Developers: | Nvidia |
Date of the premiere of the system: | May 2023 |
Branches: | Electrical and Microelectronics |
Technology: | Supercomputer |
2023: Product Announcement
At the end of May 2023, the company Nvidia introduced the DGX GH200 supercomputer, which, according to the developers, is unique in its kind, since it is tailored for the work of generative models (artificial intelligence these are the basis of the popular neural network). ChatGPT
The novelty is based on the NVLink system, which connects 256 Grace Hopper superchips into one graphic one. processor Thanks to this technology, a record computer power is achieved, as well as an incredible amount of memory of 144 TB. This is 500 times more than one DGX A100 system provides.
GH200 Grace Hopper chips can work as a single GPU (each of these superchips combines a 72-core Grace CPU based on Arm, H100-class GPU, 96 GB of HBM3 and 512 GB of LPDDR5X memory). Nvidia claims that this configuration allows DGX GH200 to develop 1 exaflop performance and have 144 terabytes of total memory. The company claims that this is almost 500 times more than in the previous generation supercomputer solution, the DGX A100.
By comparison, the latest supercomputer ranking by May 2023 named Frontier Top500 only known exaflop system, which achieved nearly 1.2 exaflops in the Linmark benchmark.
The DGX GH200 architecture provides 48 times more NVLink bandwidth than the previous generation, providing massive power supercomputer AI with simple programming of a single GPU.
The first companies using DGX GH200 will be Microsoft and Google. Supercomputers will be used to train large language models and other computations related to neural networks and artificial intelligence. The cost of DGX GH200 was not reported by the end of May 2023.[1]