Ethernet platform purpose-built for AI, from single-cluster fabrics to multi-site AI factories with Spectrum-XGS.
NVIDIA Spectrum-X is an end-to-end Ethernet platform built for multi-tenant generative AI factories. Spectrum-X switches and SuperNICs work together to deliver the bandwidth, low latency and performance isolation that AI needs on standards-based Ethernet.
Spectrum-XGS extends the same fabric across data centers, so geographically distributed sites can operate as one large AI factory.
Compared with off-the-shelf Ethernet for AI workloads.
Two-tier topologies reach hundreds of thousands of GPUs; Multiplane extends to 128,000 GPUs.
Telemetry-based congestion control keeps noisy neighbours from slowing other jobs.
Switch and SuperNIC coordinate to maximise effective bandwidth and resilience.
Multi-site NCCL collectives run up to 1.9x faster across data centers.
Up to 5x better power efficiency than transceiver-based networks.
| SN5000 series | Spectrum-4 Ethernet switches for AI factories |
|---|---|
| SN6600 | 128× 800G OSFP ports, Spectrum-6 ASIC, 102.4 Tb/s |
| SN6800 | 512× 800G ports in 5U |
| SerDes | 200G per lane |
| BlueField-3 | RDMA / RoCE-optimised SuperNIC for Hopper-era systems |
|---|---|
| ConnectX-8 | 800 Gb/s total (2× 400G), PCIe Gen6, for Blackwell |
| ConnectX-9 | 1,600 Gb/s per GPU, PCIe Gen6, for Vera Rubin NVL72 |
End-to-end 800 Gb/s InfiniBand platform for the largest AI training clusters.
SuperNIC delivering up to 800 Gb/s for InfiniBand or Ethernet GPU-to-GPU networking.
Data processing unit that offloads networking, storage and security from host CPUs using NVIDIA DOCA.
Tell us your workload and timeline for the NVIDIA Spectrum-X Ethernet. We reply within one business day with a configuration, pricing and lead time.