First Batch of US AMD Rack-Scale AI Systems Installed at Meta

en.Wedoany.com Reported - On October 27, US-based AMD confirmed that the first batch of Helios rack-scale artificial intelligence computing systems has been installed and entered the testing phase at Meta, with the two parties advancing subsequent large-scale deployment. The collaboration covers EPYC server processors, Instinct AI accelerators, Pensando networking technology, and the ROCm software platform. Meta has already deployed AMD hardware for AI model training, content ranking, recommendation systems, and running the Muse personal AI agent, while simultaneously carrying out adaptation work for next-generation computing hardware and software.

In server processor deployment, Meta has cumulatively installed millions of AMD EPYC processors across different generations, some of which are used in dedicated computing environments for the Muse personal AI agent. Meta has also become a major customer of sixth-generation EPYC processors, and the two parties are jointly developing a follow-on processor product codenamed Verano.

GPU deployment involves both existing production systems and custom chip development. AMD Instinct accelerators already handle AI training, content ranking, and recommendation tasks in Meta's data centers. Meta previously participated in the early product definition of the MI450 architecture and plans to deploy custom Instinct GPUs optimized for its own computing workloads. The collaboration also extends to the MI500 series accelerators and next-generation rack-scale systems.

Software adaptation work covers ROCm, PyTorch, and Triton. AMD has completed first-day adaptation of Instinct GPUs for Meta's Muse Spark Open model and supports the Muse Glimmer model running on Ryzen AI personal computers and Radeon GPU platforms. The two parties have also carried out large-scale recommendation model training optimization for the Instinct MI350X accelerator and submitted related software improvements to the PyTorch, Triton, and ROCm projects.