Dell exec reveals Nvidia has a 1,000-watt GPU within the works • The Register
For those who thought Nvidia’s 700W H100s have been sizzling and power-hungry machines, simply wait till the GPU slinger’s B100 arrives later this yr.
In line with Dell Applied sciences COO Jeff Clarke, Nvidia’s newest AI accelerator will devour 1,000 watts – 42 % greater than its predecessor. However don’t fret, he is fairly positive liquid cooling received’t be required to tame the beast.
“We’re enthusiastic about what’s taking place with the H200 and its efficiency enchancment,” Clarke advised buyers on Dell’s earnings call [PDF] final week, earlier than including he feels the identical emotion about Nvidia’s forthcoming B100 accelerator and one other he known as the B200.
He opined that direct liquid cooling will not be wanted to deal with GPUs that devour 1,000 watts apiece – a degree he stated “occurs subsequent yr with the B200.”
It is not totally clear what card Clarke is referring to with the “B200,” since no chip by that moniker seems on the roadmap Nvidia shared with buyers final fall. Nevertheless, we suspect Clarke is definitely referring to the GB200 Superchip which, just like the GH200, is anticipated to mix Nvidia’s Grace CPU with its B100 GPU.
In line with an investor presentation launched this month, Nvidia plans to shift to a one-year launch cadence – Click on to enlarge
Based mostly on what we all know of the Grace CPU within the GH200, and assuming no main adjustments in energy consumption, that might put the GB200’s thermal design energy (TDP) someplace within the neighborhood of 1,300 watts – 30 % increased than its predecessor.
It is also doable that Nvidia has one other card up its sleeve that we do not find out about but. Particulars of the GPU big’s next-gen Blackwell structure stay scanty.
Nomenclature apart, Clarke instructed the forthcoming chip would supply a possibility to showcase Dell’s experience in different types of liquid cooling at scale. He referred to “issues in fluid chemistry and efficiency, our interconnect work, the telemetry we’re doing, the facility administration work we’re doing” as steps towards options to direct liquid cooling, even for very dense chips.
Nvidia declined to remark – as you’d count on, given its annual GTC convention is just a few weeks away. The Register will probably be onsite on the occasion to deliver you all the small print after they drop.
The B100 is not anticipated to launch till late 2024 after Nvidia’s bandwidth juiced H200 GPUs debut within the first half of the yr.
Introduced in late 2023, the H200 is a refresh of the H100 with as much as 141GB of HBM3e reminiscence that is good for a whopping 4.8TB/sec of bandwidth. Nvidia claims the system can double the efficiency of huge language fashions together with Llama 70B, due to the chip’s HBM3e reminiscence stacks.
Even with two new accelerators slated to hit the market this yr, analysts warn Nvidia’s provide of GPUs will stay provide constrained. That is regardless of reports predicting Nvidia might transfer greater than triple shipments of GPUs in 2024.
Past its new accelerators, Nvidia’s roadmap additionally requires sooner, extra succesful InfiniBand and Ethernet NICs and switches able to 800Gb/sec of bandwidth per port earlier than the yr is out. ®