Buyers looking for current NVIDIA Mellanox models can review the Top 20 NVIDIA Mellanox products for AI and HPC shortlist before preparing a bill of materials. That page focuses on current product availability and model selection, while this article explains how to match each product type to an AI or HPC network design.
The shortlist covers InfiniBand switches, Spectrum Ethernet switches, ConnectX server adapters, and high-speed optical transceivers. You can also review the NVIDIA Mellanox networking product category when expanding the selection beyond the 20 models listed here.
- Understand the difference between a demand-led shortlist and a universal ranking.
- Compare switches, adapters, and optical transceivers by deployment role.
- Match product families to 800G Ethernet, 400G RoCE, NDR, HDR, and 25G access scenarios.
- Verify protocol, connector, PCIe, firmware, cable, and airflow requirements.
- Confirm optical type, distance, lane structure, polarity, and protocol coding.
- Prepare the switch, NIC, transceiver, cable, software, and support requirements together.
- Answers to common NVIDIA Mellanox AI and HPC procurement questions.
- Use the complete fabric and compatibility requirements to finalize the model list.
Start With the Network Fabric
The first decision is not the switch model. It is the network fabric that the servers and GPUs must use. NVIDIA Mellanox products are commonly deployed in InfiniBand and Ethernet environments, but those designs are not interchangeable simply because they use a similar data rate or connector.
| Fabric or Role | Typical Product Type | What Must Be Decided |
|---|---|---|
| InfiniBand AI or HPC fabric | Quantum InfiniBand switch and InfiniBand or VPI adapter | Fabric speed, subnet management, topology, cable type, and adapter generation |
| Ethernet or RoCE fabric | Spectrum Ethernet switch and Ethernet or VPI adapter | Ethernet speed, PFC, ECN, routing, congestion control, and NOS support |
| Server connectivity | ConnectX adapter | PCIe generation, port count, connector type, Ethernet versus VPI operation, and bracket |
| Physical interconnect | OSFP, QSFP56, QSFP112, QSFP-DD, DAC, or AOC | Protocol coding, lane structure, fiber type, reach, breakout mode, and airflow |
A 400G InfiniBand switch should not be treated as a direct substitute for a 400GbE Spectrum switch without checking the adapter mode, software stack, transceiver coding, cable type, and fabric management requirements. The same caution applies when comparing 200G HDR, 400G NDR, and 800G Ethernet products.
How This Product Shortlist Is Organized
“Top 20” in this article does not mean a universal performance ranking. The shortlist is organized from existing AI and HPC BOM activity, product visibility, and channel inquiry or circulation signals. It is intended to help buyers start a technical shortlist, not to replace a model-level compatibility review.
- Switches are grouped by InfiniBand or Ethernet fabric role.
- Adapters are separated by protocol, port count, PCIe interface, and speed.
- Transceivers are evaluated by connector, optical medium, reach, and link type.
- Current stock, pricing, and lead time must be confirmed against the exact SKU before purchase.
- Airflow direction, power configuration, and rack depth should be checked before installation.
Compare current switches, adapters, and transceivers before preparing your AI or HPC BOM.
The 20 NVIDIA Mellanox Models
InfiniBand and Ethernet Switches
Switch selection should start with protocol and network position. Quantum models are intended for InfiniBand fabrics, while Spectrum models are used for Ethernet, RoCE, cloud-scale, and data-center switching designs.
| Model | Product Role | Key Selection Point |
|---|---|---|
| SN5610 | Spectrum-4 Ethernet switch | 800GbE OSFP platform for high-density Ethernet, RoCE, spine, and super-spine designs. |
| MQM9700-NS2F | Managed NDR InfiniBand switch | 400Gb/s NDR fabric with onboard management features and P2C airflow. |
| MQM9790-NS2F | Unmanaged NDR InfiniBand switch | 400Gb/s NDR fabric requiring external UFM or Subnet Manager planning. |
| MSN4700-WS2FC | Spectrum-3 Ethernet switch | 400GbE QSFP-DD platform for high-bandwidth leaf and spine networks. |
| MSN4600-CS2FC | Spectrum-3 Ethernet switch | High-density 100GbE switching for spine, super-spine, and cloud-scale designs. |
| MSN3700-CS2FC | Spectrum-2 Ethernet switch | 32-port 100GbE QSFP28 platform for 100G leaf, spine, and top-of-rack networks. |
| MSN2700-CS2FC | Spectrum Ethernet switch | 100GbE QSFP28 switching with supported breakout options for 25G or 50G connectivity. |
| MSN2410-CB2FC | 25GbE/100GbE Ethernet switch | 48 SFP28 access ports with QSFP28 uplinks for dense server or storage access. |
| MSN2010-CB2FC | 25GbE/100GbE Ethernet switch | Compact 25GbE access and 100GbE uplink option for smaller data-center deployments. |
| MQM8700-HS2F | Managed HDR InfiniBand switch | 200Gb/s HDR InfiniBand option for AI, HPC, and high-performance storage fabrics. |
ConnectX Network Adapters
The adapter determines how the server connects to the fabric. Two products with a similar port speed may still differ in Ethernet-only operation, VPI support, PCIe generation, port count, security features, and bracket type.
