[Reading Science] They Called It an 'IT Powerhouse'... Yet Can't Build a Supercomputer

National Supercomputer No.6 Bidding Announcement This Week
Three Foreign Companies Competing... Domestic Purchase Impossible
Small Market, No Technology or Market Formation
"4th Industrial Era Infrastructure Requires Localization for National Security"

South Korea prides itself as a powerhouse in information and communication technology (IT) and semiconductors. Yet, it cannot self-sufficiently produce the fundamental infrastructure of the Fourth Industrial Revolution era: supercomputers. The procurement of the nation's sixth supercomputer, scheduled for this year, is dominated solely by foreign companies. This is because domestic demand is insufficient, and independent industries and technologies have not developed. However, the demand for data processing speed and capacity is rapidly increasing with the rise of ultra-large artificial intelligence (AI) represented by ChatGPT, big data, autonomous driving, and more. The global economy is becoming bloc-oriented amid the US-China technological hegemony competition and supply chain restructuring. There are growing calls for the localization of supercomputers not only to meet the surging domestic demand but also from a national security perspective.

Surging Demand for Supercomputers

High-performance computers (HPC) consist of a physical assembly (hardware) of hundreds or thousands of high-performance computers connected via high-speed interconnect networks, along with operating systems (system software). They include over 10,000 central processing units (CPUs) and graphics processing units (GPUs), as well as large-scale hard disks and memory. Software, such as interconnecting technology that efficiently utilizes resources, is as important as hardware. The use of supercomputers has greatly expanded across academic research, public services, and various industrial fields. Supercomputers are essential research infrastructure in massive computation fields such as weather and climate prediction, particle physics, astronomy, and biotechnology. They are also widely used in defense, security, and energy sectors where experiments like nuclear fusion and nuclear testing are impossible. Recently, demand has surged across everyday life areas such as precision medicine, agricultural technology, smart cities, renewable energy, and autonomous vehicles, as big data and AI converge. Services like ChatGPT, which sparked the generative AI craze, would be impossible without supercomputer utilization.

According to data released by the Korea Institute of Science and Technology Information (KISTI) in January, the demand for high-performance computing operations has surged 49-fold over the past five years, but resource capacity has only expanded 3.2 times, supplying only 1/480 of total demand as of 2020. In 2015, annual computational demand was 311 petaflops (Peta FLOPS, PF; 1PF = 1,000 trillion operations per second), with 9.7PF (1/32) supplied. By 2020, demand reached 15.1 exaflops (Exa-FLOPS, EF; 1EF = 100 quintillion operations per second = 1,000PF), but only 31.5PF (1/480) was met. A 1-exa-scale supercomputer has the performance to process in one second the calculations that 8 billion people worldwide would take four years working 24 hours a day to complete. This is equivalent to about 5 million general desktop computers (200 gigaflops each).

Determining National Future Competitiveness

Abroad, supercomputing capabilities are already seen as a factor determining national future competitiveness, prompting active expansion. Over the past decade, the world’s top high-performance computer computational power improved 32-fold, from 34PF in 2013 to 1.1EF in 2022. As of November last year, the total global computational power was 4.9EF, with the US at 2,132PF, Japan 624PF, China 514PF, Finland 321PF, and Italy 253PF. South Korea ranked 8th with 89PF. Notably, the US is leading by introducing the world’s first 1-exa-scale supercomputer. The ‘Frontier’ installed at Oak Ridge National Laboratory under the Department of Energy (DoE) recorded 1.1 exa-scale performance in the first half of last year’s International Supercomputing Conference (ISC) evaluation. Additional exa-scale supercomputers such as ‘Aurora’ (2EF) at Argonne National Laboratory and ‘El Capitan’ at Oak Ridge are also being pursued. South Korea’s national supercomputer No. 5, ‘Nurion,’ had a performance of 25.7PF, ranking 11th globally in 2018, but has since fallen to around 40th place. Nevertheless, due to increasing demand, its monthly average utilization rate reached 99.44% (2022), with a usage rate of 74.2%, indicating full operation.

"Need for Independent Technology Development"... A Long Road to Localization

As demand grows, the government plans to procure the sixth national supercomputer with a total budget of 292.9 billion KRW from this year through 2024. The Korea Institute of Science and Technology Information (KISTI) National Supercomputing Center, the lead agency, plans to announce a bid through the Public Procurement Service in the third week of March (20?24). The goal is to purchase hardware and software and establish system development tools and operational support to introduce a 600PF-class high-performance supercomputer by the first half of 2024. Upon completion, the target is to rank between 5th and 10th globally in performance. An official from the Ministry of Science and ICT’s Fundamental Technology Division, Ahn Hee-nam, explained, "Building a higher-performance supercomputer requires dedicated building construction and additional operating costs. We decided to build a supercomputer at an appropriate level that fits demand and domestic circumstances rather than aiming for the highest performance."

The problem is that domestic procurement is impossible. Due to US-China trade conflicts and technological hegemony competition, Japan’s semiconductor materials, parts, and equipment export restrictions, and global supply chain restructuring, the need for self-sufficiency in supercomputers for national security is increasing. However, the domestic high-performance computer market accounts for only about 2% of the global market, so companies are reluctant to enter. Not only must CPUs and GPUs be fully imported from foreign companies like Intel and AMD, but there are also no companies capable of creating parallel processing environments for large-scale computations or developing applications. Therefore, only three foreign companies are expected to participate in this bid: the traditional leader US-based HPE Cray, Lenovo (which inherited IBM’s technology but is Chinese capital), and the European leader Atos from France.

Efforts toward localization have not been absent. In the 1990s, a ‘mainframe’ development project involving four private companies including Samsung and LG was underway but was abandoned due to the small domestic market. In 2011, the High-Performance Computing Act was enacted to lay the foundation. From 2016, a 100 billion KRW investment was declared to develop HPC independently. The goal was to develop 1PF-class by 2020 and 30PF-class by 2025. As a result, 33 billion KRW was invested in research to build system construction capabilities and customization skills. Additionally, from 2020, 49 billion KRW has been spent over five years on independent CPU development.

Official Ahn said, "The US and Japan invest trillions of won when developing and using independent supercomputers," adding, "Demand for supercomputers in South Korea is growing, and companies are increasing their own construction efforts." This indicates that market conditions for commercializing domestic supercomputers are maturing. Lee Sik, head of KISTI’s National Supercomputing Center, noted, "In the past, there was a strong public opinion that it was fine to just buy supercomputers like cars without developing them independently." However, he pointed out, "Demand is increasing, domestic market conditions are improving, and from a national security perspective in preparation for economic bloc formation, long-term technological independence for supercomputer self-sufficiency is necessary." He added, "While South Korea produces memory and hard disks, CPUs and GPUs must be imported, and capabilities in interconnection technology and software construction are weak. Since it has become clear that 100% dependence on foreign countries is risky, it is time for independent technology development."