NVIDIA is promoting Spectrum-X Ethernet as an open, Artificial Intelligence-native networking fabric for large-scale training infrastructure, aimed at supporting the performance, resilience and scale required by the biggest Artificial Intelligence factories. OpenAI, Microsoft and Oracle are identified among the organizations deploying the platform in environments where network efficiency and availability are critical to keeping large model training on track.
A central addition is Multipath Reliable Connection, an RDMA transport protocol introduced through collaboration among NVIDIA, Microsoft and OpenAI and released as an open specification through the Open Compute Project. MRC enables a single RDMA connection to distribute traffic across multiple network paths, improving throughput, load balancing and availability for large-scale Artificial Intelligence training fabrics. NVIDIA says the protocol was first proven in production and optimized on Spectrum-X Ethernet hardware, where purpose-built hardware, telemetry and intelligent fabric control helped move it from concept into large-scale deployment.
The design is intended to keep GPU utilization high by spreading traffic across available paths and dynamically steering around congestion. When data loss occurs, intelligent retransmission is designed to speed recovery and reduce disruption to long-running jobs. Administrators also get more detailed visibility into traffic paths, which can simplify operations and accelerate troubleshooting across large environments.
NVIDIA also highlights resilience features in Spectrum-X Ethernet running MRC. Its failure bypass technology can, in just microseconds, detect a network path failure and reroute traffic automatically in hardware. The company argues that this is especially important in Artificial Intelligence training clusters where thousands of GPUs must remain synchronized, because even brief network interruptions can slow or halt an entire job.
Another key element is support for multiplanar network designs, which OpenAI deploys with Spectrum-X Ethernet together with MRC. NVIDIA says its Spectrum-X Multiplane capability adds hardware-accelerated load balancing across independent network planes, improving resiliency and scale while maintaining predictable latency and supporting expansion to hundreds of thousands of GPUs. Spectrum-X Ethernet also supports multiple RDMA transport options, including Adaptive RDMA, MRC and custom protocols, running across NVIDIA ConnectX SuperNICs and Spectrum-X Ethernet switches.
NVIDIA frames MRC as part of a broader push toward open, flexible networking for modern Artificial Intelligence infrastructure. The company says Spectrum-X Ethernet gives customers a choice of transport models while integrating across large cluster deployments. NVIDIA collaborated on MRC development with AMD, Broadcom, Intel, Microsoft and OpenAI.
