5 Key Insights into OpenAI's MRC Networking Protocol for AI Supercomputers

When training the world’s most advanced AI models, the biggest bottleneck isn’t just computing power—it’s the network connecting thousands of GPUs. OpenAI, in partnership with AMD, Broadcom, Intel, Microsoft, and NVIDIA, has spent two years developing MRC (Multipath Reliable Connection), a new open networking protocol published through the Open Compute Project (OCP). This listicle breaks down the five most crucial aspects of MRC, from the core networking challenges it solves to the three innovative mechanisms that make it work.

1. Why Networking Is the Hidden Bottleneck in AI Training

Training a frontier AI model involves countless data transfers between GPUs and CPUs every second. A single delayed packet can stall an entire training step, causing expensive GPUs to sit idle. As cluster sizes grow, network congestion, link failures, and device errors become more frequent and harder to manage. With over 900 million weekly ChatGPT users, every second of GPU idle time translates into real cost and lost capability. OpenAI’s goal with MRC is not merely to build a fast network, but one that delivers predictable performance even when failures occur, ensuring training jobs keep moving smoothly.

2. MRC Builds on Proven Standards and Extends Them

MRC is not invented from scratch. It extends RDMA over Converged Ethernet (RoCE), an IBTA standard that allows direct memory access between machines over Ethernet, bypassing the CPU for maximum throughput. Additionally, MRC incorporates techniques from the Ultra Ethernet Consortium (UEC) and introduces SRv6-based source routing. In SRv6 (Segment Routing over IPv6), the sending machine embeds the exact route into the packet header, so switches no longer need complex calculations. This reduces processing load and saves power—critical at data-center scale. MRC essentially takes the best of existing protocols and layers on new capabilities to handle the demands of massive AI clusters.

3. Mechanism 1: Adaptive Packet Spraying Eliminates Congestion

Traditional RoCEv2 sends each transfer over a single network path, leading to congestion and hot spots. MRC tackles this with adaptive packet spraying: packets from the same transfer are spread across hundreds of paths simultaneously. If one path becomes unusable due to congestion or failure, packets automatically reroute via other available paths. This intelligent load balancing reduces core network congestion and minimizes the chance of packet loss, keeping training jobs running at full speed. As a result, MRC ensures that even in highly loaded clusters, individual flows achieve consistent low latency.

4. Mechanism 2: Multipath Reliable Delivery with Selective Retransmission

Packet losses are inevitable at scale, but how a protocol handles them makes all the difference. MRC uses a multipath reliable delivery mechanism that detects lost packets and retransmits only the missing data, not entire streams. By combining per-packet acknowledgments over multiple paths, MRC avoids the overhead of traditional TCP-style retransmissions. This selective retransmission, coupled with the ability to use alternative paths for lost packets, dramatically reduces the impact of link or device failures. The result is that training jobs can continue almost uninterrupted, with retransmission delays kept to a minimum.

5. Mechanism 3: End-to-End Flow Control and Congestion Detection

Beyond spraying and retransmission, MRC features advanced end-to-end flow control that continuously monitors network conditions. It uses per-flow pacing to prevent any single transfer from overwhelming a path, and it can dynamically adjust sending rates based on real-time congestion signals from switches. Additionally, MRC integrates with SRv6 source routing to quickly reroute around failed links or switches without waiting for global routing convergence. This holistic approach ensures that even under heavy load or component failures, the network maintains predictable, low-jitter performance—essential for synchronizing thousands of GPUs during large-scale training.

Back to top

MRC represents a significant step forward in open, scalable networking for AI. By making the protocol available through the Open Compute Project, OpenAI invites the entire industry to adopt, test, and improve it. For anyone building or operating large-scale AI training clusters, understanding MRC’s three core mechanisms—adaptive packet spraying, multipath reliable delivery, and end-to-end flow control—is essential to unlocking maximum GPU utilization and keeping training on track.