US Supers Maintain Grip On Top500 List As China Seemingly Hides Its Powers

ISC American supercomputers continued to dominate the Top500 ranking of the world's mightiest publicly known silicon machines this northern spring, with Oak Ridge National Laboratory's Frontier still the only exascale system on the list.

But while US and European supers held their positions as they prepare to unleash new exascale systems, China remains conspicuously absent from this conversation.

China's last truly competitive system to grace the Top500 was the 61.4 petaflop Tianhe-2A, which debuted in 2018. The more powerful 93 petaflop Sunway TaihuLight is even older, having made its appearance on the Top500 ranking back in 2016.

Yet, it's an open secret that the Chinese have several exascale-class systems capable of going toe-to-toe with the US's number-one ranked Frontier system. China just isn't talking about them.

Over the past few years, the Top500 has seen fewer and fewer submissions from China. Those that are made come largely from smaller industrial players in the single digit petaflop range.

China remains one of the biggest players, with 134 systems in the latest rankings. However, that number has shrunk. In 2023, China has added just one new system. Holding the 185th spot, the Geely Wise Star Dubhe is a CPU-only system built by Lenovo for Geely Auto Group. By our count, it features roughly 1,280 32-core Sapphire Rapids Xeon Scalable processors good for 3.5 petaflops of performance in the Linpack benchmark.

So, what gives?

China lurks in the shadows

A year ago, the US Department of Energy's Frontier supercomputer was the first exaflop system to grace the Top500, unseating Japan's 442 petaflop Fugaku in the process. It was heralded as a major step forward and a return to form for the US, which has long held a leadership position on the international ranking.

However, the "first" was purely a paper one. We've known since 2021 that China's Sunway Oceanlite and Tianhe-3 systems had exceeded the exaflop barrier in the Linpack benchmark. Needless to say neither of these systems has ever appeared on the Top500 ranking.

There are a few ways to look at this. One is that US trade restrictions are having their intended effect. Many of China's national supercomputing centers have found themselves on the US Entities List and are thus subject to stiffer export controls on sensitive technologies like GPUs used in HPC and AI/ML applications.

It's also worth noting that the aforementioned Oceanlite and Tianhe-3 systems aren't using chips from Intel, AMD, or Nvidia. As we understand it, they're based on homegrown chip architectures. That's not to say US restrictions haven't made things harder for Chinese supercomputer development.

Those chips are almost certainly not being fabbed in China, where the country's most advanced foundries have only just gained the ability to produce chips on a 14nm process. That means they're subject to US trade restrictions on the sale of chipmaking equipment and intellectual property that extend to foreign fabs like TSMC or Samsung Electronics.

For example, in order to have TSMC build its GPU, fabless Chinese chipmaker Biren was forced to rework its designs to comply with US limits on transfer rates that went into effect last fall. So designing its own HPC-centric chips is by no means a silver bullet.

• Aurora exascale system gets 'mini-me' testbed for researchers
• Fujitsu's A64FX successor will be an Arm-based datacenter chip
• SambaNova's AI paired with Fugaku supercomputer to develop 'digital twins'
• Unless things change, first zettaflop systems will need nuclear power, AMD's Su says

There is also evidence to suggest China's Academy of Engineering Physics is using backchannels and shell companies to obtain US chips to build supercomputers used in war games and nuclear weapons sims.

The question becomes: What does submitting two or more Chinese exascale systems to the Top500 achieve, other than to piss off the US Commerce Department and risk even stiffer sanctions from the Biden administration?

At the end of the day, these machines are tools, tasked with all manner of HPC workloads – some for the betterment of mankind, and others for ensuring the world-ending efficacy of nuclear arsenals. The machines don't have to be listed on the Top500 to do their jobs – so the Chinese have largely gone silent.

Europe seizes its moment, while US awaits Aurora

While China's supercomputers lurk off-list, over the past few years Europe has made steady gains on the Top500 ranking.

A year after first appearing on the ranking, Finland's LUMI supercomputer remains the planet's third most powerful system. Meanwhile, last northern Autumn, Italy's Leonardo system displaced America's Summit to claim the Number 4 spot. In the six months since, the Leonardo system has grown even more powerful, now claiming 238 petaflops – up from 174 petaflops last fall.

Europe is expected to further enhance its status position on the Top500 following the completion of the Jupiter system at the Jülich Supercomputing Centre in Germany. Last we heard, the system could come online as early as late 2023 or early 2024. Whether it'll be ready to run Linpack in time for the Supercomputing Conference in Denver this coming November remains to be seen.

While Europe awaits its first exascale system, the US is eagerly awaiting the completion of Argonne National Laboratory's Aurora Supercomputer. The system has been delayed since 2018 by Intel's failure to ship its Sapphire Rapids CPUs and Ponte Vecchio GPU blades. The final piece of that puzzle finally arrived with the launch of Intel's 4th-gen Xeon Scalable systems in January.

While Argonne missed the deadline to participate in this spring's Top500, it seems likely that the system – which is expected to top two exaflops in the Linpack benchmark – will make its first appearance in the second half of 2023. ®