NVIDIA evokes a processor
for data center integrating 144 cores in order to compete directly with EPYC and Xeon.
During its 2022 GPU Technology Conference (GTC), NVIDIA had some new GPUs to showcase and, notably, the Hopper-generation H100, but the American also had a CPU surprise in store, with a chip that we assume could have benefit from the takeover of the ARM company.
144 ARM Neoverse cores
If the acquisition in question did not succeed, NVIDIA is still working on the famous ARM architecture with the Grace CPU Superchip processor. It is based on the ARM Neoverse and is intended for data centers.
Of course, NVIDIA is full of praise for its new creation stating ” deliver the highest level of performance, twice the memory bandwidth and twice the power efficiency of today’s best server chips “.
As presented by NVIDIA, the Grace CPU Superchip is the assembly of two Grace CPUs connected by an NVLink C2C. It is a question of integrating 144 ARM Neoverse cores and having a memory bandwidth of 1 TB/s associated with LPDDR5x equipped with ECC error correction. The total consumption would be 500 Watts.
NVIDIA compares the performance of its Grace CPU Superchip to a DGX A100 solution integrating two AMD EPYC Rome 7742 64-core processors. On the SPECrate_2017 benchmark, NVIDIA would then be 1.5x faster… but the EPYC Milan are coming to AMD.
Just introduced by NVIDIA, the Grace CPU Superchip still has a long way to go. The firm of Jensen Huang, founder and CEO, estimates that the release should take place in early 2023, but there will obviously be no question of a consumer processor.
On the same subject :
NVIDIA: a first shot of the Hopper H100, the huge GPU for servers
Source: NVIDIA
2