Ultra Ethernet 1.1 Expansion for AI Scale-Up Networks

The Ultra Ethernet Consortium recently updated its networking standards to better accommodate the massive data demands of artificial intelligence and high-performance computing. Released on October 12, the UEC 1.1 specification introduces critical features such as scale-up network support, in-network collectives, and programmable congestion control to improve how data moves across modern clusters.

Enhancing Performance for Scale-Up Architectures

The shift toward massive artificial intelligence models requires networking that can handle short-range, high-speed connections between accelerators like GPUs. While scale-out networking focuses on connecting different server nodes across a large data center, scale-up networking handles the communication between components inside a single node or small cluster. These environments typically involve small data messages sent at extremely high rates, making the reduction of per-packet overhead a top priority for engineers.

To address these needs, the 1.1 update introduces the Unified Local profile, which provides a streamlined transport protocol specifically for short-range interconnects. This profile reduces the amount of state information and acknowledgment packets required, making it ideal for fabrics that use priority or credit-based flow control. By simplifying the communication process, the consortium aims to provide a more efficient alternative to traditional Ethernet for tightly coupled systems.

Another significant addition is the Unified Forwarding Header, an optional feature that can replace the standard IP header in specific environments. This change allows addressing to be handled within the existing 12-byte Ethernet media access control header, effectively cutting out the 20 bytes typically required for an IPv4 header. This reduction in overhead ensures that more of the available bandwidth is dedicated to the actual payload rather than administrative data.

Furthermore, the specification introduces an idle reduction feature that shrinks the required spacing between frames. By reducing the average gap from 84 bytes to a constant 80 bytes, the protocol allows for a higher density of data transmission. These technologies are being developed in coordination with other industry groups, including the Open Compute Project, to ensure they integrate well with broader hardware standards for scale-up networking.

Advanced Congestion Signaling and Programmable Management

Network congestion remains a primary bottleneck in high-performance environments where multiple senders may attempt to transmit data to the same destination simultaneously. When queues fill up, packets are delayed or dropped, which can significantly hinder the training of large AI models. The previous 1.0 version of the specification established basic transport and control mechanisms, but it lacked the flexibility to adapt to rapidly changing workloads.

One of the major limitations of previous standards was that congestion signals provided very little detail. A typical explicit congestion notification only indicates that a queue is filling up somewhere on the path, but it does not specify the location or the severity. The new Congestion Signaling feature addresses this by adding a small tag to packets that every switch along the path can update. This tag tracks specific metrics, such as the longest delay or the minimum available bandwidth encountered.

When the receiver returns this information to the sender, the sender gains a precise view of the worst conditions on the network path. This level of visibility allows for much more accurate adjustments to transmission speeds. This feedback loop is essential for maintaining high throughput without overwhelming the network hardware, especially during bursts of intense activity common in scientific computing and machine learning.

Perhaps the most significant change in version 1.1 is the introduction of Programmable Congestion Management. Traditionally, the algorithms that dictate how a network interface card reacts to congestion are fixed in hardware. The new specification allows these algorithms to run as software programs loaded onto the endpoint. This flexibility means that as AI workloads evolve, operators can update their congestion control strategies without needing to replace physical hardware.

Moving Collective Operations into the Network Fabric

In distributed computing, accelerators often need to work together as a group to solve complex mathematical problems. This involves collective operations, such as combining values from multiple sources into a single result and sharing that result back with every participant. Traditionally, these tasks are handled by the endpoints, which requires significant bandwidth as data is sent back and forth between servers.

The UEC 1.1 specification introduces In-Network Collectives to move these reduction operations directly into the network switches. By having the switch combine the data as it passes through the fabric, the system can drastically reduce the amount of traffic on the wire. For example, in a large-scale reduction task, an endpoint might only need to send half as much data as it would under a traditional setup to achieve the same result.

This process is managed by a specialized daemon that organizes various endpoints into communication groups. When each endpoint sends its data, the switch waits until all members of the group have contributed before processing the final result and returning it. This approach not only saves bandwidth but also reduces the latency associated with waiting for multiple rounds of point-to-point communication between servers.

As the industry moves from the specification phase to actual product development, the consortium is focusing on the management tools necessary for vendors to build compliant hardware. This includes defining data models for controllers and establishing clear test checklists for self-attestation. While many of these features require hardware-level support, many existing devices may gain compatibility through firmware updates because these concepts have been under development for a significant period.

The transition to version 1.1 represents a concerted effort to modernize Ethernet so it can compete with proprietary, high-speed interconnects. By integrating management planes and clarifying compliance standards, the consortium is paving the way for a more open and flexible networking ecosystem. This ensures that data center operators have access to high-performance tools that are not tied to a single vendor’s closed architecture.