ISC This northern spring, US supercomputers continued to dominate the top 500 rankings of the world’s most powerful silicon machines, with Oak Ridge National Laboratory’s Frontier being the only exascale system on the list.
But while the U.S. and European superpowers are gearing up to introduce new exascale systems, China is clearly not participating in the dialogue.
The last truly competitive system for China to enter the top 500 is the 61.4 petaflops Tianhe 2A, which debuted in 2018. The more powerful 93 petaflops Sunway TaihuLight is even older, appearing in the Top 500 rankings in 2016.
But it’s no secret that China has several exascale-class systems that can compete with the top US frontier systems. China just isn’t talking about them.
Over the past few years, Chinese entries to the Top500 have been declining. Many of the products manufactured are made by smaller industrial players in the single-digit petaflops range.
China remains one of the biggest players, with 134 systems in the latest rankings. However, their numbers have decreased. In 2023, China added only one new system. The No. 185 Geely Wise Star Dubhe is a CPU-only system built by Lenovo for Geely Automobile Group. With around 1,280 32-core Sapphire Rapids Xeon Scalable processors, according to our calculations, he delivers 3.5 petaflops of performance on the Linpack benchmark.
So what is given?
China lurk in the shadows
A year ago, the U.S. Department of Energy’s frontier supercomputer was the first exaflops system to crack the top 500, ahead of Japan’s 442 petaflops Fugaku. It was hailed as a major step forward and a return to normal for the United States, which has long held a leading position in the international rankings.
However, the “first” was purely paper. It has been known since 2021 that China’s Sunway Ocean Lite and Tianhe 3 systems have surpassed the exaflops barrier on his Linpack benchmark. Needless to say, neither of these systems have ever appeared in his Top500 rankings.
There are several ways to check this. One is that US trade restrictions are having their intended effect. Many of China’s national supercomputing centers are on the U.S. Entity List, which tightens export controls on sensitive technology such as GPUs used in HPC and AI/ML applications.
Also note that the Oceanlite and Tianhe-3 systems mentioned above do not use chips from Intel, AMD, or Nvidia. As we understand them, they are based on a homegrown chip architecture. That’s not to say US regulations aren’t making China’s supercomputer development more difficult.
These chips are almost certainly not made in China, and the country’s most advanced foundries have just acquired the capability to manufacture chips on the 14nm process. In other words, the sale of chip manufacturing equipment and intellectual property to overseas factories of TSMC, Samsung Electronics, etc. is subject to US trade restrictions.
For example, in order to get TSMC to build its GPUs, fabless Chinese chipmaker Biren was forced to redesign to comply with US transfer rate limits that went into effect last fall. So designing his own HPC-centric chip is by no means a silver bullet.
There is also evidence to suggest that China’s Engineering Physics Society is using backchannels and dummy companies to obtain US chips to build supercomputers used in war games and nuclear weapon simulations.
The question is, what can submitting two or more Chinese exascale systems to the Top 500 accomplish, other than anger the U.S. Department of Commerce and risk even harsher sanctions from the Biden administration?
After all, these machines are tools for all sorts of HPC workloads, some to better humanity, some to ensure the effectiveness of world-ending nuclear weapons. The Chinese are mostly silent because the machine doesn’t have to be listed in the top 500 to work.
Europe seizes the moment while America waits for the Northern Lights
While China’s supercomputers are off the list, Europe has climbed steadily up the Top 500 rankings in recent years.
Finland’s LUMI supercomputer remains the third most powerful system in the world, one year after it first appeared in the rankings. Meanwhile, Italy’s Leonardo system overtook the America’s Summit in the northern part of last fall to take fourth place. In the six months since then, the Leonardo system has gone from strength to strength, going from 174 petaflops last fall to 238 petaflops today.
Europe is expected to further advance its position in the Top500 following the completion of the Jupiter system at the Jupiter Supercomputing Center in Germany. Last heard, the system could go live as early as late 2023 or early 2024. It remains to be seen if Linpack will be ready to run in time for the Supercomputing conference in Denver next November.
While Europe awaits its first exascale system, the United States awaits the completion of Argonne National Laboratory’s Aurora supercomputer. The system has been delayed since 2018 due to Intel’s failure to ship his Sapphire Rapids CPU and Ponte Vecchio GPU blades. With the launch of Intel’s 4th Gen Xeon Scalable systems in January, the final piece of that puzzle is finally complete.
Argonne missed the deadline to join the Top500 this spring, but the system, which is expected to exceed 2 exaflops on the Linpack benchmark, looks likely to first appear in late 2023.
