400G Optical Transceivers: Types, Form Factors and Cabling

1 400G Electrical Interfaces: 8 × 50G and 4 × 100G Lanes

1.1 Port Speed and Lane Rate

A 400G port moves data between the switch or NIC chip and the transceiver over parallel electrical lanes. The port speed is the total: 400 Gb/s. The lane rate is the speed of each lane. The same 400 Gb/s can be carried as eight lanes at 50 Gb/s or as four lanes at 100 Gb/s, and the two arrangements are not interchangeable. The lane rate is set by the switch or NIC chip, not by the transceiver.

1.2 The 8 × 50G Interface

The eight-lane interface was standardized in 2017 (IEEE 802.3bs; interface name 400GAUI-8). Each lane carries 50 Gb/s using PAM4 signaling, which encodes two bits per symbol. This interface is used by the QSFP-DD switches sold from 2019 to 2024, built on Broadcom Tomahawk 3 and Tomahawk 4, Trident 4 and Cisco Cloud Scale chips. The 400G ports installed in enterprise and cloud data centers between 2019 and 2024 use this interface.

1.3 The 4 × 100G Interface

The four-lane interface was standardized in 2022 (IEEE 802.3ck; 400GAUI-4). Each lane carries 100 Gb/s, again with PAM4. This is the interface on every 400G AI server NIC (NVIDIA ConnectX-7, the Ethernet ports of ConnectX-8, and BlueField-3), on NVIDIA's OSFP switches, and on the 51.2 Tb/s switch generation built on Broadcom Tomahawk 5 and Cisco Silicon One. A third interface, two lanes at 200 Gb/s (400GAUI-2), is in the IEEE P802.3dj draft and already appears on NVIDIA's ConnectX-9 NIC.

1.4 Effects of the Lane Rate: Copper Reach, Breakout and Compatibility

The lane rate does not change the optics. A 400G DR4 module sends four 100 Gb/s optical signals over the fiber whether its electrical side has eight lanes or four; Section 4 explains how. The lane rate changes three other things:

• Passive copper reach. Higher lane rates lose more signal per meter of copper. The IEEE target is at least 3 m at 50G per lane and at least 2 m at 100G per lane (Section 5).
• Breakout options. An eight-lane port splits into four two-lane 100G ports. A four-lane port splits into four single-lane 100G ports (Section 8).
• Module compatibility. A module built for one lane rate does not work in a port running the other (Section 2.5).

2 400G Form Factors: QSFP-DD, QSFP112, OSFP and OSFP-RHS

A form factor is the mechanical and electrical package of a transceiver: its size, its electrical connector, the cage it plugs into on the switch or NIC, and how it is cooled. Four form factors carry 400G. QSFP-DD and OSFP each exist at both lane rates, so the form factor alone does not identify the module to order.

2.1 QSFP-DD

QSFP-DD (Quad Small Form-factor Pluggable, Double Density) extends the QSFP family from four electrical lanes to eight. It is the form factor of the 2019–2024 400G switch fleet, where it runs 8 × 50G. The same shell is used on 800G switch ports (QSFP-DD800), where the lanes run at 100G and the port can be split into 2 × 400G. QSFP-DD cages accept older QSFP28 and QSFP56 modules, which protects existing 100G and 200G optics during an upgrade.

2.2 QSFP112

QSFP112 is the four-lane member of the QSFP family, with each lane at 100G. At 400G it is always 4 × 100G. It is the NIC-side form factor: NVIDIA ConnectX-7 and BlueField-3 cards and the dual-port ConnectX-8 use it, as do 128-port 400G QSFP112 switches built on Tomahawk 5. It is the smallest 400G module and is cooled by the heat sink on the host cage.

2.3 OSFP and OSFP-RHS

OSFP (Octal Small Form-factor Pluggable) is a larger shell with eight electrical lanes, used from 400G to 1.6T. At 400G it exists at both lane rates: 8 × 50G on Tomahawk 3-era switches such as the Arista 7060PX4, and 100G lanes on NVIDIA's Quantum-2 and Spectrum-4 switches and on Tomahawk 5 systems.

