NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

100,000 Domestic Accelerators Won the Supercomputing Nobel

Divide 1.456 by 0.589 and you get 2.47. The units are EFLOP/s.

2026-11-22 · 4482 words · NeuroAI
100,000 Domestic Accelerators Won the Supercomputing Nobel

Divide 1.456 by 0.589 and you get 2.47.

The units are EFLOP/s. The first number is the sustained double-precision performance the winning team reported on 20,480 Lingxi nodes. The second is the share of peak that number represents: 58.9%.

The division is ours. It puts the machine's peak near 2.47 EFLOP/s, and it shows what the award was actually measuring.

The prize

The 2026 Gordon Bell Prize was announced on the afternoon of November 19, US time, at SC26 in Chicago. The prize was established in 1987 and is widely called the Nobel Prize of supercomputing.

The winning project was "Bridging the 'Quantum Advantage' Gap in FeMoCo Computational Chemistry with an Exascale Supercomputer," a quantum chemistry simulation of the iron-molybdenum cofactor.

Ten people

The team has 10 members from four institutions: the University of Science and Technology of China, the Institute of Physics at the Chinese Academy of Sciences, the National High Performance Computer Engineering Technology Research Center, and the National Supercomputing Center. The names on the citation are Shang Honghui, Xia Zhuozhao, Wu Hao, Ma Huan, Li Zhenyu, Yang Jinlong, Xiao Zhiyu, Chen Yongqing, Li Bin and Lu Yutong.

Ten people, 20,480 nodes. That works out to 2,048 nodes per person, a division with no technical meaning and one useful implication: at exascale, the work is not done by adding hands. It is done by cutting the problem into pieces that can run at the same time.

The number that took the prize

The team produced the first exascale classical solver for FeMoCo. On 20,480 Lingxi nodes it sustained 1.456 EFLOP/s in double precision, 58.9% of the machine's peak, and 5.626 EFLOP/s in mixed precision.

The mixed-precision figure is 3.86 times the double-precision one. That multiple is not free. Lower precision buys speed, and correctness then depends on error control. Reporting both numbers means handing in speed and accuracy together.

58.9% is not a snapshot of peak utilization. It is the average across a full problem. A run that starts fast and drifts, or one that only looks good in a short benchmark, does not produce this ratio. And the denominator can only be the machine's own peak, not another machine's measured result.

The machine

The platform is the Sugon 8000, also called Dengfeng. It is a fully domestic 100,000-accelerator AI supercluster, deployed in Zhengzhou in July and connected to the national supercomputing internet. By the official description it is the country's first fully domestic cluster at that scale.

Three qualifiers, three separate jobs. 100,000 is size. Fully domestic is provenance. First is sequence, and it covers only the fully domestic 100,000-accelerator category, not supercomputers as a class.

The unit is accelerators, not a chip model. Writing a specific part number into that sentence would be filling in a list nobody published.

And 100,000 is the cluster. The run used 20,480 Lingxi nodes. Those are different quantities: one describes the building, the other describes the job.

The distance from July to November 19 is under five months. In that window the machine had to be installed, tuned and adapted, while the result had to be written up and reviewed.

More than one machine

The project was also ported and tuned on two domestic exascale systems.

That reads as a footnote and weighs a lot in engineering. One code base across two machines means redoing compilation, communication and parallel strategy. The two may differ in architecture, interconnect and toolchain, and matching efficiency on both is real work.

Code that runs on one machine belongs to that machine. Code that runs on two belongs to a method. For a domestic stack, portability matters more than a single fast run.

300 applications

The cluster is not dedicated to this problem. By the official count, more than 300 key applications have been deeply optimized on the Sugon 8000, spanning more than 30 research and industrial fields.

A prize is a ceiling. 300 applications are a workload. The first decides how others see the machine. The second decides whether anyone uses it. A cluster a few months old with 300 optimized applications did not assemble that list for a benchmark; it arrived with a queue.

A finalist is not a winner

In the same round, SD3, a global kilometer-scale numerical weather system from Zhongke Tianji Meteorological Technology, reached the finals of the Gordon Bell climate modeling prize with a 1.5-kilometer global resolution.

Finals and winners are two lists. SD3 is on the first. FeMoCo is on the second. Merging them hands out a trophy that was not awarded.

The fourth since 2016

Chinese supercomputing application teams have now won the Gordon Bell Prize four times since 2016, and this is the first since 2021.

Four in ten years averages one every two and a half years. The average describes output, not spacing: the previous win was five years ago.

That is the useful frame for this award. 100,000 accelerators answer whether the machine can run. 58.9% answers whether it runs well. The prize answers whether anyone else accepts the result. The first two can be asserted. The third is voted on at SC26 in Chicago, and it is the only one of the three that cannot be self-certified.

*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*

Related coverage

Cite this article

NeuroAI (脑机网), "100,000 Domestic Accelerators Won the Supercomputing Nobel," 22 November 2026, https://neuroai.site/a/na-chips-sugon-gordon-bell-1122

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Chips & Compute” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories