Nokia looks to be extending its hand to Multipath Reliable Connection (MRC), the OpenAI-developed transport protocol for large-scale AI training networks.
A company blog post describes the protocol as “an important evolution in Ethernet transport” and that foundational capabilities for MRC-based architectures are already found in its portfolio.
Incorporating some of the design principles found in Ultra Ethernet, MRC makes use of IPv6 segment routing (SRv6)-based source routing to allow senders to choose the optimal path packets should take through networks with up to millions of links, even across disparate locations.
Nokia already supports open networking, with the vendor having joined the Software for Open Networking in the Cloud (SONiC) Foundation last December as well as its own SR Linux network operating system (NOS).
Alfred Nothaft, senior director of product management for Nokia’s service router product group, said the vendor believes its customers should have the option to pick a preferred operating model, adding: “Whether an AI fabric is built on SONiC or SR Linux, the goal is the same: provide a high-performance, resilient network that can scale efficiently to meet the growing demands of AI workloads.”
While the Finnish firm hasn’t explicitly said it’s supporting the concept, Nothaft’s posts saw the exec tout the need for AI-focused networks to make use of means like MRC that are capable of ensuring GPUs in a distributed computing system can coordinate as one. He said MRC “brings a fresh approach to how AI traffic is distributed across the fabric, helping improve path utilization, respond more effectively to transient congestion, and make better use of the path diversity already built into modern Ethernet networks.”
OpenAI is the mind behind MRC, with Open Compute Project (OCP) members including Broadcom, Nvidia, and AMD contributing to the effort. Microsoft is among the hyperscalers already making use of it across hardware from Nvidia and Broadcom, with the cloud provider inadvertently revealing the concept to the world when detailing the ‘AI wide area network’ supporting its Fairwater data centers last November, a whole six months prior to its official unveiling.