El Capitan and U.S. National Security

Let’s start with some specs around El Capitan, which only dates back to November of 2024. Here’s how it was described in internal coverage of its unveiling, as a machine that would serve Los Alamos and other labs, and was built: ”to advance nuclear weapon science and scientific discovery, providing the vast computational power necessary to ensure the safety, security and reliability of the nation's nuclear deterrent without nuclear testing.”

Although there were brief reports of the White House wanting to return to above-ground nuke testing this year, none of that came to pass, so you might imagine that, domestically, El Capitan is doing its job. But if its job is to tower over Chinese operations, well, that has not panned out.

Let’s also be specific about El Capitan’s relative performance: technically, the system runs at 1.742 exaFLOPs on the High Performance Linpack — the benchmark used by the people who measure these things, with a total peak performance of around 2.79 exaFLOPs. The LineShine system is rated at around 2.2 exaflops standard, so, it outpaces El Capitan.

Here are more specs on the U.S. mainframe, from Copilot:

“El Capitan combines 11,039,616 CPU and GPU cores, including 43,808 AMD fourth-generation EPYC ‘Genoa’ CPUs (1,051,392 cores) and 43,808 AMD Instinct MI300A GPUs (9,988,224 compute units) integrated with 24 Zen4 CPU cores and CDNA3 GPU cores per Each MI300A GPU has 128GB of HBM3 memory, and the system uses a dragonfly network topology with HPE Slingshot 64-port switches providing 12.8 terabits/second Total cabling spans 145 km, optimized for minimal latency.”

The Chinese Answer

What about LineShine?

Well, for starters, it has no GPUs. No Blackwells, no Hoppers, no B200s or H100s, no Nvidia products of any kind. Again, the U.S. worked, on an incredibly piecemeal basis, to put these products out of reach for the Chinese.

Specifically, LineShine is reportedly using Armv9‑based LX2 processors, each with 304 CPU cores, across 20,480 nodes, totaling over 2.45 million CPU cores. It starts to sound like the Cerebras “dinner plate” model, in other words, swapping quality for quantity. But it works, in a robust display of what James Titcomb at The Telegraph calls “full-stack independence” – and it seems that, under export pressure, this is what the Chinese were going for.

Some Thoughts on Cooling

Let’s make a little detour here to talk about some of the logistics of these monumental computing systems, with the understanding that ENIAC, built in the middle of the twentieth century, ran at about 500 flops.

I went down a bit of a rabbit hole here, noting that El Capitan has a direct cooling method using cold plates, and color coded conduits for intake and outtake of cooling components.

By contrast, we actually don’t know what LineShine uses, as its engineers are being tight-lipped, in general.

There’s a new method known as immersion cooling that’s coming online, that might become the norm in newer systems. In general, even cutting-edge stuff is becoming obsolete at a spectacular rate.

U.S. Private Sector Efforts

Detractors of LineShine’s “top spot” according to the TOPS500 reviewers who came up with the analysis suggest that many top U.S. private-owned systems are probably beyond both El Capitan and the Chinese contender, but are not submitting their data to TOPS500 for consideration. That may include, for example, the xAI “Colossus” network, and systems owned by Amazon and other top U.S. tech firms.

So maybe take the above with a grain of salt. What this illustrates, though, is that nations vying for AI power understand that there’s more than one way to skin a cat. Stay tuned.