Ultra Ethernet: Addressing AI/HPC Interconnect Challenges Beyond the InfiniBand Monopoly#

Andreas Roeder (Nokia)

• Why: Ultra Ethernet is the emerging standard for AI/HPC fabrics — directly relevant to anyone working on RDMA, lossless Ethernet, and next-gen DC interconnect. The PM's interest in RDMA/lossless fabrics makes this the single most strategically important talk at the conference.
• What: A comprehensive technical walkthrough of the Ultra Ethernet Transport (UET) specification as published by the Ultra Ethernet Consortium (June 2025 spec), covering how it differs from RoCEv2 and InfiniBand. Covers the InfiniBand lock-in problem and how Ethernet's ecosystem advantage (multi-vendor NICs, optics, cables, tooling) positions UET for AI/HPC build-outs.
• How: UET reuses what works from RoCEv2 (RMA verbs, RDMA semantics via LibFabric API) and redesigns the transport: a Packet Delivery Sublayer (PDS) with four delivery modes (RUD, ROD, RUDI, UUD), packet-spraying ECMP (no long-lived flows), per-packet memory semantics (each packet carries its own memory address — no reorder buffer needed), trimmed-packet NACK for selective retransmission, and dual congestion control (NSCC for network signals + RCCC receiver-credit for incast). Runs on standard Ethernet (lossy OK — no PFC required), IPv4/IPv6, with Day-1 security.
• Where applicable: Any operator or architect evaluating AI/HPC fabric technology choices. The spec is vendor-neutral and multi-vendor by design. Relevant to hyperscaler-adjacent and enterprise DC operators planning GPU/XPU interconnects in the next 2–3 years. Nokia-employee speaker but content is spec-grounded, not a pitch.