1. Establish the deployment boundary
Start with the exact compute system and network role. A rack-scale NVLink domain, an external GPU fabric, a storage network and management links are separate connection inventories. Record the platform generation, installed adapters, switch models and intended protocol for each domain.
- Name both endpoints, including adapter or line-card part numbers and software or firmware releases.
- Record whether each connection serves scale-up, scale-out, storage or management.
- Keep logical links, physical modules and cable assemblies in separate columns. A multi-port package can represent more than one link.
Read DGX B300 and GB300 NVL72 and GB200 interconnect architecture to see where those boundaries differ.
2. Resolve the actual interfaces
For each endpoint pair, record the selected port mode, electrical lane arrangement, optical interface and FEC requirements. Then check the package, thermal interface and permitted module power. An adapter family name or an aggregate rate cannot fill these fields.
ConnectX-7/8 optics and OSFP thermal interfaces provides the mechanical and host context. Use host modes for the protocol record and optical interfaces for the media-side comparison.
If a port is split, document every destination and supported lane grouping. Map the optical lanes through the complete fiber path; a breakout harness is only one part of that design.
3. Check the behavior above the link
Document the topology, uplink assumptions and operational policy alongside the physical link. For Ethernet RDMA, the congestion-management configuration belongs to a specific switch and NIC software combination. Preserve that context instead of applying a generic queue recipe.
Use RoCEv2 planning for the transport and congestion record. Use Tomahawk 6 to distinguish ASIC capacity from the ports and features an actual system exposes.
4. Keep the connection record reviewable
Complete one endpoint requirements worksheet for each materially different connection class. Retain the exact source revision, both endpoint configurations, planned media and unresolved requirements.
- Separate documented support from observations made on the actual configuration.
- Attach the test scope and error observations to any qualification result.
- Keep missing thermal, lane-mode or interoperability evidence open until it is resolved.
For phased deployment planning, carry the approved requirements into the deployment and supply record. Export the working records before leaving the browser; their local recovery is not a shared system of record.
Sources and applicability
This path organizes the source-linked engineering references below. Manufacturer examples retain their original platform and release scope; they do not establish qualification of a WBN module.
-
Platform architecture references
NVIDIA hardware and network-fabric documentation, with platform scope.
-
ConnectX endpoint references
Adapter, LinkX and thermal-interface source documents.
-
Qualification evidence
How to retain a test result without extending it beyond its configuration.
