CEO Jensen Hiang made a string of bulletins throughout his Computex keynote, together with particulars concerning the firm’s subsequent DGX supercomputer. Given the place the business is clearly heading, it shouldn’t come as a shock that the DGX GH200 is essentially about serving to corporations develop fashions.
The supercomputer makes use of a brand new NVLink Swap System to allow 256 GH200 Grace Hopper superchips to behave as a single GPU (every of the chips has an Arm-based Grace CPU and an H100 Tensor Core GPU). This, in response to NVIDIA, permits the DGX GH200 to ship 1 exaflop of efficiency and to have 144 terabytes of shared reminiscence. The corporate says that is practically 500 instances as a lot reminiscence as you’d discover in a single DGX A100 system.
For comparability, the of the High500 supercomputers lists as the one identified exascale system, having reached a efficiency of practically 1.2 exaflops on the Linmark benchmark. That is over twice the height efficiency of the second-placed system, Japan’s .
In impact, NVIDIA claims to have developed a supercomputer that may stand alongside essentially the most highly effective identified system on the planet (Meta is constructing one which it claims would be the quickest AI supercomputer on the planet as soon as it’s totally constructed out). NVIDIA says the structure of the DGX GH200 affords 10 instances extra bandwidth than the earlier era, “delivering the facility of an enormous AI supercomputer with the simplicity of programming a single GPU.”
Some massive names have an interest within the DGX GH200. Google Cloud, Meta and Microsoft needs to be among the many first corporations to achieve entry to the supercomputer to check the way it can deal with generative AI workloads. NVIDIA says DGX GH200 supercomputers needs to be obtainable by the top of 2023.
The corporate can also be constructing its personal supercomputer, Helios, that mixes 4 DGX GH200 methods. NVIDIA expects Helios to be on-line by the top of the yr.
Huang mentioned different generative AI developments throughout his keynote, together with one on the gaming entrance. NVIDIA Avatar Cloud Engine (ACE) for Video games is a service builders will have the ability to faucet into with a purpose to create customized AI fashions for speech, dialog and animation. NVIDIA says ACE for Video games can “give non-playable characters conversational expertise to allow them to reply to questions with lifelike personalities that evolve.”