LEVIATHAN SYSTEMS

Cabling

Spectrum-X vs Quantum: What the Network Choice Actually Specifies

Technical contentPublished Updated

Spectrum-X is NVIDIA’s Ethernet networking platform for AI; Quantum identifies its InfiniBand switch family. That choice affects the network architecture and supported components, but neither name alone specifies a cable plant. The physical package needs exact endpoint hardware, supported interconnect parts, port modes, routes and acceptance boundaries.

Compare equivalent levels of the system

NVIDIA’s Network Operator documentation describes Spectrum-X as a combination of Spectrum switches and supported adapter/SuperNIC families, with software that exposes and configures the networking resources. Its LinkX product-family guide separately identifies Quantum InfiniBand switches, Spectrum Ethernet switches and the interconnect products between endpoints.

Those are different naming levels. A useful comparison specifies the intended Ethernet or InfiniBand architecture, then identifies the actual switch, server adapter and interconnect combination. It does not compare a platform-wide capability on one side with an isolated transceiver on the other.

References: [N1][N2]

Separate the platform name from the component selection

The table is a reading guide to NVIDIA’s product names. It is not an interoperability matrix or a recommendation to choose one network for a workload. Generation, software and exact part support must be verified for the purchased configuration.

Scroll the table horizontally on smaller screens.

Names that belong at different levels
NameWhat it identifies hereWhat still needs specificationSource
Spectrum-XNVIDIA AI Ethernet networking platformApplicable reference architecture, switch, adapter and software combination[N1][N2]
QuantumNVIDIA InfiniBand switch familyExact switch generation/model and supported fabric components[N1][N2]
LinkXNVIDIA cable and transceiver product familyExact assembly/module part and both endpoints[N1][N2]
Ethernet or InfiniBandNetwork technology choiceActual physical topology, port modes and implementation records[N1][N2]

Original comparison of naming levels, based on the cited NVIDIA documentation. It does not establish cross-generation support.

Ask what the network design owner has actually approved

Request the switch and adapter identifiers, applicable software/firmware support references, logical port plan and approved cable/optic schedule. Include the exact end-to-end component relationship where patch panels, trunks or breakout assemblies are involved. A line that says 800G optical link leaves the physical interfaces and mapping unresolved.

Do not assume that parts with matching connector shapes will implement the same electrical or optical interface. Conversely, avoid declaring every Ethernet-labelled optical component incompatible with every InfiniBand endpoint without checking the exact supported product combination. The OEM’s applicable compatibility evidence must settle the actual selection.

Translate the choice into physical records

Map each released logical connection to two physically locatable endpoints, its cable/channel identifiers and the approved route. Keep fabric, rail or plane labels supplied by the design owner separate from cable identifiers. A cable tag should let operations trace the installed connection; it cannot create a network configuration or certify a topology.

For a proposed change from one fabric implementation to another, compare the actual endpoint list and material schedule. Some physical routes may remain useful while modules, assemblies, adapter positions or patch relationships change. Do not assume either complete reuse or complete replacement solely from the platform names.

Separate AI compute Ethernet from the management LAN

An Ethernet label does not make every connection in an AI rack the same installation task. NVIDIA’s HGX H100/H200/B200 reference distinguishes a 400GbE compute fabric, 200GbE converged connections and bulk 1Gb RJ45 out-of-band management. These are roles and interfaces in that reference, not a universal recipe for an enterprise or an AI site.

That distinction is more useful than saying that AI Ethernet simply needs faster LAN cables. An enterprise LAN can also use optical or high-speed interfaces. Compare the actual endpoint and function: a GPU-facing SuperNIC path, an upstream customer connection and a switch’s management port can require different assemblies, labels, channel evidence and accepting owners despite all carrying Ethernet. Do not apply an office access-port cabling allowance to an entire compute fabric or assume the management connection is included in the compute link’s record.

The dense compute-port population also makes service access and record population important. The selected model’s cable face, supported module/assembly form, accessible identification and permitted service space belong in the work package. Exact clearances, bend protection and field handling come from the applicable parts and site instructions. Ethernet protocol identity supplies none of those physical decisions.

References: [N5]

Scroll the table horizontally on smaller screens.

