Key takeaways: Validate NVIDIA DGX/HGX MPO cabling against the approved architecture, fabric role, both endpoints, and exact module or complete-cable part numbers. A label such as
DGX,HGX,OSFP,400G, or800Gdoes not establish the MPO interface, fiber type, connector, mapping, or cable quantity.
A platform name helps you find the right documents. It does not specify a passive cable assembly. NVIDIA documents copper, active optical cables (AOCs), multimode passive fiber, and single-mode passive fiber across different products and reference designs. Use the platform and speed to locate the controlling sources, then use the exact endpoints and modules to specify the cable. See NVIDIA's cable technology overview and NDR connection guide.
Procurement summary: Do not ask a supplier to select an MPO assembly from the NVIDIA platform name or speed alone. Freeze the architecture title and revision, identify the fabric role, record both endpoint devices and port modes, then confirm the exact module or complete-cable part number at each end. Classify unresolved links as open, unsupported, or conflicting instead of filling gaps with catalog defaults. Only after those records agree should procurement release the cable schedule for polarity, loss-budget, route, testing, sample, quotation, and delivery review.
Decision rules:
- The platform name identifies a source set, not a passive cable.
- The fabric role defines the link domain; GPU count does not define external cable quantity.
- The module line-side interface, not the host cage, controls the passive connector.
- The approved project P2P record closes the gap between a reference design and an order.
- A source or revision conflict holds the cable schedule until a named owner resolves it.
Technical notice: Each architecture example below is a sourcing reference only. It applies to the cited document and revision, not every NVIDIA or OEM design. Verify the port count, medium, connector, and cable construction against the project's own controlling sources. This guide relies on the linked NVIDIA documents; it does not report a BWNFiber deployment case or certify product compatibility.
Preparing a supplier request? Keep the approved architecture title and revision, both endpoint records, and exact module or complete-cable part numbers beside this guide. BWNFiber can review source completeness before a passive-cable configuration is discussed.
Contents:
- Why the NVIDIA platform name is not a cable specification
- Three checks before you approve an MPO order
- What current NVIDIA documents establish, and what remains open
- How to interpret rail-optimized link counts
- When MPO is appropriate in an NVIDIA fabric
- How to build the evidence packet
- Where downstream MPO decisions belong
- How procurement should build a comparable RFQ
- Procurement shortcuts that should stop the order
- NVIDIA DGX/HGX MPO cabling FAQ
- Request an evidence review
Why the NVIDIA platform name is not a cable specification
NVIDIA names refer to different layers of a design. An order can go wrong when a platform or network family is treated as a passive-cable specification.
| Name in the project | What it helps identify | What it does not identify by itself |
|---|---|---|
| DGX | An NVIDIA system family and its platform documentation. DGX documentation index | The optical receptacle on every selected module or the installed route |
| HGX | A platform used across integrated OEM server patterns. NVIDIA enterprise reference architectures | One universal NIC, DPU, switch, optic, or cable configuration across OEM systems |
| BasePOD / SuperPOD | A solution/reference-architecture context with different degrees of prescription and flexibility. NVIDIA DGX SuperPOD FAQ | A site-specific approved BOM without revision and component reconciliation |
| Spectrum-X | An Ethernet networking architecture and component family. HGX Spectrum-X logical architecture | The passive interface on every permitted endpoint combination |
| Quantum-2 / NDR | An InfiniBand switch and interconnect context. NDR connection guide | One mandatory medium. NVIDIA documents several NDR connection types |
| NVLink / NVSwitch | Scale-up GPU communication within the supported platform architecture. DGX H100/H200 architecture | The external scale-out cable plant between compute nodes and network switches |
Use BWNFiber's data-center fiber connectivity hub when you need to compare broad product and solution families. Stay with this guide when the open question is which NVIDIA or OEM evidence must be confirmed before MPO selection begins.
Three checks before you approve an MPO order
1. Separate scale-up from scale-out
NVLink and NVSwitch describe the supported scale-up architecture. The cluster network reaches other nodes through external network interfaces and switches. For DGX H100/H200, NVIDIA lists eight GPUs and four NVSwitch components, while four OSFP openings serve eight single-port ConnectX-7 interfaces for external cluster networking. DGX H100/H200 system architecture
Count structured-cabling links from the enabled external endpoints, not from the internal GPU links.
