This chapter is being written in March 2026. The UEC deployment state is honest: the specification is ahead of production silicon availability.

UEC 1.0 specification: Ratified Q4 2024. Published on ultra-ethernet.org. Complete normative specification for UET transport, retransmit, congestion control, and multipath.

Switch silicon:

• Broadcom Tomahawk 5 (StraDNX): ECN marking and basic UEC forwarding support in firmware Q4 2024 release. NPM support in beta SDK as of early 2026.
• Mellanox/NVIDIA Spectrum-4: ECN forwarding for UEC traffic (all Spectrum-4 switches support standard ECN). UET-specific extensions under active development. No public NPM announcement as of March 2026 from NVIDIA for Spectrum-4.
• Intel Tofino2 / P4-based platforms: Programmable. UEC reference P4 program available from Intel P4 Studio as of H1 2025. Deployment requires P4 expertise.

NIC/endpoint silicon:

• AMD Pensando Salina (DPU): UEC 1.0 compliant silicon, announced February 2025. First production-available UEC endpoint silicon.
• Intel Gaudi 3 (networking ports): UEC-capable per Intel roadmap. GA availability announced for 2025. Gaudi 3 AI accelerators include integrated 24-port 200GbE UEC-capable networking.
• Broadcom Astera (formerly Emulex): UEC NIC silicon in qualification with hyperscale customers Q1 2026. Not yet shipping at volume.
• NVIDIA ConnectX-8 / BlueField-4: No public UEC announcement from NVIDIA as of March 2026. NVIDIA's Spectrum-X platform (ConnectX-7 + SN5600) represents their current AI fabric strategy. The absence of a NVIDIA UEC announcement is the significant industry observation as of this writing.

DGX platform compatibility: DGX H100/H200 use ConnectX-7 NICs. ConnectX-7 does not implement UEC. DGX B200 uses BlueField-3 NICs. BlueField-3 does not implement UEC. A DGX cluster cannot run UEC today without replacing the NICs. This is the key constraint for the current DGX install base.

Hyperscale deployments: Meta, Microsoft, and Google have all publicly referenced UEC in their AI networking roadmaps. Meta published internal UEC evaluation results showing PFC-free operation with comparable throughput to RoCEv2-with-PFC in controlled lab conditions. No publicly disclosed large-scale production deployment of UEC (vs RoCEv2) exists as of March 2026.

The honest summary for a network engineer

UEC solves real problems. The PFC elimination is operationally valuable at scale. The retransmit model is more resilient than RoCEv2's QP-error recovery. The native packet spraying is cleaner than RSHP hacks.

But in March 2026, if you are deploying a DGX H100 or B200 cluster, you are deploying RoCEv2. The ConnectX-7 and BlueField-3 NICs in those nodes do not speak UEC. The SN5600/QM9700 switches are tuned for RoCEv2 with Spectrum-X lossless extensions.

UEC is the right reference point for what the next generation of AI networking looks like. It tells you what to evaluate in a 2026/2027 purchase decision when buying new NIC silicon. It tells you what to ask vendors: "Is this UEC 1.0 compliant?" "Does your switch support NPM?" "What is the retransmit overhead at 0.01% loss rate?"

Right now, the curriculum you have been building -- PFC, DCQCN, lossless fabric, RSHP, BGP unnumbered -- is what runs production AI clusters. UEC is where they are going.