Disaggregated Scheduled Fabric

Overview

The Disaggregated Scheduled Fabric (DSF) is Meta's AI scale-out network architecture that takes the switching fabric originally located inside a chassis and distributes it across an entire data hall. This disaggregation enables a shared, high-radix switching domain for AI training clusters at the building scale.

The original DSF (v1) targeted approximately 4,000 GPUs. It required heavy control-plane tuning for AI workloads and was deployed in production across multiple accelerator generations. In summer 2025, Meta announced a 2-stage DSF expansion that scales to approximately 20,000 GPUs per building. The 2-stage DSF is deployable in combination for even larger clusters and is already in production at Meta's Ohio facilities.

The DSF was eventually complemented by the Non-Scheduled Fabric (NSF) — a simpler, lower-latency architecture based on shallow-buffer Ethernet switches with adaptive routing — for Meta's largest gigawatt-scale clusters.