2. Read the module interface behind the OSFP cage
OSFP describes the host-side module form factor. The installed module determines whether the line side exposes MPO, duplex connectors, an integrated AOC, or an electrical cable. An OSFP count is therefore useful for finding endpoint documentation, but not sufficient for ordering the passive cable.
3. Reconcile the reference architecture with the installed project
A reference architecture can define topology, supported components, and a bounded cable count. The project still needs the approved version, selected component part numbers, enabled port modes, and point-to-point implementation. An OEM HGX server must be reconciled with both the approved OEM design and the applicable NVIDIA material; an NVIDIA example cannot silently replace either one.
| Boundary | Unsafe inference | Evidence needed before proceeding |
|---|---|---|
| Scale-up vs. scale-out | "Eight GPUs means eight external MPO cables" | Approved external fabric and enabled port map |
| Host cage vs. line interface | "OSFP means MPO-16" | Exact installed module or complete-cable part number and current interface document |
| Reference design vs. project BOM | "The NVIDIA component table is our purchase order" | Approved architecture revision, project deviations, endpoint schedule, and owner sign-off |
What current NVIDIA documents establish, and what remains open
Read each source on its own terms. The table separates documented facts from project decisions that still need evidence.
| Document-scoped example | What the source establishes | What remains unproven |
|---|---|---|
| DGX H100/H200 user guide | Four OSFP openings serve eight ConnectX-7 cluster interfaces, with InfiniBand or Ethernet operation up to 400 Gb/s in the cited guide. Source | One universal module, fiber type, connector, mapping, or route |
| DGX H100 SuperPOD reference architecture | Its listed networking components are described as multimode fiber. Source | A rule that every H100 deployment must use multimode fiber |
| DGX B200 BasePOD reference architecture | Four OSFP openings serve eight ConnectX-7 adapters, with operation up to 400 Gb/s in the cited design. Source | The line-side connector or medium of another B200 design |
| HGX H100/H200/B200 Spectrum-X reference architecture | Each server in the named design has eight 400 Gb/s BlueField-3 B3140H SuperNIC connections in a rail-optimized compute fabric. Source | A universal HGX or OEM server configuration |
| DGX B300 best-practices example | The described compute fabric mainly uses MPO-12/APC passive single-mode fiber with DR4 optics. Source | Permission to transfer the B300 arrangement to H100, H200, B200, or another HGX system |
How to interpret rail-optimized link counts
A rail-optimized topology groups corresponding compute interfaces into leaf planes. It explains the endpoint relationships and link count within a named architecture. The topology alone cannot select DAC, AOC, multimode MPO, single-mode MPO, or another supported interface.
The NVIDIA HGX H100/H200/B200 Spectrum-X reference architecture provides a bounded example. Its 32-node design lists 256 compute-to-leaf cables and 256 switch-to-switch cables. Use those figures only for the named topology and network scope. Another 256-GPU cluster may have different adapters, enabled ports, or fabric domains. NVIDIA HGX network architecture and component-count table
Separate these quantity domains when reading a component table:
| Count domain | Controlling input | Why it cannot be inferred from GPU count alone |
|---|---|---|
| Compute node to leaf | Enabled compute-fabric ports in the approved rail map | Server and adapter configurations can differ |
| Leaf to spine / super-spine | Named topology and switch configuration | Oversubscription and fabric scale affect the count |
| Storage / converged network | Separate storage and north-south architecture | It may use different endpoints and media |
| Management | Platform and management design | It is operationally separate from the compute fabric |
Reconcile the total against GPU count only after the endpoint schedule exists.
When is MPO appropriate in an NVIDIA fabric?
