Nvidia announces 144-core Arm-based 'Superchip'

Nvidia claims the Grace CPU Superchip has double the memory bandwidth and energy efficiency than today's leading server chips.

Nvidia has announced its first Arm Neoverse-based discrete data centre chip - named 'Grace CPU Superchip' - designed for AI and high performance computing (HPC) applications.

Announced at its annual GTC conference for AI developers on Tuesday, the Superchip is based on the latest data centre architecture, Arm v9, and comprises two CPU chips on a single socket (or package). The dual chips are connected over NVLink-C2C, a chip-to-chip interconnect with high speed and low latency.

The Grace CPU packs 144 cores, and - Nvidia says - offers double the memory bandwidth and energy efficiency of today's leading server chips. Specifically the company is aiming at AMD's Epyc 7742 processors, which pack 128 cores and 256 threads in total but are based on the Zen 2 core architecture.

The CPU achieves a projected score of 740 on the SPECrate 2017_int_base benchmark; a 1.5x improvement over the performance of the dual AMD CPUs shipping with the DGX A100.

The chip has a memory subsystem with up to 1TB of LPDDR5x memory and error correcting code capabilities, balancing power consumption and performance.

When it debuts, Nvidia expects the Grace CPU Superchip - with a 500-watt TDP package inclusive of memory - to provide twice the performance-per-watt of top-of-the-line CPUs.

Nvidia says the Grace CPU will run all of the company's computing software stacks, including Nvidia AI, Nvidia HPC, Nvidia RTX and Omniverse.

"A new type of data centre has emerged - AI factories that process and refine mountains of data to produce intelligence," said Jensen Huang, founder and CEO of Nvidia.

"The Grace CPU Superchip offers the highest performance, memory bandwidth and Nvidia software platforms in one chip and will shine as the CPU of the world's AI infrastructure."

Nvidia says the Grace CPU Superchip 'complements' the Grace Hopper Superchip, Nvidia's first CPU-GPU integrated module announced last year. The Grace Hopper chip uses one Grace CPU and a GPU based on the new Hopper architecture for AI and large-scale HPC applications.

'Both [the Grace CPU Superchip] and the Grace Hopper Superchip are expected to be available in the first half of 2023,' the company added.

Additionally, Nvidia has demonstrated its next-generation Hopper architecture, which will be released later this year.

This architecture is designed for Nvidia's data centre accelerators, and the firm is introducing it in the H100 GPU, which will replace Nvidia's previous A100.