The company is also preparing the Helios supercomputer, which will be four times more powerful.
Nvidia today announced a new class of AI supercomputer with large memory capacity, the Nvidia DGX supercomputer based on Nvidia GH200 Grace Hopper superchips and Nvidia NVLink switching system, designed to develop new generative AI applications, recommender systems, and big data processing.
The Nvidia DGX GH200 uses NVLink Switch System technology to combine 256 GH200 superchips, allowing them to operate as a single GPU. This delivers 1 EFLOPs of performance and 144TB of total memory—nearly 500 times more memory than the previous generation Nvidia DGX A100 introduced in 2020.
Nvidia’s latest supercomputer unveiled: 256 GH200 superchips
“Generative AI, large language models, and recommender systems are the digital engines of today’s economy,” said Jensen Huang, founder and CEO of Nvidia. “The DGX GH200 AI Supercomputer combines Nvidia’s most advanced accelerated computing and networking technologies to push the boundaries of AI.”
The GH200 superchips eliminate the need for traditional PCIe CPU-GPU connectivity by combining an Arm-based Nvidia Grace CPU with an Nvidia H100 Tensor Core GPU in a single package using Nvidia’s NVLink-C2C chip. This increases the throughput between GPU and CPU by 7x compared to the latest PCIe technology, reduces interconnect power consumption by more than 5x, and provides a 600GB Hopper architecture GPU building block for DGX GH200 supercomputers.
The DGX GH200 is the first supercomputer to combine Grace Hopper superchips with Nvidia’s NVLink switching system, which allows all GPUs in the DGX GH200 system to work together as one. The previous generation system had only eight GPUs, which could be combined with NVLink into a single GPU without sacrificing performance.
The DGX GH200 architecture delivers 48 times the throughput of NVLink than the previous generation, delivering the power of a massive AI supercomputer with the ease of programming a single GPU.
Nvidia is building its own DGX GH200-based AI supercomputer to power its research and development teams. The supercomputer, called Nvidia Helios, will be powered by four DGX GH200 systems. Each will be connected to an Nvidia Quantum-2 InfiniBand network to increase data throughput for training large AI models. Helios will include 1,024 Grace Hopper superchips and is expected to be launched by the end of the year.