NVIDIA documents several physical paths: copper, AOC, multimode passive fiber, and single-mode passive fiber. The endpoint modules and their approved operating mode decide which path applies. The table records what each source supports and where that evidence stops.
| NVIDIA source context | Media or assembly types documented | Procurement limit |
|---|---|---|
| NDR connection guidance | DAC/ACC copper, AOC, multimode passive fiber, and single-mode passive fiber appear across the cited options. Source | NDR or InfiniBand does not select one passive medium |
| NDR components | Both single-mode and multimode passive-fiber support is described for the cited family. Source | The exact transceiver pair and reach still control the interface |
| DGX H100 SuperPOD components | The listed networking components are multimode fiber. Source | The source does not support a claim that all NVIDIA clusters are single-mode |
| DGX B300 best practices | The named compute-fabric example is mainly MPO-12/APC passive single-mode DR4. Source | A B300 example is not a default for another generation |
Move to the cable-type and channel-design workflow only after the exact optic exposes an MPO interface. Until then, MPO remains a candidate rather than an approved specification.
How to build the NVIDIA DGX/HGX MPO cabling evidence packet
Complete these six steps before opening the generic MPO worksheet. Each step should point to a project-controlled record.
Which source controls each decision?
Use the narrowest source that actually owns the decision. A higher-level architecture document can establish the topology or fabric role without establishing the passive interface.
| Decision | Controlling record | What to verify |
|---|---|---|
| Platform, topology, and fabric role | Project-approved NVIDIA or OEM architecture source | Exact system, architecture title, revision, topology, and approved deviations |
| Host port, card, and cage mapping | Platform and OEM hardware or integration documents | Real port designation, adapter or DPU, enabled mode, and host form factor |
| Supported device or complete-cable combination | Applicable adapter firmware release and its validated support list | Exact adapter, firmware, cable or module part number, and supported endpoint combination |
| Line-side connector, medium, reach, and lane mapping | Exact installed module or approved complete-cable document | Both endpoints, operating mode, connector, fiber medium, reach conditions, and mapping |
| Quantity, length, bundles, labels, and route | Reviewed project point-to-point cable map | Every endpoint pair, rack and port location, physical distance, route, and label |
| Passive-channel polarity, loss, and acceptance | Project cabling design and applicable standards after the interface is fixed | Channel components, test method, limits, references, and approval owner |
For DGX H100/H200, NVIDIA states that ConnectX-7 firmware determines which network cables and adapters are supported and directs users to the firmware-specific validated list. This makes firmware-specific cable support a separate gate from connector fit. A part that can be inserted is not automatically approved for the installed system. DGX H100/H200 supported network cables and adapters
How should each candidate link be classified?
Do not convert a source search directly into a cable line item. Give each link one of four review states and record the source or owner that justifies the state.
| Review state | Use it when | Procurement action |
|---|---|---|
| Supported for handoff | The approved architecture scope, fabric role, both endpoints, operating mode, and exact module or complete-cable part numbers agree | Transfer only the documented interface facts to the MPO channel worksheet |
| Project decision open | The controlling sources identify the link but leave route, length, channel design, mapping, acceptance, or another project-owned field unresolved | Keep the field open, assign an owner, and withhold the final cable schedule |
| Unsupported inference | The proposed cable follows only from a platform name, GPU count, speed, cage, connector fit, or another shortcut that the controlling source does not support | Remove the inferred value and request the missing endpoint or component evidence |
| Source conflict: hold | NVIDIA, OEM, module, firmware-support, or project records disagree or refer to different revisions | Stop the order until the named technical owner records which source governs and why |
The output of this review is a decision record, not a substitute BOM. It should show which facts can move downstream, which fields remain open, and which inference was rejected.
Six-step evidence handoff
- Identify the exact DGX platform or HGX OEM system and hardware generation.
- Classify the link as scale-up, scale-out compute, storage/converged, or management.
- Freeze the architecture title, revision, topology, and project-approved deviations.
- Record both endpoint ports and the exact installed modules or approved complete cables.
- Name the person who will resolve conflicts between NVIDIA, OEM, component, and project sources.
- Transfer only supported interface facts into the existing MPO channel worksheet. Keep unsupported or project-owned decisions open.
