Enter for the full results page · Esc to close
Data Centre & AI Networking
An industry effort to redesign Ethernet's transport for AI and HPC — keeping the Ethernet ecosystem, replacing the parts RDMA struggles with.
The Ultra Ethernet Consortium is a multi-vendor effort to give Ethernet the properties AI and HPC need without leaving the Ethernet ecosystem. Its premise is that the physical layer, the switch silicon and the operational tooling are all fine, and that the thing that never fitted is the transport RDMA was bolted onto.
It is aimed squarely at the operational pain of RoCE: the tuning burden, the PFC hazards, and the fragility of holding a lossless invariant across hundreds of switches through firmware upgrades and hardware refreshes.
Ultra Ethernet Transport tolerates out-of-order delivery as normal and recovers from real loss selectively, instead of demanding that the network never lose or reorder anything. That single change unlocks packet spraying: an entropy value carried in the UDP source port is varied per packet, so one flow uses every available path rather than being pinned to one by a hash.
It also drops the per-peer connection state that limits scale. Delivery contexts are ephemeral — established with a lightweight message and discarded — which is how the specification reaches for millions of endpoints. Three profiles exist: AI Base for collective libraries, AI Full above it, and HPC with the tag matching MPI expects.
Specification 1.0 was published in June 2025, running to roughly 560 pages across the transport, the link layer, congestion management and security. Maintenance releases have followed steadily — 1.0.1 that September, 1.0.2 in January 2026 and 1.0.3 in July 2026 — while the working groups extend programmable congestion management, congestion signalling and in-network collectives.
Silicon and adapters are the gating item rather than the document. The question worth asking a vendor is not whether they support Ultra Ethernet — everyone says yes — but whether the specific switch ASIC and adapter being quoted can run UET through a firmware update or need new hardware. Ultra Ethernet vs RoCE sets out how to ask it.
Nothing, and that is the point worth making to anyone specifying a plant. UEC deliberately leaves the PHY and the link layer alone, so the same 800G SR8 modules, the same passive DAC and the same MPO trunks serve a UEC fabric and a RoCE fabric identically. Cabling bought today is not stranded by the transition.
The second-order effect is real, though. If spraying raises usable fabric capacity toward installed capacity, a given job needs fewer links to run at the same speed — which changes the port count, not the port type. That is a TCO question, not a compatibility one.
| Specification 1.0 | published June 2025, about 560 pages |
|---|---|
| Maintenance releases | 1.0.1 Sept 2025, 1.0.2 Jan 2026, 1.0.3 July 2026 |
| Transport | Ultra Ethernet Transport (UET) |
| Connection model | ephemeral packet delivery contexts, not long-lived queue pairs |
| Spraying mechanism | an entropy value carried in the UDP source port |
| Profiles | AI Base, AI Full, and HPC |
| Link layer options | link-level retry and credit-based flow control |
| Stated scale target | millions of endpoints |
| Optics impact | none — same PHY, same modules, same fibre |
RoCE, InfiniBand, lossless Ethernet, ECN, InfiniBand vs RoCE vs Ethernet.
Send the specification and an engineer answers it — no obligation, and no sales sequence. If the answer is that you do not need what we sell, that is the answer you get.
Usually answered within one working day.