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

For years, we barely heard a glimpse of Nvidia on the CPU front, after the sluggish arrival of the Project Denver CPU and the accompanying mobile processors of Tegra K1 in 2014. But now the company is starting a big deal in CPUs again. with: the new Nvidia Grace, an arm-based processing chip specifically designed for AI data centers.

This is a good time for Nvidia to bend its arm: it is currently trying to buy Arm itself for $ 40 billion, setting it up specifically as an attempt to ‘create the world’s leading computer age for AI’, and this chip is perhaps the first proof point. Arm also has a moment in the consumer space for consumers, where Apple’s M1 chips recently improved our concept of laptop performance. Of course, this is also more competition for Intel, whose shares fell after the Nvidia announcement.

The new Grace is named after computer pioneer Grace Hopper, and it comes in 2023 to ‘deliver 10x the performance of today’s fastest servers on the most complex AI and high performance computing’, according to Nvidia. This will of course make it attractive for research companies building supercomputers, which the Swiss National Supercomputer Center (CSCS) and the Los Alamos National Laboratory have also already signed to build in 2023.

A Grace Next is also already on the roadmap for 2025. Here’s a slice from Nvidia’s GTC 2021 presentation where the news was announced:

I would recommend reading what our friends are after AnandTech has to say about where Grace can fit into the data center market and Nvidia’s ambitions. It is worth noting that Nvidia does not yet announce much in the form of specifications, but Nvidia does say that it contains a fourth-generation NVLink with a record connection of 900 GB / s between the CPU and the GPU. “This is critically larger than the CPU’s memory bandwidth, which means that NVIDIA’s GPUs will have a cohesive cache link to the CPU that has access to the memory of the full bandwidth system, and also the entire system. can give a single shared memory. address space, ”writes AnandTech.