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

Andreas Roeder (Nokia)

• Why: This is the clearest operator-level walkthrough of the published Ultra Ethernet (UE 1.0, June 2025) spec — the standard most likely to reshape RDMA/lossless fabric design over the next few years. For anyone whose remit covers lossless Ethernet, AI/HPC fabrics, or DC interconnect strategy, this is the session to attend.
• What: A structured technical introduction to the Ultra Ethernet Transport (UET) layer: how it compares to RoCEv2, what problems it solves (multipathing, relaxed ordering, rapid loss recovery, modern congestion control), and why the InfiniBand lock-in problem (switch/NIC/optics/toolchain/management all single-vendor) motivates an Ethernet-based alternative with a multi-vendor ecosystem at every layer.
• How: UET runs on IPv4/IPv6 and standard Ethernet (lossy operation is fine — no PFC required). The Packet Delivery Sublayer (PDS) supports four delivery modes (RUD/ROD/RUDI/UUD), selective ACK/NACK with compact SACK encoding (64 sequence numbers in 8 bytes vs TCP's 4 in 36), packet spraying instead of flow-based ECMP, per-packet memory semantics (each packet carries its own target memory address, enabling direct commit without a reorder buffer), and dual congestion control (NSCC reacting to trimmed packets/ECN/latency plus RCCC, a receiver-credit scheme aimed at incast). Ephemeral asymmetric connections replace TCP-style sessions.
• Where applicable: Broadly applicable to AI/HPC fabric planning in any operator or enterprise environment — the spec is vendor-neutral by design. The talk focuses on DC-internal technologies that will also influence metro/regional/WAN design. Nokia-employed speaker, but the content is grounded in the UEC specification, not a product pitch.