Electronics Guide

Articles in This Category

Where Copper Runs Out

A copper channel loses energy through two mechanisms that both worsen with frequency. Conductor loss rises with the square root of frequency as skin effect confines current to an ever-thinner surface layer, and it worsens further when rough copper foil lengthens the path that current must follow. Dielectric loss rises roughly in proportion to frequency, because the laminate's loss tangent converts a fixed fraction of the stored field energy to heat on every cycle. The combined attenuation curve is steep, and it is the slope rather than the absolute value that does the damage: a channel that attenuates high-frequency content far more than low-frequency content spreads each symbol across its neighbors as intersymbol interference.

Equalization repays part of that loss, but not for free. A linear equalizer that boosts the attenuated high frequencies boosts the noise and crosstalk sitting at those frequencies along with the signal. A decision-feedback equalizer avoids that noise amplification but can propagate its own errors. Both consume power, and the power grows with the loss they must undo. Beyond roughly thirty to forty decibels at Nyquist, the combination of equalization complexity, power, and residual noise stops being attractive.

Distance translates that loss ceiling into a reach. At 200 Gb/s per lane, where PAM4 signaling places the Nyquist frequency in the vicinity of 50 GHz, passive twinaxial direct-attach copper is generally limited to about one to two meters. Active electrical cables, which place retiming or equalizing silicon in the connector shell, extend the reach to several meters at the cost of power and a component that can fail. Fiber begins where those options end. Standard single-mode fiber attenuates roughly 0.32 decibels per kilometer near 1310 nm and about 0.2 decibels per kilometer near 1550 nm, and that attenuation does not depend on the symbol rate. A kilometer of fiber costs less loss than a meter of twinax.

Two properties beyond loss favor optics at distance. Fiber is a dielectric, so it neither radiates nor picks up interference, which removes crosstalk between links and breaks the ground loop between chassis. It is also far denser per unit of bandwidth: a ribbon of sixteen fibers thinner than a pencil carries what would otherwise require dozens of shielded copper pairs. In cabinets where cable bulk obstructs airflow, that density is sometimes the deciding argument rather than the reach.

Latency deserves a more careful statement than it usually receives, because optics is often credited with an advantage it does not possess at short reach. Light in silica travels at roughly five nanoseconds per meter, given the fiber's group index near 1.47. A signal on a printed circuit board stripline travels at roughly six to seven nanoseconds per meter. The media are therefore comparable, and fiber's modest edge is easily erased by the transceiver: serialization, modulation, detection, and especially the digital signal processing inside a retimed optical module each add latency that a passive copper cable does not. Optics wins on latency over long distances, where copper would require repeaters or would not close at all. Over a meter, a passive copper cable is usually the faster path, which is why short scale-up fabrics inside a rack still favor copper wherever the geometry permits.

Anatomy of an Optical Link

Every optical interconnect performs the same sequence: generate light, impress data on it, guide it to the far end, convert it back to current, and recover the data. The design choices at each stage determine reach, power, and cost.

Sources and Modulation

Two source strategies dominate. Vertical-cavity surface-emitting lasers emit near 850 nm, are cheap, are tested at the wafer level, and are directly modulated by switching the drive current. They pair with multimode fiber and serve short reaches. Distributed-feedback lasers emitting near 1310 nm run continuous-wave and hand their light to a separate modulator, which is either an electro-absorption device or a Mach-Zehnder or ring modulator built in silicon photonics. Separating the source from the modulator avoids the frequency chirp that direct current modulation imposes on a laser, and chirp interacting with fiber dispersion is what limits directly modulated links at longer reach.

The modulator choice is itself a trade-off. A Mach-Zehnder modulator is broadband, tolerant of temperature, and comparatively large and power-hungry. A ring modulator is compact and needs far less drive energy, because it exploits a resonance rather than an interferometric path-length difference, but that resonance drifts with temperature and must be held in place by a control loop. Recent co-packaged designs have favored rings precisely because their lower drive voltage translates directly into lower energy per bit, and have accepted the thermal tuning burden that comes with them.

Fiber and Wavelength

Multimode fiber has a core near 50 micrometers, which makes alignment tolerant and connectors cheap. Its limitation is modal dispersion: many spatial modes propagate at slightly different group velocities and arrive smeared in time. OM4 fiber is specified for an effective modal bandwidth of 4700 MHz·km at 850 nm, which sets a bandwidth-distance product rather than a simple reach. Single-mode fiber has a core near 9 micrometers and supports one spatial mode, so modal dispersion disappears and reach is set by attenuation and chromatic dispersion instead. Alignment tolerances tighten accordingly, which is why single-mode connectors and the assembly steps behind them cost more.

