Nvidia has introduced a brand new AI supercomputer known as the DGX GH200. Whereas the moniker of ‘supercomputer’ is usually simply assigned, if the specs shared by the corporate are to be believed, it might truly stay as much as the title. The most recent providing from Nvidia combines its 256 Grace Hopper superchips, which have been introduced in March 2022. It packs a 144TB GPU with confirmed 900 GBps GPU-to-GPU bandwidth. The corporate claims that the supercomputer is able to 1-exaflop of FP8 AI efficiency.
“Generative AI, massive language fashions and recommender techniques are the digital engines of modernity economic system. DGX GH200 AI supercomputers combine NVIDIA’s most superior accelerated computing and networking applied sciences to broaden the boundaries of AI,” stated Jensen Huang, founder and CEO of Nvidia. assertion.
Nvidia launched a brand new AI supercomputer
The corporate revealed that the DGX GH200’s giant shared reminiscence area makes use of NVLink Interconnect expertise to attach 256 GH200 Superchips with an NVLink swap system, permitting them to perform as a single GPU. It presents 1 exaflop of efficiency and 144TB of shared reminiscence. That is virtually 500x extra reminiscence than the earlier era Nvidia DGX A100, which was launched in 2020.
As a substitute of taking the standard CPU-to-GPU route, the GH200 Superchips join the Grace Hopper CPU and Tensor Core GPY collectively utilizing the NVLink C2C chip interconnect. Based on Nvidia, it will increase bandwidth between GPU and CPU by 7x and reduces interconnect energy consumption by greater than 5x in comparison with the most recent PCIe expertise.
The DGX GH200 is the primary supercomputer to mix Grace Hopper superchips with Nvidia NVLink. Change on system, a brand new interconnect that permits all GPUs within the DGX GH200 system to work concurrently, the corporate stated.
Moreover, Nvidia claims that the DGX GH200 structure gives 48x extra NVLink bandwidth than the earlier era making it a constant unit that delivers efficiency with optimized energy consumption.