OSFP comes in two mechanical versions. The finned version carries its own heat sink and is used in air-cooled switches. The flat-top version, called OSFP-RHS (Riding Heat Sink), has no heat sink of its own and is cooled by a heat sink on the host cage. OSFP-RHS is used on NICs (ConnectX-7, ConnectX-8 and ConnectX-9 OSFP cards), on GB200 NVL72 compute trays, and inside DGX H100, H200 and B200 servers. The two are not interchangeable: a finned module is too tall for a flat-top cage, and a flat-top module is not supported in a cage built for finned modules.

2.4 Form Factor and Lane Rate Combinations

QSFP112 is the only form factor with a fixed lane rate. A QSFP-DD or OSFP port can run either lane rate, and the module must match the port.

2.5 Mechanical Fit and Electrical Compatibility

Mechanical fit and electrical compatibility are two separate tests. A QSFP112 module fits a QSFP-DD cage, but a QSFP-DD port on a Tomahawk 4 switch runs 50G lanes, so a 4 × 100G module plugged into it fits mechanically and does not link. The port runs at the lane rate of its chip, not at the lane rate of the module. Three rules follow:

1. Identify the switch or NIC model and its port mode before choosing a module.
2. Check finned versus flat-top on every OSFP order. NVIDIA's ConnectX OSFP cards accept flat-top modules only. DGX H100, H200 and B200 internal cages accept twin-port flat-top modules, sold with an -FLT suffix.
3. Treat "400G OSFP" on a product listing as incomplete. For a Quantum-2 or SN5600 switch it means a twin-port finned module, which is an 800G-class part. For a ConnectX-7 card it means a single-port flat-top module. For an Arista 7060PX4 it means an 8 × 50G module.

3 400G Optical Types: VR4, SR4, DR4, FR4, LR4 and Beyond

The optical type defines what a transceiver puts on the fiber: the number of optical lanes, their wavelengths, the fiber type, the connector, the reach and the loss budget. IEEE 802.3 and the industry MSAs define more than fifteen 400G optical types; Figure 4 at the end of this section lists them. Five cover almost every link in a data center: VR4, SR4, DR4, FR4 and LR4.

3.1 Multimode Types: VR4, SR4, SR8 and SR4.2

Multimode types use VCSEL lasers at 850 nm (SR4.2 adds a second wavelength at 910 nm) on OM3, OM4 or OM5 multimode fiber. They are the lowest-cost optics but have the shortest reach and the tightest loss budgets.

• VR4 (IEEE 802.3db): four optical lanes on eight fibers, MPO-12 connector. 50 m on OM4, 30 m on OM3. Loss budget 1.7 dB.
• SR4 (IEEE 802.3db): same structure as VR4. 100 m on OM4, 60 m on OM3. Loss budget 1.8 dB.
• SR8 (IEEE 802.3cm): eight optical lanes on sixteen fibers, MPO-16 connector. Built for the 8 × 50G electrical interface. 100 m on OM4, 70 m on OM3.
• SR4.2 (IEEE 802.3cm): two wavelengths on each of eight fibers, MPO-12 connector. 100 m on OM4, 150 m on OM5. Used to reuse MPO-12 multimode cabling installed for 40G and 100G.

Read the reach on the datasheet, not the name. NVIDIA's 400G SR4 modules, and others sold under the SR4 name, are specified to 50 m on OM4, which is VR4 reach.

3.2 Parallel Single-Mode Types: DR4 and DR4-2

Parallel single-mode types send each optical lane on its own fiber at 1310 nm.

• DR4 (IEEE 802.3bs): four optical lanes at 100G on eight fibers, MPO-12 APC connector. 500 m. Loss budget 3.0 dB. DR4 is the main optical type for AI server links, and a DR4 port can be broken out into four 100G DR links.
• DR4-2 (IEEE 802.3df), also sold as DR4+: the same fiber layout with 2 km reach and a 4.0 dB loss budget.

3.3 Duplex Single-Mode Types: FR4 and LR4

Duplex types multiplex four wavelengths onto one fiber pair, so they run on standard duplex LC cabling.

• FR4 (IEEE 802.3cu): four CWDM wavelengths between 1271 and 1331 nm. 2 km. Loss budget 4.0 dB. Because the four lanes share one fiber pair, FR4 cannot be broken out.
• LR4-6 (IEEE 802.3cu): 6 km. Loss budget 6.3 dB.
• LR4-10 (100G Lambda MSA): 10 km. Loss budget 6.3 dB. At 10 km the fiber alone uses about 4.7 dB of that budget, so LR4-10 is comfortable at 6–8 km.

