South Korea Unveils Hangang Supercomputer With 8,496 GPUs, Targeting a Top-10 TOP500 Spot

The South Korean government is moving to a new generation of national supercomputing. On November 2, 2026, the Ministry of Science and ICT and the Korea Institute of Science and Technology Information (KISTI) held a launch ceremony for the sixth national supercomputer, "Hangang," at KISTI headquarters in Daejeon. The new system is equipped with 8,496 GPUs and is billed as delivering 614 petaflops of theoretical computing performance, 205 petabytes of storage, and 400 Gbps of network bandwidth. Compared with its predecessor, the fifth-generation "Nurion," its theoretical computing performance is said to be 24 times higher.

What matters about this upgrade is not only the increase in computing power. The design philosophy has shifted from a CPU-centered supercomputer to a computing platform that carries large numbers of GPUs and can run AI workloads alongside conventional scientific simulation. However, what took place on November 2 was a launch ceremony; the official service for general researchers is scheduled to begin on December 14. In addition, the announced 614 petaflops is the theoretical peak performance of the hardware, not the processing speed of actual scientific programs or a measured value from the benchmarks used for the TOP500 ranking.

Launch ceremony on November 2; official service scheduled for December 14

The ceremony held at KISTI headquarters on November 2 was an official event marking progress in building and trial-running the new system. It does not mean that computing resources have been opened to all domestic researchers. According to the Ministry of Science and ICT and KISTI, recruitment of user researchers for the new system began in September 2026, and a beta service is under way ahead of formal operation.

In the call for participants published on September 9, KISTI described the beta service as a preliminary operation to verify how the new computing environment behaves and whether research programs can be applied to it. Access is limited to research groups selected through advance recruitment, and researchers can port their programs to the new system and test them before official resource provision begins. The official service is scheduled to start on December 14, 2026.

For researchers, what counts is not just the size of the theoretical performance but how much computing resource they will actually be able to use once official service begins. Job queue wait times, the maximum number of nodes a single job can use at once, and how GPUs are allocated, for example, all shape real research plans. Running existing scientific programs stably also requires a mature software environment, including compilers, communication libraries, and job schedulers.

KISTI describes the beta service as an opportunity to use the new computing resources early and verify how well research programs can be applied. Even when hardware installation is complete, legacy programs will not necessarily deliver high performance when ported as they are. The usage rules and resource allocation terms to be released when official service starts on December 14 will be the first indication of how usable the new system is in research settings.

From a CPU-centered design to an integrated GPU platform

Hangang's defining feature, beyond the large jump in theoretical performance, is its heterogeneous configuration centered on GPUs. Whereas the fifth-generation Nurion was built mainly around running many CPUs in parallel, the new system carries large numbers of GPU accelerators and is designed to handle both conventional scientific computing and AI processing.

According to the Korean government, Hangang has 8,496 GPUs, including NVIDIA's integrated CPU-GPU processor, the "GH200 Grace Hopper Superchip," along with 9,936 CPUs. Storage capacity is 205 petabytes and network bandwidth is 400 Gbps. Compared with the fifth-generation system, the government says theoretical computing performance is 24 times higher, storage capacity 10 times higher, and network performance 4 times higher.

{
  "type": "bar",
  "title": "Performance comparison: South Korea's fifth and sixth national supercomputers",
  "unit": "Multiple (fifth-generation system = 1)",
  "categories": [
    "Theoretical computing performance",
    "Storage capacity",
    "Network bandwidth"
  ],
  "series": [
    {
      "name": "Fifth-generation: Nurion",
      "values": [1, 1, 1]
    },
    {
      "name": "Sixth-generation: Hangang",
      "values": [24, 10, 4]
    }
  ],
  "caption": "Spec-based multiples over the fifth-generation system as announced by the Korean government. This is not a comparison of real-world application performance.",
  "source": "Press materials from South Korea's Ministry of Science and ICT"
}

These are, however, comparisons on published specifications, and they do not mean that actual research programs will run 24 times faster. Moreover, the published materials consulted here do not give the absolute values for the fifth-generation system's specifications, so the multiples shown by the government are presented as they are. Theoretical computing performance, storage capacity, and network bandwidth are all important indicators of a system's capability, but each measures something different.

The earlier Nurion emphasized parallel computing across many CPUs to support computational science such as physics simulation, lattice quantum chromodynamics (lattice QCD), and fluid dynamics. Hangang, by contrast, anticipates "AI for Science" as a major use, bringing deep-learning training and inference into scientific research alongside these simulations.

This difference signals that South Korea is trying to move the emphasis of its national computing infrastructure from a configuration centered on conventional numerical simulation to one that combines AI processing with scientific computing. In research that, for example, uses simulation-generated data to train AI models, or uses trained models to approximate computationally heavy processing, GPU performance could be put to direct use. The aim is not to replace conventional scientific computing but to combine AI with it as a new research method.

