NVIDIA Spectrum-X is the world's first end-to-end, AI-optimized Ethernet platform for multi-tenant generative AI factories. It delivers high effective bandwidth and low latency for hyperscale AI infrastructure.
The platform comprises Spectrum-X Ethernet switches (SN5000/SN6000 series, up to 102.4 Tb/s) and SuperNIC network adapters (BlueField-3, ConnectX-8, ConnectX-9) for GPU-to-GPU communication. Key mechanisms include adaptive routing (coordinating switches and SuperNICs), advanced congestion control with noise isolation between tenants, and RDMA over Converged Ethernet (RoCE).
Spectrum-X improves network performance by 1.6x versus standard Ethernet (and Spectrum-XGS delivers 1.9x higher NCCL performance across multiple data centers), while Multiplane technology scales networks to 128,000 GPUs across two tiers.
AI networking platform · serves as: Connectivity / data interconnect.
Which group NVIDIA Spectrum-X belongs to and how it is built
A subcategory covering networking components for AI infrastructure — Ethernet switches, NICs and SuperNICs delivering high effective bandwidth and low latency for training and inference.
A hardware component type covering complete networking platforms designed for AI workloads — combining switches and network adapters (NICs/SuperNICs) with mechanisms such as adaptive routing, congestion control and RDMA/RoCE.
