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 and ConnectX-9 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 recommended for new multiplane deployments; select Software Multiplane when the deployment requires a separate workload interface for each plane. Match the mode to the supplied RA2.4 or RA3.1 profile and switch fabric configuration.

Hardware Multiplane requires DOCA xPlane to manage plane failover inside OVS-DOCA. Software Multiplane and single-plane deployments do not use DOCA xPlane. See Architecture and Components.

Spectrum-X Kubernetes deployments fall into the following network architectures, distinguished by the number of planes per rail and the load-balancing mechanism. The table describes configuration layouts; use the release-specific qualification guidance below when selecting a platform:

Architecture	NICs	GPU platforms	Multiplane mode
Single-Plane	BlueField-3 SuperNIC, ConnectX-7 NIC, ConnectX-8 SuperNIC	H100/H200/B200, GB200	`none` (1 plane)
Dual-Plane	ConnectX-8 SuperNIC	B300, GB300	`hwplb` (2 planes) `swplb` (2 planes)
Quad-Plane	ConnectX-8 SuperNIC	B300, GB300	`hwplb` (4 planes) `swplb` (4 planes)
Eight-Plane	ConnectX-9 SuperNIC	Vera Rubin (VR)	`hwplb` (8 planes, RA3.1)

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

Spectrum-X RA3.1 for Vera Rubin (VR) and RA2.4 for all other platforms are released together. Network Operator 26.10 integrates both through the same configuration workflow with a matching doSPCX profile ConfigMap and CRDs. Select the architecture for the target platform. The release mapping below distinguishes this workflow from earlier releases; see Platform Support for qualification requirements and validation of earlier releases.

Spectrum-X RA Version	NVIDIA Network Operator Release	Configuration contract
Spectrum-X RA 2.4	26.10.x	All platforms other than VR; doSPCX data-bundle ConfigMap
Spectrum-X RA 3.1	26.10.x	Vera Rubin (VR); same doSPCX contract with `version: RA3.1` and `platformType: vr`
Spectrum-X RA 2.3	26.7.x	Legacy `data.profile` ConfigMap; version selects the ConfigMap
Spectrum-X RA 2.1	26.4.x	Earlier reference-architecture workflow
Spectrum-X RA 2.1	26.1.x	Earlier reference-architecture workflow

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
Network Operator Helm values	Enable SR-IOV Network Operator and Maintenance Operator and configure their coordination.	Quick Start
`NicClusterPolicy`	Cluster-wide Network Operator configuration: enables the Spectrum-X Operator, NIC Configuration Operator, NV-IPAM, and Multus.	Architecture and Components
Spectrum-X profile ConfigMap	Cluster-wide Spectrum-X configuration data. Exactly one current-format profile may be selected.	NIC Configuration
`NicConfigurationTemplate`	NIC-level firmware and PF configuration: link type, `numVfs`, multiplane mode, architecture version, and platform type.	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 applying the profile ConfigMap or writing 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.