FR4 is the choice when only a fiber pair is available or the cabling is already duplex LC. DR4 is the choice when breakout matters or when the fiber should be reused for 800G parallel optics. Section 6 compares the fiber requirements.

3.4 Long-Reach and Coherent Types: ER4 and 400ZR

ER4-30 (100G Lambda MSA) reaches 30 km with a 15 dB loss budget and is the last conventional 400G optic. Beyond that, 400ZR (OIF) uses coherent transmission: up to 120 km on an amplified point-to-point link, or an unamplified link with an 11 dB budget, which corresponds to about 40 km. These types connect data centers to each other and are not used inside the AI fabric.

3.5 Linear Pluggable Optics: FR4-LPO

FR4-LPO (LPO MSA) is an FR4 module with the DSP chip removed; the host's SerDes performs the signal conditioning instead. It uses less power than a conventional FR4 but is specified to 500 m rather than 2 km, and it must be qualified for each host platform because the host does part of the work. Section 4.2 explains the role of the DSP.

3.6 Summary Table of 400G Optical Types

4 400G Transceiver Architecture: Optics, DSP and Power

A 400G transceiver has three functional blocks: the optical engine that converts electrical signals to light and back, the DSP that conditions the electrical signals, and the control circuitry that manages the module and reports its status.

4.1 The Optical Engine

On the transmit side, lasers and modulators turn four electrical lanes into four optical signals. DR4 modules use 1310 nm lasers: either four separate lasers or one laser feeding four silicon-photonics modulators. FR4 modules use four lasers at different wavelengths between 1271 and 1331 nm. VR4 and SR4 modules use four 850 nm VCSELs. On the receive side, four photodiodes convert the light back to current, and transimpedance amplifiers (TIAs) turn that current into electrical signals. In a DR4 module the four lanes connect to eight fibers of an MPO-12 connector. In an FR4 module a multiplexer combines the four wavelengths onto one duplex LC fiber pair.

4.2 The DSP: Retimer and Gearbox

The DSP (digital signal processor) cleans up and re-times the electrical signals between the host and the optical engine. In a 4 × 100G module (QSFP112, OSFP, OSFP-RHS) it works as a retimer: four lanes in, four lanes out. In an 8 × 50G QSFP-DD module it works as a gearbox: eight 50G lanes in, four 100G lanes out to the optics. The gearbox is the reason a QSFP-DD DR4 and a QSFP112 DR4 can connect to each other over the same fiber although their electrical interfaces differ.

Linear pluggable optics (LPO) remove the DSP and rely on the host's SerDes. This saves power and latency but makes the module dependent on the host, which is why LPO modules are qualified per platform.

4.3 Module Power and Cooling

Module power is listed on each datasheet as a maximum in watts. Current single-port 400G modules draw 8.5–10 W: a QSFP112 DR4 up to 10 W, an OSFP-RHS DR4 up to 9 W, a QSFP-DD DR4 up to 9 W, and VR4 modules 8.5–9 W. Longer-reach types draw up to about 12 W. Twin-port 800G-class modules, which carry two 400G links, draw 15–17 W.

Heat leaves a module through its heat sink. Finned OSFP modules carry their own and depend on the switch airflow over the fins. QSFP-DD, QSFP112 and OSFP-RHS modules have no heat sink of their own and depend on the heat sink built into the host cage. A module that runs hotter than its rated case temperature raises a temperature alarm and is outside specification, so the thermal rating of the port matters as much as the wattage of the module.

4.4 EEPROM and Management

Every module carries a small memory, the EEPROM, that stores its identity: vendor, part number, serial number, optical type, supported lane rates and other coded fields. The host reads this memory when the module is inserted. A host that enforces a list of accepted modules refuses, or warns about, a module whose coding it does not recognize, and a refused module will not carry traffic even if its optics are perfect. Modules also report live monitoring values, such as temperature, supply voltage, and transmit and receive power per lane, through the CMIS management interface. Section 10 covers how this affects compatibility testing.

