NVIDIA Unveils 1-Exaflop AI Supercomputer | | | | Turtles AI

  While its stocks continue to skyrocket, #NVIDIA has announced a new #AI #supercomputer called the #DGX #GH200 which combines 256 of the company's Grace Hopper #superchips to enable massive AI and #machinelearning models. The supercomputer provides up to 1 #exaflop of performance and 144 terabytes of shared memory, 500 times more than NVIDIA's previous DGX A100 system. The DGX GH200 links the Grace Hopper superchips together using NVIDIA's #NVLink switch technology which allows all 256 #GPUs to work together as a single massive #GPU. This means developers can program the system as a single GPU even though it provides an exaflop of computational power and 144TB of memory. NVIDIA says the DGX GH200 will enable the next generation of generative AI, large language models and recommender systems. Google Cloud, Meta and Microsoft are among the first customers that plan to test the system for such workloads. NVIDIA is also building its own GH200-powered supercomputer, named NVIDIA Helios, to power the AI research and development work of its teams. The DGX GH200 will be a turnkey system for developers that includes NVIDIA software for managing and optimizing AI workloads. NVIDIA aims to make the GH200 design available as a blueprint for cloud providers and others to customize. The DGX GH200 is expected to be available by the end of the year.