OpenAI, in a landmark collaboration with industry giants including AMD, Broadcom, Intel, Microsoft, and NVIDIA, has officially announced the release of Multipath Reliable Connection (MRC). This new open network protocol is designed specifically for GPU interconnects within large-scale artificial intelligence training clusters. By addressing the critical "compute crunch" facing the industry, MRC aims to optimize data transmission across hundreds of thousands of processors, ensuring maximum hardware utilization for the next generation of frontier AI models.
Technical Innovation: Beyond Standard RoCE
The MRC protocol represents a significant evolution of the Remote Direct Memory Access over Converged Ethernet (RoCE) standard. By integrating SRv6 (Segment Routing over IPv6) source routing, the protocol effectively eliminates traditional network bottlenecks. Unlike standard transmissions that rely on a single path, MRC "sprays" data packets across hundreds of available routes simultaneously.
Key technical features of the protocol include:
- Microsecond Failure Bypass: The ability to detect and bypass link or switch failures in microseconds, preventing long-running training jobs from stalling.
- Hardware-Accelerated Load Balancing: Dynamic distribution of traffic to maintain high bandwidth even during periods of heavy core congestion.
- Reduced Infrastructure Footprint: The capacity to connect over 100,000 GPUs using only two layers of switches, significantly lowering power consumption and hardware costs.
Production Deployment and Industry Adoption
OpenAI has already validated the effectiveness of MRC through real-world deployments in massive supercomputing environments. The protocol is a core component of the Stargate supercomputer, developed in partnership with Oracle Cloud Infrastructure (OCI), as well as the Microsoft Fairwater facility. These "AI factories" represent the pinnacle of current infrastructure, designed to support the training of models with trillions of parameters.
MRC delivers high levels of GPU utilization by load-balancing traffic across all available paths, enabling every GPU to get the bandwidth it needs throughout a training run.
To encourage widespread adoption and ensure interoperability across different vendor ecosystems, the MRC specification has been submitted to the Open Compute Project (OCP). This move allows the broader tech industry to develop Spectrum-X compatible networking stacks, fostering a more open environment for high-performance computing (HPC) and blockchain-integrated AI services.
In conclusion, the introduction of the MRC protocol marks a strategic shift toward open standards in the high-stakes race for AI dominance. By resolving the networking inefficiencies inherent in massive GPU clusters, OpenAI and its partners are providing the foundational infrastructure necessary for the reliable scaling of artificial general intelligence and complex decentralized computing networks.
Frequently Asked Questions
Quick answers to the most common questions about this topic.
For educational purposes only. Nothing here constitutes financial or investment advice. Crypto markets are highly volatile — always DYOR before making any decisions.
Fact-checked per our editorial policy. We maintain strict independence from advertisers. Spot an error? Let us know.