Evidence fields and stop conditions
| Evidence item | Record | Stop condition |
|---|---|---|
| Platform identity | DGX/HGX model, OEM system where applicable, and hardware generation | Stop if the project uses only a generic "GPU server" label |
| Fabric role | Scale-up, scale-out compute, storage/converged, or management | Stop if link domains are mixed in one quantity |
| Architecture source | Document title, URL or controlled file, revision/date, and selected topology | Stop if the source cannot be tied to the approved project |
| Endpoint pair | Source device/port and destination device/port | Stop if either endpoint is a role name rather than a real port |
| Installed module or cable | Exact part number at both ends, or approved complete-cable part number | Stop before deciding MPO, fiber type, mapping, or reach if absent |
| Point-to-point implementation | Rack/port locations, physical distance, route, bundles, and labels | Stop final quantity/length handoff if the project map is absent |
| Source conflict | Conflicting revision, OEM guidance, substitution, or interface description | Assign an owner; do not choose the more convenient source silently |
NVIDIA's deployment guidance recommends documenting logical port pairs, physical distances, endpoints, bundling, labels, rack positions, and exact port locations. NVIDIA point-to-point cable documentation
The same guidance warns that cables ordered too short can force reordering and delay deployment, while conservative overlength can create cable-management and long-term maintenance problems. Treat length as an output of the reviewed point-to-point map and route, not as a catalog default.
Do not use publication date alone to settle a source conflict. The project owner must identify the governing architecture revision, the OEM source must match the installed server, and each component document must match the exact part number. Record the resolution before releasing a cable schedule.
Use one evidence-chain line per link class:
[platform/OEM + revision] -> [fabric role] -> [endpoint A port] -> [endpoint B port]
-> [exact module/cable part numbers] -> [controlling source] -> [open decision + owner]
Classify each link as supported for handoff, a project decision that remains open, an unsupported inference, or a source conflict that holds the order. A cable supplier should return an incomplete request for clarification instead of filling the gaps with a catalog default.
Where to continue after the architecture check
A confirmed passive MPO interface does not settle the cable type, channel map, loss budget, deployment method, or test plan. Use the owner page below for the next decision so the same requirement is not resolved twice.
| Decision still to make | BWNFiber owner page | Why it stays there |
|---|---|---|
| Product and solution landscape | data-center fiber connectivity hub | Broad AI/data-center product selection |
| Trunk, patch cord, harness, breakout, or loopback | MPO cable-type guide | Cable taxonomy and Base-8/12/16 teaching |
| Endpoint mating, channel map, lane map, approvals, and changes | MPO polarity and channel-map worksheet | Type A/B/C tutorial and generic BOM-release workflow |
| Backbone architecture and channel budget | MPO trunk and loss-budget guide | Trunk construction and loss-grade tables |
| Route measurement, pulling, labeling, and delivery acceptance | pre-terminated MPO deployment guide | Installation procedure |
| Test instruments and measurement workflow | fiber-optic cable testing guide | Inspection, insertion-loss, OTDR, and continuity procedure |
Standards and testing boundary. IEC 61754-7-1 defines interface dimensions for one-row MPO connector variants; it does not select a finished assembly or prove system compatibility. IEC 61300-3-35 covers visual inspection and states that inspection does not replace attenuation or return-loss measurement. IEC 61280-4-5 covers attenuation, polarity, and length measurements for installed MPO-terminated cabling within its stated scope. The project must still approve the applicable method, limits, test cords, instrument setup, reporting format, and acceptance owner after the exact interface and channel are fixed.
How should procurement move from evidence to a comparable RFQ?
