Nvidia unveils new AI Supercomputer DGX GH200; Features 256 Grace Hopper Superchips

tech.hindustantimes.com:

Nvidia has announced a new AI supercomputer called DGX GH200. While the moniker of ‘supercomputer' is sometimes handed out easily, if the specifications shared by the company are to be believed, it may actually live up to the title. The latest offering by Nvidia connects 256 of its Grace Hopper Superchips, which was first announced in March 2022. It packs together a 144TB GPU which proved 900 GBps of GPU-to-GPU bandwidth. The company claims that the supercomputer is capable of 1-exaflop of FP8 AI performance.

“Generative AI, large language models and recommender systems are the digital engines of the modern economy. DGX GH200 AI supercomputers integrate NVIDIA's most advanced accelerated computing and networking technologies to expand the frontier of AI,” said Jensen Huang, founder and CEO of Nvidia in a statement.

The company has revealed that the DGX GH200's large shared memory space uses NVLink interconnect technology with the NVLink Switch System to combine 256 GH200 Superchips, allowing them to perform as a single GPU. This provides 1 exaflop of performance and 144TB of shared memory. This is nearly 500x more memory than the previous generation Nvidia DGX A100, which was introduced in 2020.

Instead of taking the traditional CPU-to-GPU route, the GH200 Superchips combines the Grace Hopper CPU and Tensor Core GPY together using NVLink C2C chip interconnects. According to Nvidia, it increases the bandwidth between GPU and CPU by 7x compared with the latest PCIe technology and reduces interconnect power consumption by more than 5x.

DGX GH200 is the first supercomputer to pair Grace Hopper Superchips with the Nvidia NVLink Switch System, a new interconnect that enables all GPUs in a DGX GH200 system to work together as one,