Nvidia’s Grace is an ARM-based processor with high bandwidth

A few years ago Protected trademark in the name of Nvidia Hopper, Named after programming pioneer Grace Hopper, then rumors of naming a new architecture have emerged. At this year’s GPU Technology Conference (GTC) Presents Nvidia Grace, The company’s first processor for a data center and server environment.

Nvidia Grace is based on ARM and aims to meet the needs and performance requirements of the most advanced and challenging accounts. The focus is on research in artificial intelligence (AI), for example in Natural language processing (NPL). NPL is used to teach artificial intelligence and conversation interfaces a contextual understanding of human language.

The new processor has been specially developed to be integrated with Nvidia graphics chips. At the moment, there is a missing product for deep learning purposes that the company owns, and the use of PCI Express to communicate between the processor and the graphics card creates a bottleneck in terms of bandwidth. Nvidia graphics cards already use Nvlink technology in between to get around it.

Nvlink is implemented for the same reason at Grace and allows for speeds of up to 900 GB / s between the processor and the graphics card. This means Grace-powered systems should be able to deliver 10 times higher performance than Nvidia’s current DGX series with x86 architecture and AMD processor. With Nvlink, it is also possible to have a cache coherent memory address space.

In addition to Nvlink, Grace was revealed to be equipped with an LPDDR5X with an Error Correction Code (ECC) which according to Nvidia doubles memory bandwidth and provides ten times more energy efficiency compared to the memories used by today’s servers. Adding an Error Correction Code (ECC) also reduces the risk of data corruption.

It takes two years before the first Grace processors are in use and Nvidia is relatively little with details. Nothing has been said about performance other than using ARM’s Neoverse cores. Among other things, it is still not known what the current manufacturing technology or the number of cores is, however According to Anandtech påstår Nvidia at “Cores must break 300 points on the SPECrate2017_int_base”.

Interesting in this context is that Nvidia has chosen the codename Grace for the wizard. Traditionally, Nvidia has always used the titles of famous inventors and researchers in its graphics structures. As Hopper has appeared in several previous reports, it is possible that the processor in addition to the first name Grace will be accompanied by a new graphics architecture with the last name Hopper.

Grace is a niche product dedicated to a few special slides. The first to gain access to the devices were researchers at CSCS and Los Alamos National Laboratory in 2023.

Read more about the codename Hopper: