When Nvidia announced its expansion into designing and manufacturing Arm-based CPUs last year, it was a huge surprise, and the addition of the Hopper GPUs just added to the surprise. Nvidia’s efforts are paying off, as the company announced that it’s Grace CPU Superchip and Grace Hopper Superchips would power two future supercomputers, even though the silicon has yet to be shipped.
The flagship Grace-only Venado supercomputer, which will give up to ten ‘AI ExaFlops’ of performance, will be deployed by the Department of Energy’s (DOE) Los Alamos National Laboratory.
Grace CPU Superchips will be installed in HPE Cray EX cabinets at the Swiss National Computing Center’s current 20 AI ExaFlops Alps supercomputer. This supercomputer is already in use by the research community as a general-purpose machine.
However, there is one caveat: ‘AI ExaFlops’ uses lower-precision math (FP32 or lower, such as INT8) than the official rankings, which use FP64, and we have yet to receive a confirmation on the data type used for the the the these projections. As a result, it’s unclear whether Venado will be able to deliver an ExaFlop of performance when measured with FP64, as Frontier, the first true Exascale-class supercomputer, did. Frontier has a 6.88 HPL-AI ExaFlops result, and it will be interesting to see if the Nvidia-powered systems can match or beat that.
The Grace supercomputer announcements follow Nvidia’s recent announcement that it has provided reference CGX, OVX, and HGX system designs for six major OEMs to ship in the first half of 2023, launching the company into the broader AI server market.
Nvidia and the Los Alamos National Laboratory (LANL) haven’t revealed many specifics about Venado yet, but we do know that it will have a mix of dual-CPU 144-core Grace CPU Superchips and the Grace Hopper Superchip, which is a single CPU + Hopper GPU combo. This will be the first Grace supercomputer to be built in the United States.
The Grace CPU Superchip is Nvidia’s first CPU-only Arm chip designed for the data center, and it comes as two chips on one motherboard, whereas the Grace Hopper Superchip combines a Hopper GPU and the Grace CPU on the same board. The Neoverse-based CPUs also support the Arm v9 instruction set, and systems include two chips fused using Nvidia’s newly branded NVLink-C2C interconnect technology.
Nvidia promises to deliver the fastest CPU Superchip in the market
Nvidia promises that when the Grace CPU Superchip delivers in early 2023, it will be the fastest processor on the market for a wide range of applications, including hyper-scale computing, data analytics, and scientific computing.
Nvidia GPUs are still the preferred accelerator for supercomputers and HPC, but Intel and AMD are dominating the CPU market with x86 as IBM fades away. With the release of its Arm-based Grace CPUs, Nvidia hopes to change that. Nvidia’s recent focus has been on systems that boast significant power in AI computing, which means performance in lower-precision and mixed-precision numerical representations, rather than the usual FP64 utilized in traditional HPC and supercomputing.
The two Grace-equipped supercomputers will join a slew of other Nvidia GPU-powered supercomputers, including NERSC’s Perlmutter system with 3.8 AI ExaFlops, Nvidia’s EOS with 18 AI ExaFlops, and CINECA’s Leonardo supercomputer with 10 AI ExaFlops.
Nvidia is currently developing some public-facing Grace Superchip-powered supercomputers. As the business gets closer to shipping the Grace Superchip silicon in early 2023, we expect that number to grow dramatically.
also read:
AYA Neo Air is to be launched with an OLED display and a price of $549