Still, simply installing large numbers of GPUs does not ensure that computing performance is fully used. CPU-oriented MPI parallel programs written in Fortran or C++ will not necessarily run efficiently on GPUs as they are, and algorithms, data placement, and communication handling may need to be reworked. If GPU support does not progress, actual utilization could remain low relative to the theoretical computing capability.

The announced 614 petaflops is also a theoretical peak based on hardware specifications and does not indicate processing speed in real applications. In evaluating supercomputer performance, it is necessary to distinguish between Rpeak, which expresses theoretical computing performance, and Rmax, the effective performance measured with the HPL (High Performance Linpack) benchmark. The announced figure cannot be equated with actual computing performance, and an official TOP500 evaluation requires confirmation of benchmark results.

Priority support for AI drug discovery and climate prediction; software GPU readiness is the challenge

The Korean government plans to use Hangang as a computing foundation supporting national strategic R&D programs such as the "K-Moonshot." Intended users include not only domestic universities and government-funded research institutes but also private-sector R&D. The aim is to support AI-driven scientific research and industrial technology development in addition to conventional large-scale simulation.

The priority support areas cited include the following research:

• Advanced AI research, including large language models and scientific foundation models
• Screening of drug candidate compounds and molecular dynamics simulation
• Design of advanced materials such as those for secondary batteries and semiconductors
• Climate change simulation, localized weather forecasting, and disaster prediction
• R&D for manufacturing that combines fluid analysis and structural analysis

In its announcement of the fiscal 2026 computing resource support program, KISTI sets the eligibility requirement as researchers affiliated with domestic institutions who wish to use computing resources. The policy is to provide computing resources to a broad research community that includes industry as well as universities and public research institutions. However, these describe future usage plans and priority support areas, and they do not demonstrate that the new system has already improved research efficiency in any field.

In fields such as drug discovery and climate prediction in particular, the scientific computing programs developed and refined over many years are often optimized for CPUs. Moving them to a GPU environment can take more than simply recompiling. The computation must be restructured for GPUs, and the efficiency of data transfer between CPU and GPU, memory access, and parallel processing must be improved.

The NVIDIA GH200 connects a Grace CPU and a Hopper GPU through the high-speed, cache-coherent NVLink-C2C. Its design, in which CPU and GPU memory can be used cooperatively, suits large-scale scientific computing and AI processing, but the programs themselves must be adapted to draw out that performance. In simulations handling large volumes of data in particular, execution speed depends heavily not just on GPU computing performance but on memory bandwidth, communication handling, and the optimization of data placement.

Improvements in research productivity from the national supercomputer upgrade will therefore not be determined by hardware performance alone. How smoothly researchers can move existing programs to the new environment, and how far GPU libraries and development support environments are built out, also matter. Results from the beta service, GPU utilization after official operation begins, and run times of representative scientific programs will be the basis for judging the return on investment in the national computing infrastructure.

TOP500 "top-10 range" outlook: official ranking and measured performance not yet confirmed

Another point the Korean government stressed in this announcement is the outlook for Hangang's position in the TOP500, the global supercomputer ranking. The Ministry of Science and ICT expects Hangang to enter the top 10 worldwide in the TOP500 list to be released alongside the international conference SC2026, held in the United States in November 2026.

This is, however, the government's projection and does not mean an official TOP500 ranking has been settled. Because rankings are determined by each system's HPL benchmark results, position cannot be judged from the hardware's theoretical computing performance alone. The actual rank will depend not only on Hangang's measured results but also on the performance of other countries' supercomputers registered at the same time.

The key metrics to check in the official TOP500 list are Rmax, the measured HPL computing performance, and Rpeak, the theoretical peak. Their ratio, Rmax ÷ Rpeak, shows how much of the hardware's theoretical performance the HPL benchmark was able to draw out. Even if theoretical computing performance is 614 petaflops, low execution efficiency would leave measured HPL performance far below it. Checking the system scale used in the computation and power consumption can also reveal characteristics beyond simple computing performance.

Even a high TOP500 ranking, though, does not mean every scientific computation will speed up. HPL is a benchmark centered on solving dense linear systems, and its workload characteristics differ from those of computations used in actual research, such as fluid dynamics, climate prediction, molecular dynamics, and AI model training. For example, for programs constrained by memory access or inter-node communication, high HPL efficiency does not guarantee the same level of performance.

The government announcement does not include measured effective performance using representative scientific computing programs. The top-10 outlook is therefore a sign of South Korea's expanding computing infrastructure, but it does not guarantee the processing performance that research institutions and companies will see in actual use. For research settings, it is essential to check not only the TOP500 rank but also the GPU readiness of the programs in question and their parallel scaling performance.

Hangang marks an important turning point in extending South Korea's national computing infrastructure from conventional CPU-centered scientific simulation to GPU-powered, AI-driven scientific research. Whether that translates into actual research results, however, cannot be judged from the hardware's theoretical performance alone. Attention now turns to the official TOP500 ranking at SC2026, the usage terms after service begins on December 14, and verification of effective performance with research programs.