ISC American supercomputers continued to dominate the Top500 rating of the world’s mightiest silicon machines this northern spring, with Oak Ridge Nationwide Laboratory’s Frontier remaining the one exascale system on the listing.
However whereas US and European supers held their positions as they put together to unleash new exascale techniques, China stays conspicuously absent from this dialog.
China’s final really aggressive system to grace the Top500 was the 61.4 petaflop Tianhe-2A, which debuted in 2018. The extra highly effective 93 petaflop Sunway TaihuLight is even older, having made its look on the Top500 rating again in 2016.
But, it is an open secret that the Chinese language have a number of exascale-class techniques able to going toe-to-toe with the US’s number-one ranked Frontier system. China simply is not speaking about them.
Over the previous few years, the Top500 has seen fewer and fewer submissions from China. These which might be made come largely from smaller industrial gamers within the single digit petaflop vary.
China stays one of many largest gamers, with 134 techniques within the newest rankings. Nevertheless, that quantity has shrunk. In 2023, China has added only one new system. Holding the 185th spot, the Geely Sensible Star Dubhe is a CPU-only system constructed by Lenovo for Geely Auto Group. By our depend, it options roughly 1,280 32-core Sapphire Rapids Xeon Scalable processors good for 3.5 petaflops of efficiency within the Linpack benchmark.
So, what provides?
China lurks within the shadows
A yr in the past, the US Division of Vitality’s Frontier supercomputer was the primary exaflop system to grace the Top500, unseating Japan’s 442 petaflop Fugaku within the course of. It was heralded as a significant step ahead and a return to type for the US, which has lengthy held a management place on the worldwide rating.
Nevertheless, the “first” was purely a paper one. We have identified since 2021 that China’s Sunway Oceanlite and Tianhe-3 techniques had exceeded the exaflop barrier within the Linpack benchmark. Evidently neither of those techniques has ever appeared on the Top500 rating.
There are just a few methods to have a look at this. One is that US commerce restrictions are having their supposed impact. A lot of China’s nationwide supercomputing facilities have discovered themselves on the US Entities Checklist and are thus topic to stiffer export controls on delicate applied sciences like GPUs utilized in HPC and AI/ML purposes.
It is also price noting that the aforementioned Oceanlite and Tianhe-3 techniques aren’t utilizing chips from Intel, AMD, or Nvidia. As we perceive it, they’re based mostly on homegrown chip architectures. That is to not say US restrictions have not made issues more durable for Chinese language supercomputer improvement.
These chips are nearly actually not being fabbed in China, the place the nation’s most superior foundries have solely simply gained the flexibility to provide chips on a 14nm course of. Meaning they’re topic to US commerce restrictions on the sale of chipmaking tools and mental property that stretch to international fabs like TSMC or Samsung Electronics.
For instance, in an effort to have TSMC construct its GPU, fabless Chinese language chipmaker Biren was compelled to remodel its designs to adjust to US limits on switch charges that went into impact final fall. So designing its personal HPC-centric chips is certainly not a silver bullet.
There’s additionally proof to counsel China’s Academy of Engineering Physics is utilizing backchannels and shell corporations to acquire US chips to construct supercomputers utilized in battle video games and nuclear weapons sims.
The query turns into: What does submitting two or extra Chinese language exascale techniques to the Top500 obtain, aside from to piss off the US Commerce Division and danger even stiffer sanctions from the Biden administration?
On the finish of the day, these machines are instruments, tasked with all method of HPC workloads – some for the betterment of mankind, and others for guaranteeing the world-ending efficacy of nuclear arsenals. The machines do not need to be listed on the Top500 to do their jobs – so the Chinese language have largely gone silent.
Europe seizes its second, whereas US awaits Aurora
Whereas China’s supercomputers lurk off-list, over the previous few years Europe has made regular beneficial properties on the Top500 rating.
A yr after first showing on the rating, Finland’s LUMI supercomputer stays the planet’s third strongest system. In the meantime, final northern Autumn, Italy’s Leonardo system displaced America’s Summit to say the Quantity 4 spot. Within the six months since, the Leonardo system has grown much more highly effective, now claiming 238 petaflops – up from 174 petaflops final fall.
Europe is anticipated to additional improve its standing place on the Top500 following the completion of the Jupiter system on the Jülich Supercomputing Centre in Germany. Final we heard, the system may come on-line as early as late 2023 or early 2024. Whether or not it will be able to run Linpack in time for the Supercomputing Convention in Denver this coming November stays to be seen.
Whereas Europe awaits its first exascale system, the US is eagerly awaiting the completion of Argonne Nationwide 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 ultimate piece of that puzzle lastly arrived with the launch of Intel’s 4th-gen Xeon Scalable techniques in January.
Whereas Argonne missed the deadline to take part on this spring’s Top500, it appears seemingly that the system – which is anticipated to high two exaflops within the Linpack benchmark – will make its first look within the second half of 2023. ®