New Chinese System Takes Top Supercomputer Spot

El Capitan and U.S. Nationwide Safety

Let’s begin with some specs round El Capitan, which solely dates again to November of 2024. Here’s how it was described in inside protection of its unveiling, as a machine that might serve Los Alamos and different labs, and was constructed: ”to advance nuclear weapon science and scientific discovery, offering the huge computational energy essential to make sure the security, safety and reliability of the nation’s nuclear deterrent with out nuclear testing.”

Though there have been temporary experiences of the White Home desirous to return to above-ground nuke testing this yr, none of that got here to go, so you may think that, domestically, El Capitan is doing its job. But when its job is to tower over Chinese language operations, effectively, that has not panned out.

Let’s even be particular about El Capitan’s relative efficiency: technically, the system runs at 1.742 exaFLOPs on the Excessive Efficiency Linpack — the benchmark utilized by the individuals who measure this stuff, with a complete peak efficiency of round 2.79 exaFLOPs. The LineShine system is rated at round 2.2 exaflops commonplace, so, it outpaces El Capitan.

Listed below are extra specs on the U.S. mainframe, from Copilot:

“El Capitan combines 11,039,616 CPU and GPU cores, together with 43,808 AMD fourth-generation EPYC ‘Genoa’ CPUs (1,051,392 cores) and 43,808 AMD Intuition MI300A GPUs (9,988,224 compute items) built-in with 24 Zen4 CPU cores and CDNA3 GPU cores per Every MI300A GPU has 128GB of HBM3 reminiscence, and the system makes use of a dragonfly community topology with HPE Slingshot 64-port switches offering 12.8 terabits/second Complete cabling spans 145 km, optimized for minimal latency.”

The Chinese language Reply

What about LineShine?

Properly, for starters, it has no GPUs. No Blackwells, no Hoppers, no B200s or H100s, no Nvidia merchandise of any type. Once more, the U.S. labored, on an extremely piecemeal foundation, to place these merchandise out of attain for the Chinese language.

Particularly, LineShine is reportedly utilizing Armv9‑based mostly LX2 processors, every with 304 CPU cores, throughout 20,480 nodes, totaling over 2.45 million CPU cores. It begins to sound just like the Cerebras “dinner plate” mannequin, in different phrases, swapping high quality for amount. But it surely works, in a sturdy show of what James Titcomb at The Telegraph calls “full-stack independence” – and plainly, below export strain, that is what the Chinese language have been going for.

Some Ideas on Cooling

Let’s make just a little detour right here to speak about a number of the logistics of those monumental computing methods, with the understanding that ENIAC, built in the middle of the twentieth century, ran at about 500 flops.

I went down a little bit of a rabbit gap right here, noting that El Capitan has a direct cooling methodology utilizing chilly plates, and coloration coded conduits for consumption and outtake of cooling elements.

In contrast, we truly don’t know what LineShine makes use of, as its engineers are being tight-lipped, basically.

There’s a brand new methodology generally known as immersion cooling that’s coming on-line, that may develop into the norm in newer methods. Normally, even cutting-edge stuff is turning into out of date at a spectacular charge.

U.S. Personal Sector Efforts

Detractors of LineShine’s “prime spot” in keeping with the TOPS500 reviewers who got here up with the evaluation recommend that many prime U.S. private-owned methods are probably beyond each El Capitan and the Chinese language contender, however aren’t submitting their knowledge to TOPS500 for consideration. That will embrace, for instance, the xAI “Colossus” community, and methods owned by Amazon and different prime U.S. tech companies.

So perhaps take the above with a grain of salt. What this illustrates, although, is that nations vying for AI energy perceive that there’s multiple technique to pores and skin a cat. Keep tuned.