Ultra Ethernet Consortium (UEC)
The Ultra Ethernet Consortium (UEC) is developing a groundbreaking open standard (Ultra Ethernet Transport, UET) specifically engineered to optimize
Source: mortalapps.com- The Ultra Ethernet Consortium (UEC) is developing a groundbreaking open standard (Ultra Ethernet Transport, UET) specifically engineered to optimize Ethernet for AI and HPC workloads.
- UET deliberately replaces legacy transport protocols with advanced multi-path packet spraying, extremely fast congestion reaction, and robust out-of-order packet delivery handling.
- It introduces standardized In-Network Computing (INC) over Ethernet, enabling switch-level offloads for collective operations, functioning similarly to InfiniBand's proprietary SHARP technology.
- The initiative aims to provide highly interoperable, hyperscale-friendly alternatives to proprietary fabrics while meticulously maintaining fundamental Ethernet backward compatibility.
Why This Matters
As AI clusters aggressively scale to hundreds of thousands of GPUs, hyperscalers face immense vendor lock-in and precarious supply chain risks associated with proprietary interconnects like InfiniBand. Legacy Ethernet transports (like RoCEv2 and TCP) rely on fragile flow control mechanisms (PFC) and single-path routing algorithms (ECMP) that buckle under massive, synchronized AI incast workloads. UEC provides a rigorous open-standard roadmap to achieve lossless, high-performance scale-out networking utilizing commodity Ethernet silicon, democratizing exascale infrastructure.
Core Intuition
Traditional Ethernet protocols were architected for the Internet—general-purpose, strictly ordered, and reliable delivery across highly unpredictable networks. AI workloads, however, are highly predictable but intensely dense. If a single packet is delayed in a standard network, Head-of-Line (HoL) blocking stalls the entire pipeline. UET acts like a highly coordinated logistics company that intentionally breaks a massive shipment into tiny pieces, sprays them across every available highway simultaneously (multi-pathing), and allows the destination to assemble them out-of-order. This prevents any single traffic jam from delaying the overall transfer.
Technical Deep Dive
The UET protocol operates directly over standard L2/L3 Ethernet infrastructure but completely revamps the Transport layer stack. It is rigorously segmented into four distinct sub-layers. First, the Packet Delivery Sub-layer (PDS) defines packet types and supports four delivery modes, heavily prioritizing Reliable Unordered Delivery (RUD) to facilitate packet spraying and eliminate HoL blocking. Second, the Semantic Sub-layer (SES) handles advanced memory semantics and atomic operations crucial for distributed memory access. Third, the Congestion Management Sub-layer (CMS) replaces slow TCP/DCQCN algorithms with high-speed telemetry, relying on Network Signal Congestion Control (NSCC) at the sender, and Receiver Credit Congestion Control (RCCC) to manage in-cast scenarios. Finally, the Transport Security Sub-layer (TSS) embeds encryption directly into the protocol to ensure strict multi-tenant isolation.
Furthermore, UEC explicitly standardizes In-Network Collectives (INC), defining a lightweight, interoperable protocol for Ethernet switches to actively perform mathematical reductions on AI data streams, establishing the first open standard of its kind for Ethernet.