Intel, Microsoft, Other Tech Giants Want To Boost AI, HPC With ‘Ultra Ethernet’ Solution Stack
The newly formed Ultra Ethernet Consortium—whose members also include AMD, Broadcom, Cisco and Hewlett Packard Enterprise—seeks to ‘build a complete Ethernet-based communication stack architecture’ for the fast-growing market of high-speed data center networking.
Intel and Microsoft are among several tech giants planning to develop an open “Ultra Ethernet” solution stack that will address the growing network demands of AI and high-performance computing workloads.
The companies announced their plan Wednesday through a new group called the Ultra Ethernet Consortium, which seeks to “build a complete Ethernet-based communication stack architecture for high-performance networking” though “industry-wide cooperation.”
[Related: The ChatGPT-Fueled AI Gold Rush: How Solution Providers Are Cashing In]
Established by The Linux Foundation, the consortium’s other founders include network chip giant Broadcom, networking hardware vendors Cisco and Arista Networks, Intel chip rival AMD, server behemoth Hewlett Packard Enterprise, Facebook parent company Meta and Eviden, a cloud, big data and cybersecurity business owned by solution provider powerhouse Atos.
The group is tackling a fast-growing area that could represent major opportunities for channel partners. According to research firm 650 Group, the data center AI networking market is expected to grow to $3 billion this year and reach $10 billion by 2027.
“This isn’t about overhauling Ethernet,” said Dr. J Metz, chair of the Ultra Ethernet Consortium, in a statement. “It’s about tuning Ethernet to improve efficiency for workloads with specific performance requirements. We’re looking at every layer—from the physical all the way through the software layers—to find the best way to improve efficiency and performance at scale.”
A Possible Swipe At Nvidia?
A solution provider executive told CRN that one possible aim of the Ultra Ethernet Consortium is to provide an alternative to Nvidia’s InfiniBand networking solutions, which are targeted for AI and HPC workloads. He found it noteworthy that Nvidia isn’t a member of the group.
“They’re trying to come up with an alternative way of handling these fabrics here,” said Dominic Daninger, vice president of engineering at Nor-Tech, a Burnsville, Minn.-based HPC system integrator.
Nvidia didn’t respond to a request for comment.
If successful in building an interoperable and faster Ethernet-based solution stack, the group could chip away at Nvidia’s dominance in the AI and HPC space, according to Daninger. That’s because Nvidia pushes InfiniBand solutions, which the company gained through its 2021 acquisition of Mellanox, as the best low-latency solution for connecting GPU clusters to perform such workloads.
“It could be that companies like Intel, AMD and Broadcom are interested in getting away from being so tied to Nvidia for accelerators and the fabrics that tie those together,” he said.
Group Could Influence Products Arriving In 2024
With the goal of developing an Ethernet-based architecture that exceeds the performance of “today’s specialized technologies,” the group is putting an emphasis on functionality, performance, total cost of ownership, interoperability and a friendly environment for users and developers.
From a technical perspective, the Ultra Ethernet Consortium is looking to develop specifications, APIs and source code that define several aspects in the Ethernet-based solution stack.
These aspects in the stack include the software, storage, management and security constructs needed to enable a variety of workloads and operating environments. They also cover link and transport protocols as well as electrical and optical signaling characteristics plus application program interfaces and data structures for Ethernet communications.
The Ultra Ethernet Consortium expects the first products using its specifications to arrive in 2024.