Wavelength Division Multiplexing
# Wavelength Division Multiplexing (WDM) in Silicon Photonics
Wavelength division multiplexing (WDM) is the technique of sending multiple independent signals through a single fiber by assigning each signal a distinct optical wavelength (or "color") — a natural fit for silicon photonics, which can separate wavelengths using resonant filters at the micrometer scale. A single-mode fiber 9 μm in diameter can carry dozens of independent channels simultaneously without crosstalk, each occupying a different part of the spectrum, each independently modulated and detected. This is the only practical way to pack terabit-per-second data rates into the space and power budget of a chip I/O interface: you cannot make a single modulator or photodiode fast enough; you must parallelize across wavelengths instead.
The simplest WDM system sends just two channels — one wavelength in, a different wavelength out on the same fiber (CWDM or coarse WDM, channel spacing often 20 nm). Industrial transceivers use dense WDM (DWDM), where spacing drops to 0.8 nm or smaller, packing 80–100+ independent channels into the C-band (1530–1565 nm), each with its own laser, modulator, filter, and photodiode on the silicon chip. The channel count is fundamentally limited only by how tightly the filters can be tuned (laser wavelength locking, id 972, becomes critical here) and how much optical return loss each channel can tolerate (id 973) without coupling noise back into its laser.
At the system level, WDM is fundamentally a trade-off between channel count and per-channel bandwidth. A single wavelength modulated at 25 Gbaud (gigabauds per second, ~25 Gbit/s in most encodings) is the limit for a reasonable single silicon modulator and photodiode pair. To hit 400 Gbit/s aggregate throughput, use 16 independent wavelengths, each at 25 Gbaud. To hit 800 Gbit, use 32 channels. This is why silicon photonics matured as a technology only after the datacenter industry pushed the per-lane bandwidth requirement up to 25–50 Gbaud — below that, a handful of electrical lanes was enough, and optical parallelization bought nothing.
The silicon photonics implementation cost scales with channel count. Each wavelength needs its own laser (or a tunable laser + wavelength-stabilization feedback, which is complex), its own modulator, its own photodiode, and its own filter (a microring resonator or Fabry-Pérot etalon tuned to a specific wavelength). A 16-channel transceiver is 16× the on-die laser + modulator + photodiode area, though the filters are tiny and the mux/demux optics scale sublinearly. The channel-to-channel crosstalk arises if the filters are imperfect — if λ₁'s filter leaks even a few percent of λ₂'s power into the λ₁ photodiode, it adds noise. This is why the wavelength-locking entry (id 972) and the return-loss entry (id 973) both matter: λ₁'s laser must stay locked on its assigned wavelength within the filter's bandwidth (often ~1 GHz optical, a frequency precision tighter than most RF systems), and λ₁'s modulator and photodiode must not scatter light back into any other wavelength's path.
The alternative to WDM is time-division multiplexing (TDM): use a single wavelength and send data at higher and higher baud rates on that one channel — push one modulator to 100 Gbaud, 200 Gbaud. Silicon is approaching its limits here (200 Gbaud electrical signaling requires sub-100-fs electronics, on the edge of what today's silicon processes support), so WDM is the natural next step. A hybrid approach is often best: 8 wavelengths × 50 Gbaud each = 400 Gbit/s, a balanced point where the per-channel circuit speeds remain reasonable and the channel count stays manageable.
| Multiplexing technique | Per-channel max | Channel count | Implementation cost | Cross-coupling risk |
|---|---|---|---|---|
| Single wavelength TDM | 100–200 Gbaud (approaching silicon limit) | 1 | minimal (one laser, one modulator, one PD) | none |
| CWDM (coarse, 20 nm spacing) | 50–100 Gbaud | 4–8 | moderate (filters simpler) | low (wide filter spacing) |
| DWDM (dense, 0.8 nm spacing) | 25–50 Gbaud | 16–100 | high (many on-die components, tight filters) | moderate–high (must lock wavelengths, manage return loss) |
| Hybrid WDM + TDM | 50–100 Gbaud | 8–16 channels | moderate–high | moderate (requires both per-channel and per-wavelength discipline) |
WDM matured in telecom networks first, where the economic constraint (cost per bit per kilometer of fiber laid) made parallelism across the spectrum worth any amount of complexity. In datacenter silicon photonics, the constraint is different: cost per I/O pin on the chip, which makes dense WDM appealing (many channels on one fiber), though the on-die component density required is pushing against the limits of what micrometer-scale photonics can deliver. The future is a tighter integration between the optical and electrical: lasers tunable by integrated heaters, photodiodes with integrated transimpedance amplifiers, filters with integrated phase shifters for dynamic wavelength routing — all designed to work together as a single WDM system, not as independent off-chip modules adapted to silicon.