This guide owns the architecture check at the start of the buying process. It does not need to replace product, channel-design, installation, or testing pages to support the full procurement chain. Each later stage should begin with the same approved evidence set.
| Buying stage | Question to close | Release condition or next action |
|---|---|---|
| 1. Recognize the problem | Is the request based only on DGX, HGX, speed, GPU count, or cage type? | If yes, keep the request at architecture review; do not ask for a passive-cable price yet |
| 2. Understand the eligible connection type | Do the exact endpoint modules call for copper, an integrated cable, duplex fiber, or passive MPO? | Use the current endpoint and support documents; if passive MPO is confirmed, continue through the relevant MPO owner page |
| 3. Define the application | Is the link scale-out compute, storage/converged, management, or another documented domain? | Record the fabric role, topology, endpoints, and any approved project deviation |
| 4. Confirm technical inputs | Are connector, medium, mapping, reach conditions, route, length, quantity, labels, channel budget, and acceptance ownership controlled? | Keep any source-owned or project-owned field open until the correct document or owner resolves it |
| 5. Compare suppliers | Are all suppliers answering the same cable schedule and revision? | Compare line-item compliance, declared deviations, document scope, and exception handling rather than marketing claims |
| 6. Evaluate price and lead time | Do the quotations use the same configuration, quantity, document package, sample or inspection scope, destination, and target schedule? | Compare commercial terms only on that common basis; record validity, assumptions, exclusions, and proposed substitutions |
| 7. Request documents or a sample | Is the offered assembly identified by a revision-controlled datasheet or drawing? | Request the applicable drawing, datasheet, and approval evidence. Add a sample or pre-production review only when the project plan requires it, and confirm availability and terms separately |
| 8. Define quality evidence | Does the request state the inspection or test method, limits, frequency, report format, and approval owner? | Reject an unlabeled promise to "test" without the agreed scope; use the downstream testing workflow after the channel is fixed |
| 9. Send the inquiry | Does the inquiry include the architecture source, endpoints, part numbers, P2P record, quantity, destination, target schedule, required documents, and open issues? | Send one controlled packet so technical and commercial replies refer to the same requirement |
| 10. Release the order or project discussion | Are all Project decision open and Source conflict: hold items closed by the authorized project owner? | Approve one schedule revision and require the quotation, drawing, sample decision, and purchase order to reference it |
What should every shortlisted supplier return?
- The offered assembly or complete-cable identifier and document revision.
- A line-by-line response to the approved cable schedule, including every exception.
- The applicable datasheet or drawing for connector, fiber, mapping, length, labels, and construction fields that the supplier is responsible for.
- The source for any endpoint-support or compatibility statement the supplier makes.
- The proposed inspection, test-record, and traceability scope for the offered configuration.
- A written list of substitutions, assumptions, exclusions, and buyer-owned decisions.
- Price, quotation validity, quantity basis, destination basis, and configuration-specific lead-time confirmation.
- Sample or pre-production approval terms when that stage is requested.
Compare completeness before unit price. Two quotations are not equivalent if they use different drawings, mappings, document packages, test scopes, or unresolved assumptions.
When should a custom cable request begin?
Start customization only after the architecture evidence supports the physical interface. Project-specific length, breakout geometry, channel mapping, labels, packaging, or documentation may then be defined in the downstream cable schedule. Require a controlled drawing and revision for every accepted change. Customization does not prove NVIDIA or OEM compatibility.
How should an international buyer localize the RFQ?
A destination region is not a cable specification. North America, Europe, the Middle East, Southeast Asia, Latin America, and Africa each contain different authorities, project standards, tender rules, import processes, and site conditions. Record the exact project requirement instead of assigning one standard or construction to an entire region.
| Procurement context | What the buyer should identify | What the supplier must not assume |
|---|---|---|
| North American data-center project | The approved edition of the project cabling standard, any use of ANSI/TIA-942-C, the applicable building/fire authority, and owner-specific acceptance requirements | That a reference to TIA proves the offered assembly, fire rating, channel performance, or local-code acceptance |
| European or EEA building installation | The project cabling hierarchy, any use of ISO/IEC 11801-5, and whether construction-product rules, fire classification, CE marking, or a Declaration of Performance apply to the exact cable and intended use. Check the current EU Construction Products Regulation framework and national requirement | That LSZH, a fiber type, or a catalog family proves a required European fire class or regulatory document |
| Middle East data-center or EPC package | The employer specification, local authority/fire requirements, document language, submittal format, inspection stage, and import-document owner | That one Gulf or Middle East project standard applies across countries or clients |
| Southeast Asian project | The site and tender standard hierarchy, installation environment, local approval owner, labeling language, import records, and packing constraints | That humidity, heat, or a regional label automatically selects a jacket, armor, or test requirement |
| Latin American project | The country and site specification, Spanish or Portuguese document needs, import/labeling owner, required submittals, and commercial basis | That a regional market label determines compliance, tax treatment, connector, or cable construction |
| African project | The country, employer/tender specification, authority and import requirements, installation environment, document set, and delivery handoff points | That one standard, certificate, packing method, or lead-time assumption covers the continent |
For any destination, add the standard title and edition, governing-document hierarchy, installation space, route segments, and approval owner to the RFQ. A controlled data-hall route is not the same as an exterior campus segment. If a link leaves the controlled building space, separate that segment and confirm its environmental and mechanical requirements through the correct outside-plant design process. Do not select UV resistance, armoring, water blocking, rodent protection, jacket material, fire class, storage limit, or transport limit from climate alone.
