It’s strange that it’s not a Chinese supercomputer leading the top 500 list

American supercomputers dominate the Top500 ranking of the world’s most powerful silicon machines, but Europe keeps coming up.

A Top500.org publishes a list of the world’s fastest supercomputers every six months, so they did most is. The first is the Frontier system at Oak Ridge National Laboratory (Tennessee, USA), and it is the only one with a computing power exceeding one exaflops. In general, American and European supercomputers have maintained their positions, but China has dropped noticeably from the list. China’s last competitive system to top the Top500 list was the 61.4 petaflop Tianhe-2A, which debuted in 2018. The higher-powered, 93-petaflop Sunway TaihuLight is even older, as it was already included in the Top500 ranking in 2016.

Of open secret, that the Chinese have several exascale-class systems capable of competing with the American Frontier system, only China is not talking about them. In the last few years, the Top500 has received fewer and fewer applications from China. Those that have been included are largely from smaller industry players in the single-digit petaflop range. China is still one of the biggest players, with 134 systems participating in the latest ranking, but this number is constantly decreasing. In 2023, only one new system was reported, the Geely Wise Star Dubhe takes the 185th place. It’s built by Lenovo for automaker Geely (they also own the Volvo and Lotus brands) and has 1,280 32-core Sapphire Rapids Xeon Scalable processors.

Frontier appeared on the list a year ago in June 2022, the US Department of Energy’s supercomputer ahead of Japan’s 442-petaflop Fugaku. This was called a significant step forward and the return of the USA, which for a long time held a leading position in the international ranking. However, the “first place” is only on paper, because we have known since 2021 that the Chinese Sunway Oceanlite and Tianhe-3 systems crossed the exaflop limit in the Linpack benchmark. Neither system is included in the Top500 ranking.

There are several reasons for the Chinese retreat. Perhaps this shows that US trade restrictions are having the desired effect. Many of China’s national supercomputing centers have been placed on US sanctions lists, subjecting them to stricter export controls on sensitive technologies. It is also worth noting that the aforementioned Oceanlite and Tianhe-3 systems do not use chips from Intel, AMD or Nvidia, but are based on proprietary chip architectures. That’s not to say that US restrictions haven’t hampered China’s supercomputer development: these chips are almost certainly not made in China, where the country’s most advanced factories have only recently acquired the ability to produce chips using the 14nm process.

Factory-less Chinese chip maker Biren, for example, was forced to revise its designs to comply with U.S. restrictions on maximum transmission rates that came into effect last fall. Designing your own HPC-focused chips is therefore by no means a solution. There is also evidence that the Chinese Academy of Engineering Physics uses back channels and front companies to procure US chips to build supercomputers used in war games and nuclear weapons simulations.

The main reason may be that there is no point in publicly listing two or more Chinese exascale systems in the Top500 other than to anger the US Commerce Department and risk even tougher sanctions from the Biden administration. Ultimately, these machines are tools that perform all sorts of computationally intensive tasks – some for the benefit of humanity, others for ensuring the worldwide effectiveness of nuclear arsenals. The machines don’t need to be in the Top500 list to do their job – so the Chinese don’t even hype their results.

The first European supercomputer is expected to break the exaflop mark next year
But while China’s supercomputers are hiding, Europe has steadily advanced in the rankings in recent years. Celebrating its first anniversary, the Finnish supercomputer LUMI is the third most powerful system in the world. Last fall, the Italian Leonardo system displaced the American Summit and took the 4th place. In the six months that have passed since then, Leonardo has achieved even greater performance, and after last fall’s 174 petaflops, it can now claim 238 petaflops. After the completion of the Jupiter system at the Jülich Supercomputing Center in Germany, Europe is expected to be even further ahead, with the system already in operation at the end of 2023 or the beginning of 2024.

While Europe is waiting for its first exascale system, the United States is eagerly awaiting the completion of Argonne National Laboratory’s Aurora supercomputer. The system has been delayed since 2018 due to Intel not shipping Sapphire Rapids CPUs and Ponte Vecchio GPU servers. The last piece of the puzzle finally arrived with the introduction of Intel’s 4th generation Xeon Scalable systems in January. Although Argonne missed the deadline for participation in the current Top500, it seems likely that the system – which is expected to achieve two exaflop performance in the Linpack benchmark – will appear in the second half of 2023.