IntelligentHQ
IntelligentHQ
Skip to content
Platform AI Networking & ConnectivityAI Factories & Infrastructure

NVIDIA Spectrum-X Ethernet

NVIDIA Spectrum-X Ethernet is a networking platform for AI clusters that pairs Spectrum-X switches with SuperNICs in the GPU servers, so congestion control, adaptive routing and telemetry work end to end on standards-based Ethernet.1

Also known as Spectrum-X, Spectrum-X Ethernet Networking Platform, Spectrum-XGS Ethernet, Spectrum-X Ethernet Photonics, Spectrum-X Photonics, Spectrum-X Multiplane

Official NVIDIA product page (opens in a new tab) Official documentation (opens in a new tab)

At a glance

What is it?
Spectrum-X is NVIDIA's Ethernet platform for AI factories. It is built from two coupled parts: Spectrum-X Ethernet switches (the SN5000 and SN6000 lines, the latest on the Spectrum-6 ASIC) and Spectrum-X Ethernet SuperNICs in each GPU server (BlueField-3, ConnectX-8 or ConnectX-9, depending on the GPU generation). It runs open network operating systems, SONiC or NVIDIA Cumulus, with NVIDIA NetQ for visibility, and it can be simulated in NVIDIA DSX Air. Extensions include Spectrum-XGS Ethernet for linking data centers, Spectrum-X Multiplane for larger two-tier fabrics, and Spectrum-X Ethernet Photonics, which places the optics on the switch package.1
What does it do?
Spectrum-X carries GPU-to-GPU traffic for training and distributed inference using RDMA over Converged Ethernet (RoCE), and NVIDIA says it and storage partners are extending it to AI storage fabrics. The switch and SuperNIC cooperate on adaptive routing, which spreads traffic dynamically, and on telemetry-based congestion control, which isolates tenants from each other's traffic. NVIDIA states the platform scales to hundreds of thousands of GPU connections in two-tier leaf and spine designs. Spectrum-XGS adds distance-aware congestion control so sites hundreds of kilometers apart can act as one cluster.1
Who needs it?
NVIDIA aims Spectrum-X at cloud providers and large companies that share AI clusters among many tenants. It fits teams that want an Ethernet-based AI fabric rather than InfiniBand, need performance isolation between tenants, or are building clusters with thousands of GPUs or more across one or several sites.1
What does it need?23
  • Spectrum-X switches and a supported SuperNIC (BlueField-3, ConnectX-8 or ConnectX-9) in each GPU server
  • Component versions aligned with the Spectrum-X Validated Solution Stack
  • A network operating system: SONiC or NVIDIA Cumulus
  • DOCA software versions tested with the Spectrum-X reference architecture release
  • RoCE network design and operations skills
What it is not
Spectrum-X is not simply a faster Ethernet switch. Its AI features depend on the switch and SuperNIC working together, so buying the switches alone with other network cards does not deliver the platform's adaptive routing as designed. It is also not InfiniBand: NVIDIA sells Quantum InfiniBand as a separate family. NVIDIA's headline figure of 1.6x network performance over off-the-shelf Ethernet is a vendor claim without published conditions on the product page. For Spectrum-X Ethernet Photonics, NVIDIA's photonics page gives availability in the second half of 2026 and also headlines a ramp to full production, and NVIDIA's May 31, 2026 Vera Rubin release says it is now in production.145

Availability and licensing. Spectrum-X switches and SuperNICs are sold through NVIDIA and partners. NVIDIA's photonics page states Spectrum-X Ethernet Photonics switches are available in the second half of 2026; NVIDIA's 31 May 2026 Vera Rubin announcement describes them as now in production. The Spectrum-6 generation targets Vera Rubin AI factories.45

The problem it solves

Distributed training and inference move large amounts of data between GPUs in synchronized bursts. On a standard Ethernet fabric, congestion and uneven path use slow the whole job down, because every GPU waits for the slowest transfer. In shared clouds, one tenant's traffic can also degrade another's.

