AI & HPC networks
Which HGX connections belong to a SuperNIC and which to a DPU?
Keep GPU communication, converged services and management distinct in the node and fabric BOM.
An Ethernet AI fabric can still contain several different network roles. NVIDIA’s HGX reference distinguishes GPU communication from storage and customer access, with dedicated management requirements that a generic 400G BOM can miss.
The short answer
Map the role before the adapter: the selected reference uses B3140H SuperNICs for compute E-W and B3220 DPUs for converged N-S traffic. Preserve the reference configuration and validate any proposed substitution as a design change.
The selected HGX AI Factory reference assigns different roles to B3140H SuperNIC and B3220 DPU paths; OOB management remains separate. Conceptual illustration; equipment shapes are illustrative.
Official technical sources · Independent analysis by SwitchInfra
This is a documented design analysis, not a customer deployment claim. Candidate roles require verification for your configuration.
400GbE — B3140H reference port
2 × 200GbE — B3220 reference ports
E-W — compute fabric role
N-S — converged service role
What the technical reference establishes
The HGX H100/H200/B200 networking hardware page identifies a single-port 400GbE B3140H for compute and a dual-port 200GbE B3220 for in-band management and storage. The logical architecture keeps traffic roles explicit.
This is a useful procurement comparison because two BlueField-branded devices do not automatically serve the same purpose. A quote should preserve the node’s reference configuration, connection ownership and expected software behavior.
HGX AI Factory networking hardware
Compute — B3140H serves the documented E-W role.
Services — B3220 serves the documented converged N-S role.
Fabric — Spectrum-X Ethernet underpins the reference networks.
Management — OOB connectivity has a distinct infrastructure role.
Scope of the analysis: Independent selection analysis of official NVIDIA documentation. Proposed examples are not a deployed customer project or a SwitchInfra delivery.
Work through the engineering choices
Start with the documented constraint, then state the conditions for the proposed equipment role.
Preserve the GPU-to-network mapping
The reference connects compute traffic through SuperNICs.
Design exercise: Proposed task: document the intended GPU, adapter and fabric-rail relationship.
Do not start by replacing several reference adapters with one faster-looking card. Identify which server resources each connection serves, then preserve the intended path through the certified node configuration. A different slot placement can change the relationship even if the external port rate remains the same.
Request the exact server configuration, adapter OPNs and supported software combination. Describe the planned training or inference workload and its scale. Use application-level tests to assess a proposed deviation; a working Ethernet link does not establish equivalent GPU communication behavior.
HGX AI Factory networking hardware · HGX AI Factory logical architecture
B3140H compute E-W reference role
Exact crypto-enabled B3140H OPN. Server integration, software and supply require confirmation.
Budget the service fabric independently
The reference gives the converged N-S fabric a separate role.
Design exercise: Proposed task: allocate storage, customer and management-service connections explicitly.
List data ingestion, storage access and external users against their actual endpoints. Determine which traffic can share the proposed fabric and what isolation or resource policy it requires. Keep capacity calculations tied to those services rather than assuming that unused compute ports cover every other need.
Where dual connections are proposed, identify their separate switch peers and the supported recovery behavior. A two-port device alone does not establish independent power, paths or application continuity. Include storage endpoints and customer-facing uplinks in the port budget before declaring the service fabric sized.
HGX AI Factory logical architecture
B3220 converged N-S reference role
Exact crypto-enabled B3220 OPN; do not substitute the 00SV variant without review.
Treat Spectrum-X as an integrated configuration
The reference combines switches, adapters and software.
Design exercise: Proposed procurement decision: retain the reference stack or document a supported deviation.
A RoCE-capable NIC or an Ethernet switch by itself does not reproduce the whole reference. Ask the integrator to identify the software, congestion-management configuration and validation required for the offered combination. Keep component capability separate from proof that this complete fabric has been qualified.
