NVIDIA has announced its DGX GH200 supercomputer oriented to artificial intelligence systems and made up of 256 “superchips” NVIDIA GH200 GPUs+CPUs with Grace Hooper architecture working as a single GPU thanks to the interconnection system NV Link.
These 256 chips working together manage to offer 1 exaflop of power together with nothing more and nothing less than 144 Terabytes of shared memory, 500 times more than the last generation NVIDIA DGX A100.
Each of their GPUs have a ARM NVIDIA Grace CPU together with the GPU with Tensor Cores NVIDIA H100, in this way both the CPU and the GPU are interconnected in the same package and the need to use PCI Express is avoided, slower (7 times slower in particular) and with higher consumption . Each of these chips is unified with the NVLink-C2C system making them work as a single GPU.
In fact, the NVIDIA DGX GH200 is the world’s first supercomputer to combine Grace Gooper chips with the NVLink interconnect switch system, achieving 48 times more NVLink bandwidth than previous generations.
Its main orientation is to execute generative Artificial Intelligence systems such as GPT and other similar ones. They will initially be available on Google Cloud, Meta (Facebook) and Microsoft systems, although NVIDIA will also offer the DGX GH200 as a base for other cloud service providers.
The first units will be available later this year.
End of Article. Tell us something in the Comments!
Anthony Delgado
Computer Engineer by training, writer and hardware analyst at Geeknetic since 2011. I love gutting everything that comes my way, especially the latest hardware that we receive here for reviews. In my free time I mess around with 3d printers, drones and other gadgets. For anything here you have me.
