During his Computex keynote, NVIDIA CEO Jensen Huang unveiled several key developments, prominently featuring the DGX GH200 supercomputer, designed to accelerate the creation of generative AI models. This innovative supercomputer employs a new NVLink Switch System, enabling 256 GH200 Grace Hopper superchips to function as a unified GPU. Each chip integrates an Arm-based Grace CPU and an H100 Tensor Core GPU, allowing the DGX GH200 to achieve an impressive performance of 1 exaflop, coupled with 144 terabytes of shared memory—nearly 500 times the memory of a single DGX A100 system.
In the current landscape, the Top500 supercomputers list only one acknowledged exascale system, Frontier at Oak Ridge National Laboratory, which boasts a performance nearing 1.2 exaflops. This places it significantly ahead of the second-ranked Fugaku in Japan. NVIDIA aims to position the DGX GH200 in direct competition with the world’s most powerful supercomputers, including a forthcoming project by Meta that promises to be the fastest AI supercomputer upon completion. With a new architecture that offers ten times the bandwidth of its predecessor, the DGX GH200 simplifies the complex programming typically associated with supercomputers, making advanced AI capabilities more accessible.
Notable companies such as Google Cloud, Meta, and Microsoft are among the first expected to utilize the DGX GH200 to explore its generative AI capabilities. NVIDIA anticipates the availability of the DGX GH200 supercomputers by the end of 2023. Additionally, NVIDIA is developing Helios, a supercomputer that will integrate four DGX GH200 systems, with plans for it to be operational by year-end.
Huang also highlighted NVIDIA's advancements in gaming technology, introducing the Avatar Cloud Engine (ACE) for Games. This service will enable developers to create tailored AI models for speech, conversation, and animation, enhancing the conversational abilities of non-playable characters to interact with players in a more lifelike manner.