OpenAI、スーパーコンピュータ向けネットワーク技術「MRC」をOCP経由で公開Unlocking large scale AI training networks with MRC (Multipath Reliable Connection)

OpenAI has introduced a networking technology called MRC, short for Multipath Reliable Connection, aimed at interconnecting the supercomputer-scale GPU clusters that train its largest models. The announcement matters because the network fabric has become one of the central bottlenecks in modern AI training, where tens of thousands of accelerators must exchange data almost continuously, and a single congested or failed link can stall an entire job that spans thousands of machines.

According to OpenAI, MRC is being released through the Open Compute Project (OCP), the industry consortium originally founded by Meta to share open hardware and data-center designs. Publishing the protocol through OCP rather than keeping it proprietary suggests OpenAI is seeking broader interoperability and external adoption, a path that could let hardware vendors and other operators build compatible switches, network interface cards, and software stacks. The move aligns with a wider trend of large AI labs and cloud providers contributing infrastructure specifications to open bodies in order to influence the supply chain they depend on.

The core idea behind MRC, as its name indicates, is to spread traffic across multiple network paths simultaneously while preserving reliable delivery. Large-scale training depends heavily on collective communication operations such as all-reduce, all-gather, and reduce-scatter, in which gradients and model parameters are synchronized across every participating GPU. These collectives are only as fast as their slowest path, so uneven load distribution or transient congestion can disproportionately slow down an entire training step. By using several paths at once and rerouting around hotspots or faults, a multipath approach is intended to raise effective throughput and reduce the impact of so-called stragglers and tail latency.

This addresses a known limitation of conventional data-center routing. Many AI clusters today run RDMA over Converged Ethernet (RoCE) or InfiniBand, and they often rely on equal-cost multi-path (ECMP) routing that hashes each flow onto a single path. When several large flows hash onto the same link, the result is congestion on some paths while others sit idle. Techniques that distribute packets more granularly across the fabric, sometimes described as packet spraying, attempt to balance load more evenly, but they require careful handling to reassemble data correctly and to maintain reliable, in-order delivery. MRC appears to target exactly this class of problem, combining multipath distribution with reliability mechanisms suited to the bursty, synchronized traffic that AI workloads generate.

MRC also enters a crowded and fast-moving area of the industry. The Ultra Ethernet Consortium (UEC), which includes major chip, networking, and cloud companies, is developing updated Ethernet specifications for AI and high-performance computing that similarly emphasize multipathing and modern congestion control. NVIDIA, meanwhile, promotes its InfiniBand and Spectrum-X Ethernet platforms, along with NVLink for tight intra-server GPU connections, as integrated alternatives. It is not yet fully clear from the announcement how MRC relates to these efforts, whether it complements emerging Ethernet standards or defines a distinct protocol layer, and the details that OpenAI publishes through OCP will likely determine how easily it can be implemented on existing or near-future hardware.

The context for this work is the relentless growth in cluster size. Training frontier models now involves coordinating very large numbers of accelerators across many racks, and at that scale the probability of component failures during a long run rises substantially. Resilience features that allow the network to keep operating smoothly when individual links degrade are therefore as important as raw bandwidth. OpenAI frames MRC as contributing to both goals, improving performance under normal conditions and maintaining stability when parts of the fabric falter.

For readers tracking the AI infrastructure landscape, MRC is a reminder that progress in model capability is increasingly tied to advances below the software layer, in the interconnects, switches, and protocols that bind accelerators together. The practical impact will depend on independent implementation, published benchmarks, and how the protocol performs in production deployments outside OpenAI. As an openly released specification, it offers other operators a chance to evaluate the design directly, though its long-term significance will become clearer only as adoption, tooling, and real-world results accumulate.