Leading cloud, semiconductor, and systems providers unite to form the NoLimit Foundation
San Francisco. A group of companies with long records in high-performance systems announced today that they are forming the NoLimit Foundation, an industry effort to build a complete Ethernet-based communication stack for AI and HPC at scale. The goal is not to replace Ethernet. It is to tune it, so that the networks running the largest training clusters and supercomputers can get the performance they need without leaving the ecosystem the rest of the data center already depends on.
The founding steering members are AMD, Arista, Broadcom, Cisco, Eviden, HPE, Intel, Meta, Microsoft, and Oracle. Between them they carry decades of networking, cloud, AI, and HPC deployments at scale, and each is contributing to the broader ecosystem on an equal footing.
AI and HPC workloads are moving fast and they demand best-in-class functionality, performance, interoperability, and total cost of ownership, all without sacrificing developer and end-user friendliness. Ethernet already has the ubiquity and flexibility to handle a wide range of workloads at low cost. What it has lacked is a coordinated set of enhancements aimed squarely at scale-out AI and HPC fabrics. That is the gap the foundation is here to close.
Our approach is to look at every layer, from the physical through the software, and change only what these workloads actually require. Communication stack changes will be kept to a minimum, and Ethernet interoperability will be maintained and promoted throughout.
What the foundation will define
- Protocols, electrical and optical signaling characteristics, APIs, and data structures for Ethernet communication.
- Link-level and end-to-end transport protocols that extend or replace existing ones.
- Link-level and end-to-end congestion, telemetry, and signaling mechanisms suited to AI, machine learning, and HPC environments.
- Software, storage, management, and security constructs that support a variety of workloads and operating environments.
The work is organized into modular, compatible, interoperable layers with tight integration between them. The founding companies are seeding four working groups with substantial contributions: Physical Layer, Link Layer, Transport Layer, and Software Layer.
Industry analysts who follow HPC and AI have said much the same thing in different words: users are leaving performance on the floor because of interconnect weaknesses, they do not want to learn a second fabric, and because most data centers already run on Ethernet, an extensible Ethernet-based solution makes scale more straightforward and more accessible. Several noted that the debate over which network should carry generative AI training and inference has been leaning toward Ethernet for its ease of installation and use, and that a coordinated effort like this one will be welcome.
The foundation will begin accepting applications for new members later this quarter. Details will be published on this site.