NVIDIA super GPU and super CPU combined: Also integrated “network card”

NVIDIA super GPU and super CPU combined: Also integrated “network card”

NVIDIA super GPU and super CPU combined: Also integrated “network card”

On March 22, NVIDIA released a new generation of Hopper H100 GPU core and Grace CPU processor at the same time , and also put the two on one card, named ” Grace Hopper “, which is officially called the accelerated computing system with the highest density.

The two chips are placed side by side, and they communicate with each other through the NVLink bus, with a bandwidth of 900GB/s.

The specific specifications have not been announced yet, but it should not be able to be full at the same time, especially the Hopper GPU.

The Grace CPU is designed based on the ARMv9 architecture instruction set, 72 cores , integrated 198MB cache, partnered with LPDDR5X memory, and supports ECC.

Hopper GH100 GPU 18432 CUDA cores, 576 Tensor cores , 60MB L2 cache, 6144-bit HBM2e/HBM3 video memory width.

The H100 computing card only has 16896 CUDA cores, 528 Tensor cores, 50MB L2 cache, and 5120-bit HBM3 video memory bit width. The Grace Hopper card is also likely to have this specification.

What’s the use of this combined “monster”? Obviously, it is also doing HPC and AI computing. The CPU and GPU can coordinate and distribute different loads to maximize efficiency.

At the same time, NVIDIA also released the “Grace CPU SuperChip”, which integrates two Grace CPUs in a single card.

NVLink C2C interconnect, a total of 144 cores, 396MB cache, 1TB/s LPDDR5X ECC memory, power consumption 500W.

There is also “H100 CNX”, which integrates H100 GPU and CX-7 SmartNIC network chip.

Eliminate system bandwidth by connecting GPU directly to the network, support PCIe 5.0, and support up to 7 concurrent multi-GPU instances.