Nvidia Introduces DGX GH200 Supercomputer With 256 Grace CPUs And H100 GPUs – Computer – News


Nvidia reviews at Computex that its Grace Hopper “superchips” for knowledge facilities are actually totally in manufacturing. The corporate additionally exhibits a DGX GH200 supercomputer, which consists of such chips and has 144TB of shared reminiscence.

De Nvidia DGX GH200 options 256 Grace Hopper-superchips. These are chips that mix an H100 GPU with an Nvidia Grace CPU on a single module of about 200 billion transistors. Every Grace Hopper superchip has 72 Neoverse V2 CPU cores and 16,896 CUDA cores, 96GB of HBM3 reminiscence and 480GB of Lpddr5 reminiscence. This provides your entire DGX GH200 system a complete of 18,432 CPU cores, over 4.3 million CUDA cores and 144TB of shared reminiscence.


Nvidia’s Grace Hopper-superchip

The Grace Hopper superchips are interconnected with Nvidia’s proprietary NVLink interconnect, which permits all GPUs to entry one another’s reminiscence. Which means the system features as a single GPU. That NVLink interconnect makes use of 96 L1 switches and 32 L2 switches, reviews the producer. Nvidia says the GPU-to-GPU bandwidth is 900GB/s. For instance: a PCIe 5.0 x16 GPU has a bandwidth of about 63GB/s.

Nvidia says the DGX H200 is for giant AI workloads. In line with the producer, the system has ‘1 exaflop’ of fp8 computing energy for this. The producer doesn’t share efficiency figures for different compute codecs. The corporate does share some benchmarks evaluating the DGX GH200 to a present DGX A100 system, reviews Tom’s {Hardware}. On this, the brand new DGX GH200 system is 2.2 to six.3x quicker. Nevertheless, present DGX A100 techniques have 32 to 256 A100 GPUs, relying on the benchmark. Nvidia’s DGX GH200 system has 256 GPUs in all benchmarks proven.

In line with the producer, Google Cloud, Meta and Microsoft would be the first to get entry to the DGX GH200 supercomputer. In time, Nvidia needs to make the blueprint for the system obtainable to cloud suppliers and different hyperscalers. The DGX GH200 ought to be launched on the finish of this yr. The producer doesn’t share a worth.

Nvidia itself goes to construct one other Helios supercomputer that mixes 4 DGX GH200 techniques. These 4 techniques, which collectively have 1024 Grace Hopper superchips, are related to Nvidia’s Quantum-2 InfiniBand with a bandwidth of 400Gbit/s. That system ought to come on-line on the finish of this yr.

Bron: Nvidia, Tom’s {Hardware}