Meta right now introduced its next-generation AI platform, Grand Teton, such as NVIDIA’s collaboration on structure.
As opposed to the company’s preceding era Zion EX system, the Grand Teton procedure packs in far more memory, community bandwidth and compute capability, claimed Alexis Bjorlin, vice president of Meta Infrastructure Components, at the 2022 OCP Global Summit, an Open Compute Job meeting.
AI designs are used thoroughly across Fb for services this sort of as news feed, information recommendations and dislike-speech identification, among lots of other purposes.
“We’re enthusiastic to showcase this newest relatives member listed here at the summit,” Bjorlin said in prepared remarks for the convention, incorporating her thanks to NVIDIA for its deep collaboration on Grand Teton’s style and continued assistance of OCP.
Table of Contents
Designed for Info Heart Scale
Named soon after the 13,000-foot mountain that crowns a single of Wyoming’s two national parks, Grand Teton uses NVIDIA H100 Tensor Main GPUs to educate and run AI types that are fast increasing in their measurement and capabilities, necessitating increased compute.
The NVIDIA Hopper architecture, on which the H100 is based, involves a Transformer Engine to accelerate perform on these neural networks, which are usually called foundation products mainly because they can tackle an increasing established of purposes from pure language processing to health care, robotics and more.
The NVIDIA H100 is created for general performance as nicely as electricity effectiveness. H100-accelerated servers, when related with NVIDIA networking across hundreds of servers in hyperscale data facilities, can be 300x a lot more electrical power successful than CPU-only servers.
“NVIDIA Hopper GPUs are crafted for solving the world’s challenging troubles, offering accelerated computing with better electricity performance and enhanced general performance, when introducing scale and lowering expenses,” explained Ian Buck, vice president of hyperscale and higher performance computing at NVIDIA. “With Meta sharing the H100-powered Grand Teton system, program builders about the planet will soon have entry to an open style for hyperscale details middle compute infrastructure to supercharge AI across industries.”
Mountain of a Machine
Grand Teton sports activities 2x the network bandwidth and 4x the bandwidth involving host processors and GPU accelerators when compared to Meta’s prior Zion procedure, Meta claimed.
The additional community bandwidth permits Meta to produce larger sized clusters of methods for schooling AI models, Bjorlin said. It also packs more memory than Zion to retail store and run much larger AI versions.
Simplified Deployment, Amplified Reliability
Packing all these abilities into one particular built-in server “dramatically simplifies deployment of systems, allowing us to put in and provision our fleet a great deal much more rapidly, and raise dependability,” said Bjorlin.