NVIDIA Spectrum Ethernet Platform
Official vendor reference for an end-to-end AI Ethernet networking stack.
ARCHITECTURE GUIDE · AI ETHERNET
AI Ethernet Optical Interconnect: Server → Leaf → Spine
An AI Ethernet fabric is not one optic. It is a chain of host ports, network tiers, media and configurations. Start from each endpoint and link role before choosing the module, AOC or DAC.
Server EdgeNIC / DPU / SuperNIC to leaf
Leaf LayerHost aggregation and breakout context
Spine LayerHigh-radix fabric interconnect
ValidationPort mode, firmware, FEC and fiber path
Fabric Path
Treat every hop as a separate link requirement
A useful BOM separates server-to-leaf and leaf-to-spine requirements. The two link classes may use different rates, form factors, media, reach and breakout plans.
01GPU / AI serverIdentify the exact NIC / DPU and host-side port.
02Leaf switchConfirm server-facing and fabric-facing port modes.
03Spine switchConfirm uplink rate, density and reach.
04Return pathValidate the complete endpoint-to-endpoint topology.
Link-by-Link Checklist
Eight fields to freeze before selecting optics
01 / ENDPOINTSExact device modelsServer NIC / DPU, leaf switch and spine switch.
02 / PORTPhysical form factorOSFP, QSFP-DD, QSFP112 or another exact port implementation.
03 / RATENative link rateConfirm the required rate on each hop, not only the aggregate fabric target.
04 / BREAKOUTLane and breakout planDocument straight-through or breakout operation at each endpoint.
05 / REACHActual cable routeUse deployed distance and patch path rather than rack-to-rack guesswork.
06 / MEDIADAC, AOC or optical moduleSelect media from distance, routing, serviceability and platform support.
07 / FIBERConnector and fiber architectureParallel or duplex path, polarity and patch-panel context.
08 / SOFTWAREFirmware / FEC / port policyFreeze the configuration used for project validation.
Procurement Model
Build the BOM by link role, not by one global “800G” label
Server → Leaf
- Endpoint port on NIC / DPU and leaf switch
- Short-reach routing and cable serviceability
- Native / breakout mode
- Rack and patching constraints
- Sample validation before scale
Leaf → Spine
- Fabric-facing port rate and density
- Optical reach and fiber plant
- Parallel / duplex optical architecture
- Breakout or native uplink topology
- Volume quantity and staged delivery
Architecture boundary: leaf-spine is a useful procurement model, not a claim that every AI Ethernet deployment uses the same topology. Final selection follows the actual network design and platform documentation.
Topology & Plane Model
Do not assume every AI Ethernet fabric is one classic leaf-spine plane
Modern AI Ethernet designs can use single-plane, dual-plane, quad-plane or flatter two-tier structures. Procurement should identify the link role and network plane for every BOM line.
Single-plane / conventional two-tier
- Server-facing links terminate at the access / leaf layer.
- Fabric-facing links connect leaf and spine tiers.
- Each hop can have a different rate, media and optical architecture.
- Useful as a simple procurement decomposition when the actual design matches it.
Multiplane AI Ethernet
- One GPU server can attach across multiple independent network planes.
- NIC / SuperNIC generation and rail mapping become part of the BOM evidence.
- Two-tier fabrics can scale without adding another switching tier when the validated platform architecture supports it.
- Never collapse multiple planes into one generic “800G link” line item.
Current platform example: NVIDIA's current Spectrum-X reference architecture documents single-plane designs with ConnectX-7 / BlueField-3 and dual- or quad-plane designs with ConnectX-8. Treat this as platform-specific evidence, not a universal Ethernet topology rule.
BOM by Link Role
Add plane, tier and endpoint fields to every interconnect line
| BOM field | Why it matters | Example evidence |
|---|---|---|
| Plane / rail | Prevents multiple network planes from being merged into one requirement | Plane A / B / C / D or documented rail ID |
| Source endpoint | Defines host port, protocol and module/cable policy | NIC / DPU / SuperNIC model + port |
| Destination endpoint | Defines peer cage, port mode and link compatibility | Leaf / spine switch model + port |
| Link role | Separates server-edge from fabric-uplink requirements | Server→Leaf, Leaf→Spine, peer/fabric link |
| Physical path | Determines media, optics, fiber and reach | DAC / AOC / pluggable optics + route length |
| Validation state | Prevents an untested link from entering the volume BOM as approved | PLANNED / SAMPLE / PASS / HOLD / RE-TEST |
Reference Context
Use vendor platform documentation to anchor the real topology
For example, NVIDIA describes its Spectrum Ethernet platform as an end-to-end Ethernet stack spanning switches, SuperNICs, DPUs, cables and networking software. That is a platform reference—not a universal compatibility claim for every optic.
Spectrum-X Optical Selection
For Spectrum-X projects, use the platform-specific guide and NVIDIA validated configuration evidence.
Building an AI Ethernet BOM?
Send the server NIC / DPU, leaf and spine models, port modes, link distances and quantity by hop.
Planning the adapter-side optical path?
Use the ConnectX guide to verify the exact adapter endpoint, peer switch and NVIDIA-supported interconnect evidence.