The commercial request should identify currency, quotation validity, quantity basis, destination, target date, and the selected Incoterms 2020 rule with its named place. It should also assign ownership for import records, customs classification, broker instructions, partial shipments, packing units, labels, barcodes, and document language. If raw-wood pallets, crates, or dunnage are proposed, confirm the destination's implementation of ISPM 15 and record who owns that packaging evidence.
Treat private-label, OEM, or ODM language as a requested scope, not as an assumed supplier capability. State the drawing owner, artwork and label files, packaging format, document language, revision approval, and change-control process; require the supplier to confirm each item for the offered configuration. Likewise, request only the declarations, certificates, inspection records, or test reports that the exact product, intended use, project specification, and destination require.
The BWNFiber rule: source before schedule
BWNFiber's published project-review process begins with a BOM, topology, drawing, transceiver list, or node schedule before engineering review. For this NVIDIA-specific page, the proposed output is deliberately narrower: a source-reconciliation record showing what is supported for handoff, what remains a project decision, what is an unsupported inference, and what conflict holds the order. Sample, quotation, production, and test-evidence discussions remain configuration-dependent. See how BWNFiber describes its project-review process.
Copy this block into the cover of the technical inquiry:
Project and destination:
Platform, OEM system, and hardware generation:
Approved architecture title, URL/file, and revision:
Fabric role and topology:
Endpoint A device, port, module/cable PN, and operating mode:
Endpoint B device, port, module/cable PN, and operating mode:
Firmware version and validated-support source, where applicable:
P2P schedule or drawing reference:
Quantity, destination, and target schedule:
Required datasheet, drawing, sample, inspection, and test evidence:
Open decision, conflicting source, and named owner:
If the architecture fields are available but the downstream cable schedule is not, send the packet for a BWNFiber evidence review. If only the platform or speed label is available, request source reconciliation rather than a quotation.
Procurement shortcuts that should stop the order
- Do not reuse the B300 arrangement for B200 or H100/H200. The cited B300 details apply to that source and generation.
- For an HGX system, reconcile the OEM integration documents with the applicable NVIDIA architecture.
- Count enabled external endpoints. Eight GPUs do not automatically produce eight passive cables.
- Check the installed module behind each OSFP opening. A visible cage does not specify an MPO assembly.
- Read the exact interface document for 400G or 800G links. Speed does not select MPO-12 or MPO-16.
- Use the rail map to count endpoint relationships, not to choose the cable medium.
- Reconcile component totals with project deviations, switch modes, storage and management scope, and the physical route before purchasing.
- Do not treat an NVIDIA example as proof of BWNFiber product compatibility. That requires evidence for both the exact endpoint and the offered assembly.
- If sources conflict, hold the order until the network or optics owner records the resolution.
NVIDIA DGX/HGX MPO cabling FAQ
Do all NVIDIA DGX or HGX systems use passive MPO cabling?
No. NVIDIA documents copper, AOC, multimode passive fiber, and single-mode passive fiber. Passive MPO enters the design only when the approved endpoint modules call for it.
Does NVLink use the data-center MPO plant?