5 Copper and Active Cables at 400G: DAC, ACC, AEC and AOC

For short links, a cable with the connectors built in replaces the pair of transceivers and the fiber between them. Four cable classes exist at 400G, distinguished by what is inside the plug.

5.1 Passive DAC

A passive direct-attach copper cable (DAC) contains no electronics; the host's signal travels straight through the copper. Reach depends on the lane rate and the wire gauge. On Vitex datasheets, QSFP-DD DACs with 50G lanes run to 2.5 m, using 28 AWG cable at 2 m and longer, and QSFP112 DACs with 100G lanes run to 3 m, using thicker 26 AWG cable at 2 m and longer. Passive DACs draw no power beyond the host port and add no retiming delay. They are used inside a rack, from a server to a top-of-rack switch.

5.2 ACC and AEC

Active copper cables add electronics in each plug to extend reach. An ACC (active copper cable) uses an analog equalizer and reaches 3–5 m, drawing about 1.5–3 W per cable. An AEC (active electrical cable) uses a DSP in each plug and reaches 6.5 m at 100G per lane on 26 AWG cable, per the Vitex 800G AEC datasheet. An AEC draws about 4.5 W per end at 8 × 50G and 7–12 W per end at the 800G class, and it is thinner than a passive cable of the same length. Because the plug electronics interact with the host, active copper cables are qualified per platform.

5.3 AOC

An active optical cable (AOC) has a short-reach optical transceiver permanently attached to each end of a multimode fiber. On Vitex datasheets, a 400G AOC with four 100G lanes reaches 30 m on OM3, and one with eight 50G lanes reaches 70 m on OM3 and 100 m on OM4; both draw under 10 W per end. The ends and the fiber cannot be separated, so an AOC is replaced as a unit. NVIDIA does not offer AOCs for its 100G-per-lane systems. Third-party 400G AOCs and 800G-to-2 × 400G breakout AOCs exist; on NVIDIA platforms they should be qualified on the exact platform and port mode and installed with strain relief so that the weight of the OSFP ends does not load the fiber.

5.4 Reach and Power Comparison

6 Fiber and Loss Budgets

Beyond the reach of a cable, a 400G link is two transceivers and the fiber between them. The fiber type, the connector type and the losses along the path decide whether a given optical type works on a given route.

6.1 Multimode and Single-Mode Fiber

Multimode fiber (OM3, OM4, OM5) has a wide core and works with low-cost 850 nm optics, but it supports 400G only to 100 m, or 150 m for SR4.2 on OM5. Single-mode fiber (OS2) has a narrow core and supports 400G from 500 m with DR4 to 120 km with 400ZR. At short distances either fiber can work. The decision then depends on the fiber already installed, the number of fibers available, whether breakout is needed, and whether the row may grow past 100 m. A site with OS2 trunks already in place can run single-mode DR4 even on 30 m links and avoid re-cabling when the row grows.

6.2 Parallel and Duplex Fiber: MPO-12 and LC

Parallel optics (VR4, SR4, DR4, DR4-2) use eight fibers, four to transmit and four to receive, on an MPO-12 connector. For single-mode types the connector is angle-polished (APC). For multimode types the polish varies by vendor and is read from the datasheet. Duplex optics (FR4, LR4, ER4, 400ZR) use one fiber pair on an LC connector. A parallel link needs four times the fiber of a duplex link but can be broken out into four 100G links. A duplex link cannot.

6.3 Loss Budget Calculations

A transceiver's loss budget is the maximum optical loss, in dB, that the link tolerates between transmitter and receiver. The fiber, each connector pair and each splice consume part of it. In the DR4, FR4 and SR4 rows below, the connectors consume more of the budget than the fiber; only in the 8 km LR4-10 row does the fiber take the larger share.

The planning values are 0.47 dB per km of single-mode fiber, 0.5 dB per connector pair for reference-grade connectors, 0.75 dB per connector pair at the TIA worst case, and 0.3 dB per splice. For a real deployment, use measured loss or the installer's guaranteed values.

Multimode budgets are the tightest: 1.7 dB for VR4 and 1.8 dB for SR4 leave room for about three connector pairs at full reach. A link whose total loss exceeds its budget is outside specification. It may still come up, but with no margin for connector aging, contamination or temperature.

