inference/research
verifiedmedium · 0.80updated 2026-10-03

Ultra Ethernet (UEC) and Ethernet scale-out baseline

None GB/s

Fields

id
interconnect/ethernet-uec
Kind
ethernet
Version
UE Specification 1.0.3
Bandwidth basis
NULL - deliberately, and this is the whole point of the record. Ethernet's per-port rate is set by IEEE 802.3 physical standards (currently to 800 Gb/s per the IBTA's statement that link speeds reach 800 Gb/s) and by the NIC/switch product generation, not by the Ultra Ethernet Consortium. UEC specifies the communications STACK - transport, congestion control, packet processing - to make that commodity bandwidth usable for AI and HPC. No per-link rate is recorded because any number would belong to an 802.3 or vendor record, and inventing one here would double-count.
Topology
Switched Ethernet, classically leaf-spine. UEC's design goal is precisely to keep this topology while fixing the collective-communication behaviour that makes commodity Ethernet awkward for AI.
Scale-up
no
Scale-out
yes
Switching
Standard Ethernet switching, with UEC specifying multi-pathing, congestion control and packet handling so that fat-tree oversubscription and PFC-induced problems are addressed in the stack rather than by traffic engineering.
notes
The scale-out baseline, and the counterweight to InfiniBand. Current specification state: UEC publishes UE Specification 1.0.3 alongside the 1.0 white paper, under the Linux Foundation, with a member base spanning AMD, Arista, Broadcom, Cisco, HPE, Intel, Meta, Microsoft, Alibaba Cloud, Google Cloud, Huawei and ZTE - broad enough that Ethernet scale-out is not a single-vendor path. Inference consequence: this is the tier that decides whether a model can be sharded across NODES. Its practical limits are fabric-level, not link-level: uplink-to-downlink ratio caps bisection bandwidth (2:1 halves all-to-all), and tail latency under incast shows up as inter-token latency variance in decode. UEC exists to attack the second problem without giving up the first one's economics. Compare nvswitch-4/5 (scale-up, where bandwidth per direction is ~450-900 GB/s and topology is non-blocking by construction) with this record (scale-out, where oversubscription and tail latency dominate). 2026-10-10: set null_bandwidth=definitional. The record's own basis/notes already establish the reason; the field only names the case the schema now distinguishes. No number changed.
null bandwidth
definitional

Notes

The scale-out baseline, and the counterweight to InfiniBand. Current specification state: UEC publishes UE Specification 1.0.3 alongside the 1.0 white paper, under the Linux Foundation, with a member base spanning AMD, Arista, Broadcom, Cisco, HPE, Intel, Meta, Microsoft, Alibaba Cloud, Google Cloud, Huawei and ZTE - broad enough that Ethernet scale-out is not a single-vendor path. Inference consequence: this is the tier that decides whether a model can be sharded across NODES. Its practical limits are fabric-level, not link-level: uplink-to-downlink ratio caps bisection bandwidth (2:1 halves all-to-all), and tail latency under incast shows up as inter-token latency variance in decode. UEC exists to attack the second problem without giving up the first one's economics. Compare nvswitch-4/5 (scale-up, where bandwidth per direction is ~450-900 GB/s and topology is non-blocking by construction) with this record (scale-out, where oversubscription and tail latency dominate).

2026-10-10: set null_bandwidth=definitional. The record's own basis/notes already establish the reason; the field only names the case the schema now distinguishes. No number changed.

Cross-references

This record references no other records — it is a leaf. It is cited by 11.