Compare the specified role before comparing installation needs
Connection roleQuestion that changes the physical packageEvidence that remains separateSource
GPU compute pathWhich physical adapter/cage, port mode and supported channel implement the released compute edge?Network configuration and compute-fabric behavior are not established by a cable-ID record.Planning analysis
Customer, storage or converged connectionWhich service associations use this physical connection, and is it shared or separate in this design?A service-role count is not automatically a cable count or permission to combine fabrics.Planning analysis
OOB or management connectionWhich actual management port and cable are intended, separately from data and console ports?Management access/security configuration and permissions are not supplied by physical identification.Planning analysis
Enterprise LAN connection used for comparisonWhat exact endpoint, media, route and required result does that actual LAN design specify?Neither all-copper LAN assumptions nor an AI bandwidth headline is a complete channel specification.Planning analysis

Source-informed buyer comparison, not a new topology, product selection, material allowance or field method.

Resolve the Ethernet model before releasing materials

The SN5000 hardware manual provides a concrete counterexample to a generic “Spectrum switch” cable line: SN5400 data interfaces are QSFP-DD 400GbE, while SN5600 data interfaces are OSFP 800GbE; RJ45 management and console interfaces are separately identified. Those source distinctions do not establish a particular module, cable or channel as compatible. They show which information a material request still needs.

Fictional comparison EX-NE120 contains one compute connection request naming SN5600 and one management connection request naming its RJ45 management interface. A returned material sheet lists the same unspecified “Ethernet patch cable” for both. That is two inadequately specified request lines, not evidence that two identical supplied items will work. Return the compute line for its exact supported endpoint/part/mode definition, and the management line for its own actual endpoint/media definition. Do not change a connector or port assignment to make the generic description fit.

The resulting physical package should identify both endpoints, actual approved interconnect or channel components, branch relationships where relevant, routed length basis, identifiers and required checks. Mounting and cable installation follow their own applicable equipment/site instructions. The physical completion record then identifies what was installed and checked; it does not certify congestion behavior, routing, RDMA configuration or application performance.

References: [N6]

Keep the acceptance decision at the right boundary

The physical team can document identified components, routes, labels, accessible conditions and separately agreed verification evidence. The network team establishes configuration and fabric behavior. Workload acceptance adds another layer. A mounted switch and a completed cable schedule do not prove congestion control, routing or application performance.

Leviathan can discuss the bounded physical installation and documentation package for an approved fabric design. Materials supply, network software configuration and full cluster commissioning are not implied. Use the companion installation-SOW resource to assign the interfaces before requesting a comparable field proposal.

References: [L1][L6]

Common planning questions

Is Spectrum-X a different type of fiber?

No. It is a networking platform. The actual switch/adapter combination and selected optical modules determine the supported physical media and channel.

Can a cable installer decide which fabric a cluster should use?

Architecture selection and compatibility approval belong to the responsible network/OEM owners. The physical team needs their released design and a clear installation boundary.

Does AI Ethernet require a completely different cabling process from every enterprise LAN?

Compare the actual hardware, connection roles and required evidence. Both can use high-speed optical or copper interfaces, while a named AI reference can have distinct compute, converged and management fabrics. The AI label alone does not determine media, cable quantities, service access or an installation method.

Primary sources

  1. [N1] NVIDIA Spectrum-X platform in Network Operator 26.7.0Checked 2026-10-15. September 9, 2026 documentation. Platform and software-role definitions only; no configuration commands or universal supported-system matrix reproduced.
  2. [N2] NVIDIA Quantum, Spectrum and LinkX product familiesChecked 2026-10-15. The 100G-PAM4 document family distinguishes Quantum InfiniBand, Spectrum Ethernet and LinkX interconnects. It does not establish compatibility across every generation.
  3. [L1] Leviathan structured cabling scopeChecked 2026-10-15. Approved physical cabling scope; exact customer design, materials, access and verification remain engagement-specific.
  4. [L6] Leviathan physical-layer verification scopeChecked 2026-10-15. Formal testing requires confirmed personnel, equipment, method and criteria. No blanket test or system-commissioning capability is asserted.
  5. [N5] NVIDIA HGX H100/H200/B200 reference: compute, converged and OOB fabricsChecked 2026-10-19. Named reference roles only:400GbE compute,200GbE converged and1GbRJ45 OOB, with distinct adapter roles. No reference topology/count table, configuration, performance/security guarantee, customer design or acceptance adopted.
  6. [N6] NVIDIA SN5000 hardware manual: model interfaces and management rolesChecked 2026-10-19. Named SN5400 QSFP-DD400GbE versus SN5600 OSFP800GbE and separate RJ45 management/console roles only. No actual part compatibility, port mode, populated cable count, installation, electrical or site-fit approval.

Define the physical fabric package

Share the approved switch/adapter list, endpoint schedule, selected interconnects and required records.

Discuss a ProjectEmail Project Scope