Nvidia’s enlargement into designing and producing Arm-based CPUs was fairly the stunning revelation when the corporate introduced it final yr, and the addition of the Hopper GPUs into the design upped the ante. Regardless that Nvidia has but to ship the silicon, these efforts proceed to repay as the corporate introduced that its Grace CPU Superchip and Grace Hopper Superchips would energy two future supercomputers. The Division of Power’s (DOE) Los Alamos Nationwide Laboratory will deploy the flagship Venado supercomputer that may ship as much as ten ‘AI ExaFlops’ of efficiency.
There’s a caveat, although: ‘AI ExaFlops’ encompass lower-precision math (FP32 or decrease, like INT8) than we see with the official rankings that use FP64, however we’ve got but to obtain a affirmation on the information kind used for these projections. As such, it’s unclear if Venado will ship an ExaFlop of efficiency as measured with FP64, as we see with Frontier, the primary true Exascale-class supercomputer. Notably, Frontier has achieved a 6.88 HPL-AI ExaFlops end result, and will probably be attention-grabbing to see if the Nvidia-powered methods can match or exceed that benchmark.
The Swiss Nationwide Computing Middle may even deploy a brand new as-yet-unnamed system with the Grace CPU Superchip in HPE Cray EX cupboards. This new supercomputer shall be a general-purpose machine for the analysis neighborhood, however efficiency projections have not been shared.
Nvidia’s Grace supercomputer bulletins come scorching on the heels of the corporate’s current announcement that it has offered reference CGX, OVX, and HGX system designs that may ship from six main OEMs within the first half of 2023, springboarding the corporate into the broader AI server market.
Nvidia and the Los Alamos Nationwide Laboratory (LANL) aren’t sharing many particulars about Venado but, however we do know that it’ll include a mix of the dual-CPU 144-core Grace CPU Superchips and the only CPU + Hopper GPU combo that comes within the type of the Grace Hopper Superchip. This would be the first US-based Grace supercomputer.
As a fast reminder, Nvidia’s Grace CPU Superchip is the corporate’s first CPU-only Arm chip designed for the information heart and comes as two chips on one motherboard, whereas the Grace Hopper Superchip combines a Hopper GPU and the Grace CPU on the identical board. As well as, the Neoverse-based CPUs help the Arm v9 instruction set and methods include two chips fused along with Nvidia’s newly branded NVLink-C2C interconnect tech.
General, Nvidia claims the Grace CPU Superchip would be the quickest processor available on the market when it ships in early 2023 for a variety of purposes, like hyperscale computing, knowledge analytics, and scientific computing. (You possibly can learn the deep-dive particulars concerning the Grace Superchip silicon right here.)
As you possibly can see within the album above, Nvidia GPUs nonetheless function the go-to accelerator for supercomputers and in HPC, however Intel and AMD dominate the CPU panorama as IBM recedes. So naturally, Nvidia needs to vary that with the arrival of its Arm-based Grace CPUs. Nevertheless, it is vital to notice that Nvidia’s focus of late has been on methods that tout great energy in AI computing, that means efficiency in lower-precision and mixed-precision numerical codecs, not the usual FP64 that is utilized in conventional HPC and supercomputing. As such, Nvidia touts its efficiency in ‘AI ExaFlops,’ and above, we are able to see a roster of supercomputers that Nvidia says already meet this criterion.
The 2 new Grace-based supercomputers will be a part of a plethora of different Nvidia GPU-powered supercomputers, together with the three.8 AI ExaFlops Perlmutter system at NERSC, Nvidia’s personal EOS with 18 AI ExaFlops, and the ten AI ExaFlops Leonardo supercomputer at CINECA.
Nvidia additionally powers the 20 AI ExaFlops Alps system at CSCS (opens in new tab), and like we’ll see with Venado, this method makes use of Nvidia’s Grace CPU. Together with the Swiss system additionally introduced right now, Nvidia has three public-facing Grace Superchip-powered supercomputers within the works. We anticipate that quantity to increase quickly as the corporate works nearer to transport the Grace Superchip silicon in early 2023.