6.4 Reusing 400G Fiber for 800G

Single-mode fiber installed in eight-fiber groups on MPO-12 APC connectors for 400G DR4 can be reused for 800G and 1.6T DR8 modules that have two MPO-12 receptacles: each 800G link uses two groups. It can also be reused for twin-port 800G modules, which are two 400G links. It cannot be reused unchanged for DR8 modules with a single MPO-16 receptacle; those need MPO-16 trunks or conversion harnesses. The coming 400G DR2 type uses four fibers on 200G lanes.

Two planning rules follow. Check the connector of the next optic before promising reuse. Install spare eight-fiber paths, because an 800G DR8 link consumes twice the fiber of a 400G DR4 link.

7 400G in the AI Data Center

Inside a server, or across a whole GB200 NVL72 rack, GPUs connect to each other over NVLink on copper, with no pluggable optics. Engineers call this the scale-up side. Between servers, on the scale-out side, a cluster runs four separate networks, and 400G links are used in three of them. Speed belongs to each link, not to the cluster: one cluster often runs 400G, 800G and 1.6T links at the same time.

7.1 The Four Networks in an AI Cluster

7.2 Rails, Planes and Scalable Units

Three terms describe how servers connect to the compute fabric.

A rail is the set of NICs with the same slot number on every server, all cabled to the same leaf switch. A server with eight NICs connects to eight leaf switches, one per rail. This layout, called rail-optimized, keeps traffic between matching GPUs on different servers one switch hop apart. Because each leaf switch serves one rail across many servers, leaf switches sit in network racks rather than at the top of each server rack. Most server links therefore cross racks and run on optics.

A plane is a complete, independent copy of the fabric. NVIDIA's HGX B300 reference design recommends two planes: each ConnectX-8 NIC is configured as two 400G ports, and each port connects to a leaf switch in a different plane. The single-plane alternative uses one 400G port per NIC at half the bandwidth.

A scalable unit (SU) is the repeating block of servers that shares one set of rail leaf switches; an H100 SuperPOD SU has 32 servers. Spine switches join the SUs. A fabric with as much uplink as downlink bandwidth is non-blocking; one with less uplink is oversubscribed.

Rails and planes are the reason four counts differ in a cluster: GPUs, NICs, cages and links. Section 9 works through them.

7.3 Typical Link Speeds by GPU Generation

One cluster runs several speeds at once. A B300 Ethernet cluster, for example, runs 400G links from the servers and to storage, and 800G links from leaf to spine.

7.4 The Transition to 800G and 1.6T

Each tier changes speed on its own schedule. The leaf-to-spine tier moves first, to 800G and then 1.6T. The GPU server link follows about one generation later. Storage and front-end links carry less bandwidth per server and move last. That is why operators still buy 400G modules for new clusters after the GPU links have moved on.

7.5 Co-Packaged Optics

Co-packaged optics (CPO) builds the optical engines into the switch chip package instead of into pluggable modules. CPO switches exist at 400G: switches built on Broadcom's Tomahawk 5 "Bailly" part present 128 ports of 400G FR4 on duplex LC connectors. The server end of each link is unchanged; it still needs a pluggable transceiver of the same optical type. CPO changes who supplies the switch-side optics, not whether the server needs a module.

8 800G to 2 × 400G Breakout: Modules and Cables

Breakout is the configuration of one switch port as two or more lower-speed ports. In NVIDIA's Quantum-2 and Spectrum-4 fabrics, and on 51.2 Tb/s Ethernet switches, each switch cage carries two 400G links, so one cage is shared by two servers: the 800G port is configured as two 400G ports. Three types of hardware do this. In all three, the switch port must be configured for 2 × 400G.

8.1 Twin-Port Modules

A twin-port module is an 800G-class OSFP module with two independent 400G optical engines and two MPO-12 APC connectors in one shell. Each engine runs its own eight-fiber cable to a separate server or switch. NVIDIA's Quantum-2 and Spectrum-4 SN5600 switches use twin-port cages: a 1U QM9700 presents 64 NDR 400G ports from 32 cages, and a 2U SN5600 has 64 cages that run as 64 × 800GbE or 128 × 400GbE. The two engines are independent links, but they share one module, so replacing the module takes down both links. Single-port 400G modules cannot be used in twin-port switch cages.

