An NVIDIA DGXTM supercomputer powered by NVIDIA GH200 Grace Hopper Superchips and the NVIDIA NVLink Switch System was unveiled as a new type of large-memory AI supercomputer.
It was designed to enable the creation of enormous, next-generation models for generative AI language applications, recommender systems and data analytics workloads, according to a May 28 news release.
“Generative AI, large language models and recommender systems are the digital engines of the modern economy,” NVIDIA Founder and CEO Jensen Huang said in the release. “DGX GH200 AI supercomputers integrate NVIDIA’s most advanced accelerated computing and networking technologies to expand the frontier of AI.”
The NVIDIA DGX GH200's enormous shared memory area combines 256 GH200 superchips into one GPU using the NVLink connection technology and NVLink Switch System, the release reported. This offers 144 terabytes of shared memory and 1 exaflop of performance, which is roughly 500 times more memory than the 2020-released NVIDIA DGX A100.
Using NVIDIA NVLink-C2C chip interconnects, GH200 superchips combine an Arm-based NVIDIA GraceTM CPU and an NVIDIA H100 Tensor Core GPU in a single package, doing away with the need for a conventional CPU-to-GPU PCIe connection, according to a news release. This offers a 600GB Hopper architecture GPU building block for DGX GH200 supercomputers and boosts interconnect bandwidth between GPU and CPU by seven times, compared to the most recent PCIe technology. It also reduces interconnect power consumption by more than five times.
The NVIDIA NVLink Switch System, a revolutionary connection that enables all GPUs in a DGX GH200 system to function as one, is paired with Grace Hopper Superchips in DGX GH200, making it the first supercomputer to do so, the release reported. Only eight GPUs could be integrated with NVLink into a single GPU in the previous generation's architecture without sacrificing performance.
By offering 48 times more NVLink bandwidth than the previous generation, the DGX GH200 architecture combines the simplicity of programming a single GPU with the power of a powerful AI supercomputer, the release said.
