NVIDIA Spectrum-X sets gigascale AI standard with open AI-native Ethernet fabric, now supporting MRC

NVIDIA Spectrum-X Ethernet has strengthened its networking standard for gigascale AI factories by adding MRC (Multipath Reliable Connection). OpenAI, Microsoft, and Oracle are using this line of networking, and MRC allows a single RDMA connection to spread traffic across multiple paths, improving throughput, load balancing, and availability.

Also released as an open specification through the Open Compute Project (OCP), MRC is presented as a case of combining Spectrum-X hardware, deep telemetry, and intelligent fabric control. Large AI data centers such as Microsoft's Fairwater and Oracle Cloud Infrastructure's Abilene also use MRC to align performance, scale, and efficiency.

The key effects of MRC are as follows.

• Spreads traffic across every available path to boost GPU utilization.
• Maintains bandwidth by steering away from overloaded paths in real time when congestion occurs.
• Reduces jitter in long-running training jobs through precise retransmission when data loss occurs.
• Gives administrators finer visibility and control over traffic paths.

In addition, failure bypass technology detects network path failures at the microsecond level and automatically reroutes around them in hardware. In training clusters where thousands of GPUs are synchronized, even brief failures can slow down the entire job, which is why this hardware-level response is described as important.

Spectrum-X Multiplane provides hardware-accelerated load balancing for multi-plane networks made up of multiple independent fabrics, increasing scalability while maintaining low latency up to hundreds of thousands of GPUs. Spectrum-X natively supports Adaptive RDMA, MRC, and other custom protocols on ConnectX SuperNICs and Spectrum-X Ethernet switches, and MRC was developed together with AMD, Broadcom, Intel, Microsoft, and OpenAI.