Nvidia Announces Grace, Its First ARM-Based CPU for Data Centers

When Nvidia announced its plans to acquire ARM last year, many viewed the move as Nvidia's attempt to seize market share in the segment smartphone chipset. But at that time, Jen-Hsun Huang (CEO of Nvidia) explained that their focus in the near future would be in the field of data center and cloud.

He didn't seem to be talking. Nvidia has just introduced a new ARM-based CPU which they dubbed Grace, after one of the pioneers in the world of computer programming, Grace Hopper. Grace is Nvidia's first CPU designed for use in computers server on a data center, more or less the same as the Intel Xeon and AMD EPYC CPU lines.

The reason they designed the Grace is actually quite simple. Nvidia needs CPU server which supports interface NVLink, which allows communication between the CPU and GPU at very high speeds (at least 900 GB/s), well above that interface PCI Express offers today (approximately 30x faster).

With bandwidth At the same time, the very high speed that Grace offers, Nvidia also sees it as the most ideal CPU to pair with the next generation of GPU class. server made by them. To describe computer performance server which uses Grace in its entirety, Nvidia uses the scenario train system natural-language processing with one trillion parameters.

According to him, this work can be done at a speed 10x faster than if using a line of computers server made by them at this time, namely Nvidia DGX which relies on a CPU based on the x86 architecture.

Let's just say that now we need about a month to train a system natural language processing. With Grace, the time required can be reduced to just three days. It's no wonder that Nvidia immediately got a big client, even though Grace itself will only be released in 2023.

The client in question is the Swiss National Supercomputing Center (CSCS), which is currently building an AI supercomputer called Alps. The prediction is that Alps will become a supercomputer with the fastest AI performance (20 exaflop) when it is completed in 2023.

In addition, Nvidia also has plans to use Grace in Atlan, a chipset Nvidia's new design for self-driving cars, which is expected to arrive in 2025.

What Nvidia does is basically not much different from Apple. Apple, as we know, decided to design your own processor laptop ARM based because they were not satisfied with the limitations they encountered on processors based on the x86 architecture. Nvidia is the same way, it's just that the context is for computers server rather than consumer.

Sumber: Engadget and AnandTech.