Include OOB connections, management addressing and recovery procedures in the delivery package. Confirm the exact switch model and physical interconnect before counting ports, because an aggregate cage rate and an exposed network interface are different units. Freeze the agreed configuration with the acceptance workload.
HGX AI Factory networking hardware · HGX AI Factory logical architecture
SN5600 / SN5610 family
Spectrum-X fabric inquiry
Require exact switch, software and port allocation for the chosen reference scale.
SN2201 family
OOB connectivity inquiry
Allocate infrastructure management endpoints separately.
Compare quotations by network role
The meaningful comparison is the delivered and supported configuration for each traffic class. A common vendor logo or aggregate bandwidth number cannot demonstrate functional equivalence.
HGX AI Factory networking hardware
Node
Reference node configuration, GPU mapping, adapter OPNs and slot placement.
Fabric
Per-role endpoints, switch ports, software and congestion policy.
Validation
Workload, isolation, recovery and management-access acceptance evidence.
Candidate models and selection boundaries
These are conditional model or product-family inquiries for the documented equipment roles. Confirm the required rate and complete channel before finalizing any line.
| Model / family and status | Role and condition |
|---|---|
| 900-9D3D4-00EN-HA0 Listed OPN; conditional candidate | B3140H compute E-W reference role Exact crypto-enabled B3140H OPN. Server integration, software and supply require confirmation. HGX AI Factory networking hardware · BlueField-3 exact ordering configurations |
| 900-9D3B6-00CV-AA0 Listed OPN; conditional candidate | B3220 converged N-S reference role Exact crypto-enabled B3220 OPN; do not substitute the 00SV variant without review. HGX AI Factory networking hardware · BlueField-3 exact ordering configurations |
| SN5600 / SN5610 family Conditional model / family inquiry | Spectrum-X fabric inquiry Require exact switch, software and port allocation for the chosen reference scale. HGX AI Factory logical architecture |
| SN2201 family Conditional model / family inquiry | OOB connectivity inquiry Allocate infrastructure management endpoints separately. HGX AI Factory networking hardware |
Rows with product IDs are mapped to the exact listed model or ordering variant checked on 19 October 2026. Remaining rows are model or family inquiries. A public listing does not establish stock, offered condition, supply or project compatibility.
Continue in the relevant product family
Use the analysis to define the role before comparing catalog options.
Discuss this project
Carry the route, media, endpoint and loss requirements into the inquiry. Leave unknown quantities blank; reference port and fiber counts are not an equipment BOM.
Follow the next selection question.
Compare related decisions while keeping each source’s scope, hardware and link conditions separate.
Browse more engineering analyses →
BlueField-3: include management and recovery in the BOM
Separate data ports, OOB access, RShim and server power requirements.
Read the engineering analysisOfficial technical source
ConnectX-6 Dx: qualify the card at both interfaces
Compare OPN, PCIe resources and complete Ethernet channels for a 25G or 100G refresh.
Read the engineering analysisOfficial technical source
ConnectX-7: a 400G upgrade starts inside the server
Qualify the host slot, exact connector and current feature support before buying interconnects.
Read the engineering analysisOfficial technical source
Before you finalize the channel
Are B3140H and B3220 interchangeable?
The reference assigns different roles; any substitution needs configuration review.
Does RoCE support prove Spectrum-X equivalence?
No. The offered hardware, software and fabric behavior must be assessed together.
Can N-S demand be inferred from GPU count?
Not alone. Storage and customer traffic requirements remain project inputs.
Is this a SwitchInfra customer case?
No. It is independent analysis of an official reference architecture.
Trace the project facts and technical limits
- HGX AI Factory networking hardware
- HGX AI Factory logical architecture
- BlueField-3 exact ordering configurations
Source check: 19 October 2026. Catalog status is stated per candidate.
Official sources read 19 October 2026. Project-specific engineering review, firmware support, exact offered parts and supply remain unconfirmed.