Understanding the 400g Optical Module

 

Why 400G Happened When It Did

 

There's a certain inevitability to bandwidth progression that industry veterans have come to expect. The jump from 10G to 40G felt significant at the time. Then 100G arrived and suddenly everyone was talking about spine-leaf architectures and east-west traffic patterns. But 400G? That transition has been different.

The NRZ modulation scheme that served us well from 1G through 25G simply couldn't scale economically beyond 100G. The physics got expensive. Signal integrity became a nightmare. You could technically push NRZ harder, sure-but the cost curves made no sense for volume deployment. So the industry pivoted to PAM4.

What PAM4 does-and this is worth understanding if you're spec'ing infrastructure-is encode two bits per symbol instead of one. Four amplitude levels instead of two. Twice the data throughput without doubling the baud rate. The tradeoff? Your signal-to-noise ratio takes about a 10dB hit compared to NRZ. That's not trivial. It's why every 400G module ships with forward error correction baked in, and why the DSP (digital signal processor) has become such a critical component in these transceivers.

 

 

The Form Factor Wars

 

I've watched the QSFP-DD versus OSFP debate play out at trade shows and in procurement meetings for years now. Both sides have valid arguments. Neither has definitively won.

QSFP-DD came out of the QSFP-DD MSA alliance with backward compatibility as its killer feature. Got a bunch of QSFP28 modules you're not ready to replace? They'll slot right into a QSFP-DD cage. The dimensions-18.35mm width, 89.4mm length-keep the form factor familiar. You can fit 36 ports in a 1U front panel. That's 14.4Tbps of aggregate throughput if you populate every slot. For operators upgrading incrementally, this matters enormously.

OSFP took a different approach. The Octal Small Form-factor Pluggable group said: forget backward compatibility, let's optimize for thermal management and future scalability. At 22.58mm wide and 107.8mm long, OSFP modules have more surface area for heat dissipation. They support power envelopes up to 15-20 watts versus QSFP-DD's 12-15W ceiling. When you're running coherent optics or planning for 800G, that headroom becomes relevant.

NVIDIA went all-in on OSFP for their Quantum-2 InfiniBand platform. That's not nothing. But enterprise switches from Cisco and Arista still predominantly ship QSFP-DD ports.

 

What the Letters Actually Mean

 

If you've ever stared at a spec sheet wondering why DR4 costs less than FR4 despite both being "400G modules," you're not alone. The nomenclature follows patterns, but those patterns have exceptions and the exceptions have their own logic.

• SR (Short Range): Multimode fiber, 850nm wavelength. The 400G-SR8 uses eight parallel lanes of 50G PAM4 over an MPO-16 connector. Range tops out around 100 meters on OM4 fiber-70 meters on OM3. There's also SR4, which packs 100G into each of four lanes using higher-speed VCSELs. Same reach, fewer fibers. The 400G-SR4.2 variant (sometimes called BIDI) gets clever with bidirectional transmission, running two wavelengths in each direction to achieve 400G over just four fibers.
• DR4: Single-mode fiber, 1310nm, 500 meters. This is the workhorse for intra-datacenter connections longer than SR can handle. Each of the four optical lanes carries 100G PAM4 over a dedicated fiber pair. MPO-12 connector. The breakout capability here is significant-one DR4 can split into four independent 100G-DR links, which helps when connecting legacy 100G equipment.
• FR4: Two kilometers, single-mode. Here's where the gearbox architecture earns its keep. The module takes eight 50G electrical lanes, converts them to four 100G optical lanes via DSP, then wavelength-multiplexes all four onto a single fiber pair using CWDM4 spacing (1271, 1291, 1311, 1331nm). Duplex LC connector. Much tidier cabling than DR4's parallel approach.
• LR4 and beyond: Same wavelength scheme as FR4, but optimized for 10km reach. ER4 pushes to 40km. ZR4 hits 80km but requires coherent detection-different technology entirely, different price point, different use case. The 400ZR standard from OIF specifically targets metro DCI applications where you need pluggable coherent optics in a switch faceplate.

 

 

The DSP Question

 

Every 400G transceiver contains a digital signal processor. Every single one. This isn't optional-PAM4 modulation simply doesn't work without sophisticated signal conditioning.

What does the DSP actually do? Feed-forward equalization to compensate for channel loss. Decision feedback equalization for inter-symbol interference. Clock and data recovery to extract timing from the received signal. FEC encoding on transmit, FEC decoding and error correction on receive. In coherent modules, add chromatic dispersion compensation and polarization mode dispersion management to that list.

The DSP burns power. A lot of it. In many 400G modules, the DSP accounts for more than half the total power consumption. Marvell, Broadcom, and Inphi (now part of Marvell) have been locked in competition to shrink process nodes and improve efficiency. The jump from 7nm to 5nm DSPs has been meaningful-power savings on the order of 20% for equivalent functionality.

