WDM Links and Transceivers
Pull any fibre out of a datacenter switch and you are holding one end of the machine this lesson
dissects. The previous
lesson argued that beyond the crossover distance light carries bits for less; now we
build the thing that actually does it. An optical link is a little assembly line: a
laser supplies a pure carrier, a modulator stamps bits onto it, a
multiplexer merges several such carriers — each a different wavelength — onto one
fibre, the fibre hauls them, a demultiplexer sorts them back out,
and for each wavelength a photodetector, a transimpedance amplifier
(TIA) and a clock-and-data-recovery (CDR) circuit turn faint photocurrent back into
clean digital bits. That trick of stacking wavelengths — wavelength-division multiplexing,
WDM — is the single biggest lever photonics owns: one strand of glass, many independent channels,
because light of different colours passes through the same fibre without interacting.
The engineering question for any link is brutally simple: does enough power survive the journey
for the receiver to tell a 1 from a 0? The bookkeeping that answers it — the link
budget — is done in decibels, and by the end of this lesson you will do it in your head.
Anatomy of a link
Step through the assembly line. Note what is per lane (laser, modulator, detector,
TIA, CDR — one set per wavelength) and what is shared (mux, fibre, demux — one of each,
however many lanes ride inside). That asymmetry is why WDM pays: the expensive shared medium is
amortised over every extra colour.
You have already met every part: the modulator is typically a Mach–Zehnder or microring from
Module 3, and the mux/demux is an interference device — an arrayed-waveguide grating or cascade of
ring filters, the integrated descendant of the
diffraction grating,
steering each colour to its own port because path differences that are a whole number of one
wavelength are a fraction of another.
The link budget: accounting in decibels
Decibels turn multiplication into addition. Power in dBm is
10\log_{10}(P/1\,\text{mW}) — so 1 mW is 0 dBm, 10 mW is +10 dBm,
0.1 mW is −10 dBm — and every loss along the path is a subtraction:
- received power = transmitted power − the sum of all losses:
P_{\text{rx}} \;=\; P_{\text{tx}} - \sum_i L_i \quad \text{(all in dB/dBm)};
- the link closes if P_{\text{rx}} exceeds the
receiver sensitivity — the minimum power at which it still achieves the
target bit-error rate;
- the surplus, P_{\text{rx}} - P_{\text{sens}}, is the
margin — the engineering headroom for ageing, temperature, dirty connectors
and bad days; healthy links keep ≥ 3 dB;
- every 3 dB is a factor of 2 in power; every 10 dB a factor of 10.
A worked example, in the style of a 10 km "LR4"-class link. Per lane: transmit power
+2 dBm; mux insertion loss 1.5 dB; two connectors at 0.5 dB each;
10 km of fibre at 0.4 dB/km = 4 dB; demux insertion loss 1.5 dB. Then
P_{\text{rx}} = 2 - 1.5 - 1 - 4 - 1.5 = -6\ \text{dBm}.
Against a receiver sensitivity of -11 dBm, the margin is a comfortable
5 dB. Notice what would break it: another 12 km of fibre (≈ 5 dB) would eat the margin exactly —
this link's reach is budget-limited to roughly 22 km. Run the same arithmetic over a sweep of
lengths:
const txDbm = 2.0; // per-lane launch power
const muxDb = 1.5, demuxDb = 1.5;
const connectorsDb = 2 * 0.5;
const fibreDbPerKm = 0.4; // O-band single-mode
const sensitivityDbm = -11.0;
function marginDb(km: number): number {
const rx = txDbm - muxDb - demuxDb - connectorsDb - fibreDbPerKm * km;
return rx - sensitivityDbm;
}
for (const km of [2, 10, 20, 22, 25]) {
const m = marginDb(km);
console.log(km + " km: margin " + m.toFixed(1) + " dB " + (m >= 0 ? "(closes)" : "(FAILS)"));
}
The chart below makes the same story graphical: received power slides downhill with distance at the
fibre's loss slope, and the link dies where the line crosses the sensitivity floor. Raise the launch
power and watch the reach stretch — and note how much further the same budget goes at 1550 nm,
where glass is at its clearest:
Rate arithmetic: wavelengths × baud × bits
A transceiver's headline number is a product of three factors you can now read like a part number:
\text{aggregate rate} \;=\; N_\lambda \times B \times b
— the number of wavelengths N_\lambda, the symbol rate
B in gigabaud, and the bits per symbol b.
Modern links use PAM4 — four amplitude levels, so b = 2 —
doubling the rate at the same baud, at the price of squeezing three smaller eyes into the old
one's voltage swing (≈ 9.5 dB of SNR penalty, paid for with DSP and forward error correction). An
"800G-DR4" pluggable is exactly this arithmetic: 4 lanes × 100 GBd × 2 bits = 800 Gb/s (the real
baud is 106.25, the excess carrying FEC overhead). The physical package that wraps all this — a
pluggable transceiver, in form factors like QSFP28, QSFP-DD and OSFP — snaps into
the switch faceplate and presents plain electrical SerDes lanes to the host. The suffix tells you
the reach and lane plan: DR (500 m, parallel fibres), FR (2 km,
WDM), LR (10 km, WDM). One tidy consequence of the product form: to double a
link, you may double any factor — more colours, faster symbols, or denser modulation —
and the industry has historically rotated between all three.
The classic units blunder: dBm is an absolute power (referenced to 1 mW);
dB is a ratio. You may subtract dB from dBm (a loss applied to a power → a power),
and subtract dBm from dBm (two powers → a ratio in dB). But adding dBm to dBm is
meaningless — +3 dBm plus +3 dBm is not +6 dBm; two 2 mW lasers make 4 mW ≈ +6 dBm only by
converting to milliwatts first, adding, and converting back. The same trap bites when combining
WDM lanes: forty lanes at 0 dBm each total 40 mW = +16 dBm on the shared fibre, not "+0 dBm × 40".
Keep the type discipline of a programmer: dBm + dB → dBm ✓; dBm − dBm → dB ✓; dBm + dBm →
compile error.
Every datasheet mentions the same two magic wavelengths, and neither is arbitrary — they are the
two clear windows in silica glass. Attenuation in fibre is a tug-of-war between Rayleigh
scattering (falling as 1/\lambda^4 — the same physics that blues the
sky) and infrared absorption by the glass itself (rising steeply past ~1.6 µm). The trough between
them bottoms out near 1550 nm at a miraculous ≈ 0.17 dB/km — light circles a
16-km spool and still keeps half its power — which is why all long-haul telecom lives there,
helped by erbium-doped fibre amplifiers that happen to amplify exactly that band.
1310 nm is the fibre's zero-dispersion wavelength, where different
spectral components travel at the same speed and pulses stay sharp without correction — worth the
slightly higher loss for short, unamplified datacenter hops, which is why your DR and FR optics
are O-band. Two windows, two philosophies: lowest loss for going far, lowest dispersion for going
fast and cheap.
Where this goes next
The pluggable transceiver is a triumph with a flaw: it lives at the switch faceplate, a full
circuit-board-length of hostile copper away from the ASIC whose bits it carries — and at 200 Gb/s
per lane, that last electrical foot now costs as much engineering as the ten optical kilometres
that follow. The obvious move is to shorten it: pull the optics out of the faceplate and set them
down millimetres from the die. That is
co-packaged
optics — the next lesson.