Electronics Guide

What a Pluggable Module Actually Is

A pluggable transceiver is a contract as much as a component. The multi-source agreement that defines a form factor fixes the mechanical outline, the connector pinout, the insertion and retention forces, the electrical lane count and signaling, a management interface, and a set of power classes that the host cage must be prepared to dissipate. Any module that honors the contract will mate with any compliant host. The specification deliberately says nothing about what happens between the electrical connector and the optical connector, which is why one cage can accept a hundred-meter multimode module, a two-kilometer single-mode module, a coherent module reaching eighty kilometers, and a passive copper cable with no optics at all.

Inside a conventional module the block diagram is stable across generations. On the transmit side, the host's electrical lanes enter the connector, pass to a digital signal processor that recovers timing and equalizes the host channel, then to a driver that swings a laser or a modulator. On the receive side, a photodiode feeds a transimpedance amplifier, whose output the same processor equalizes and retimes before driving the electrical lanes back toward the host. Around that path sit a microcontroller, a nonvolatile memory holding identification and calibration data, temperature and optical power monitors, and control loops that hold the laser bias and modulation current steady as the device ages and heats. The monitors are exposed to the host as digital diagnostics, and they are the first thing an operator reads when a link misbehaves.

The digital signal processor deserves particular attention because it is the component that linear drive removes. Its function is to make the module independent of the host. Because it slices the incoming symbols and regenerates them, impairments accumulated on the host board do not propagate into the optical transmitter, and impairments accumulated on the fiber do not propagate onto the host board. Compliance can therefore be defined separately at the electrical interface and at the optical interface, and a module qualified against the electrical specification will work with any host qualified against the same specification. That separation is what makes multi-vendor interoperability tractable, and it is exactly what linear drive gives up.

The Form Factor Family

Form factors evolve by adding lanes and, reluctantly, by adding volume for heat. The small form-factor pluggable family began with a single lane and a power budget on the order of one watt, suitable for gigabit and ten-gigabit links. The quad small form-factor pluggable added a second dimension, carrying four lanes in a body only modestly larger, and became the workhorse of forty- and hundred-gigabit networking as lane rates climbed from ten to twenty-five gigabits per second.

The current generation doubles lane count again. QSFP-DD, where the suffix denotes double density, carries eight electrical lanes in a body that remains backward compatible with the four-lane quad form factor: a legacy module drops into a QSFP-DD cage and uses the first row of contacts. That compatibility was a decisive commercial advantage during the transition, because it let operators reuse cages and cabling. OSFP is the competing eight-lane definition, and it made the opposite trade. It is physically larger and abandons backward compatibility, but the extra volume buys thermal headroom, and the heat sink is part of the module itself rather than a riding heat sink attached to the host cage, which is the arrangement QSFP-DD uses. An OSFP-RHS variant exists for hosts that prefer the cage-mounted heat sink; it keeps the same electrical definition and omits the integrated one. As module power climbed through the 800 Gb/s generation, that headroom mattered, and OSFP became common on high-power switch and accelerator front panels while QSFP-DD retained a strong position where power is lower. Single-lane and double-lane form factors persist for access and aggregation.

The two families reached 1.6 Tb/s by different routes, and the difference is instructive. QSFP-DD kept its eight lanes and doubled the lane rate: the QSFP-DD1600 variant, introduced with hardware specification revision 7.0 in 2023, carries eight lanes at 200 Gb/s in the same mechanical envelope, raises the contact current rating to widen the module's power headroom, and stays backward compatible with the whole QSFP and QSFP-DD family. OSFP added lanes instead. OSFP-XD carries sixteen electrical lanes, roughly doubling the bandwidth density of the faceplate, and reaches 1.6 Tb/s at 100 Gb/s per lane and 3.2 Tb/s at 200 Gb/s per lane. A host designer therefore chooses between a denser connector and a faster lane, and the choice propagates back into the serializer-deserializer, the package, and the board.

Two neighbors share the same cages and deserve mention because they compete for the same ports. Direct-attach copper cables terminate in a module shell containing no active circuitry and reach a meter or two. Active electrical cables place equalizing or retiming silicon in the shell to stretch that reach to several meters. Both present themselves to the host through the same management interface as an optical module, and the choice among passive copper, active copper, and optics is made per link on the basis of reach, power, and cost rather than by any change to the chassis.

Reach classes fill out the matrix. Short-reach modules use vertical-cavity surface-emitting lasers on multimode fiber for roughly a hundred meters. Parallel single-mode modules reach several hundred meters over one fiber per lane, terminating in a multi-fiber push-on connector. Wavelength-multiplexed modules combine several wavelengths on a duplex pair for two to ten kilometers. Beyond that, coherent pluggables built to the industry's ZR implementation agreements carry a full coherent digital signal processor and reach eighty kilometers or more over amplified links, at power levels near the top of what any pluggable cage can dissipate.