There's ongoing debate about whether DSPs should move into the switch ASIC itself (what some call "linear pluggable optics" or LPO). The argument goes: if you're already doing signal processing on the switch, why replicate it in every transceiver? The counter-argument involves module interoperability and the practical challenges of qualifying optics across different switch platforms. This one's going to play out over the next few years.

 

Silicon Photonics Enters the Picture

 

Remember when everyone assumed InP (indium phosphide) lasers would dominate 400G? The narrative shifted.

Intel and Cisco bet early on silicon photonics-integrating optical components onto silicon substrates using standard CMOS fabrication processes. The promise was always about cost at scale. Traditional discrete optics require manual assembly of laser chips, modulators, photodetectors, each from different materials. Silicon photonics lets you build much of the optical engine on a single die.

The 400G-DR4 silicon photonics modules shipping today offer compelling economics for hyperscale deployments. They're not universally cheaper than EML-based alternatives-yet-but the cost trajectory favors silicon as fab yields improve. Power consumption benefits too, particularly in the modulator section.

That said, silicon makes a mediocre laser. The indirect bandgap problem hasn't been solved. So even silicon photonics modules typically use an external InP or GaAs gain chip, hybrid-integrated onto the silicon platform. It's elegant engineering, but "silicon photonics" remains somewhat aspirational terminology.

 

Power and Thermal Realities

 

Air cooling is reaching its practical limits for these densities. Liquid cooling-cold plates on switch ASICs, potentially immersion for entire racks-is no longer exotic. It's just expensive infrastructure that facilities organizations are still learning to specify and maintain.

 

Breakout Scenarios

 

One capability that doesn't get enough attention: 400G modules can often be configured for breakout operation, presenting as multiple lower-rate interfaces to the host system.

A 400G-SR8 can become two 200G-SR4 links, or two 100G-SR4 links running at half rate, or even eight independent 50G channels (the "channelized" or SR8-C variant). A 400G-DR4 can break out to four 100G-DR connections-useful when you need to connect a 400G switch port to four separate 100G servers.

The cabling gets interesting here. An MPO-12 to 4xLC duplex breakout harness takes a single DR4 port and fans it out to four independent SMF pairs. Network architects love this flexibility, but the cable management implications are real. Your structured cabling plan needs to account for breakout scenarios from day one, or you'll be running ad-hoc patch cables six months after deployment.

 

What 800G Means for 400G

 

The industry moves fast. 800G transceivers are shipping now-mostly SR8 and DR8 variants for AI cluster interconnects. Does that make 400G obsolete? Not even close.

The 400G ecosystem has matured. Module costs have dropped substantially. Interoperability across vendors is well-established. For the majority of enterprise and cloud networking needs, 400G represents the sweet spot of performance, cost, and operational familiarity. It'll remain the volume play for leaf-spine fabrics and general-purpose datacenter connectivity for years.

800G and eventually 1.6T will dominate in AI/ML environments where GPUs need to move training data at absurd rates. Different market, different requirements, different budget conversations. Most networks won't need to chase that curve.

 

 

Practical Considerations Nobody Writes About

 

Some things learned the hard way:

Module EEPROM compatibility matters more than vendors admit. "Compatible" transceivers that work fine in one switch model may throw errors in another with the same ASIC but different firmware. Build in test time when qualifying third-party optics.

The LC connector on FR4 and LR4 modules is duplex-two fibers total-but the MPO connector on DR4 and SR8 uses APC (angled physical contact) polish. Mixing APC and UPC connectors will get you 20dB+ return loss and intermittent errors. Color coding exists for a reason.

PAM4 modules from different manufacturers can have subtly different FEC implementations. The standards leave room for interpretation. If you're seeing inexplicably high corrected error counts on a link, try swapping one end for a same-vendor module before blaming the fiber plant.

Temperature matters. These modules are rated for case temperatures up to 70°C typically, but performance degrades before you hit that ceiling. Keep them cool if you want consistent behavior.

 

The Road Forward

 

400G optical modules have transitioned from leading edge to mainstream infrastructure. The technology decisions-QSFP-DD versus OSFP, parallel versus WDM, silicon versus discrete optics-no longer carry the same risk they did three years ago. Robust supply chains exist. Multiple qualified vendors compete on price and features. Standards bodies have ironed out most interoperability edge cases.

For network architects planning deployments today, the choice framework is straightforward: match the form factor to your switch platform strategy, select the transceiver type (SR/DR/FR/LR) based on actual reach requirements, and don't over-spec. A 400G-LR4 costs substantially more than a 400G-DR4-if your longest runs are 300 meters, you're burning budget for no operational benefit.

The next few years will bring incremental improvements: lower power DSPs, better silicon photonics yields, perhaps some standardization around linear pluggable architectures. But the fundamental technology platform has stabilized. 400G is no longer emerging. It's just infrastructure now-the kind you can plan around with confidence.

And honestly? After the chaos of the early 100G era, that predictability is worth appreciating.