Birds of a Feather Archive

Ultra Ethernet: Next-Generation Networking for HPC Supercomputers and AI Factories


Authors: Torsten Hoefler (ETH Zürich, Microsoft Corporation), Eric Spada (Broadcom), Keith Underwood (HPE), Rip Sohan (Advanced Micro Devices (AMD)), Dhabaleswar K. (DK) Panda (The Ohio State University), Ron Brightwell (Sandia National Laboratories)

Abstract: Ultra Ethernet has emerged as a major effort to create a scalable, open networking architecture for both AI factories and high-performance computing systems. As AI clusters continue to grow while HPC workloads demand ever-higher communication performance and efficiency, many networking requirements are beginning to converge. This BoF provides a forum for researchers, system architects, operators, application developers, and network vendors to discuss experiences, emerging requirements, unresolved challenges, and future directions for large-scale networking. The discussion aims to identify common needs across HPC and AI and foster collaboration toward open, interoperable networking solutions for next-generation computing systems.

Long Description: The rapid growth of large-scale AI systems has transformed datacenter networking requirements. At the same time, HPC systems continue to push the limits of communication performance, scalability, resilience, and efficiency. Historically, these domains have evolved somewhat independently, often relying on different hardware and software ecosystems. Today, however, the distinction is becoming less clear as traditional supercomputers increasingly support AI workloads and AI infrastructures confront challenges long familiar to the HPC community.

Ultra Ethernet forms a significant effort to address the networking requirements of both communities using an open Ethernet-based ecosystem. As deployments begin to scale and software ecosystems mature, the broader community has an opportunity to discuss which requirements are shared, which remain unique, and what technical challenges still need to be addressed.

This BoF will bring together stakeholders from academia, national laboratories, hyperscalers, cloud providers, communication-library developers, network vendors, and standards organizations to discuss:

- Networking requirements across HPC and AI workloads - Scalability challenges in future AI factories and supercomputers - Experiences from production and pre-production deployments with Ethernet - Congestion control, collectives, load balancing, and resilience - Software ecosystem support and communication libraries - Opportunities for open standards and interoperability - Outstanding research problems and future directions

The objective is not to present a specific technology roadmap but to facilitate community discussion about the future of large-scale networking and the role that open Ethernet-based solutions may play in that future.

Expected outcomes include:

- Improved understanding of shared networking requirements across HPC and AI. - Identification of open technical challenges and research opportunities. - Stronger collaboration among networking researchers, practitioners, vendors, and system operators. - Community feedback on priorities for future networking infrastructure and software ecosystems.

Because networking technology is rapidly evolving and AI workloads continue to reshape system architecture, this discussion is particularly timely for the SC26 community.


Back to Birds of a Feather Archive Listing