Spectrum-X tackles this by coordinating the switch and the server network card: traffic is routed adaptively across paths, congestion is controlled using telemetry, and tenants are isolated. Optional extensions address scale across sites and the power used by optical transceivers.1

How it works

A Spectrum-X fabric has three working parts.

  • Switches: Spectrum-X switches form a two-tier leaf and spine network. The Spectrum-6 generation offers 102.4 Tb/s per switch chip with 200G SerDes; the SN6600 has 128 ports of 800G and the SN6800 has 512 ports of 800G.
  • SuperNICs: each GPU server uses a SuperNIC that offloads collective communications and returns telemetry for adaptive routing: BlueField-3 for Hopper systems, ConnectX-8 for Blackwell and ConnectX-9 for Vera Rubin NVL72.
  • Software: SONiC or Cumulus runs on the switches, NetQ traces flows from GPU to SuperNIC to switch port, and DSX Air simulates the fabric before deployment.

Multiplane splits each SuperNIC across two or more independent network planes for larger fabrics. Spectrum-XGS adapts congestion control for long distances between sites. Photonics switches replace pluggable transceivers with co-packaged optics.

NVIDIA publishes a Validated Solution Stack listing component versions tested together.13

NVIDIA Spectrum-X Ethernet architecture: components by layer and how they connectOperations &orchestrationAcceleratedcomputingNetworking, power &facilitiesSONiC or NVIDIA Cumulus: Switch operating systemSONiC or NVIDIA CumulusNVIDIA NetQ: Flow-level visibility and troubleshootingNVIDIA NetQNVIDIA DSX Air: Simulation of the fabric before deploymentNVIDIA DSX AirGPU servers (Hopper, Blackwell, Vera Rubin NVL72): Endpoints of the AI fabricGPU servers (Hopper,Blackwell, Vera Rubin NVL72)Spectrum-X SuperNICs (BlueField-3, ConnectX-8, ConnectX-9): RoCE connectivity, collective offload, routing telemetrySpectrum-X SuperNICs(BlueField-3,…Spectrum-X switches (SN5000, SN6000): Two-tier leaf and spine fabric with adaptive routingSpectrum-X switches(SN5000, SN6000)Spectrum-X Ethernet Photonics switches: Co-packaged optics switching (H2 2026)Spectrum-X EthernetPhotonics switchesSpectrum-XGS Ethernet: Links fabrics across data centersSpectrum-XGS Ethernet
Diagram as a list

  1. Operations & orchestration

    • SONiC or NVIDIA CumulusSwitch operating systemConnects to Spectrum-X switches (SN5000, SN6000)
    • NVIDIA NetQFlow-level visibility and troubleshootingConnects to Spectrum-X switches (SN5000, SN6000), Spectrum-X SuperNICs (BlueField-3, ConnectX-8, ConnectX-9)
    • NVIDIA DSX AirSimulation of the fabric before deploymentConnects to Spectrum-X switches (SN5000, SN6000)

  2. Accelerated computing

    • GPU servers (Hopper, Blackwell, Vera Rubin NVL72)Endpoints of the AI fabricConnects to Spectrum-X SuperNICs (BlueField-3, ConnectX-8, ConnectX-9)

  3. Networking, power & facilities

    • Spectrum-X SuperNICs (BlueField-3, ConnectX-8, ConnectX-9)RoCE connectivity, collective offload, routing telemetryConnects to Spectrum-X switches (SN5000, SN6000)
    • Spectrum-X switches (SN5000, SN6000)Two-tier leaf and spine fabric with adaptive routingConnects to Spectrum-X SuperNICs (BlueField-3, ConnectX-8, ConnectX-9), Spectrum-XGS Ethernet
    • Spectrum-X Ethernet Photonics switchesCo-packaged optics switching (H2 2026)Connects to Spectrum-X switches (SN5000, SN6000)
    • Spectrum-XGS EthernetLinks fabrics across data centers
Components and connections as documented by NVIDIA.1

Capabilities

  • Coupled switch and SuperNIC1

    Spectrum-X switches and SuperNICs work together on adaptive routing, with the SuperNICs offloading collective communications and supplying telemetry.

