May 30, 2023 11:24:00

NVIDIA announces large-scale supercomputer 'DGX GH200' for generation AI with processing capacity of 1 ExaFLOPS and memory of 144 TB

NVIDIA CEO Jensen Huang will be presenting large-scale AI such as generative AI training and large-scale language model workloads at

COMPUTEX TAIPEI 2023 held in Taipei, Taiwan from May 30, 2023 (Tuesday). We announced the DGX GH200 supercomputer for workloads. The processing performance of the DGX GH200 reaches 1 exaFLOPS (1000 petaFLOPS), and it is planned to be operated on a trial basis with cloud computing such as Google and Microsoft.

DGX GH200 for Large Memory AI Supercomputer | NVIDIA
https://www.nvidia.com/en-us/data-center/dgx-gh200/

Announcing NVIDIA DGX GH200: The First 100 Terabyte GPU Memory System | NVIDIA Technical Blog
https://developer.nvidia.com/blog/announcing-nvidia-dgx-gh200-first-100-terabyte-gpu-memory-system/

Nvidia Unveils DGX GH200 Supercomputer, Grace Hopper Superchips in Production | Tom's Hardware
https://www.tomshardware.com/news/nvidia-unveils-dgx-gh200-supercomputer-and-mgx-systems-grace-hopper-superchips-in-production

First, CEO Huang announced that the NVIDIA GH200 Grace Hopper Superchip, a chipset for AI and high-performance computing (HPC) applications that combines NVIDIA's own Arm CPU ' Grace ' and GPU ' Hopper ' for AI, has entered production. announced that

This NVIDIA GH200 Grace Hopper Superchip connects Grace and Hopper with

NVIDIA NVLink-C2C technology, realizing a bandwidth of 900 GB / s, which is about seven times that of PCI Express 5.0 connection. The computing performance is 4 Peta FLOPS, and the memory is equipped with 96 GB of HBM3 and 512 GB of LPDDR5X .

'DGX GH200' is a platform that combines this NVIDIA GH200 Grace Hopper Superchip and

NVLink / NVSwitch for multi-GPU communication, integrates up to 256 Hopper GPUs and can be executed as a single GPU. By integrating 256 NVIDIA GH200 Grace Hopper Superchips, a total of 144 TB of shared memory, which is about 500 times that of the previous generation DGX A100, can be used as GPU memory, and the computing power reaches 1 ExaFLOPS. .

Ian Buck, vice president of hyperscale and HPC at NVIDIA, said, ``When performing arithmetic processing using a huge model such as generative AI or a large-scale language model, the memory capacity has already reached its limit. AI researchers need huge memory capacity of terabyte size or more, and DGX GH200 can meet such needs by allowing up to 256 Hopper GPUs to be integrated and run as a single GPU. It will be possible, ”he said, appealing that the DGX GH200 will be a big presence to eliminate the bottlenecks of generative AI and large-scale language models.

Already Google Cloud, Meta and Microsoft will have early access to the DGX GH200 as they figure out its capabilities in generative AI workloads.

“Building advanced generative models requires an innovative approach to AI infrastructure,” said Mark Romeyer, vice president of computing at Google Cloud. Large shared memory eliminates a major bottleneck that can occur in large-scale AI workloads, and we look forward to Google Cloud exploring its capabilities.'

'NVIDIA's Grace Hopper design allows us to explore new approaches to solving our researchers' biggest challenges,' said Alexis Bjorlin, vice president of infrastructure, AI systems and accelerated platforms at Meta. 'The ability of the DGX GH200 to handle terabyte-sized datasets will help developers to allows us to conduct larger, faster and more advanced research.'

In addition, NVIDIA is proposing a supercomputer 'NVIDIA Helios' that combines four DGX GH200s. In NVIDIA Helios, each DGX GH200 is interconnected by the NVIDIA Quantum-2 InfiniBand platform . So in total you will be able to use 1024 NVIDIA GH200 Grace Hopper Superchips as one unit. This NVIDIA Helios is scheduled to be online by the end of 2023.

You can see the full story of CEO Huang's keynote speech at COMPUTEX TAIPEI 2023 below.

NVIDIA Keynote at COMPUTEX 2023-YouTube