NVIDIA has launched its next-generation of AI supercomputer chips that may seemingly play a big function in future breakthroughs in deep studying and huge language fashions (LLMs) like OpenAI's GPT-4, the corporate introduced. The know-how represents a major leap over the past generation and is poised for use in information facilities and supercomputers — engaged on duties like climate and local weather prediction, drug discovery, quantum computing and extra.
The important thing product is the HGX H200 GPU based mostly on NVIDIA's "Hopper" structure, a alternative for the favored H100 GPU. It's the corporate's first chip to make use of HBM3e reminiscence that's sooner and has extra capability, thus making it higher fitted to massive language fashions. "With HBM3e, the NVIDIA H200 delivers 141GB of reminiscence at 4.8 terabytes per second, practically double the capability and a pair of.4x extra bandwidth in contrast with its predecessor, the NVIDIA A100," the corporate wrote.
In phrases of advantages for AI, NVIDIA says the HGX H200 doubles inference velocity on Llama 2, a 70 billion-parameter LLM, in comparison with the H100. It'll be out there in 4- and 8-way configurations which can be appropriate with each the software program and {hardware} in H100 methods. It'll work in each sort of information middle, (on-premises, cloud, hybrid-cloud and edge), and be deployed by Amazon Net Providers, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure, amongst others. It's set to reach in Q2 2024.
NVIDIA's different key product is the GH200 Grace Hopper "superchip" that marries the HGX H200 GPU and Arm-based NVIDIA Grace CPU utilizing the corporate's NVLink-C2C interlink. It's designed for supercomputers to permit "scientists and researchers to sort out the world’s most difficult issues by accelerating complicated AI and HPC purposes working terabytes of information," NVIDIA wrote.
The GH200 might be utilized in "40+ AI supercomputers throughout world analysis facilities, system makers and cloud suppliers," the corporate mentioned, together with from Dell, Eviden, Hewlett Packard Enterprise (HPE), Lenovo, QCT and Supermicro. Notable amongst these is HPE's Cray EX2500 supercomputers that may use quad GH200s, scaling as much as tens of 1000’s of Grace Hopper Superchip nodes.
Maybe the largest Grace Hopper supercomputer might be JUPITER, positioned on the Jülich facility in Germany, which can change into the "world’s strongest AI system" when it's put in in 2024. It makes use of a liquid-cooled structure, "with a booster module comprising near 24,000 NVIDIA GH200 Superchips interconnected with the NVIDIA Quantum-2 InfiniBand networking platform," in response to NVIDIA.
NVIDIA says JUPITER will assist support scientific breakthroughs in a quantity of areas, together with local weather and climate prediction, producing high-resolution local weather and climate simulations with interactive visualization. It'll even be employed for drug discovery, quantum computing and industrial engineering. Many of these areas use customized NVIDIA software program options that ease growth but additionally make supercomputing teams reliant on NVIDIA {hardware}.
The brand new applied sciences might be key for NVIDIA, which now makes most of its income from the AI and information middle segments. Final quarter the corporate noticed a report $10.32 billion in income in that space alone (out of $13.51 billion complete income), up 171 p.c from a yr in the past. It little doubt hopes the brand new GPU and superchip will assist proceed that pattern. Simply final week, NVIDIA broke its personal AI coaching benchmark report utilizing older H100 know-how, so its new tech ought to assist it prolong that lead over rivals in a sector it already dominates.
This text initially appeared on Engadget at
