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

In large-scale training, a single delayed data packet can idle thousands of GPUs, turning minor network jitter into a massive productivity drain. Adopting MRC allows infrastructure teams to maintain predictable performance despite the inevitable hardware failures inherent in massive clusters. While this protocol minimizes downtime for synchronous pretraining, it requires broad industry alignment to move beyond proprietary networking silos. Teams that ignore these architectural shifts risk constant, costly restarts as their compute clusters grow in size and complexity.

MRC extends RDMA over Converged Ethernet (RoCE) by integrating SRv6-based source routing to support 800Gb/s network interfaces. By moving away from traditional routing, the protocol eliminates core congestion and prevents single-point failures from stalling entire training jobs. The following table highlights the shift in network management:

Currently, the primary limitation remains the requirement for hardware support across the networking stack, including compatible network interfaces and switches.

This week

• Audit existing supercomputer networking stacks for compatibility with Open Compute Project standards.
• Review the 'Resilient AI Supercomputer Networking' paper to assess the impact of SRv6 routing on current cluster latency.

This quarter

• Evaluate the feasibility of migrating to MRC-compliant 800Gb/s hardware during the next scheduled infrastructure refresh.
• Engage with hardware vendors to determine their roadmap for supporting the MRC specification in upcoming switch deployments.

This year

• Integrate MRC-based networking into the design phase for new large-scale training cluster builds.
• Establish performance benchmarks for network-related job stalls to measure the ROI of implementing multi-plane network topologies.