Power, Heat, and the Faceplate

Thermal design is the binding constraint on pluggable optics, and it operates at two levels. At the module level, the specification defines power classes, and a compliant host must either supply and remove the power a given class demands or refuse to bring the module fully up. The module announces its class through the management interface during initialization, and the host grants or withholds high-power operation. This negotiation exists because a cage designed for a modest airflow cannot safely energize a module that dissipates twice what the cage was built for.

At the system level the arithmetic is stark. A one-rack-unit switch presenting thirty-two eight-lane ports at 800 Gb/s per port carries 25.6 Tb/s across its faceplate; a two-rack-unit design with sixty-four such ports reaches 51.2 Tb/s. Vendor datasheets for retimed 800 Gb/s modules quote figures in the region of eleven to sixteen watts, depending on the reach class and on the process node of the module's processor. Sixty-four of them therefore draw somewhere between seven hundred watts and a kilowatt, which is of the same order as the switching ASIC behind them and frequently exceeds it. That power must leave through a faceplate already perforated by sixty-four optical connectors and their cages, in a duct shared with the ASIC's own heat sink. Front-panel area, not silicon area, has become the scarce resource.

Heat also degrades the optics it passes through. Laser threshold current rises and slope efficiency falls with temperature, so a hot transmitter delivers less optical modulation amplitude for the same drive. Resonant devices such as ring modulators shift their resonance wavelength with temperature and require active tuning, which itself consumes power. Wavelength-multiplexed transmitters must hold each channel within its grid slot. Every one of these effects means that a thermally marginal port does not fail cleanly; it drifts, raising the pre-correction error rate until the forward error correction can no longer keep up, often intermittently and often only under load.

Management and the Host Interface

The host talks to a module over a low-speed two-wire serial bus into a paged memory map. The Common Management Interface Specification, developed alongside the double-density form factors and now maintained by the Optical Internetworking Forum, defines that map and the state machine behind it: how a module identifies itself, advertises its capabilities and power class, is commanded through low-power and high-power states, applies a datapath configuration to each lane group, and reports faults. The specification continues to move: current work adds an optional I3C two-wire interface alongside the original one and the management support that 200 Gb/s per lane Ethernet requires, including in-band link training.

Digital diagnostic monitoring rides on the same interface and is the practical instrument of optical troubleshooting. A module continuously reports its case temperature, supply voltage, laser bias current, transmitted optical power, and received optical power per lane, with configurable warning and alarm thresholds. Because these values are read without disturbing traffic, they turn many field problems into arithmetic. A received power far below the link budget points at the fiber plant or a dirty connector. A received power within budget alongside a high pre-correction error rate points at the transmitter, the receiver, or the electrical path rather than at the fiber.

The management interface also carries the vendor identification fields that some equipment makers use to restrict which modules a host will accept. That practice is a commercial rather than a technical matter, but it shapes procurement, sparing, and the economics of second sourcing, and it belongs in any honest account of what a pluggable module is.

Why the Module DSP Became the Problem

The digital signal processor inside a module exists because of the electrical channel in front of it. Between the host serializer and the module's connector lie package routing, several inches of printed circuit board, a connector, and the cage. At 100 Gb/s per lane with four-level pulse amplitude modulation, the symbol rate is about 53 GBd and the Nyquist frequency sits near 27 GHz. A few inches of printed circuit board are not transparent at that frequency: the chip-to-module specifications written for such channels budget insertion losses in the double-digit decibels at Nyquist, and the Optical Internetworking Forum's 112 Gb/s half-retimed agreement, for one published example, is written to channels of up to sixteen decibels. At 200 Gb/s per lane the symbol rate doubles to roughly 106 GBd and the Nyquist frequency to roughly 53 GHz, where the same physical geometry costs far more. The processor's first job is to undo that damage.

Undoing it is expensive. A module processor implements feed-forward and decision-feedback equalization, clock and data recovery, and transmitter pre-compensation on every lane, in both directions, at line rate. Eight lanes of that work in a leading-edge process node consume a large share of the module's power budget. The processor also adds latency in each direction, which matters in AI training fabrics, whose collective operations wait on the slowest link, and it is a complex integrated circuit in a hot, small, mechanically abused package, which makes it a leading contributor to module failure rates.

The awkward part is that the work is largely duplicated. The host ASIC already contains a sophisticated serializer-deserializer with its own equalization, built to drive channels considerably worse than the path to a nearby faceplate. In a retimed link, that host equalizer corrects the board, and then a second full equalizer inside the module corrects essentially the same channel again. Removing the second one is the entire idea behind linear drive.

