Ethernet has long been driven forward by standards organizations, notably the IEEE, across its 50-year history and it is now getting (yet another) organization to advance the pervasive networking standard for the artificial intelligence (AI) era.
The Ultra Ethernet Consortium (UEC) is officially launching today, as a new organization hosted by the Linux Foundation and its Joint Development Foundation initiative. The goal of the UEC is to go beyond existing Ethernet capabilities such as remote direct memory access (RDMA) and RDMA over Converged Ethernet (RoCE) to provide a high performance, distributed and lossless transport layer that is optimized for high performance computing (HPC) and AI. It's an effort that is taking direct aim at rival transport protocol InfiniBand.
Founding members of the UEC include AMD, Arista, Broadcom, Cisco, Eviden , HPE, Intel, Meta and Microsoft,
“We wanted to create an approach to solving some of the problems that high performance computing and AI have been bringing into the networking realm,” J Metz, chair of the steering committee for the Ultra Ethernet Consortium told SDxCentral. Metz is also a technical director, system design at AMD.
What the Ultra Ethernet Consortium plans to doThe consortium plans to make changes at multiple layers of the Ethernet stack, from the physical layer up through the software layer.
Metz noted that there is no shortage of network standards and organizations working on Ethernet today. While the IEEE has a primary role to play, Metz said that the UEC is looking at more than just the physical transport layer that the IEEE is typically concerned with.
The goal with the UEC is to look at all the elements needed to improve Ethernet and then work with the relevant standards bodies and technical groups to get those improvements implemented.
“The intent was that we are going to get our own ducks in order and then we were going to go work with these different organizations where appropriate,” he said.
RDMA vs. UEC TransportAmong the first efforts from the UEC to improve Ethernet is a technology currently known as UEC Transport.
Metz said that UEC Transport is being developed to help provide better Ethernet transport than what is currently possible with RDMA. While RDMA will still be supported, UEC aims to provide a more flexible transport that does not require a lossless network. This will allow for features like multi-pathing and out-of-order packet delivery needed for many-to-many AI workloads.
“UEC Transport is a new form of approaching the transportation layer that has some semantic adjustments, congestion notification protocols, and some security enhancements,” Metz said.
The road ahead for the UECThe consortium was incorporated under the Linux Foundation's Joint Development Foundation to leverage an existing legal and governance framework.
Uri Elzur, chair of the technical advisory committee for the UEC told SDxCentral that ensuring interoperability through plugfests and compliance testing will be a focus going forward. Elzur said the goal is for UEC to make sure it is open and interoperable.
“To meet our goals, there's going to be a set of things that must change, while we preserve the goodness of the Ethernet environment and we will build a stack on top of that,” Elzur said.
Comments