Data communication concentrated on 1310 nm rather than the lower-loss 1550 nm window for a specific reason: standard G.652 fiber has its zero-dispersion point near 1310 nm. Inside a data center, where a few kilometers is a long link, dispersion rather than attenuation is the binding constraint, so operating at the dispersion null buys more than the extra 0.1 decibel per kilometer costs. Long-haul transport, where attenuation accumulates over hundreds of kilometers, makes the opposite choice and compensates dispersion separately.

Capacity is then multiplied in one of two ways. Parallel fiber assigns each lane its own strand and terminates them in a multi-fiber push-on connector, which is simple and favors short reach where fiber is cheap. Wavelength-division multiplexing assigns each lane a different wavelength and combines them onto a single strand, which costs multiplexers and wavelength-stable sources but conserves fiber over distance. The naming of Ethernet optical interfaces reflects the split directly, as the next section describes.

Detection and Recovery

At the far end a photodiode converts optical power to current, and a transimpedance amplifier converts that current to a voltage the receiver can slice. Two consequences follow from the physics. First, a photodiode responds to optical power rather than to field amplitude, so the electrical signal is proportional to the square of the optical field. A three-decibel loss of optical power costs six decibels of electrical signal, which is why optical power budgets are unforgiving. Second, sensitivity is a power, not a voltage: the receiver is characterized by the least optical power at which it still meets the target error rate, and that floor is set by photodiode shot noise and by the thermal noise of the transimpedance amplifier's feedback resistance.

Because the signal is intensity, the transmitter is characterized by its extinction ratio, the ratio of power in a one to power in a zero, and by optical modulation amplitude, the difference between those two levels. A large extinction ratio opens the eye but pushes the laser toward its threshold, where it turns on slowly and noisily. Practical designs settle at a finite extinction ratio and accept the resulting penalty, and control loops hold both average power and extinction ratio steady as the laser ages and its temperature changes.

Reach Classes, Form Factors, and Standards

IEEE 802.3 organizes optical Ethernet interfaces into reach classes whose names encode the medium and distance. The suffix identifies the class and the trailing number gives the count of lanes. At 800 Gb/s the family illustrates the pattern well. 800GBASE-SR8 uses eight 850 nm VCSEL lanes on multimode fiber and reaches at least 100 meters. 800GBASE-DR8 uses eight 1310 nm lanes on parallel single-mode fiber and reaches at least 500 meters. The FR class reaches about two kilometers, typically by multiplexing four wavelengths onto a fiber pair, and the LR class extends to about ten kilometers. Each step in reach adds cost in the source, the fiber plant, or both, so choosing the class is an architectural decision made early.

Modules are packaged into pluggable form factors that trade faceplate density against thermal capacity. QSFP-DD and OSFP dominate current switch and accelerator front panels, with OSFP offering a larger thermal envelope for the higher-power modules that 800 Gb/s and 1.6 Tb/s demand. Active optical cables occupy an intermediate position: they present an electrical interface at each end and hide the optics and the fiber inside a permanently terminated assembly, which removes connector cleanliness from the operator's concerns at the cost of a fixed length.

The standards themselves advance in lane-rate steps. IEEE 802.3df defined 800 Gb/s Ethernet built on 100 Gb/s lanes. The successor project, IEEE P802.3dj, defines 200 Gb/s per lane and the 200, 400, 800 Gb/s and 1.6 Tb/s interfaces built from it; as of mid-2026 the task force was still working through ballot toward completion, so implementations ahead of publication track draft specifications and multi-source agreements. Alongside IEEE, the Optical Internetworking Forum's Common Electrical Interface projects specify the electrical lanes that connect host silicon to the optical module, and it is those electrical specifications, not the optics, that most often determine whether a given host board can support a given module.

Forward error correction is now assumed rather than optional. Reed-Solomon RS(544,514), commonly called KP4, is the workhorse for PAM4 lanes, and the optical budget closes against the raw error rate the code can repair rather than against an unaided target. This changes the character of the budget: the link is permitted to be far noisier than a pre-FEC design would tolerate, and the coding gain supplies the remaining reliability. It also introduces latency and a power cost of its own, which matters in latency-sensitive fabrics.

The Energy Argument

Modern interconnect design is judged in picojoules per bit as much as in gigabits per second, because at machine scale the interconnect's power competes directly with the compute it feeds. Retimed 800 Gb/s pluggable modules typically land in the range of ten to fifteen picojoules per bit. A substantial share of that budget is spent not on the optics but on the electrical channel between the host silicon and the module faceplate, and on the digital signal processing inside the module that undoes the damage that channel causes.

That observation drives the two main lines of attack. The first shortens or simplifies the electrical path. Linear pluggable optics remove the module's digital signal processor entirely and rely on the host SerDes to equalize the whole path, which cuts power and latency but demands a well-controlled host channel and complicates interoperability, since the module no longer regenerates the signal it receives. Linear receive optics keep a transmit-side digital signal processor while running the receive path linearly, a compromise that gives back some efficiency in exchange for easier integration and thermal behavior at 1.6 Tb/s.