Linear Pluggable Optics

A linear pluggable optics module contains no digital signal processor. On the transmit side, the electrical lane runs from the connector through a linear driver directly to the modulator or laser. On the receive side, the photodiode current passes through a linear transimpedance amplifier and out to the host connector. Nothing in the module makes a decision about a symbol; the module is an analog stage that converts between the electrical and optical domains and does nothing else. The host serializer-deserializer must therefore equalize a composite channel that now includes the host board, the connector, the module's internal electrical path, the driver, the optical device, the fiber, the photodiode, and the transimpedance amplifier.

The gains follow directly, though their size deserves care. Module makers publishing datasheets for 800 Gb/s parallel single-mode parts quote linear modules at roughly seven to eight watts against roughly eleven to fourteen watts for retimed modules of the same reach class. The widely repeated headline of a fifty percent saving sits at the optimistic end of that comparison, and every figure of the kind is a supplier's number for its own product under its own conditions rather than an independent measurement. The defensible statement is that linear drive removes several watts per module, enough to move an 800 Gb/s module into a range where the cage's thermal design stops being the limiting factor. Latency falls by the processor's contribution in each direction. Bill-of-materials cost falls because the most expensive silicon in the module is gone. Reliability tends to improve for the same reason.

The costs are structural rather than incidental. The clean separation between host compliance and module compliance disappears, because there is no regeneration point at which to define it. Link performance becomes a property of the specific pairing of host and module, so qualification turns into a matrix exercise across every combination an operator intends to deploy, and a module that performs well with one switch may not close the link with another. The host's forward error correction, ordinarily asked to repair only the optical span, must now absorb errors from the whole analog chain, which consumes budget that a retimed design reserved elsewhere. Host channel requirements tighten considerably, since loss that the module processor would have absorbed is now loss the host equalizer must pay for. Automatic gain control, transmitter output swing, and the linearity of the transimpedance amplifier become interface parameters rather than internal implementation details.

Those constraints explain where linear drive is being deployed. It suits short-reach classes, principally parallel single-mode links of a few hundred meters and wavelength-multiplexed links to about two kilometers, where the optical impairments are modest and the analog budget can accommodate them. It suits operators who own both ends of the link and the switch design as well, which describes hyperscale data center builders. It does not suit long-reach or coherent applications, where dispersion compensation and carrier recovery require digital processing that no host serializer performs, and it fits poorly in multi-vendor environments where a module must work with equipment the buyer does not control.

Linear Receive Optics and the Middle Ground

Removing the processor entirely is not the only option, and the industry has converged on a compromise. Linear receive optics, also called a half-retimed module and standardized by the Optical Internetworking Forum under the name retimed transmitter, linear receiver, keeps a digital signal processor on the transmit path while running the receive path linearly. The reasoning is asymmetric in a useful way. On the transmit side, digital pre-compensation is worth its power because it can be shaped to the specific bandwidth limitations and nonlinearity of the particular laser or modulator in that module, which no host equalizer can know about. On the receive side, the host's own equalizer is well suited to correcting a channel it already characterizes, and the receive path is where a large share of the processor's power is spent.

The result recovers a meaningful fraction of the power saving while restoring a regeneration point in the direction that is hardest to specify, which eases interoperability. Because the transmitter remains under module control, its optical output stays within a conventional optical specification, and the host is spared responsibility for the optical device's idiosyncrasies. This middle position attracted substantial support during the move to 200 Gb/s lanes and 1.6 Tb/s modules, where a fully linear path is harder to close. It is also the first of the two linear architectures to acquire a published interface specification of its own: the Optical Internetworking Forum released implementation agreement OIF-EEI-112G-RTLR in November 2025, covering the chip-to-module electrical interface, the optical parameters, and the test methods for a 100 Gb/s per lane link whose transmitter is retimed and whose receiver is not. A companion project at 200 Gb/s per lane is under way.

Taken together, the architectures form a ladder rather than a set of rivals. Passive copper is the cheapest and lowest-power option where reach permits. Active copper extends it. Fully retimed pluggable optics maximizes interoperability and reach. Linear receive optics trades a little of that interoperability for power. Fully linear pluggable optics trades more. Near-package and co-packaged optics abandon the pluggable form factor altogether in exchange for the largest efficiency gain and the loss of field replaceability. A single large system commonly contains several rungs of this ladder at once, chosen link by link.

Standards and Interoperability

Standardization of linear drive has proceeded differently from standardization of conventional optics, because the thing to be standardized is a relationship rather than an interface. IEEE 802.3 specifies physical medium dependent sublayers at the optical boundary and says little about a module's internals, so a linear module can in principle meet an existing optical specification while behaving very differently at the electrical connector. The Optical Internetworking Forum's common electrical interface projects define those very short reach electrical lanes, and their host and module output specifications, compliance boards, and reference receivers are where linear drive has the most to say.

