UEC model 1.0
Work on the UEC specs is following what the group calls a really aggressive timeline, with model 1.0 slated to be launched within the third quarter of 2024. The UEC 1.0 Overview explains a number of the group’s priorities for the forthcoming specification.
“Even when contemplating the benefits of utilizing Ethernet, enhancements can and ought to be made,” the UEC said. “Networks should evolve to raised ship this unprecedented efficiency for the elevated scale and better bandwidth of networks of the long run. Paramount is the necessity to have the community assist supply of messages to all taking part endpoints as shortly as doable, with out lengthy delays for even a number of endpoints.”
For instance, the UEC cites the necessity to reduce “tail latency” within the coaching of AI fashions: “Coaching consists of frequent computation and communications phases, the place the initiation of the subsequent part of the coaching relies on the completion of the communication part throughout the suite of GPUs. The final message to reach gates the progress of all GPUs. This tail latency – measured by the arrival time of the final message within the communication part – is a vital metric in system efficiency.”
To attain low tail latency, the UEC specification will tackle vital networking necessities for the subsequent technology of purposes, together with:
- Multi-pathing and packet spraying
- Versatile supply order
- Fashionable congestion management mechanisms
- Finish-to-end telemetry
- Bigger scale, stability, and reliability
“This final level locations an additional burden on the entire earlier ones,” the UEC said. “Excessive-performance techniques go away little margin for error, which compounds in a bigger community. Determinism and predictability change into harder as techniques develop, necessitating new strategies to realize holistic stability.”
One other of the challenges UEC is working to handle for AI and high-performance networks is establishing the power to assist a number of pathways for communications between clusters.
