Ultra Ethernet

Overview

Ultra Ethernet is an industry standard developed by the Ultra Ethernet Consortium that extends Ethernet for AI backend (scale-out and scale-across) networking. It addresses four structural weaknesses of standard RoCEv2 at AI cluster scale: ECMP hash collisions that underutilize fabric links; PFC-based lossless transport that creates deadlock/livelock problems; suboptimal go-back-N retransmission under any packet loss; and slow failover on path failures.

Ultra Ethernet introduces: intelligent load balancing to eliminate hash-collision-induced link imbalance; lossy-network congestion management that removes the need for PFC/DCQCN; selective acknowledgement to limit retransmission scope under packet loss; and fast failover for path failures. Under 0.1% packet loss — realistic at million-GPU cluster scale — UEC fast failover and selective acknowledgement deliver 50% improvement in LLM training job completion time versus standard RDMA. With UEC intelligent load balancing alone, AMD observed 25% better RCCL collective performance on the Pollara 400.

AMD's Pensando Vulcano NIC (2026) ships with Ultra Ethernet as the first-day transport.