Nvidia strikes Hopper GPUs for AI into full manufacturing

Nvidia introduced in the present day that the Nvidia H100 Tensor Core graphics processing unit (GPU) is in full manufacturing, with international tech companions planning in October to roll out the primary wave of services primarily based on the Nvidia Hopper structure.

Nvidia CEO Jensen Huang made the announcement at Nvidia’s on-line GTC fall occasion.

Unveiled in April, H100 is constructed with 80 billion transistors and has a spread of know-how breakthroughs. Amongst them are the highly effective new Transformer Engine and an Nvidia NVLink interconnect to speed up the biggest artificial intelligence (AI) fashions, like superior recommender methods and large language models, and to drive improvements in such fields as conversational AI and drug discovery.

“Hopper is the brand new engine of AI factories, processing and refining mountains of information to coach fashions with trillions of parameters which might be used to drive advances in language-based AI, robotics, healthcare and life sciences,” mentioned Jensen Huang, founder and CEO of Nvidia, in a press release. “Hopper’s Transformer Engine boosts efficiency as much as an order of magnitude, placing large-scale AI and HPC inside attain of corporations and researchers.”

Along with Hopper’s structure and Transformer Engine, a number of different key improvements energy the H100 GPU to ship the subsequent huge leap in Nvidia’s accelerated compute data center platform, together with second-generation Multi-Occasion GPU, confidential computing, fourth-generation Nvidia NVLink and DPX Directions.

“We’re tremendous excited to announce that the Nvidia H100 is now in full manufacturing,” mentioned Ian Buck, basic supervisor of accelerated computing at Nvidia, in a press briefing. “We’re able to take orders for cargo in Q1 (beginning in Nvidia’s fiscal yr in October). And beginning subsequent month, our methods companions from Asus to Supermicro will likely be beginning to ship their H100 methods, beginning with the PCIe merchandise and increasing afterward this yr to the NVLink HDX platforms.”

A five-year license for the Nvidia AI Enterprise software program suite is now included with H100 for mainstream servers. This optimizes the event and deployment of AI workflows and ensures organizations have entry to the AI frameworks and instruments wanted to construct AI chatbots, suggestion engines, imaginative and prescient AI and extra.

International rollout of Hopper

H100 permits corporations to slash prices for deploying AI, delivering the identical AI efficiency with 3.5 occasions extra vitality effectivity and 3 times decrease whole value of possession, whereas utilizing 5 occasions fewer server nodes over the earlier era.

For patrons who need to strive the brand new know-how instantly, Nvidia introduced that H100 on Dell PowerEdge servers is now out there on Nvidia LaunchPad, which gives free hands-on labs, giving corporations entry to the most recent {hardware} and Nvidia AI software program.

Clients may also start ordering Nvidia DGX H100 methods, which embody eight H100 GPUs and ship 32 petaflops of efficiency at FP8 precision. Nvidia Base Command and Nvidia AI Enterprise software program energy each DGX system, enabling deployments from a single node to an Nvidia DGX SuperPOD, supporting superior AI improvement of huge language fashions and different huge workloads.

H100-powered methods from the world’s main pc makers are anticipated to ship within the coming weeks, with over 50 server fashions out there by the top of the yr and dozens extra within the first half of 2023. Companions constructing methods embody Atos, Cisco, Dell Applied sciences, Fujitsu, GIGABYTE, Hewlett Packard Enterprise, Lenovo and Supermicro.

Moreover, among the world’s main larger training and analysis establishments will likely be utilizing H100 to energy their next-generation supercomputers. Amongst them are the Barcelona Supercomputing Heart, Los Alamos Nationwide Lab, Swiss Nationwide Supercomputing Centre (CSCS), Texas Superior Computing Heart and the College of Tsukuba.

In comparison with the prior A100 era, Buck mentioned the prior system had 320 A100 methods in a datacenter, however with Hopper an information middle would solely want 64 H100 methods to match that throughput of the older information middle. That’s a 20% discount in nodes and an enormous enchancment in vitality effectivity.