Scheduled Ethernet Fabric

Overview

Scheduled Ethernet Fabric is an OCP-compliant networking architecture that uses receiver-based scheduling to guarantee congestion-free, collision-free AI training fabric operation. Unlike traditional Ethernet where switches make independent forwarding decisions, scheduled fabric requires all traffic to get explicit permission from the receiver before traversing the fabric, enabling perfect load balancing without ECMP hashing. The technology uses Broadcom Jericho (leaf) and Ramon (spine) silicon families. ByteDance deployed the first production scheduled Ethernet fabric in June 2024 with 1,280 xPU cards, achieving 22-119% performance improvement over typical Ethernet on collective operations.

The Open Scheduled Fabric (OSF) initiative formalizes this architecture as an open standard, with six vendors (ByteDance, H3C, DriveNets, Ruijie, Accton, Broadcom) submitting an IETF draft. Meta separately announced a Disaggregated Scheduled Fabric (DSF) platform using the same Broadcom Jericho3-AI and Ramon3 silicon, indicating convergence around this approach among major hyperscalers.