The SC24 Nvidia’s newest HPC and AI chip is a unmarried supercomputer packing 4 Blackwell GPUs, 144 Arm Neoverse cores, as much as 1.3 terabytes of HBM, and a 5.4 kilowatt TDP warmth sink. In some ways, the brand new GB200 NVL4 chip, unveiled at Supercomputing in Atlanta this week, seems like two of Nvidia’s upcoming Grace-Blackwell Superchips put in combination. Then again, not like the two.7 kilowatt GB200 forums now we have noticed prior to, the short NVLink connectivity present in Nvidia’s NVL4 replace is restricted to 4 Blackwell GPUs and two Grace CPUs on board. All I/O to and from the board will probably be treated via your Ethernet or InfiniBand NICs.
The GB200 NVL4 is in reality the GB200 Superchips attached in combination minus the exterior NVLink – Click on to amplify Even supposing this will likely appear absurd, it suits smartly with the collection of HPCs constructed. For instance, the Cray EX Blades to be had at Frontier had a third-generation Epyc CPU in conjunction with 4 MI250X accelerators. This additionally implies that primary HPC machine developers akin to HPE, Eviden, and Lenovo don’t seem to be all the time tying along with Nvidia so as to add out and in. Each HPE and Eviden have their very own networking generation. In reality, HPE has already teased the brand new EX methods, anticipated in overdue 2025, which is able to use Nvidia’s GB200 NVL4 forums. The EX154n, which used to be introduced remaining week, will pack as much as 56 of the bigger Superchips – one according to blade – in its huge water-cooled HPC cupboards. On this configuration, a unmarried EX cupboard can ship 10 petaFLOPS of FP64 vector or matrix compute. This may occasionally sound like so much, but when high-precision clinical computing is what you care about, AMD’s HPE methods be offering the best efficiency. The MI300A APUs present in Cray’s EX255a portfolio every boast 61.3 teraFLOPS of vector FP64 or 122.6 teraFLOPS of matrix FP64 in comparison to 45 teraFLOPS of high-precision vector/matrix efficiency noticed in every Blackwell GPU. For extra AI-centric packages, the efficiency distinction is minimum as every MI300A can ship 3.9 petaFLOPS of restricted FP8 efficiency. So, on a rack filled with EX, you would be having a look at round 2 exaFLOPS of FP8 grunt in comparison to round 2.2 exaFLOPS from a Blackwell machine the usage of part as many GPUs – double that if you’ll make the most of FP4 datatypes no longer supported via Symbol for the MI300A Even supposing HPE Cray is likely one of the first to announce enhance for Nvidia’s NVL4 shape issue, we do not be expecting it to be lengthy prior to Eviden, Lenovo, and others get started freeing their very own forums and servers in accordance with the design. H200 PCIe playing cards get NVL improve Together with Nvidia’s GB200 NVL4 dual-core, Nvidia has additionally introduced the supply of a PCIe-based H200 NVL config. However prior to you get too excited, very similar to the H100 NVL we had firstly of 2023, the H200 NVL is in reality a bunch of 2 PCIe playing cards – as much as 4 this time – attached in combination via an NVLink bridge. As with Nvidia’s major SXM platforms from the DGX and HGX, the NVLink bridge lets in GPUs to make use of the compute and reminiscence assets to accomplish huge duties with out blocking off the gradual PCIe 5.0 x16 connection that maxes out at 128 GBps of bandwidth in comparison to 900 GBps. for NVLink. Moreover, the H200 NVL can enhance as much as 564 GB of HBM3e reminiscence and 13.3 petaFLOPS of excessive FP8 efficiency and sparsity. Once more, it is because they’re 4 H200 PCIe playing cards attached via a high-speed connection bridge. Then again, this all comes all the way down to power and therms. Every H200 card in a stack of 4 is rated for as much as 600 W of energy or 2.4 kilowatts overall. I say this, the process has its benefits. For something, those playing cards will also be deployed in nearly any 19-inch server with enough room, energy, and air go with the flow for cooling. ®
