Nvidia Enters the Arms Race with Homegrown "Grace" CPUs
takyon writes:
Nvidia Enters The Arms Race With Homegrown "Grace" CPUs:
Everything that Nvidia and the market has learned about CPUs in the past decade can be brought to bear in Grace CPUs, and now Nvidia itself is big enough to be the primary contractor for exascale-class machines or at the very least supply the CPU and GPU compute as well as the networking and DPU offload engines to a supplier such as Hewlett Packard Enterprise, which really doesn't mind making custom variants of its "Shasta" EX supercomputers. And that is precisely what will be happening with the future "Alps" supercomputer going into the Swiss National Supercomputing Center [(CSCS)] in early 2023, featuring a hybrid compute module that has the homegrown Grace Arm server CPU very tightly couple with a future Nvidia GPU accelerator.
[...] CSCS is not alone in getting a Grace-Hopper supercomputer. Los Alamos National Laboratory, one of the Department of Energy national labs in the United States, will also be getting a Grace-Hopper system supercomputer system HPE in early 2023.
[...] Nvidia estimates that a Grace-Hopper system will offer 10X the performance on training natural language models, reducing it from one month to three days, and will allow for real-time inference on a 500 billion parameter model on a single node. And on the HPC front, computational fluid dynamics and weather applications are expected to see performance on the Grace-Hopper systems as well. Other HPC applications will no doubt follow.
See also: For Arm-Driven Supercomputing, Nvidia is Right on Time
Intel CEO: 'We Are Playing Offense, Not Defense' In Response To NVIDIA Grace ARM CPU Announcement For Servers
Read more of this story at SoylentNews.