Buzz
(Homeland) - Nvidia has recently introduced the next-generation supercomputer platform named DGX GH200.
Recently, Nvidia introduced the next iteration of supercomputing technology, the DGX GH200, at the ongoing Computex 2023 event. According to Nvidia, the DGX GH200 AI supercomputing platform is poised to help customers craft success stories similar to ChatGPT.
Thanks to the integration of 256 supercharged Grace Hopper Superchips with a shared memory of 144TB, DGX GH200 can efficiently tackle the most resource-intensive tasks in artificial intelligence training.
The prowess of each superchip Grace Hopper is derived from a 72-core Grace ARM CPU, a Hopper GPU, 96GB HBM3 memory, and 512GB LPDDR5X memory all on the same chip, amounting to a total of 200 billion semiconductor balls.
The DGX A100 system, a powerhouse in artificial intelligence computing that contributed significantly to the success of OpenAI's ChatGPT, can only interconnect 8 A100 GPUs into a unified processor with a shared memory of 320GB.
To link the 256 Grace Hopper chips, DGX GH200 utilizes the innovative NVLink Switch System, allowing them to operate as a colossal GPU with superior processing capabilities – featuring a shared memory of 144TB, more than 500 times that of the DGX A100 system (used by ChatGPT).
Nvidia reveals that the entire DGX GH200 system employs up to 240 km of high-speed optical fiber cables, weighing over 18 tons but operating as a single GPU.
This supercomputer system brings 'artificial intelligence performance' up to 1 exaflop – nearly equivalent to the Frontier supercomputer at the Oak Ridge National Laboratory in Tennessee, boasting a performance score reaching close to 1.2 exaflops.
