Keith D Underwood of HPE presented the Ultra Ethernet Consortium's "Escaping the Chaos of Order" talk at the OCP Global Summit in October 2025, arguing that ordered network transports are architecturally self-defeating: each ordering assumption cascades into tangled connections, inefficient multipath use, and brittle error models. UEC's response is a clean-slate full-stack architecture—built on libfabric, Ethernet, and IP—that embraces out-of-order packet delivery, ephemeral Packet Delivery Contexts, ECMP packet spraying, and pervasive AES-GCM encryption. The resulting specification 1.0, released June 2025 and already spanning over 100 member companies, reduces bandwidth-taper impact from 2x to 25% and eliminates false contention entirely.
Key Takeaways
- UEC Specification 1.0 was released in June 2025 at just over 500 pages, organized across Software, Transport, Link, and Physical working groups, and stands as an ISO and JDF standard.
- The root cause of RoCEv2's limitations is that it was designed in 2000 when transistors were expensive: it lacks multipathing, integrated security, and handles ordering in ways that make congestion management nearly impossible.
- Ordered networks create three distinct performance problems: bandwidth taper (2x impact when flows collide on fewer uplinks), false contention (flows stuck on the same uplink despite alternatives), and incast (indistinguishable from the other two without extra protocol support).
- UEC's packet-level ECMP spraying with QuickAdapt congestion control reduces bandwidth-taper overhead from 2x down to approximately 25% and eliminates false contention entirely.
- Packet Delivery Contexts (PDCs) replace persistent connections: they are invisible to software, created on demand, use zero-round-trip pipelined startup, and close in one RTT—dramatically reducing connection-state scaling problems.
- UEC uses libfabric as the API layer, meaning PyTorch, TensorFlow, MPI, CCLs, and SHMEM require no modification to run over Ultra Ethernet.
- UEC encrypts the full transport with AES-GCM using shared key infrastructure (one master key per application, optional per-peer key derivation), and includes a novel anti-replay solution compatible with pipelined PDC startup.
- UEC has grown to over 100 member companies and more than 1,300 active participants as of October 2024, with members including AMD, Intel, Meta, Microsoft, Nvidia, Broadcom, Cisco, HPE, Google, and many others.
Sign in to read the full article.
Sign In