Dell Exec Confirms NVIDIA’s Monster Blackwell AI Accelerator Is Coming

hero nvidia gh200 photo
NVIDIA's been sitting pretty with its Ada Lovelace GPUs for consumers and professionals, while the HPC market is served by its upgraded H200, which is a fully-enabled Hopper GPU paired with super-fast HBM3e memory giving it some 5TB/second of memory bandwidth. Few parts of the technology industry sit still for long, though, and NVIDIA's certainly not resting on its laurels. Blackwell is coming, and we have the first specific mention of the product from a source outside of NVIDIA.

Speaking to industry analysts on an investor call, Dell's Chief Operating Officer Jeff Clarke remarked that he's "excited" about "what happens at the B100 and the B200". This is interesting, because NVIDIA hasn't actually announced products called "B200" yet. These would be the actual names of shipping products, not codenames for GPUs, as we do indeed already know that "GB200" is the codename of a GPU product on the way. Meanwhile, B100 was mentioned on NVIDIA's last roadmap as being the successor to the H200 for x86 training and inference. From that, we can gather that it's likely the largest configuration of Blackwell GPU.

The full quote from Dell's Jeff Clarke was this:
We're excited about what happens at the B100 and the B200, and we think that's where there's actually another opportunity to distinguish engineering confidence. Our characterization in the thermal side, you really don't need direct liquid cooling to get to the energy density of 1,000 watts per GPU. That happens next year with the B200.
He went on to talk about how B200 will be an opportunity for Dell to showcase its engineering expertise, particularly with regard to liquid cooling implementations, which is an interesting thing to talk about after saying that you "don't need direct liquid cooling" for a 1,000-watt GPU. Wait, a 1,000-watt GPU?

That's what the man said. This isn't quite as insane as it sounds; the H200 Tensor Core GPU already has a maximum thermal design power of 700W. Blackwell is still apparently going to be a bump of around 42% in power consumption, though. We wonder if Blackwell is going to be as much as 50% faster than H200; NVIDIA CEO Jensen Huang is famously known for stating "Moore's Law is dead," and in that context, increasing power consumption is really the only way to get these massive generational leaps in performance.

nvidia ai roadmap
Slide: NVIDIA

The other interesting takeaway from Clarke's remarks is that Blackwell apparently won't arrive until next year. That wasn't exactly clear from NVIDIA's roadmap, which seems to place B100 and other Blackwell-derived products in the latter half of 2024. It's always possible that he's mistaken about that, but if anyone outside NVIDIA was going to know, it would probably be Dell's Chief Operating Officer. There's a good chance that NVIDIA will talk about Blackwell at GTC 2024, which starts on March 18th, so stay tuned for further details.