Spectrum-X
NVIDIA's Ethernet platform tuning RoCE for AI collectives with adaptive routing and congestion control.
Current numbers
98–99% vs ~95%InfiniBand effective throughput vs Spectrum-X/tuned-Ethernet; untuned RoCEv2 far lower
~95%effective throughput on Spectrum-X Ethernet at xAI Colossus 100k–200k-GPU scale, zero flow-collision loss
~95%Spectrum-X / tuned-RoCE effective throughput vs raw link rate under collective load; untuned RoCE can collapse to ~60%