| Model | Adapter Type | Key Selection Point |
|---|---|---|
| MCX75310AAS-NEAT | ConnectX-7, 400GbE/NDR, single OSFP | PCIe 5.0 x16 server adapter for 400GbE or NDR InfiniBand deployments. |
| MCX755106AS-HEAT | ConnectX-7, dual-port QSFP112 | High-speed ConnectX-7 option requiring PCIe Gen5, server airflow, and cable validation. |
| MCX653105A-HDAT | ConnectX-6 VPI, 200Gb/s HDR IB/200GbE | Single-port QSFP56 adapter for flexible HDR InfiniBand or Ethernet deployments. |
| MCX623106AN-CDAT | ConnectX-6 Dx Ethernet adapter | Dual-port 100GbE QSFP56 adapter for Ethernet and RoCE networks. |
| MCX556A-EDAT | ConnectX-5 Ex VPI adapter | 100GbE and EDR InfiniBand option for existing server fleets and legacy high-speed fabrics. |
Optical Transceivers
Optics should be selected after the switch port and adapter have been identified. The connector alone does not prove compatibility. Check the protocol, lane mapping, optical medium, wavelength, reach, and vendor coding together.
| Model | Optical Role | Key Selection Point |
|---|---|---|
| MMA4Z00-NS | 800G twin-port NDR OSFP transceiver | Dual NDR design with MPO12 APC multimode connectivity for short-reach links. |
| MMS4X00-NM | 800G twin-port NDR optical transceiver | Single-mode option for longer 800G NDR links. |
| MMS1X00-NS400 | 400G NDR QSFP112 transceiver | Single-mode 400G option where reach and connector architecture are appropriate. |
| MMA4Z00-NS400 | 400G NDR OSFP transceiver | Short-reach multimode option using MPO12 APC connectivity. |
| MMS4X00-NS400 | 400G NDR optical transceiver | Single-mode 400G NDR option for verified optical links. |
Which Products Fit Each Deployment Scenario?
Large 400G or 800G AI Ethernet Cluster
The SN5610 is relevant when the design requires dense 800GbE OSFP connectivity, high aggregate bandwidth, and a Spectrum-based Ethernet or RoCE fabric. It should be evaluated together with the ConnectX generation, OSFP breakout plan, cable reach, and data-center power and cooling limits.
400G Ethernet or RoCE Leaf-Spine Network
A 400GbE design may use Spectrum-3 platforms such as the MSN4700-WS2FC. The practical decision is whether 400G is needed at the server-facing layer, at the spine layer, or only for uplinks.
400G NDR InfiniBand Fabric
The MQM9700-NS2F and MQM9790-NS2F are NDR InfiniBand options, but their management models differ. A managed design with onboard fabric management is not the same procurement decision as an unmanaged switch requiring an external UFM or Subnet Manager.
Existing 200G HDR InfiniBand Environment
For an existing HDR environment, the MQM8700-HS2F and MCX653105A-HDAT may be more relevant than moving directly to an NDR architecture. Confirm whether the existing servers, cables, firmware, and fabric management platform support the proposed upgrade.
25G Server Access With 100G Uplinks
A platform such as the MSN2410-CB2FC is better aligned with dense 25GbE server access and 100GbE uplinks than an 800G spine switch. This type of design is often used where server NICs and storage endpoints do not yet require 200G or 400G connections.
