UEC 1.0 Aligned

Ultra Ethernet — UEC 1.0 for AI Fabrics

Ultra Ethernet is the industry's open answer to the question: can Ethernet replace InfiniBand for trillion-parameter training? The Ultra Ethernet Consortium 1.0 specification redesigns the transport — packet spray, multi-path RDMA, out-of-order delivery, modern congestion control. OcNOS provides the standards-based fabric layer that UEC NICs require.

Packet-Spray Across Every Path

A single RDMA message between two GPUs is sprayed packet-by-packet across all four spine paths simultaneously. The fabric tolerates out-of-order arrival; the destination UEC NIC reassembles in order. No flow-pinning, no hash collisions, no idle uplinks.

Ultra Ethernet packet spray across a 4-spine fabric Two GPUs connected through two leaves and four spines. A single RDMA message is split into four packets and sent across all four spine paths simultaneously. At the receiving NIC, packets arrive out of order and are reassembled in order before delivery to the GPU. Bottom band lists UEC properties: packet spray, multi-path, out-of-order delivery, NIC reassembly. P1 P2 P3 P4 Spine-1UEC fabric Spine-2UEC fabric Spine-3UEC fabric Spine-4UEC fabric Leaf-1UEC entitled Leaf-2UEC entitled GPU-AUEC NIC GPU-Breassembly UEC 1.0 · PACKET SPRAY · MULTI-PATH RDMA · OUT-OF-ORDER DELIVERY · NIC REASSEMBLY

Why UEC matters now

RoCEv2 was the first generation of "RDMA over Ethernet". It works — but it inherits the design constraints of TCP/IP over Ethernet from a different era. Single flow per (5-tuple, hash) means GPU collectives can saturate one path while parallel paths sit idle. PFC pause propagation can cause head-of-line blocking and, in pathological topologies, deadlock. NACK-based loss recovery puts a tax on tail latency.

The Ultra Ethernet Consortium — backed by AMD, Arista, Broadcom, Cisco, HPE, Intel, Meta, Microsoft, Oracle, and others — released the UEC 1.0 specification to redesign the transport for AI scale. It keeps the standard physical Ethernet you can buy from anyone, and replaces the upper layers with a transport built for 1k–100k GPU jobs.

The four UEC pillars

Packet spray

A single message is sprayed across every available path. No flow pinning. No ECMP collisions. Fabric utilisation approaches the theoretical maximum.

Out-of-order delivery

The fabric is allowed to reorder. The UEC NIC reassembles in order before delivery, so applications and the GPU never see disorder.

Modern congestion control

Sender-based and receiver-driven schemes replace coarse PFC pause as the primary tool. Pause becomes a backstop, not the front line.

Selective retransmit

Lost packets are recovered selectively rather than restarting the RDMA window — collapsing the tail-latency penalty of any in-fabric loss.

The OcNOS UEC implementation

Spec

UEC 1.0 fabric profile

OcNOS-DC implements the UEC 1.0 fabric-side profile — the configuration and behaviours that switches must support to interoperate with UEC-compliant NICs.

Spray-Friendly

OOO-tolerant forwarding

Per-packet ECMP, packet-spray-friendly QoS, and shared-buffer policies that don't penalise out-of-order delivery — the conditions UEC NICs need to operate efficiently.

Hardware

TH4 / TH5 silicon

Runs on the same Broadcom Tomahawk 4 (25.6T) and Tomahawk 5 (51.2T) platforms used for current RoCEv2 fabrics. No forklift upgrade required.

Brownfield

RoCEv2 + UEC coexistence

UEC and RoCEv2 traffic share the same fabric on different priorities. Migrate clusters incrementally as UEC-capable NICs deploy.

Telemetry

UEC-aware observability

Per-path utilisation, spray-friendly buffer counters, and reorder-tolerance metrics streamed over gNMI for closed-loop tuning during cluster bring-up.

Open Hardware

Vendor-neutral path

UEC is the open alternative to InfiniBand. Pairing UEC-compliant OcNOS-DC with open hardware means no single vendor owns the AI fabric — which is the whole point.

UEC vs RoCEv2 vs InfiniBand — where each fits

  • RoCEv2 today. Production-grade for clusters built right now. OcNOS-DC ships pre-tuned RoCEv2 on TH4 / TH5 — see the RoCEv2 deep-dive. Most fabrics in production in 2026 are RoCEv2.
  • UEC tomorrow. The path forward for scale-out clusters as UEC NICs ship in volume. OcNOS provides the fabric layer; cluster owners pick the NIC vendor.
  • InfiniBand. Single-vendor performance lock-in. Different cabling, different management, separate ecosystem. UEC closes the performance gap on a standards-based, multi-vendor Ethernet fabric.
  • Co-existence is the default. Real fabrics carry RoCEv2 traffic from existing clusters and UEC traffic from the next-gen build alongside each other. OcNOS-DC supports both on the same hardware.
  • Forward-looking GLB. The OcNOS 7.1 Global Load Balancing path-quality plane is being designed to interoperate with UEC signalling as the spec matures — investments in OcNOS-DC today carry forward.

Planning a UEC-ready AI fabric? Let's design it together.

Book an Architecture Review →