8.2 800G DR8 Modules in 2 × 400G Mode

An 800G DR8 module has eight optical lanes. With the switch port set to 2 × 400G, lanes 1 to 4 form one 400G link and lanes 5 to 8 form another. DR8 modules with two MPO-12 receptacles connect each half directly to a separate 400G DR4 module. DR8 modules with a single MPO-16 receptacle need an MPO-16 to 2 × MPO-12 harness, which adds one connector pair to the loss budget. Cisco's HGX B300 reference design uses this method at both ends: a flat-top DR8 with two MPO-12 cables in each ConnectX-8 cage, with one 400G link to each plane.

8.3 Splitter Cables

A splitter cable has one 800G OSFP plug at the switch end and two 400G plugs, QSFP112 or OSFP, at the server end. It is available as a passive DAC to 2.5 m, as an ACC to 5 m, and as an AOC to 30 m. Copper splitters serve servers in the same rack as the switch. AOC splitters need the same per-platform qualification as any AOC.

8.4 Module Selection by Server Type

A DGX H100 does not have eight single-port NIC cages. Its eight ConnectX-7 NICs connect through four internal cages that accept only twin-port flat-top modules, and each cage carries two 400G links. Single-port OSFP-RHS and QSFP112 modules are used on PCIe cards, on HGX servers from other vendors, and on GB200 trays.

In a dual-plane design, both 400G links of a GPU leave through the same NIC and the same module. Two planes protect against a switch or cable failure, not against a module failure. Spares and maintenance are planned per module.

8.5 Other Breakout Patterns

Breakout is a separate test case. A module tested in native 400G mode has not been tested in 4 × 100G mode, and a twin-port module tested switch-to-switch has not been tested feeding two servers.

9 400G Fabric Bill of Materials and Power Budget

A bill of materials for a 400G fabric starts with the number of links, not the number of ports. From the links follow the modules, the optical engines, the fiber paths and the power.

9.1 Links, Modules, Engines and Fiber Paths

Four quantities are counted separately:

• Links: one 400G connection between two endpoints.
• Modules: one at each end of every link. A twin-port module is one module carrying two links.
• Optical engines: one four-lane engine per 400G link end. A twin-port module contains two.
• Fiber paths: one eight-fiber MPO-12 path per DR4 or VR4 link.

A cage is not a link. One twin-port cage carries two links, and one DGX H100 carries eight links through four cages.

9.2 Example: One H100 SuperPOD Scalable Unit

NVIDIA's H100 SuperPOD reference design uses a rail-optimized fat tree on Quantum-2 switches, with eight NDR 400G connections per server: 32 DGX H100 servers, eight QM9700 leaf switches and four spine switches. As built, one server slot per unit is used by the UFM management node, which gives 31 DGX servers, 252 server-and-UFM cables and 256 leaf-to-spine cables. As a 32-server planning model, one unit requires:

Storage and management networks are additional. With HGX servers in place of DGX, the server side becomes 256 single-port 400G OSFP-RHS modules facing 128 twin-port leaf modules.

9.3 Example: 32 HGX B300 Servers, Dual-Plane Ethernet

Cisco's HGX B300 reference design has eight ConnectX-8 NICs per server, each configured as 2 × 400G. The 32 servers produce 512 server-to-fabric links, 256 per plane, with four leaf switches and two spine switches per plane. At each NIC the design specifies a flat-top 800G DR8 module with two MPO-12 cables, and the same module type at the switch: 256 NIC-side modules and 256 leaf-side modules on the server-facing ports. The storage and front-end network adds 64 BlueField-3 ports at 400G on QSFP112 DR4. Uplink counts depend on the leaf-to-spine speed and the oversubscription ratio chosen.

9.4 Power per Link

Module power and optical power are different quantities. Module power is the electrical load at each cage, in watts, and drives the cooling budget. Optical power is the light level at the transmitter and receiver, in dBm, and decides whether the link closes against its loss budget (Section 6.3). This section covers module power.

Power per link is counted at both ends. The table uses datasheet maximums and splits a two-link module evenly between its two links.

At cluster scale the difference matters. The 256 server links of the HGX variant in Section 9.2 draw about 4.2 kW in optics at datasheet maximum: 256 NIC modules at 9 W plus 128 twin-port leaf modules at 15 W. Switch chips, fans and power conversion are additional. A lower module wattage does not by itself prove a lower cluster power bill, and every module must also stay within the thermal limit of its cage.

