Unlocking large scale AI training networks with MRC (Multipath Reliable Connection)

2026-05-04 · Source: OpenAI News · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure · Depth: Advanced, long

Summary

OpenAI, in collaboration with AMD, Broadcom, Intel, Microsoft, and NVIDIA, has developed and released Multipath Reliable Connection (MRC), a novel network protocol designed to enhance GPU networking performance and resilience in large-scale AI training clusters. Released through the Open Compute Project (OCP) on May 5, 2026, MRC addresses critical challenges in supercomputer networking, such as congestion and frequent link/device failures, which become more pronounced with increasing cluster size. MRC enables multi-plane high-speed networks, allowing over 100,000 GPUs to connect with only two tiers of Ethernet switches, reducing power consumption and cost. Its adaptive packet spraying technique spreads single data transfers across hundreds of paths, virtually eliminating core congestion, while SRv6-based source routing bypasses failures in microseconds, ensuring predictable performance even during network disruptions. This protocol is already deployed across OpenAI's largest NVIDIA GB200 supercomputers, including those with Oracle Cloud Infrastructure (OCI) in Abilene, Texas, and Microsoft's Fairwater supercomputers.

Key takeaway

For MLOps engineers and infrastructure architects building or operating large-scale AI training environments, adopting the MRC protocol can significantly enhance network reliability and efficiency. You can reduce costly training job interruptions caused by network failures and congestion, allowing for more consistent and faster model development. Consider integrating MRC-compatible hardware and software to future-proof your supercomputing infrastructure for frontier model training.

Key insights

MRC improves large-scale AI training network reliability and performance through multipath packet spraying and static source routing.

Principles

• Redundancy improves network resilience.
• Distribute traffic to avoid congestion.
• Static routing simplifies failure recovery.

Method

MRC extends RoCE with SRv6-based source routing, splitting network interfaces into multiple planes, spraying packets across hundreds of paths, and using static routing tables for failure bypass.

In practice

• Deploy MRC for resilient AI supercomputing.
• Utilize multi-plane networks for cost/power savings.
• Implement SRv6 for simplified network control.

Topics

• Multipath Reliable Connection
• AI Supercomputer Networking
• GPU Training Clusters
• Multi-plane Networks
• Packet Spraying

Best for: MLOps Engineer, CTO, VP of Engineering/Data, AI Architect, AI Engineer, Machine Learning Engineer

Related on AIssential

• The Counterintuitive Networking Decisions Behind OpenAI’s 131,000-GPU Training Fabric — MRC inverts traditional networking principles to achieve predictable performance at extreme AI training scale.
• HPC Coding Agent - Part 3: MCP Tool for Profiling — An AI agent can autonomously profile and optimize GPU-accelerated HPC applications using a custom MCP server for tool integration.
• Meta’s Broadcom Liaison Enters Next AI Phase — Custom AI accelerators and specialized networking are critical for hyperscalers to achieve massive-scale AI compute.
• Jensen Huang Just Crowned the Next $1 Trillion AI Company — and It Doesn't Make a Single GPU — Connectivity is the emerging bottleneck and essential component in disaggregated, distributed AI computing infrastructure.
• Why Most AI Models Fail Before Training Even Begins — Most AI failures originate from data and pipeline issues, not model algorithms, requiring an operational discipline.
• The HackerNoon Newsletter: The Companies Rewiring the Future of AI (6/17/2026) — AI infrastructure's future depends on advanced "wiring" and orchestration, not just raw compute power.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by OpenAI News.