The Ultra Ethernet Consortium (UEC) tweaked its flagship standard to boost scale-up networking and improve congestion control.

Version 1.1 comes a little over a year after the initial iteration launched and largely focuses on networking inside the rack, including reducing packet sizes to improve performance. V1.1 features a reduced interpacket gap to reduce the number of idle bytes that must be inserted between packets, while a unified forwarding header (UFH) overlays atop traditional Ethernet headers to save as much as 20 bytes compared to a traditional IP packet.

UEC’s packet reduction efforts come as the sheer amount of information passing through high-end networks is soaring thanks to AI.

To counter spiraling workload sizes causing bottlenecks in the network, UEC developed “in-network collectives (INC)” for V1.1, reducing bandwidth consumption by sending into the network just once. Where a traditional network would pass data from points A to B twice across two separate phases (reduce and broadcast), the group argues that waiting for packets to reach a destination is slow. Instead, INC proposes processing on the fly done inside the switch hardware.

A UEC post explains: “To produce a one-gigabyte (GB) reduced result using this traditional approach, each endpoint must send 2 GB of data. With INC, an endpoint needs to send only 1 GB of data, yielding up to a 2x communication speedup. Network switches reduce the data as it is received and broadcast the reduced data back to the endpoints.”

Another packet-focused addition to V1.1 is congestion signaling (CSIG). Here a layer-two (L2) tag is attached to packets. As they move through the network, switches can update the tag to provide data to engineers on bottleneck congestion experienced by that specific packet. UEC argues that having direct information on telemetry from endpoints across a network allows for more accurate adjustments that can improve performance.

Also added are a set of dedicated APIs for network interface card (NIC) vendors to iterate congestion control and load balancing algorithms. Programmable congestion management (PCM) APIs are designed to be ported to any NIC with PCM support.

“We’re very excited about the advances in the v1.1 specification,” the group said upon launch. “As ever, UEC welcomes new members who would like to contribute ideas to make Ethernet networks even better.

The likes of AMD, Broadcom, Cisco, and Meta all leant support to the UEC upon its founding in 2023. Nvidia would join a year later.

NTT was among the recent big-name additions, with Nokia and Broadcom among the vendors having already extended support for the spec.