An NVIDIA SN5610 800G OSFP port can be configured as two 400G logical ports in a supported breakout mode. 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.
The main compatibility question is not simply whether the switch and server are labeled “400G.” You also need to identify whether the server uses a 400G OSFP or QSFP112 adapter, whether the link is Ethernet, RoCE, or InfiniBand, and whether the selected breakout cable or optical assembly supports the required lane mapping and FEC.
- Compare 400G OSFP and QSFP112 server adapter requirements.
- Select DAC, AOC, or optical modules according to the complete link.
- Do not treat similar bandwidth labels as proof of protocol compatibility.
- Compare 800G breakout against native 400G switching.
- Calculate physical switch ports, logical links, redundancy, and future capacity.
- Review FEC, lane mapping, PCIe, optics, cables, and BOM requirements.
- Answers to common 800G switch and 400G server-link questions.
- Validate the switch, NIC, cable, transceiver, and protocol as one complete link.
How 800G-to-400G Breakout Works
The NVIDIA SN5610 provides 64 physical 800GbE OSFP ports. With a supported breakout configuration, each physical port can operate as one 800G link, two 400G links, or four 200G links. The platform also supports lower-speed breakout modes with additional port-allocation considerations.
| Switch Port Mode | Logical Links per Physical Port | Typical Use |
|---|---|---|
| 1×800G | One 800G link | 800G spine, super-spine, or SuperNIC connection |
| 2×400G | Two 400G links | 400G server, leaf, or spine connections |
| 4×200G | Four 200G links | 200G GPU or storage connections |
| 8×100G | Eight 100G links | Lower-speed or legacy server connectivity |
In 2×400G mode, one physical 800G port becomes two independent 400G logical interfaces. The total capacity remains 800G, but it is distributed across two links. A single 800G port therefore does not become one 400G port with the other half automatically available for a different purpose.
NVIDIA documents the SN5610 breakout behavior and port restrictions in its SN5000 switch cable and breakout documentation.
When all supported physical ports are configured for 2×400G breakout, the 64-port SN5610 can provide up to 128 logical 400G ports. The logical port count is not the same as the number of physical switch cages.
What Must Match on the Server Side?
The server-side adapter determines which breakout cable or transceiver assembly is appropriate. Two adapters with the same nominal 400G speed may use different connectors, protocols, firmware modes, or host interfaces.
400G OSFP Server Adapter
A single-port 400G OSFP ConnectX-7 adapter can connect to one of the 400G logical ports created by an SN5610 800G port. The MCX75310AAS-NEAT is an example of a ConnectX-7 adapter with a single OSFP interface and PCIe 5.0 x16 host interface.
This design normally requires an 800G OSFP switch-side connection, two 400G breakout legs, and one compatible 400G OSFP server-side connection for each connected server. The protocol and optical or direct-attach specifications must also match.
400G QSFP112 Server Adapter
Other ConnectX-7 configurations use QSFP112 connectors instead of OSFP. A QSFP112 server adapter cannot be treated as mechanically identical to an OSFP adapter.
The MCX755106AS-HEAT is an example of a high-speed QSFP112 adapter. Before selecting a cable, confirm whether the design requires OSFP on the switch side and QSFP112 on the server side, two independent 400G QSFP112 links, and a supported DAC, AOC, or optical breakout assembly.
| Server Adapter Type | Connector | Selection Focus |
|---|---|---|
| 400G ConnectX-7 Ethernet or VPI adapter | OSFP | OSFP breakout leg, protocol mode, PCIe 5.0, and FEC |
| 400G ConnectX-7 Ethernet adapter | QSFP112 | OSFP-to-QSFP112 cable or optical assembly and lane mapping |
| 200G adapter | QSFP56 or supported connector | 4×200G breakout profile, cable type, and host-side speed |
OSFP Breakout Cable and Transceiver Choices
A breakout cable must match the source port, target port, lane count, encoding, speed, reach, and FEC requirements. Connector shape alone is not sufficient evidence of compatibility.
| Switch Side | Breakout Design | Server Side | Main Check |
|---|---|---|---|
| 800G OSFP | 1×800G to 2×400G | Two 400G OSFP adapters | OSFP coding, lane mapping, protocol, and cable length |
| 800G OSFP | 1×800G to 2×400G | Two 400G QSFP112 adapters | Supported OSFP-to-QSFP112 breakout cable or optical assembly |
| 800G OSFP | Native 800G | One 800G server adapter | Native 800G support on both endpoints |
| 800G OSFP | 1×800G to 4×200G | Four 200G adapters | 200G connector, FEC, lane configuration, and breakout profile |
| 800G OSFP | 1×800G to 8×100G | Eight 100G links | Adjacent-port restriction and logical-port allocation |
DAC, AOC, or Optical Transceiver?
- Use a passive DAC when the distance, power, connector, and platform support requirements are suitable for a short in-rack link.
- Use an AOC when a longer direct cable is required without separate transceivers and fiber patching.
- Use optical transceivers and fiber when the link requires greater reach, structured cabling, or a patch-panel connection.
- Confirm whether the selected module is intended for Ethernet, RoCE, or InfiniBand.
- Check vendor coding, wavelength, fiber type, polarity, FEC, and platform support before ordering.
For the current model selection, the NVIDIA Mellanox AI and HPC product shortlist can be used to identify candidate switches, adapters, and transceivers. The exact cable assembly should still be validated against the switch and server-side SKU.
Send your switch model, NIC model, connector type, and cable distance for review.