The second line of attack moves the optics into the package. Co-packaged optics places optical engines on the same substrate as the switch or accelerator die, shrinking the electrical channel from a faceplate-length path to a few millimeters of substrate routing and allowing much simpler, lower-power SerDes. Production co-packaged designs have demonstrated roughly five picojoules per bit or below, a reduction on the order of two-thirds relative to comparable pluggable modules. The deployment record is no longer theoretical: Broadcom shipped its 51.2 Tb/s Tomahawk 5 Bailly co-packaged switch in volume during 2025 and followed with a 102.4 Tb/s Tomahawk 6 Davisson generation, while NVIDIA's Quantum-X Photonics InfiniBand switches reached availability in early 2026 with Ethernet counterparts following. Foundry support matured alongside, with TSMC's COUPE platform stacking an electronic die on a photonic die entering volume production in 2026.

Co-packaged optics buys efficiency by surrendering modularity. A pluggable module that fails is replaced in seconds by a technician; a failed optical engine inside a switch package is a board-level repair. Lasers dislike heat, and the package interior beside a high-power ASIC is among the hottest places in the system, which is why many designs supply light from external laser sources and route it in through fiber, leaving only modulators and detectors in the package. Yield compounds the problem, since a package combining a large digital die with several photonic engines fails if any one of them fails, which makes known-good-die screening a precondition rather than a refinement. These are exactly the trade-offs the co-packaged optics subcategory examines in detail.

Signal Integrity Does Not Disappear

Converting to light removes the copper channel but does not remove the discipline. The electrical channel moves rather than vanishes, and several impairments arrive to replace the ones left behind.

The surviving electrical channel matters most. Between the host SerDes and the optical engine there is still a path through package, board, connector, and module, and at 200 Gb/s per lane that path is demanding even when it is short. Much of the appeal of co-packaged optics is precisely that it shortens this channel; much of the risk in linear pluggable optics is that it exposes this channel with less correction downstream.

On the optical side, impairments take unfamiliar forms. Chromatic dispersion spreads a pulse because a source has finite spectral width and different wavelengths travel at different speeds, and laser chirp widens that spectrum further under modulation. Modal dispersion limits multimode links independently of loss. Relative intensity noise describes the laser's own power fluctuation and sets a noise floor that more optical power cannot overcome. Reflections from connectors and facets return to the laser cavity and destabilize it, which is why isolators and angled connector polishes appear in link designs. Connector contamination is a genuine failure mode rather than a handling nicety, since a single particle on a single-mode core face can obscure a meaningful fraction of it.

Measurement adapts accordingly. The transmitter eye mask familiar from electrical compliance gives way, for PAM4 optical interfaces, to transmitter and dispersion eye closure for PAM4, a single figure of merit expressed in decibels that folds the transmitter's noise, distortion, and dispersion penalty into one number a reference receiver can evaluate. Sensitivity, extinction ratio, and optical modulation amplitude replace voltage swing and jitter as the headline transmitter and receiver parameters, though jitter and clock recovery remain fully present in the electrical circuits at both ends.

The Integration Roadmap

The trajectory of optical interconnects is best read as optics moving steadily closer to the transistors it serves. Front-panel pluggable modules keep the optics at the chassis boundary and maximize serviceability. Linear pluggable and linear receive variants keep that position but strip out the processing that the electrical channel forced upon the module. Near-package optics moves the engine onto the same board immediately beside the ASIC. Co-packaged optics moves it onto the same substrate. The end point under active development is optical input and output as a chiplet, in which a photonic die sits in the same package as a processor and communicates with it over a standard die-to-die interface, so that a processor's off-package bandwidth is limited by photonics rather than by pin count and channel loss.

Each step trades serviceability and supply-chain flexibility for energy and bandwidth density, and no step has retired the one before it. Pluggable optics remain the right answer for links that cross a room or a campus, for equipment that must be repaired in the field, and for any deployment where multi-vendor interoperability outweighs the last few picojoules per bit. What has changed is that the highest-bandwidth machines can no longer afford a uniform answer, so a single system now mixes copper, pluggable optics, and in-package optics according to where each link falls relative to the crossover.

Conclusion

Optical interconnects are best understood not as a replacement for electrical signaling but as the other half of a boundary that keeps moving inward. Copper remains superior at very short reach, where it is denser, cheaper, more serviceable, and often faster. Fiber becomes unavoidable once the loss slope of the copper channel outruns what equalization can economically repay. Each generation of signaling shifts that boundary closer to the die, which is why the discipline now reaches inside the package rather than stopping at the faceplate. The subcategories that follow develop the field in that order, from the conversion circuits at each end of the link, through link-level design and the photonic devices that implement it, to the packaging that places optics beside the processor.

Related Topics