Standardizing the Fabric.

Today's AI clusters are dominated by **InfiniBand**, a specialized, proprietary fabric optimized for HPC. To compete, a consortium including Microsoft, Meta, AMD, and Intel founded the **Ultra Ethernet Consortium (UEC)**.

Unlike standard Ethernet, which uses flow-based hashing (ECMP) and handles congestion poorly, UEC introduces **Packet-Level Spraying** and hardware-native reordering. This allows the network to treat every switch port as a single giant pipe, achieving near-perfect utilization for the world's largest GPU clusters.

Scalable Transport

Replaces the aging RoCE/TCP stack with a hardware-native transport that allows out-of-order delivery across multiple paths.

Multipath Flow

Automatically divides large 'Elephant Flows' into 'Mice-like Packets' and sends them simultaneously through every available switch path.

UEC vs. IB Modeler.

Simulate your TFLOPS for a 1,024-node cluster using Ultra Ethernet vs. HDR/NDR InfiniBand. Calculate your ROI based on cost-per-Gbps.

Share Article

Technical Standards & References

REF [uec-spec-whitepaper]
UEC Steering Committee (2024)
Ultra Ethernet Consortium: A Scalable Transport for AI and HPC
Published: UEC Technical Whitepapers
VIEW OFFICIAL SOURCE
REF [ieee-high-performance]
L. Zhang et al. (2023)
The Future of Low-Latency Ethernet in Distributed AI Training
Published: IEEE Communications Magazine
VIEW OFFICIAL SOURCE
Mathematical models derived from standard engineering protocols. Not for human safety critical systems without redundant validation.