Nvidia released the first CPU code-named Grace based on the ARM architecture

Nvidia recently announced the launch of its first central processing unit code-named Grace. This combined processor is based on the ARM micro-architecture and uses the Neoverse core.

This new processor is designed by Nvidia for modern data centers. The processor is composed of multiple chips including the CPU part, GPU part, and I/O.

Although the company has not revealed more details about the processor, Nvidia said that the processor can get a good score of 300 points on the SPECrate2017_int_base throughput benchmark.

Grace processor is a highly specialized central processing unit, this processor is mainly used to solve workload problems, especially ultra-high workloads.

For example, when training an NLP model with more than 1 trillion parameters, the Grace processor can provide extremely high computing performance, which helps to complete model training quickly.

When this processor is used in conjunction with the Nvidia series of high-performance graphics cards, it provides more than 10 times faster performance than traditional computers based on Nvidia graphics.

Why did Nvidia launch this self-developed processor? The reason is that the amount and scale of data used in current artificial intelligence technology and model training are growing exponentially.

Large-scale artificial intelligence systems need to train models with up to 1 billion parameters, and on average, the parameters included in model training will double every 75 days.

For developers, although the current hardware can meet their needs, they do need to consider the system bottleneck problem, and the Grace processor solves the system bottleneck.

The release of this processor does not mean that the processor is currently available. In fact, the Grace processor is still under development and has not yet been officially launched.

According to NVIDIA’s instructions, the processor uses Nvidia NVLink technology to increase the connection speed between the CPU and GPU to 900GB/s to increase throughput.

At the same time, the processor will also use an innovative LPDDR5x memory subsystem that has much higher bandwidth and performance than DDR4 memory, and the purpose is to improve performance.

Nvidia plans to cooperate with industry companies and communities to provide software packages and various library support to speed up scientists and researchers to deal with the most important issues in the world.

The first batch of processors equipped with this processor will be the Swiss National Supercomputing Centre (CSCS and the Los Alamos National Laboratory, where HPE’s Cray group will provide the laboratory with supercomputers powered by Grace in 2023.

Via: AnandTech