director Kim’s argument was not wrong. Shortly after Dream AI began operation, ChatGPT was released, sparking the generative AI revolution. Since then, GPUs have become impossible to buy even with money. Overseas big tech companies also worry about not receiving NVIDIA GPUs on time. Thanks to Director Kim’s pioneering acquisition of GPU-based supercomputers, GIST plays the role of the ‘Autonomous Driving High-Performance Computing Specialized Center’ among the Ministry of Science and ICT’s top 10 specialized supercomputing centers. This aligns with the recent shift in autonomous driving technology from sensors to AI learning based on vision information. Director Kim emphasized that while it is important for Korea, a semiconductor-producing country, to develop AI computation chips, it is equally important to conduct research that enables AI utilization. For that, the role of AI supercomputing centers is crucial.

Director Kim also has concerns. The second expansion to fill the remaining space in the Supercomputing Center is delayed. Power supply is always a concern. Currently, about 50 million KRW per month is spent, but larger costs will certainly be needed in the future. Director Kim said, “We operate the center without support,” but expressed confidence that they can continue to play the role of the AI supercomputer vanguard.

GIST President Lim Ki-chul also said, “We promised to support universities in the Jeonnam region to use the supercomputing center’s resources at low cost,” and added, “GIST will actively take the lead in supporting domestic AI research.”