Check Switch-to-NIC Compatibility
The most expensive errors usually occur when the switch and adapter are purchased separately. A port speed that looks correct in a product title does not guarantee that both devices will establish the intended link.
| Compatibility Check | Why It Matters | Information to Confirm |
|---|---|---|
| Protocol | Ethernet, RoCE, and InfiniBand have different fabric requirements. | Ethernet, VPI, or native InfiniBand operation |
| Port generation | OSFP, QSFP56, QSFP112, and QSFP-DD are not automatically interchangeable. | Connector, lane count, port profile, and breakout mode |
| Host PCIe interface | A high-speed adapter can be limited by the server slot or riser. | PCIe generation, lane width, CPU attachment, and bracket |
| Transceiver coding | Unsupported or incorrectly coded optics can prevent the link from coming up. | Vendor coding, firmware support, optical type, and cable validation |
| Software and firmware | Protocol mode and feature support may depend on firmware and NOS versions. | Adapter firmware, switch NOS, driver, and management software |
| Airflow and power | High-speed optics and switches can materially change rack-level power and cooling. | Airflow direction, PSU type, rack orientation, and thermal limits |
For Ethernet or RoCE deployments, the compatibility review should also include PFC, ECN, buffer behavior, FEC, VLAN or VXLAN requirements, and the intended congestion-control model. For InfiniBand deployments, confirm subnet management, fabric topology, link speed, and the required management platform.
Do Not Separate Optics From the Network Design
A transceiver is part of the link budget, not an accessory to be added after the switch is selected. The same nominal speed can use different lane structures, connector types, fiber media, and distances.
- Use multimode optics only when the planned distance and fiber plant support them.
- Confirm single-mode reach, wavelength, polarity, and connector type for longer links.
- Check whether the OSFP, QSFP56, QSFP112, or QSFP-DD port supports the selected module.
- Verify whether a breakout cable changes one physical port into multiple logical links.
- Do not assume that a 400G Ethernet optic is suitable for a 400G NDR InfiniBand connection.
- Confirm that both ends of the link use a supported and compatible module or cable family.
The MMS4X00-NM and MMA4Z00-NS illustrate why reach and fiber type matter. They are both associated with high-speed NDR connectivity, but their optical design, connector arrangement, and intended distance are different.
For the broader technology context, NVIDIA provides an overview of its Ethernet, InfiniBand, switch, and adapter networking portfolio in the NVIDIA networking overview.
Build the BOM Around the Complete Link
A reliable AI or HPC quotation should not contain only switch chassis and NIC quantities. It should describe the complete path from the server to the fabric and include the components needed to install, manage, and support the deployment.
- Define the number of GPU servers, storage nodes, management nodes, and spare nodes.
- Choose InfiniBand, Ethernet, or RoCE before selecting the switch family.
- Calculate server-facing ports, uplink ports, spine ports, and spare capacity.
- Match each server NIC to the required PCIe slot, bracket, firmware, and driver.
- List every DAC, AOC, transceiver, breakout cable, and fiber patch connection.
- Check airflow direction, power supply type, rack depth, and available cooling.
- Confirm software, licenses, fabric management, support, and replacement requirements.
- Request validation for the exact SKU rather than quoting only a product family.
Information to Include in a Quote Request
- Number of servers and GPUs per server
- Required fabric: InfiniBand, Ethernet, or RoCE
- Target speed: 100G, 200G, 400G, or 800G
- Number of leaf and spine switches
- Server NIC model or required adapter generation
- Port connector and cable distance
- Rack airflow direction and power environment
- Required delivery region, quantity, and deployment date
Send the exact server count, fabric design, port speeds, cable distances, and delivery requirements through the AI and HPC configuration inquiry form so the switch, adapter, transceiver, and cable quantities can be reviewed together.
Frequently Asked Questions
No. The protocol, adapter mode, switch ASIC, transceiver coding, cable type, fabric management, and software stack must all be checked. A similar data rate or connector does not prove interoperability.
Choose a VPI adapter when the deployment may use either InfiniBand or Ethernet and the adapter supports the required speed and connector. Confirm the operating mode, firmware, driver, cable, and switch platform before ordering.
Confirm the switch or adapter port profile, breakout requirement, lane mapping, optical type, fiber distance, cable polarity, firmware support, power, and cooling. Also verify whether the selected 800G product is designed for Ethernet or InfiniBand.
Sometimes, but only when the switch port, breakout mode, cable, transceiver, and adapter support the intended lower-speed operation. The final design should be checked as a complete link rather than by switch speed alone.
Provide the GPU or server count, fabric protocol, target speed, leaf and spine quantities, adapter model, port type, cable distance, rack airflow, power environment, delivery region, and required timeline. This allows the exact switch, NIC, transceiver, cable, and spare quantities to be reviewed together.
Final Takeaway
The right NVIDIA Mellanox product is the one that fits the complete fabric, not simply the model with the highest port speed. Start with InfiniBand versus Ethernet, calculate the required port density, match the ConnectX adapter to the server, and validate every optical or direct-attach link.
Use the current NVIDIA Mellanox product shortlist to identify candidate models, then submit the server count, fabric design, port speeds, cable distances, and delivery requirements for a configuration-specific quotation.