    Why it matters: Raises effective bandwidth for synchronized GPU traffic.

    Limits: Requires both NVIDIA switches and SuperNICs to work as designed.

  • Performance isolation1

    Telemetry-based congestion control provides noise isolation between tenants.

    Why it matters: Keeps one tenant's jobs from slowing another's in shared AI clouds.

    Limits: Isolation applies within the Spectrum-X fabric, not across third-party network segments.

  • Spectrum-6 switch generation1

    102.4 Tb/s per switch chip with 200G SerDes; SN6600 with 128 ports of 800G and SN6800 with 512 ports of 800G in 5U.

    Why it matters: Higher port density per switch reduces the number of switches needed.

    Limits: Targeted at Vera Rubin AI factories; earlier GPU generations pair with earlier SuperNICs.

  • Spectrum-X Multiplane1

    Splits each GPU's SuperNIC across two or more independent planes; NVIDIA states this scales to 128,000 GPUs in two tiers.

    Why it matters: Avoids adding a third network tier for large clusters.

    Limits: Vendor scale figure; multiplane designs need careful planning of planes and cabling.

  • Spectrum-XGS Ethernet1

    Extends Spectrum-X across data centers with distance-aware congestion control, latency management and end-to-end telemetry.

    Why it matters: Lets separate sites operate as one AI cluster.

    Limits: NVIDIA's NCCL speed-up figure is a vendor claim compared with standard approaches.

  • Spectrum-X Ethernet Photonics45

    Switches with co-packaged optics on the ASIC package, built on 200G SerDes, with up to 409.6 Tb/s bandwidth.

    Why it matters: Removes pluggable transceivers, which NVIDIA says improves power efficiency and resiliency.

    Limits: The photonics page gives availability in the second half of 2026 and headlines a full-production ramp, and NVIDIA's May 31, 2026 release says it is now in production; efficiency figures are vendor claims versus pluggable transceivers.

  • Open NOS and flow visibility1

    Supports SONiC and NVIDIA Cumulus, with NetQ tracing performance from GPU to SuperNIC and per-hop behavior across switch ports and RoCE queues.

    Why it matters: Operators can troubleshoot AI traffic hop by hop.

    Limits: Full visibility assumes NetQ is deployed across the fabric.

  • Simulation and validated stack36

    DSX Air simulates Spectrum-X fabrics with the wider AI factory, and the Validated Solution Stack lists tested component versions.

    Why it matters: Configurations can be tested before hardware arrives.

    Limits: NVIDIA advises running all components on the latest validated configuration.

Practical use cases

Distributed training and inference jobs slow down because Ethernet congestion delays synchronized GPU transfers.
Approach
Build the GPU compute fabric with Spectrum-X switches and SuperNICs using adaptive routing and congestion control.
Role of NVIDIA Spectrum-X Ethernet
Spectrum-X is the east-west AI fabric.
Data, infrastructure and skills
Matching SuperNICs for the GPU generation, validated software versions.
Type of benefit
More predictable network performance
Caveats
Gains depend on workload traffic patterns; NVIDIA's speed-up figure is a vendor claim.
First step
Model the fabric in DSX Air with your planned GPU count.

Sources 1

An AI cloud provider must stop one tenant's traffic from degrading another tenant's jobs.
Approach
Use Spectrum-X telemetry-based congestion control for noise isolation between tenants.
Role of NVIDIA Spectrum-X Ethernet
Spectrum-X provides performance isolation in the shared fabric.
Data, infrastructure and skills
Multi-tenant network design and tenant isolation policies.
Type of benefit
Tenant performance isolation
Caveats
Isolation does not cover parts of the path outside the Spectrum-X fabric.
First step
Define per-tenant service levels to test against.