Do not assume that it does. In the cited DGX H100/H200 system, NVSwitch handles scale-up communication and external ConnectX-7 interfaces serve cluster networking. Follow the exact platform documentation for any user-cabled interface.
Is DGX cabling the same as HGX cabling?
No. DGX has NVIDIA system documentation. An HGX server also has an OEM integration layer, so the approved OEM and NVIDIA sources must agree on endpoints, adapters, modules, and topology.
Does an eight-GPU node require eight external fabric cables?
Only if the approved design exposes and enables eight external fabric interfaces. Use GPU count to check the architecture, not to create the cable quantity.
How should four OSFP openings serving eight ConnectX-7 interfaces be interpreted?
It is a system-level port arrangement in the cited DGX H100/H200 guide. The statement does not identify the line-side connector or imply one passive cable per opening.
Does an OSFP port mean the cable is MPO-16?
No. OSFP is the host-side form factor. Read the exact installed module or complete-cable datasheet for the line-side interface.
Can the DGX B300 MPO-12/APC example be used for DGX B200 or H100/H200?
No. Use the B300 example only for the cited design. A different platform needs its own controlling source.
Do Quantum-2 InfiniBand and Spectrum-X Ethernet use the same cables?
Not necessarily. Reconcile the selected switches, adapters or DPUs, port modes, optics or complete cables, and approved architecture for each fabric.
Can the 32-node HGX reference count be used for another 256-GPU cluster?
No. Those 256 compute-to-leaf and 256 switch-to-switch counts belong to one named architecture. Another cluster may use different adapters, topology, network domains, or port modes.
Which architecture revision should be frozen?
Use the revision approved by the project owner. Record its title, date or revision, and locator. A later online page does not override the approved design until the owner resolves the conflict.
When is the design ready for the MPO polarity worksheet?
Proceed when both endpoints and their modules or approved complete cables are known, the source supports a passive MPO interface, and the route and mapping inputs are available.
Can suppliers be compared before the cable schedule is approved?
Only their review discipline can be compared at that stage. Product compliance, unit price, lead time, and sample proposals are not comparable until every supplier receives the same schedule revision, quantity, document scope, destination, and target date.
When should a buyer request a sample or test report?
After the offered configuration and drawing revision are identified. State what the sample must confirm and what inspection or test evidence is required. Availability, price, MOQ, lead time, report format, and sampling frequency must be confirmed for that specific offer.
Does a destination region determine the MPO cable specification?
No. The approved project standard, exact installation space, local authority, endpoint interface, intended use, and required evidence control the decision. A country or climate label cannot select fiber type, jacket, fire class, armor, mapping, or connector.
What should an international RFQ say about delivery and packaging?
State the destination, quantity basis, target date, Incoterms rule and named place, packing units, labels, language, import-document owner, and any sample or inspection stage. If raw-wood packaging is used, identify the applicable phytosanitary requirement. The supplier must confirm the configuration-specific terms in writing.
What should be sent for an NVIDIA architecture evidence review?
Send the platform or OEM system, architecture revision, fabric role, switch/NIC/DPU identifiers, endpoint schedule, module or complete-cable part numbers, point-to-point map, and source conflicts.
Request an NVIDIA DGX/HGX cabling evidence review
If the request says only DGX, HGX, NDR, Spectrum-X, 400G, or 800G, a cable supplier would have to guess. Hold the passive-cable recommendation until the evidence packet is complete.
BWNFiber can review the controlled source set and separate:
- facts directly supported by the recorded NVIDIA/OEM source;
- assumptions that the source does not support;
- decisions still owned by the network, optics, cabling, or procurement team;
- the point at which the project can move to the existing MPO channel worksheet.
The requested review is limited to source reconciliation. It is not a compatibility certificate or an approved cable BOM. Send the platform or OEM system, document revision, fabric role, endpoint schedule, and module or complete-cable part numbers. Review BWNFiber's company context and start a scoped discussion.
Request a configuration-specific quotation after the evidence chain closes and the project issues an approved cable schedule. Include quantity, destination, target schedule, document requirements, and any requested sample or inspection stage. Price, MOQ, sample availability, and lead time must be confirmed in the written quotation for that revision.
