SAI Specifications for Newer Technologies Introduced by Ultra Ethernet Consortium

Author: Jerkumar (Broadcom); Ravi (Marvell) Collection: 2025-OCP-Global-Summit-Networking

Summary

The Ultra Ethernet Consortium (UEC) is standardizing a set of new link-layer and telemetry technologies specifically for AI backend networks, where packet loss from buffer overflow or frame corruption causes hundreds of GPUs to stall waiting for end-to-end retransmissions. This presentation covers the Switch Abstraction Interface (SAI) modeling work for three UEC technologies: CSIG (Collective Signal — a lightweight in-band telemetry protocol for congestion control), LLR (Link Layer Retry — local recovery of corrupted frames without end-to-end retransmission), and CBFC (Credit-Based Flow Control — sender-managed receiver buffer tracking to prevent overflow, an alternative to PFC).

The SAI specifications for LLR and CBFC are already released in UEC 1.0. CSIG has been contributed to OCP at version 0.5 (with ongoing changes). All three feature SAI implementations are present in Broadcom's Tomahawk Ultra, Thor Ultra, and Tomahawk 6 ASICs. CSIG is modeled within the existing TAM (Telemetry and Monitoring) 2.0 SAI infrastructure, which provides shared primitives (packet identification, collector addressing, actions) reusable across all telemetry technologies, avoiding repetition of attributes across independently modeled telemetry features.