400GbE, RoCE, and 400G NDR Are Different
The phrase “400G server link” is incomplete without identifying the protocol. A 400GbE server link, a RoCE link, and a 400G NDR InfiniBand link may use similar bandwidth numbers but require different network designs.
| Protocol | Typical Environment | Main Compatibility Requirements |
|---|---|---|
| 400GbE | Ethernet data center, storage, or AI network | Ethernet switch, Ethernet NIC, compatible optics or cables, FEC, and NOS |
| RoCE | RDMA over Ethernet AI or storage fabric | Ethernet compatibility plus PFC, ECN, QoS, buffer, and congestion-control planning |
| 400G NDR InfiniBand | InfiniBand AI or HPC fabric | NDR switch, InfiniBand or VPI adapter, compatible cables, and subnet management |
A 400GbE Spectrum switch should not automatically be paired with a 400G NDR adapter simply because both products are labeled “400G.” The switch, adapter, cable, transceiver, firmware, and fabric management model must all match the selected protocol.
The SN5610 belongs to the Spectrum Ethernet family. Quantum-2 products such as MQM9700 are designed for NDR InfiniBand fabrics and should be evaluated through a different compatibility path.
When Should You Use an 800G Switch?
An 800G switch is most useful when the design needs high radix, future expansion, or a mixture of 800G, 400G, and 200G links. It is not automatically the lowest-cost option for every 400G server deployment.
- The spine or super-spine requires 800G ports.
- Multiple 400G server links must be connected with high physical-port density.
- The network will gradually migrate from 400G to 800G.
- The design needs a mixture of 400G and 200G breakout modes.
- Rack space and switch port density are important project constraints.
If the project only needs a limited number of 400G server links, a native 400G switch may offer a simpler bill of materials and easier cabling. The correct decision depends on current server density, uplink requirements, expansion plans, and the total cost of optics and cables.
| Requirement | 800G Switch With Breakout | Native 400G Switch |
|---|---|---|
| Current server speed | 400G through supported breakout | Native 400G |
| Port density | Higher physical-port efficiency | Simpler port mapping |
| Future 800G upgrade | Better prepared | May require a later platform change |
| Cable planning | More complex | Usually simpler |
| Best fit | Large or expanding AI fabrics | Smaller or fixed 400G deployments |
How Many 800G Ports Do You Need?
For a basic 400G server connection, one 800G physical port provides two 400G logical ports. Two 400G server links therefore consume one 800G switch port when the supported breakout design is used.
For example, 32 single-port 400G servers require at least 16 physical 800G switch ports. The final design should reserve additional ports for uplinks, redundancy, maintenance, storage, and future expansion.
For a dual-rail GPU design, count each network rail separately. A server with two 400G NIC ports may consume two logical 400G switch ports, even if both ports are installed in the same server.
Record the Complete Port Map
- Physical switch port
- Breakout mode
- Logical interface
- Server hostname or rack position
- NIC model and connector type
- Cable or transceiver model
- FEC mode
- Optical distance
- Network rail and failover path
Compatibility Mistakes and Procurement Checklist
Common Compatibility Mistakes
- Looking only at “800G” and “400G” without checking the protocol.
- Treating OSFP and QSFP112 as the same physical interface.
- Confusing 800G-to-2×400G with 400G-to-4×100G breakout.
- Ignoring FEC requirements on PAM4-based links.
- Buying the switch and NIC without calculating both ends of every link.
- Failing to document logical port numbering after breakout.
- Using transceivers with unconfirmed coding or platform support.
- Testing only whether the link is up without checking FEC errors, CRC, or BER.
Procurement Checklist
- Exact switch model and port type
- Required 800G, 400G, and 200G port counts
- Server NIC model and connector type
- Ethernet, RoCE, or InfiniBand protocol
- DAC, AOC, or optical transceiver requirement
- Fiber type and link distance
- Lane mapping and breakout profile
- FEC and auto-negotiation requirements
- Server PCIe generation and slot configuration
- Switch airflow and rack power
- Spare cables, optics, and replacement requirements
Provide the switch model, server model, NIC model, protocol, required link speed, connector type, cable length, optical reach, rack airflow, quantity, and planned deployment date before requesting a final quotation.
Frequently Asked Questions
A supported 2×400G breakout normally creates two 400G logical ports from one 800G physical port. If only one 400G link is connected, the second logical port may remain unused.
It may be possible with a supported OSFP-to-QSFP112 cable or optical assembly, but it should not be assumed. Confirm the exact switch port mode, NIC model, cable, firmware, and transceiver compatibility.
No. They use different fabric protocols and management models. The switch, adapter, cable, transceiver, and software stack must all match the selected fabric.
An 800G switch is more attractive when the network needs high port density, multiple 400G server links, 800G spine connectivity, or a future migration path. A native 400G switch may be simpler for smaller fixed deployments.
Provide the switch model, server model, NIC model, protocol, required port count, connector type, cable length, optical reach, rack airflow, quantity, and planned deployment date.
Final Takeaway
An 800G switch can efficiently connect 400G servers through supported 2×400G breakout, but the design is not determined by bandwidth alone.
The switch port profile, server NIC connector, Ethernet or InfiniBand protocol, OSFP or QSFP112 cable design, optical reach, lane mapping, FEC, firmware, PCIe, and thermal requirements must be validated as one complete link.
For a configuration-specific review, send the switch model, server NIC, required port count, cable distance, and deployment quantity through the AI and HPC networking inquiry form.















































































































