The LPO MSA, a multi-source agreement group dedicated to linear pluggable optics, launched in March 2024 with twelve founding companies drawn from switch and interface silicon, module manufacture, network equipment, and accelerator platforms, and passed fifty members within a year. Its purpose is precisely the problem identified above: to define the host and module electrical parameters, the optical specifications, and the test methods needed for modules and hosts from different suppliers to be paired with confidence rather than by exhaustive testing. The group demonstrated multi-vendor interoperability over 400 and 800 Gb/s parallel single-mode links in September 2024, published a 100 Gb/s per lane linear single-mode specification covering links up to 800 Gigabit Ethernet in March 2025, and added a 400G-FR4-LPO specification for four-wavelength links in September 2025.

Work at the Optical Internetworking Forum has run in parallel along two tracks. The half-retimed case reached publication first, as the 112 Gb/s implementation agreement noted above. The fully linear case at 200 Gb/s per lane is the subject of the CEI-224G-Linear project, which is meant to serve linear pluggable, near-package, and co-packaged optics from one set of electrical specifications, and it remains in progress, as does the 200 Gb/s per lane half-retimed project beside it. IEEE P802.3dj, which defines the 200 Gb/s per lane signaling that 800 Gb/s and 1.6 Tb/s Ethernet rest on, is likewise still in ballot rather than approved, its task force's adopted timeline pointing at completion in 2026. The practical consequence is a split: linear links at 100 Gb/s per lane now have published specifications to test against, while linear links at 200 Gb/s per lane still depend on qualified pairings and vendor cooperation rather than on the blind interoperability that conventional pluggables enjoy.

Design and Deployment Practice

For a host designer, the shift toward linear drive changes what must be budgeted. A retimed design can treat the module as an isolated load and verify the host channel against the electrical specification alone. A linear design must treat the module's internal path, the optical device, and the receiver front end as part of one channel and verify end-to-end. Several practices follow.

Place the cage as close to the host silicon as the mechanical design allows, and spend layer stack and material budget on the escape routing to it; every decibel saved on the host board is a decibel available to the optics. Model the connector and cage rather than assuming they are transparent, since at these frequencies the transition into the module is a significant discontinuity. Verify with the actual module, or with a compliance board that represents it faithfully, rather than with a generic load. Reserve forward error correction budget explicitly for the analog chain instead of assuming the optical span will consume it all. Confirm that the host serializer's equalizer has enough taps and enough adaptation range to handle a channel whose response now includes optical and transimpedance stages, and check that its adaptation converges reliably at power-up and after a module swap, not merely in steady state.

For an operator, the practices are procedural. Maintain a qualified compatibility matrix rather than assuming any module works in any port, and treat firmware versions on both host and module as part of that matrix. Keep connector cleaning and inspection in the standard workflow, since a single particle on a single-mode fiber core is a common and entirely avoidable cause of link failure. Monitor pre-correction error rate continuously rather than watching only for link loss, because a degrading link announces itself in that statistic long before it drops. Populate high-power ports with attention to the chassis airflow plan, and treat a thermal alarm as a real fault rather than a nuisance.

Troubleshooting

Most pluggable optics faults resolve into a small number of patterns. A module that does not appear at all usually indicates a management-bus or identification problem: an unsupported vendor code, an out-of-date host firmware image, or a module in low-power state because the host will not grant its power class. A module that appears but reports no received optical power points at the fiber path, a swapped transmit and receive pair, or a transmitter disabled at the far end. Received power that is present but far below the link budget indicates loss in the plant, most often a contaminated or damaged connector face, a tight bend, or an unaccounted patch panel.

Received power within budget alongside an elevated pre-correction error rate moves the investigation to the electrical domain or to the transmitter's signal quality. Check case temperature first, since a module near its thermal limit exhibits exactly this symptom. On a linear-drive link, this pattern warrants confirming that the host serializer's equalizer is configured for the module in use and that the pairing is one that has been qualified, because a linear module that closes the link with one host may not close it with another. Intermittent flapping that correlates with system load is nearly always thermal. Errors that appear only after a module swap, on a port that was previously healthy, point at the pairing rather than at either component in isolation.

Conclusion

The pluggable optical module survives because it separates concerns that operators need separated: capacity from reach, host lifetime from optics lifetime, and one vendor's silicon from another's photonics. Its weakness is that the separation is enforced by a digital signal processor whose power has grown faster than the value it adds. Linear-drive optics keeps the mechanical and operational form of the pluggable module while removing that processor, and linear receive optics keeps the transmit half of it, where it earns its keep. Both recover part of the efficiency that co-packaged optics promises without surrendering field replaceability, and both pay for it in interoperability, which now depends on the pairing of host and module rather than on a regeneration point between them. Whether that is a good trade depends on who owns both ends of the link, and the answer differs sharply between a hyperscale fabric and an enterprise network.

Related Topics