Sources 1

Power or space limits mean a single site cannot hold all the GPUs a project needs.
Approach
Connect sites with Spectrum-XGS Ethernet so they operate as one cluster.
Role of NVIDIA Spectrum-X Ethernet
Spectrum-XGS handles long-distance congestion control and latency.
Data, infrastructure and skills
Data center interconnect capacity between the sites.
Type of benefit
Multi-site capacity pooling
Caveats
Cross-site performance figures are vendor claims.
First step
Read the Spectrum-XGS technical blog and measure inter-site latency.

Sources 1

Storage traffic competes with compute traffic and slows data loading for AI jobs.
Approach
Extend Spectrum-X to the storage fabric with partners from the storage ecosystem.
Role of NVIDIA Spectrum-X Ethernet
Spectrum-X carries storage traffic for AI.
Data, infrastructure and skills
A storage partner that supports Spectrum-X.
Type of benefit
Faster data access for AI workloads
Caveats
Results depend on the storage system.
First step
Check the storage partners listed on the Spectrum-X page.

Sources 1

Who uses it

  • xAI · AI model development

    xAI Colossus: Spectrum-X Ethernet for a 100,000-GPU training cluster

    xAI's Colossus cluster in Memphis started with 100,000 NVIDIA Hopper GPUs, connected with NVIDIA Spectrum-X Ethernet (SN5600 switches and BlueField-3 SuperNICs) to train Grok models. NVIDIA reports 95 percent data throughput, compared with the 60 percent it attributes to standard Ethernet at this scale.

    In production

Works with

Optional integration

  • NVIDIA DSXDSX Air simulates Spectrum-X fabrics before deployment.

Complementary tools

Same family

Optional integration for

Relationship labels follow NVIDIA's documentation. "Alternative approaches" does not mean one is better: each profile says when it fits.

Getting started

  1. Simulate first

    Use the free DSX Air trial to build a simulated Spectrum-X topology with automation scripts.

    Check: Your planned leaf and spine design boots in simulation and passes configuration tests.

  2. Match the SuperNIC to the GPU generation

    Pick BlueField-3 for Hopper, ConnectX-8 for Blackwell or ConnectX-9 for Vera Rubin NVL72 systems.

    Check: Every GPU server in the bill of materials has a supported SuperNIC.

  3. Lock component versions

    Use the Spectrum-X Validated Solution Stack to choose switch OS, firmware and DOCA versions that were tested together.

    Check: Each component version matches the current validated configuration.

  4. Choose the NOS and visibility tools

    Decide between SONiC and NVIDIA Cumulus and plan NetQ deployment for flow-level visibility.

    Check: NetQ shows GPU-to-switch flow traces in the pilot.

Official resources

Could this technology help you?

Describe your project to the Solution Architect. It starts with NVIDIA Spectrum-X Ethernet as context but recommends independently, including when you do not need it.

Check it against my project

Sources

Each statement above links to the source it comes from. Labels say who reported it.

  1. NVIDIA Spectrum-X Ethernet Networking Platform (opens in a new tab) NVIDIA · Vendor-reported · link checked 9 Oct 2026
  2. DOCA Documentation v3.5.0 (opens in a new tab) NVIDIA · Vendor-reported · link checked 9 Oct 2026
  3. NVIDIA Spectrum-X Validated Solution Stack (opens in a new tab) NVIDIA · Vendor-reported · link checked 9 Oct 2026
  4. NVIDIA silicon photonics page (opens in a new tab) NVIDIA · Vendor-reported · link checked 9 Oct 2026
  5. NVIDIA newsroom: Vera Rubin ramps into full production (opens in a new tab) NVIDIA · Vendor-reported · link checked 9 Oct 2026
  6. NVIDIA DSX Air Platform (opens in a new tab) NVIDIA · Vendor-reported · link checked 9 Oct 2026

Fill out the form below to request your copy.

Name(Required)