Meta unveiled its next generation Grand Teton hardware platform at a recent global conference. The presentation was led by the VP of Engineering who led the development of the platform, who introduced the audience to the new features of the innovative invention.
Meta trains and deploys many AI-enabled technologies. They often contain trillions of parameters and need datasets of similar size. Training these technologies requires parallelism of both data and models, which in turn means a fleet of interconnected servers with many GPUs.
Meta began providing open access to its AI hardware projects in 2016 with the Big Sur platform. Last year, Meta introduced the iteration, Zion-EX, which consists of a cluster of thousands of compute nodes, each with 4-socket CPUs and 8 GPUs.
In addition, Grand Teton Meta released a data center design – Open Rack v3 (ORV3). Unlike previous data center designs where the power shelf is attached to a bus, with ORV3 it can be installed in any location, allowing for more flexible designs.
The improved battery backup can provide power for up to four minutes compared to the previous 90 seconds. The ORV3 also supports multiple power shelves. The larger power capacity also led Meta to develop new cooling strategies.