[ad_1]
The battle for datacenter dominance retains getting hotter. Right now, Nvidia kicked off its spring GTC occasion with new silicon, new software program and a brand new supercomputer. Talking from a digital atmosphere within the Nvidia Omniverse 3D collaboration and simulation platform, CEO Jensen Huang launched the brand new Hopper GPU structure and the H100 GPU, which can energy datacenter-scale techniques for HPC and AI workloads.
Nvidia’s first Hopper-based product, the H100 GPU, is manufactured on TSMC’s 4N course of, leveraging a whopping 80 billion transistors — 68 % greater than the prior-generation 7nm A100 GPU. The H100 is the primary GPU to assist PCIe Gen5 and the primary to make the most of HBM3, enabling 3TB/s of reminiscence bandwidth.
Named after pc scientist and U.S. Navy Rear Admiral Grace Hopper, the brand new GPU (in its SXM mezzanine kind issue) gives 30 teraflops of normal IEEE FP64 efficiency, 60 teraflops of FP64 tensor core efficiency, and 60 teraflops of FP32 efficiency. A brand new numerical format launched in Hopper, FP8 tensor core, delivers as much as 4,000 teraflops of AI efficiency, in keeping with Nvidia. See spec information and gen-to-gen comparisons beneath.
Hopper introduces built-in {hardware} acceleration for transformer fashions, that are broadly used for pure language processing. The Hopper transformer engine dynamically chooses between 8-bit and 16-bit calculations, intelligently managing precision within the layers of the transformer community to ship speedups with out lack of accuracy, in keeping with Nvidia.
“Hopper H100 is the most important generational leap ever — 9x at-scale coaching efficiency over A100 and 30x large-language-model inference throughput,” Huang stated in his keynote.
You’ll be able to learn the remainder of the story right here at HPCwire.
[ad_2]
