Skip to content

Evolution of Ethernet based switch platforms and fabrics to meet Meta's AI training clusters' scale-out AI fabric demands#

OCP Global Summit 2025 digest Program Core pick 13–16 October 2025

Abstract Session recording Materials

Speakers: Parvez Shaikh (Meta), Lingjun Wu (Meta)

  • Why: The sequel to Meta's OCP 2024 fabric talk — the ground truth on how one of the largest Ethernet AI fabrics is actually evolving from 4K-GPU cells to much larger non-blocking clusters, including the SAI/FBOSS changes required.
  • What: Meta's journey moving from disaggregated deep-buffered switch platforms (OCP 2024, up to 4K-GPU cells) to disaggregated shallow-buffer Ethernet platforms for significantly larger non-blocking fabrics and GPU clusters, with the unique scaling problems they hit along the way. (from abstract only)
  • How: Evolution of switching and routing platform features to build larger non-blocking fabrics, plus the SAI and FBOSS enhancements that adapted the software stack to the new features.
  • Where applicable: Meta-specific production architecture, but the SAI/FBOSS work is in the open OCP repos (linked materials) and the buffer-depth tradeoffs are a general design decision for anyone building AI scale-out fabrics.