UEC 1.0 · Packet spray · UET

Ultra Ethernet: UEC 1.0 for AI Fabrics

Ultra Ethernet is the open, multi-vendor path to InfiniBand-class AI networking on standard Ethernet. The UEC 1.0 specification, released in June 2025, defines the Ultra Ethernet Transport (UET) with packet spraying, multi-path RDMA, out-of-order delivery, selective retransmission, and both sender- and receiver-based congestion control. It does not require a lossless PFC fabric, and OcNOS-DC provides the standards-based fabric layer that UEC-capable NICs require.

June 2025UEC 1.0 specification
1 fabricRoCEv2 and UEC, same hardware
up to 800GOcNOS-DC on Tomahawk 5
600+operator networks on OcNOS
How UET moves a message

Packet spray across every path

A single RDMA message between two GPUs is sprayed packet-by-packet across all four spine paths at once. The fabric tolerates out-of-order arrival and 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 RDMA, out-of-order delivery, NIC reassembly. P1 P2 P3 P4 Spine-1UEC fabric Spine-2UEC fabric Spine-3UEC fabric Spine-4UEC fabric Leaf-1OcNOS-DC Leaf-2OcNOS-DC GPU-AUEC NIC GPU-Breassembly UEC 1.0 · PACKET SPRAY · MULTI-PATH RDMA · OUT-OF-ORDER DELIVERY · NIC REASSEMBLY
Inside the transport

What the Ultra Ethernet Transport adds

RoCEv2 works, but it inherits older design constraints: one hashed path per flow, network-wide PFC pause, and window-restart loss recovery. The UEC 1.0 specification keeps the standard physical Ethernet you can buy from anyone and replaces the upper layers with a transport (UET) built for 1k to 100k GPU jobs.

Path use

Packet spraying

A single message is sprayed across every available path at once. No flow pinning, no ECMP collisions, so fabric utilisation approaches the theoretical maximum during a collective.

RDMA

Multi-path RDMA

RDMA is carried over many parallel paths rather than one, so a large transfer is no longer capped by the bandwidth of a single hashed link.

Delivery

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.

Recovery

Selective retransmission

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

Control

Modern congestion control

Both sender- and receiver-based schemes replace coarse PFC pause as the primary tool. UET does not require a lossless PFC fabric; pause becomes a backstop, not the front line.

Ecosystem

Open and multi-vendor

UET runs on standard Ethernet from any vendor. The consortium (AMD, Arista, Broadcom, Cisco, HPE, Intel, Meta, Microsoft, Oracle and others) publishes the spec openly.

Where each fits

UEC, RoCEv2, and InfiniBand, axis by axis

RoCEv2 runs most AI fabrics in production today. Ultra Ethernet is the standards-based path forward as UEC-capable NICs ship, and InfiniBand remains the single-vendor specialist. Which axis matters depends on the workload and the timeline.

Axis RoCEv2Ethernet, today Ultra EthernetUET, open multi-vendor InfiniBandsingle-vendor
Path useOne hashed ECMP path per flow; parallel paths can sit idle.Packet spray across every available path in parallel.Adaptive routing built into the spec.
Loss handlingLossless via PFC + ECN + DCQCN; relies on network-wide pause.Tolerates loss and reorder; does not require a lossless PFC fabric.Lossless by credit-based flow control.
ReorderingIn-order delivery only.Out-of-order delivery, reassembled in order at the NIC.In-order delivery.
RetransmissionWindow-based recovery taxes tail latency.Selective retransmission of only the lost packets.Link-level recovery.
Congestion controlDCQCN with ECN marking.Both sender- and receiver-based control at the endpoints.Credit-based at the link.
EcosystemMulti-vendor Ethernet at every layer.Open, multi-vendor; any vendor can build to the UEC 1.0 spec.Effectively single-vendor for NIC and switch silicon.
OcNOS-DC statusProduction today on Tomahawk 4 / Tomahawk 5.Fabric layer aligned with the UEC 1.0 profile; transport engages as NICs ship.Not applicable; separate fabric and tooling.
OcNOS and UEC

The fabric layer UEC NICs require

OcNOS-DC runs a production RoCEv2 fabric today and tracks the UEC 1.0 fabric-side profile: the switch configuration and behaviours needed to interoperate with UEC-capable NICs. Full transport is a roadmap item that engages as those NICs ship. Alignment with the profile is not a certification claim.

Spec

UEC 1.0 fabric profile

OcNOS-DC tracks the UEC 1.0 fabric-side profile, the configuration and behaviours switches must support to interoperate with UEC-capable NICs as they arrive.

Spray-friendly

OOO-tolerant forwarding

Per-packet ECMP, packet-spray-friendly QoS, and shared-buffer policies that do not penalise out-of-order delivery: the conditions UEC NICs need to operate efficiently.