10 Ordering, Compatibility Testing and Troubleshooting

This section covers the information a supplier needs to deliver the right 400G part, what compatibility testing means, and the order in which to check a link that does not come up.

10.1 The Order Specification

A complete 400G order specification contains five items:

1. The switch model and the NIC or DPU model at the other end, with their firmware or software versions.
2. The protocol: InfiniBand or Ethernet.
3. The port mode at each end: for example 1 × 400G, 2 × 400G on an 800G port, or 4 × 100G breakout.
4. The route: the cable length as it will run, the fiber type and count if fiber exists, the connector polish (APC or UPC), and the number of connector pairs in the path.
5. Quantity, spares, and the date the links must be live.

With these five items a supplier can name the exact part and the qualification plan. Without them, "400G OSFP" or "400G DR4" leaves two or three possible parts.

10.2 Compatibility Testing

Compatible means two things. The module meets the optical and electrical standard for its type, and its EEPROM coding (Section 4.4) is accepted by the host platform. Both are tested on the actual platform, software release and port mode. A test record for native 400G mode does not cover 2 × 400G or 4 × 100G breakout modes, and a twin-port module tested switch-to-switch has not been tested feeding two servers.

Switch vendors support their own switches; support for a transceiver they did not sell is at their discretion, and on a fault they may ask for it to be reproduced with their own module. A test record for the exact platform, software version and port mode answers that request. Ask the supplier for it before ordering.

10.3 Troubleshooting a Link That Does Not Come Up

Check in this order. The first five checks need nothing beyond the switch command line and a visual inspection.

1. Cage: correct form factor and lane rate for the port (Section 2).
2. EEPROM: the switch recognizes the module and shows its monitoring values.
3. Port mode: lane count and forward error correction (FEC) match the module.
4. Far end: the module at the other end is the same optical type. On a twin-port link, confirm which server the fiber reaches.
5. Fiber: polarity, connector polish and clean end faces.
6. Optical power: transmit and receive power per lane compared with the datasheet.

A link that comes up but carries traffic slowly is a different problem. Check pre-FEC error rates and per-lane counters first, then congestion settings and fabric configuration, then the oversubscription ratio of the fabric.

10.4 Availability and Lead Times

400G has more suppliers than 800G or 1.6T, but single-mode 400G modules use the same lasers that are in short supply at the higher speeds. Lead time depends on the supplier's stock position more than on the standard. Ask for the delivery date and the test record together, before the design freezes.

+ Appendix: Reference Tables and Sources

A. Switch and NIC Platforms by Lane Rate

B. Standards Status (October 2026)

C. Related Vitex Guides

• 1.6T optical for AI data centers
• 800G OSFP: finned (IHS) vs flat-top (RHS) selection
• 800G transceiver compatibility for NVIDIA platforms
• 400G DR4 to 4×100G breakout planning
• 800G structured cabling: MPO-16, polarity, trunks and link budget
• LPO vs DSP optics
• InfiniBand vs Ethernet for AI clusters
• H100 SuperPOD optics: a 256-GPU case study
• Optics validation and testing handbook
• Co-packaged optics (CPO) guide

D. Sources

IEEE 802.3 amendments bs, cn, cm, cu, db, ck and df, and the IEEE P802.3dj task force record. TIA Fiber Optics Tech Consortium per-type reach and loss pages. The OSFP, QSFP-DD, 100G Lambda and LPO MSAs. OIF 400ZR and CMIS. NVIDIA: the ConnectX-7, ConnectX-8 and ConnectX-9 user manuals; the twin-port 2×DR4 transceiver page; the 100G-PAM4 DAC and LACC overview; the HGX AI Factory reference architecture; the DGX SuperPOD H100 reference architecture; the Rubin platform overview. Cisco: the HGX B300 enterprise reference design. Arista, Cisco, Juniper and Dell switch datasheets. Fluke Networks, Corning, CommScope and Cisco cabling guides. Credo AEC product briefs. Vitex product datasheets for module and cable power and reach figures. Market figures: Dell'Oro Group (December 2025) and Cignal AI (2026).