Actually, Nvidia has been building their own AI systems for years. Most people bought them from Supermicro, etc, because they were cheaper, but Nvidia made whole systems as well. For newer high-end systems Nvidia is making more of the sysem. There will still be systems with Supermicro name plates, but most of the cost will be things built by Nividia, and Nvidia specifies the design in detail. if TSMC, Foxconn, etc, are ready, there's no reason Nvidia can't make whole systems here.First, the "news" in the article is actually kind of hidden:
NVidia to begin building "supercomputers".
NVidia is a chip and tool company. They have not been a player in supercomputer building. (Running "AI" on a bunch of GPUs doesn't make a supercomputer.)
The chip plants are already in the US and are not new.
Supercomputers are *ALREADY* built in the US. Cray has built supercomputers in Wisconsin for decades. IBM builds them in NY state, and Dell in Texas.
Are they supercomputers? I depends upon your definition. Things like the top 500 are collections of systems. No one Nvidia rack will be on that list. So if you think of Nvidia producing racks of equipment, they aren't inndivdualy supercomputers, nor is anyone else's products. But the amount of computing in a rack seems to me to justify using the term informally. Incidentally, the Top 500 list shows that 5% of the systems are from Nvidia. I'd bet that more of them use Nvidia GPUs.
A few years ago people built GPUs using Nvidia chips. For high-end systems, that's no longer the case. You now have to get a whole assembly from Nvidia, and in the newest systems that includes the CPUs.
Last edited:
Upvote
0