NVIDIA announces new DGX GH200 AI supercomputer with 1 exaflop performance and 144 terabytes of shared memory.
Uses NVLink interconnect technology and NVLink Switch System, allowing 256 GH200 superchips to perform as a single GPU.
GH200 superchips eliminate the need for traditional CPU-to-GPU PCIe connection, increasing bandwidth and reducing power consumption.
First supercomputer to pair Grace Hopper Superchips with NVIDIA NVLink Switch System, providing 48x more NVLink bandwidth than the previous generation.
Google Cloud, Meta, and Microsoft among the first to gain access to DGX GH200 for generative AI workloads, with NVIDIA providing design as a blueprint for cloud service providers and hyperscalers.
NVIDIA Helios, a DGX GH200-based AI supercomputer with 1,024 Grace Hopper Superchips, to come online by the end of the year to advance research and development.