China Unveils 2 Exaflop, All-CPU 'LineShine' Supercomputer
Posted by NamelessVegetable@reddit | hardware | View on Reddit | 13 comments
Posted by NamelessVegetable@reddit | hardware | View on Reddit | 13 comments
996forever@reddit
2 exaflop cpu only?! The scale of this thing will be massive
NamelessVegetable@reddit (OP)
Indeed! It was a bit surprising given that China has its own accelerators/GPUs and has used them before (the Matrix-3000/MT-3000 APU in the Tianhe-3); and the other traditionally all-CPU supercomputer builder, Japan, is moving to Nvidia GPUs paired with a future version of the Fujitsu Monaka CPU for the FugakuNext.
996forever@reddit
Yes, even Japan caved in to gpu acceleration after their cpu only supercomputer briefly took the top spot. 2 exeflops in cpu only is absolutely massive, but the article doesn’t provide a timeframe or estimated power usage
Rodot@reddit
CPU clusters still have their purpose. There are still more things to compute than AI alone.
Different_Lab_813@reddit
Not everything computed on GPU is AI.
Exist50@reddit
Even for AI, a CPU with particularly robust matrix support can be kind of interesting. Lot of things are more bandwidth intensive than they are compute.
Rodot@reddit
Oh definitely, I do neural differential equations and adaptive solvers are a pain with the CPU-GPU communication bottleneck. Even with unified memory, like on the GH nodes, some models often run faster on the CPU than the GPU depending on the diffeq order and error tolerances because of the variable number of steps.
Exist50@reddit
Got to be a fat SME implementation.
Tone-Bomahawk@reddit
Imagine how many foreign government websites you can DDoS with this bad boy.
ttkciar@reddit
"Unveils" implies they've already built it, but the article states that it is still years away from completion. Until then it's vaporware.
NamelessVegetable@reddit (OP)
The exact quote in the article is "It [the supercomputer] ... when it comes online, which will likely be in years." (Emphasis mine). This timeframe is the article's author speculating when it comes online, not a statement of when it does (which we don't know).
Also, "vaporware" is too strong a term. It implies that the HW hasn't been designed yet, there are no realized examples, and that the HW "exists" aspirationally. The preprint linked to in the article describes performance characterization of workloads on LineShine HW. That implies that the design has been finished, and there is working HW available.
What's confusing is that the article is written in a way that implies the system has yet to begin installation, but the preprint describes some aspects of the system as existing, albeit unfinished. The quoted remarks in the article from the director of the hosting supercomputer center are even more confusing; they seem to imply that the existing HW are possibly test articles.
This isn't unusual in itself. Large supercomputers can be built in stages and take multiple years to finish. The EU's Jupiter supercomputer is technically unfinished, even though installation began in 2023-12, and it was launched incomplete in 2025-09. It appears in the TOP500 list, and is running production workloads, regardless. Why is it still unfinished? The SiPearl Rhea processors for the general-purpose module have been repeatedly delayed and are due (hopefully) this year.
Lianzuoshou@reddit
Based on the available information, it appears that the first phase of development has been completed, with testing conducted across nine application areas. The project is now transitioning from closed beta to open beta, and applications for testing are being accepted starting today.
sboyette2@reddit
You can unveil a design, or a prototype, or...
It's just a way to say that something which was being worked on, but not talked about, is now being talked about.