Nvidia announced on its blog on the first (local time) that it has launched the ‘DGX H100’, a supercomputer system for data centers originally scheduled to be released in October.

DGX H100 is a supercomputing system equipped with eight ‘H100’ GPUs and connected as one through ‘Nvidia NVLink’. It delivers AI performance of 32FP at the brand new FP8 precision, which is 6x higher than the previous generation DGX A100. Power efficiency per petaflops has also doubled in comparison with the DGX A100. It also includes your entire software stack that drives the hardware.

The H100 is a GPU based on the ‘Hopper’ architecture, and the A100 is predicated on the ‘Ampere’ architecture.

DGX H100 can speed up exascale work by connecting as much as 32, or a complete of 256 based on the H100 GPU. The dedicated Transformer engine has the capability to handle large language models with trillions of parameters.

The DGX H100 system runs on ‘AI Base Command’, a family of products for accelerating compute, storage and network infrastructure and optimizing artificial intelligence (AI) workloads.

It also includes “AI Enterprise,” software that accelerates data science pipelines and simplifies the event and deployment of generative AI, computer vision, and more.

The worth is anticipated to be at the least 4 times costlier than the prevailing DGX A100 system. The worth of the DGX A100 is around 200,000 dollars (about 260 million won).

Meanwhile, Nvidia ships the product to Boston Dynamics AI Labs, legal AI platform Cicero within the UK, Germany’s DeepL, a translation company, Schreka, a Japanese startup that manages the Djonko-1 supercomputer, Johns Hopkins University, and the Royal Institute of Sweden. he added that he was doing

Chan Park, cpark@aitimes.com