Hardware

Tomahawk 4 / 5 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 and UEC coexist

UEC and RoCEv2 traffic share one 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 OcNOS-DC with open hardware means no single vendor owns the AI fabric, which is the whole point.

The migration picture

Build on RoCEv2 now, keep a clean path to UEC

Real fabrics do not switch overnight. The practical plan is production RoCEv2 today on hardware that already tracks the UEC 1.0 profile, so the transport step needs no NOS or switch swap.

  • RoCEv2 today. Production-grade for clusters built right now. OcNOS-DC ships pre-tuned RoCEv2 on Tomahawk 4 / Tomahawk 5. Most fabrics in production in 2026 are RoCEv2.
  • UEC as NICs ship. The path forward for scale-out clusters as UEC-capable NICs arrive. OcNOS-DC provides the fabric layer; cluster owners pick the NIC vendor.
  • InfiniBand stays specialist. Single-vendor performance with separate cabling, management, and ecosystem. UEC closes the performance gap on a standards-based, multi-vendor Ethernet fabric.
  • Coexistence is the default. RoCEv2 traffic from existing clusters and UEC traffic from the next build run alongside each other on the same OcNOS-DC hardware.
  • GLB carries forward. The OcNOS 7.1 Global Load Balancing path-quality plane is being designed to interoperate with UEC signalling as the spec matures, so today's OcNOS-DC investments carry forward.
The IP Infusion view

Open Ethernet is the destination, and OcNOS is the path

Ultra Ethernet brings the transport techniques that defined InfiniBand to standard, multi-vendor Ethernet. IP Infusion is a contributing member of the consortium and OcNOS-DC provides the fabric layer, so a cluster built today carries forward.

A contributing member

IP Infusion takes part in shaping the UEC fabric-layer specification, and OcNOS-DC tracks the UEC 1.0 fabric profile as it evolves.

Production RoCEv2 now

OcNOS-DC delivers lossless RoCEv2 with PFC, ECN, and sub-millisecond DLB today on validated open hardware from vendors such as Edgecore and UfiSpace.

UEC-ready by design

Because the switch already tracks the fabric profile, engaging Ultra Ethernet as NICs ship needs no NOS or switch swap. Alignment is not a certification claim.

FAQ

Ultra Ethernet, answered

What is Ultra Ethernet (UEC)?
Ultra Ethernet is an open standard from the Ultra Ethernet Consortium (UEC) that adapts Ethernet for AI and HPC workloads. The UEC 1.0 specification, released in June 2025, defines the Ultra Ethernet Transport (UET): endpoint packet spray across every path, multi-path RDMA, out-of-order delivery with reassembly at the NIC, selective retransmission, and both sender- and receiver-based congestion control, so large training and inference jobs stop being limited by single-path hashing and network-wide pause.
Is IP Infusion a member of the Ultra Ethernet Consortium?
Yes. IP Infusion is a contributing member of the Ultra Ethernet Consortium and takes part in shaping the fabric-layer specification. OcNOS-DC tracks the UEC 1.0 fabric profile, so the switches you deploy today carry forward as UEC-capable NICs arrive. Alignment with the profile is not a certification claim.
How is Ultra Ethernet different from RoCEv2?
RoCEv2 pins each flow to one ECMP path and depends on network-wide PFC to stay lossless. Ultra Ethernet sprays a single message across every available path in parallel, reassembles it in order at the destination NIC, retransmits only the lost packets, and controls congestion at both the sender and the receiver. UET does not require a lossless PFC fabric, which reduces the reliance on PFC pause and lifts tail latency on large collectives.
Does OcNOS support Ultra Ethernet today?
OcNOS-DC runs a production RoCEv2 fabric today with PFC, ECN, and sub-millisecond DLB, and it is aligned with the UEC 1.0 fabric profile. Full Ultra Ethernet transport is a roadmap item that engages as UEC-capable NICs ship; because OcNOS-DC already tracks the fabric profile, the move needs no NOS or switch swap. Alignment with the profile is not a certification claim.
Do I need new hardware for Ultra Ethernet?
The switch is the part that carries forward. Ultra Ethernet endpoint behavior (packet spray, reassembly, selective retransmission) runs on UEC-capable NICs, while the fabric your RoCEv2 traffic uses today keeps serving as those NICs roll out. Running OcNOS-DC on HCL-listed Tomahawk 5 hardware means the fabric layer is already aligned with the UEC 1.0 profile.
What are packet spray and out-of-order delivery?
Packet spray sends the packets of one message across all parallel fabric paths at once instead of down a single hashed path, so no link becomes a hot spot during a collective. Because packets can then arrive out of order, the destination NIC reassembles them into the original sequence. Together they keep every path busy and shorten the slowest transfer in a job.

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

Tell us the workload and the GPU scale, and an IP Infusion engineer will size the leaf-spine fabric with you, or start a first-pass layout in the AI Fabric Design Suite.