Josh Norem
2024-03-06 10:51:19
www.extremetech.com
According to Nvidia’s roadmap, it’ll unveil its next-gen Blackwell architecture soon. The company always launches a new architecture with data center products first and then reveals the cut-down GeForce versions many months later, so that’s what’s expected this time as well. On that note, the company’s semi-annual GTC technology conference starts in two weeks, so we expect a lot to be revealed at the show. As proof that Nvidia is close to pulling the wraps off its new data center GPUs, a Dell executive has already shared some juicy info about next-gen Nvidia hardware, saying in a recent earnings call the company has a 1,000W data center GPU in the pipeline.
The executive who has probably already received an angry call from Jensen is Jeff Clarke, a COO at Dell. On a Feb. 29 earnings call (PDF), the executive discussed Dell’s engineering superiority and how upcoming hardware from Nvidia will give the company a chance to show it off. “We’re excited about what happens at the B100 and the B200,” he said, which are the die names for Nvidia’s next-generation data center GPU and its apparent successor. For context, Nvidia currently has the H100 as its flagship data center GPU and is just now launching the second iteration with faster HBM3e memory, dubbed H200. We all know the B100 is the Blackwell successor to this chip, so it appears the B200 will be that GPU’s second iteration, though it does not currently appear on Nvidia’s roadmap (below).
Nvidia’s “AI roadmap” makes no mention of a B200, but it’s possible the Dell executive got his die names mixed up.
Credit: Nvidia
Clark then began discussing the thermal performance of these next-gen parts, stating, “You don’t really need direct liquid-cooling to get to the energy density of 1,000W per GPU. That happens next year with the B200,” he said. We journalists like to hem and haw about what someone “allegedly” stated or implied, but this statement could not be clearer. However, it’s possible Clark got the die names mixed up, as The Register suspects he is referring to the Blackwell version of the company’s CPU+GPU product, which would be the GB200, not the B200. However, the fact that he said this 1,000W GPU is coming “next year” is certainly interesting.
The current GH200 Grace Hopper CPU+GPU already has a TDP of 450W to 1,000W, depending on the configuration, so it would be a bit of a surprise to see that number remain the same for the next-generation version. At the same time, the existing H100 is a 700W GPU, but we don’t know what its successor—B100—will require for power consumption. It seems feasible that Nvidia might crank that one up to 1,000W, but we have yet to hear anything about B100’s power consumption.
For now, we will have to wait until March 18 to see what Nvidia has in store for the data center with Blackwell, and as gamers, we might be able to glean some details from that announcement as well. Given how high the company is riding in the AI market, the world will be tuning in for GTC this year to see what Nvidia has up its sleeve. Despite the proximity to launch, we still don’t know much about Blackwell other than that it’ll use TSMC’s 3nm process and that Nvidia might go with a chiplet design for the first time. Nvidia has also revealed that demand for these chips will exceed supply in the short term.