Supercomputer “晶創26” First Showcased in SC25 and Ranked 29th on the TOP500 List

At this year’s International Conference, SuperComputing 2025 (or SC25), held in St. Louis, Missouri, USA, the eagerly anticipated TOP500 list of the world’s supercomputers was officially announced. “Nano4,” the H200-based system within “NANO4,” a supercomputer developed by the National Center for High-performance Computing (NCHC) of the National Institutes of Applied Research (NIAR), a legal entity affiliated to the National Science and Technology Council (NSTC), made its debut on the list, ranked 29th worldwide. This achievement not only marks a new milestone for Taiwan, which is committed to the independent R&D of supercomputers, but also demonstrates Taiwan’s strong international competitiveness in artificial intelligence (AI).

The NCHC-developed “NANO4” is a key national infrastructure project supporting Taiwan’s promotion of AI, high-performance computing, and research autonomy. As an integral component of both the “Forward-looking Infrastructure Development Program” and the “New Ten Major AI Projects ,” 晶創26 not only represents a major evolution in Taiwan’s hardware of computing power but also serves as a critical driver for sovereign AI and technological self-reliance. Within this system, “Nano4” achieved a measured power consumption of 2.214 megawatts (MW) and a maximum measured performance (Rmax) of 81.55 PFlops1, making it Taiwan’s fastest and highest-computing-density supercomputer to date.

“晶創26” adopts a design featuring dual computing architecture, incorporating both the “Nano4” system based on the H200 architecture and the latest-generation GB200 NVL72 node. This configuration delivers high-performance computing power that supports both general-purpose scientific simulations and AI training.

The system participating in this year’s TOP500 rankings was “Nano4.” Another system, the NVIDIA GB200 NVL72, namely NVIDIA’s flagship AI computing platform newly released in 2025, has also been deployed, with NCHC being the first in Taiwan to introduce it. Each unit is equipped with up to 72 Blackwell GPUs and 13.5 TB of memory, offering exceptionally powerful AI training capabilities. This demonstrates Taiwan’s forward-looking vision and technological strength in building cutting-edge AI infrastructure.
The overall system configuration of “Nano4” includes:

• 220 NVIDIA H200 nodes, each equipped with dual Intel Xeon Platinum processors, eight NVIDIA H200 GPUs, and 2 TB of memory.
• 2 NVIDIA GB200 NVL72 systems, each outfitted with 72 NVIDIA Grace processors, 72 Blackwell GPUs, and 13.5 TB of memory.
• A total storage capacity of 25 PB (Petabyte), supported by InfiniBand NDR high-speed interconnect technology that features 400 Gb/s. The system also adopts direct liquid cooling (DLC) and achieves a power usage effectiveness (PUE) of just 1.18, delivering a design that balances high performance, energy efficiency, and sustainability.

Looking ahead, “NANO4” will provide all-encompassing support for applications across Taiwan’s academic institutions and industrial sectors, covering key fields such as generative AI training and inference, climate modeling, life sciences, the R&D of semiconductors, and smart manufacturing. Moreover, NCHC will further integrate its other supercomputers and high-performance networking systems to build for Taiwan a research computing environment that is scalable, secure, and swift.

The development of “NANO4” not only strengthens Taiwan’s capabilities in technological autonomy but also lays the monumental groundwork for the vision “computational sovereignty and AI powerhouse.” This system will continue to expand international collaboration, enhance computing services, and empower Taiwan in the era of AI and scientific computing.

A floating-point refers to a number with a decimal point. Floating-point operation is an arithmetic used to deal with this kind of number, especially for measuring computing performance. Floating-point operations per second (FLOPS) show a computer’s computing performance. One Petaflops represents 1015 times of floating-point operations per second. The TOP500 supercomputer rankings are determined based on the number of FLOPS a system can perform.