NVIDIA Spectrum-X Ethernet Networking Platform

On this page

How Spectrum-X works on Kubernetes
Supported Reference Architectures
What you configure
Where to start

Note

This section covers NVIDIA Network Operator configuration to enable NVIDIA Spectrum-X NIC setup in Kubernetes deployments. For the full Spectrum-X platform documentation — supported topologies, NIC hardware, software components, and version-specific notes — refer to the NVIDIA Spectrum-X documentation.

NVIDIA Spectrum-X is an AI-optimized Ethernet networking platform that combines NVIDIA Spectrum switches with the BlueField-3 SuperNIC, ConnectX-7 NIC, and ConnectX-8 SuperNIC families to deliver high-bandwidth, lossless RoCE for the GPU-to-GPU compute (east-west) network. NVIDIA Network Operator provides the Kubernetes side: discovering the NICs, configuring rails, and exposing them to pods as schedulable resources.

How Spectrum-X works on Kubernetes

Spectrum-X Multiplane is the Spectrum-X capability that splits each SuperNIC across two or more independent network planes — enabling Ethernet to scale from thousands to hundreds of thousands of GPUs in a flat, two-tier topology, with improved performance and resiliency over single-plane networks. Network Operator exposes it through the multiplaneMode field on NicConfigurationTemplate, with Hardware Multiplane (hwplb) and Software Multiplane (swplb) variants alongside the single-plane mode (none). On multiplane platforms, Hardware Multiplane is the mode Spectrum-X RA 2.3 recommends; Software Multiplane is selected only for software-based multiplane deployments.

On multiplane platforms (B300, GB300), Hardware Multiplane is the recommended mode and DOCA xPlane is a required component — it manages plane failover inside OVS-DOCA. Single-plane deployments on BlueField-3 SuperNIC do not use it. See Architecture and Components.

Spectrum-X Kubernetes deployments fall into three network architectures, distinguished by the number of planes per rail and the load-balancing mechanism:

Architecture	NICs	GPU platforms	Multiplane mode	Status
Single-Plane	BlueField-3 SuperNIC, ConnectX-7 NIC, ConnectX-8 SuperNIC	H100/H200/B200, GB200	`none` (1 plane)	GA
Dual-Plane	ConnectX-8 SuperNIC	B300, GB300	`hwplb` (2 planes) `swplb` (2 planes)	GA
Quad-Plane	ConnectX-8 SuperNIC	B300, GB300	`hwplb` (4 planes) `swplb` (4 planes)	GA

Note

ConnectX-8 SuperNIC is listed in the Single-Plane row because it also supports single-plane (none) configuration. Typical Single-Plane deployments use BlueField-3 SuperNIC (HGX H100/H200/B200) or ConnectX-7 NIC (GB200).

Supported Reference Architectures

Each Spectrum-X Reference Architecture version is supported by a specific Network Operator release:

Spectrum-X RA Version	NVIDIA Network Operator Release	Support Level
Spectrum-X RA 2.3	26.7.x	GA
Spectrum-X RA 2.1	26.4.x	Tech Preview
Spectrum-X RA 2.1	26.1.x	GA

Through Network Operator 26.4.x, spectrumXOptimized.version selected one of a fixed set of RA versions built into the NIC Configuration Operator image. From 26.7.0, NIC tuning ships as a Spectrum-X profile — a versioned YAML document published for each Reference Architecture and applied to the cluster as a labeled ConfigMap — and version names that ConfigMap. The operator carries no built-in profiles, so every Spectrum-X deployment applies a profile before configuring NICs, and tuning can be revised without a new operator release. See Spectrum-X NIC Configuration for details.

What you configure

Network Operator drives Spectrum-X setup through a small set of resources. Each one is documented on the page that owns it:

Resource	Purpose	Reference
`NicClusterPolicy`	Cluster-wide Network Operator configuration: enables the Spectrum-X Operator, SR-IOV Network Operator, NIC Configuration Operator, NV-IPAM, and Multus.	Architecture and Components
Spectrum-X profile ConfigMap	Per-RA NIC tuning, referenced by `spectrumXOptimized.version`.	NIC Configuration
`NicConfigurationTemplate`	NIC-level firmware and PF configuration: link type, `numVfs`, multiplane mode, profile reference.	NIC Configuration
`NicInterfaceNameTemplate`	Predictable rail and plane based interface names, applied by udev.	NIC Configuration
`SpectrumXRailPoolConfig`	Rail topology, PF selection, IPAM binding, and rail resource exposure.	CRD API Reference
`CIDRPool`	Per-rail IP allocation, or per rail-plane in `swplb`.	CRD API Reference

For Dynamic Resource Allocation workflows, the upstream Kubernetes ResourceClaimTemplate binds Pod requests to specific GPU and VF combinations — see DRA SR-IOV Driver.

Where to start

Page	Read it when
Quick Start	You are deploying. Nine steps, from node preparation to a test Pod.
Architecture and Components	You want to know what gets deployed and which operator owns what.
NIC Configuration	You are writing the profile ConfigMap or the NIC templates.
Verify and Troubleshoot	The cluster is up and you need to confirm it, or something failed.
CRD API Reference	You need field-level detail for a Spectrum-X CRD.

For supported platforms, NICs, switches, and OS combinations, see the NVIDIA Spectrum-X Solution Stack documentation. For the Kubernetes and OS matrix and component versions, see Platform Support.