The Ultra Ethernet Consortium (UEC) has recently unveiled the first version of its specification, dubbed UEC 1.0, which aims to optimize Ethernet technology for high-performance computing (HPC) and artificial intelligence (AI) applications. Formed in mid-2023 by networking giants including Cisco, Arista, HPE, and Intel, the consortium was established to tackle the unique demands posed by AI workloads and HPC environments.
The newly released UEC 1.0 specification encompasses enhancements intended to facilitate low-latency communications and high-throughput capabilities critical for advanced networking tasks. Key features of the specification include a new approach to Remote Direct Memory Access (RDMA), updated transport protocols, and improved congestion control measures.
J Metz, chair of the steering committee for the UEC, noted that the initial focus on HPC has evolved due to the rapid advancements in AI technology, particularly following the rise of tools like ChatGPT. He emphasized that while HPC has limited applications, AI spans a broader array of industries and interests, highlighting the urgency of UEC’s efforts.
A cornerstone of the UEC 1.0 specification is its approach to congestion control, which does not depend on a lossless network setup, thereby facilitating greater efficiency in AI operations. It allows a new operational mode wherein the receiver can actively control the rate of data transmission from the sender, boosting the ability to scale and utilize larger networks effectively.
Previously, Infiniband had been seen as a more potent alternative for HPC workloads when compared to Ethernet, primarily due to its robust performance characteristics. However, Metz states that UEC aims to bridge this gap by tailoring Ethernet’s delivery mechanisms to better align with workload requirements, thereby reducing the dependency on application-level changes.
Support for the UEC 1.0 specification has grown significantly, with companies like Arista Networks and Juniper Networks expressing excitement about its implications for modern networking. Arista plans to integrate the new transport protocol across its product portfolio, while Juniper is collaborating with AMD on solutions leveraging UEC technology.
HPE, which has contributed significantly to the foundation of the UEC 1.0 specification, has indicated that around 75% of its Ultra Ethernet Transport Protocol draws from its own Slingshot technology, optimized for next-generation supercomputers. HPE is also incorporating these advancements into its networking products.
Looking ahead, the UEC plans to establish further workgroups dedicated to specific initiatives, including storage, management, compliance, and performance enhancements, ensuring ongoing innovation and support for these new standards. The consortium is committed to educating networking professionals about UEC technologies and collaborating with industry partners to develop and refine their frameworks for the future.
