Nvidia is building new Arm CPUs again: Nvidia Grace, for the data center

We’ve barely heard a peep out of Nvidia on the CPU entrance for years, after the lackluster arrival of its Venture Denver CPU and its related Tegra K1 cellular processors in 2014 — however now, the firm’s getting again into CPUs in an enormous method with the new Nvidia Grace, a Arm-based processing chip particularly designed for AI data facilities.

It’s a superb time for Nvidia to be flexing its Arm: it’s at the moment making an attempt to purchase Arm itself for $40 billion, pitching it particularly as an try “to create the world’s premier computing firm for the age of AI,” and this chip could be the first proof level. Arm is having a second in the client computing house as effectively, the place Apple’s M1 chips lately upended our idea of laptop computer efficiency. It’s additionally extra competitors for Intel, after all, whose shares dipped after the Nvidia announcement.

The new Grace is named after computing pioneer Grace Hopper, and it’s coming in 2023 to carry “10x the efficiency of as we speak’s quickest servers on the most advanced AI and excessive efficiency computing workloads,” in accordance with Nvidia. That can make it engaging to analysis corporations building supercomputers, after all, of which the Swiss Nationwide Supercomputing Centre (CSCS) and Los Alamos Nationwide Laboratory are already signed up to build in 2023 as effectively.

A Grace Subsequent is already on the roadmap for 2025, too. Right here’s a slide from Nvidia’s GTC 2021 presentation the place it introduced the information:


I’d advocate studying what our friends at AnandTech have to say about the place Grace may match into the data center market and Nvidia’s ambitions, and it’s price noting that Nvidia isn’t releasing a lot in the method of specs simply but — however Nvidia does say it includes a fourth-gen NVLink with a document 900 GB/s interconnect between the CPU and GPU. “Critically, this is higher than the reminiscence bandwidth of the CPU, which implies that NVIDIA’s GPUs could have a cache coherent hyperlink to the CPU that may entry the system reminiscence at full bandwidth, and in addition permitting the complete system to have a single shared reminiscence deal with house,” writes AnandTech.

Show More

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button