11.1 Coherent transmission
Direct-detection receivers (PIN + TIA) only see the intensity of light. The phase is thrown away. This is fine for on-off-keyed amplitude signaling but throws away three quarters of the information you can carry per Hz of optical bandwidth.
A coherent receiver mixes the incoming signal with a local-oscillator laser at the receive end. The beat between the signal's electric field and the LO's electric field appears as a low-frequency electrical signal at the photodetector, with both amplitude and phase preserved. With balanced detectors and 90-degree hybrid optics, both quadratures (I and Q) are recovered.
A coherent transponder uses:
- A tunable narrow-linewidth DFB laser at the transmitter.
- A polarization-multiplexed I-Q Mach-Zehnder modulator.
- A second tunable laser at the receiver as the local oscillator.
- Two pairs of balanced photodiodes (one pair per polarization).
- Four high-speed ADCs (one per polarization-quadrature combination).
- A massive DSP that performs chromatic dispersion compensation, polarization tracking, frequency offset correction, carrier phase recovery, and forward error correction in real time.
The result: you can transmit polarization-multiplexed 16-QAM, 32-QAM, or 64-QAM at 95 Gbaud (giving 200, 400, or 600+ Gbps) over a single channel, and 800 Gbps over slightly wider channels. Standard for any new long-haul deployment since 2013.
The DSP also undoes chromatic dispersion in software. This is why 2010-vintage long-haul lines, originally designed for direct detection with dispersion-compensating fiber spools every span, can be upgraded to multi-terabit coherent operation simply by swapping transponders at the endpoints. The optical line itself does not change.
11.2 PAM-4 and intra-data-center optics
Inside a data center the constraints are different: distances under 2 km, latency budgets in the tens of nanoseconds, cost per port matters far more than spectral efficiency. The dominant modulation here is PAM-4 (4-level pulse-amplitude modulation), which sends two bits per symbol at the same baud rate as NRZ. A 53 Gbaud PAM-4 signal carries 100 Gbps; 106 Gbaud PAM-4 carries 200 Gbps; eight lanes of either give 800 GbE or 1.6 TbE in a single pluggable.
PAM-4 has half the SNR margin of NRZ at the same baud rate, so it requires more careful equalization, but the bandwidth savings are worth it. Every modern QSFP-DD and OSFP module relies on PAM-4 in the optical lanes and the electrical host interface.
11.3 Silicon photonics
Building optical components in standard CMOS-compatible silicon waveguides is the holy grail of integration: combine the modulators, detectors, and electronics on a single chip, in the same fab process as the host SoC.
Silicon photonics platforms (Intel, Cisco/Acacia, Marvell/Inphi, GlobalFoundries 45SPCLO, AIM Photonics) provide:
- Silicon waveguides, 220 nm tall and 500 nm wide, single-mode at 1550 nm.
- Mach-Zehnder modulators using carrier injection or depletion (the plasma dispersion effect).
- Germanium-on-silicon photodetectors that work at 1310 and 1550 nm.
- Edge couplers or grating couplers for fiber-to-chip light transfer.
- Heterogeneous-integrated InP lasers (silicon does not lase by itself, so the laser has to be flip-chipped or wafer-bonded onto the silicon platform).
Co-packaged optics (CPO) replace pluggable transceivers with optical engines mounted directly next to the switch ASIC, with very short electrical traces from chip to optical engine. CPO promises 10x lower power per Gbps than pluggable optics, with deployment expected to scale through the second half of the 2020s. Cisco's Silicon One, Broadcom's Tomahawk, and Marvell's Teralynx switch ASICs all have CPO roadmaps.
11.4 Free-space optical (FSO)
When fiber is impossible (across a river, between two buildings without a duct, or between the ground and a satellite) light can also propagate through air. FSO links use collimated laser beams, often 1550 nm because that is eye-safe and atmospheric absorption is low, with telescope optics at each end.
Practical issues: atmospheric scintillation, fog, rain, alignment drift from thermal expansion. Range from a few hundred meters in city air to thousands of kilometers in vacuum (for inter-satellite links).
The most consequential current FSO deployment is the inter-satellite laser links in SpaceX's Starlink constellation, where each satellite has multiple laser terminals talking to neighboring satellites at 100+ Gbps. The same physics is used by Amazon's Project Kuiper and the European IRIS² constellation. NASA's Lunar Communications Relay Demonstration successfully demonstrated 1.2 Gbps from lunar orbit to Earth in 2014.
11.5 FTTH and PON architectures
Fiber to the home (FTTH) brings fiber all the way to a residential customer. The dominant architecture is the passive optical network (PON): one fiber from a central office goes through a passive splitter (1:32 typically) to up to 32 homes. Downstream is broadcast (every home receives every packet, encrypted; only the addressed home decrypts). Upstream is time-multiplexed (each home gets its own time slot).
PON generations:
- GPON (G.984): 2.5 Gbps down, 1.25 Gbps up, shared. Standard since 2008.
- XGS-PON (G.9807): 10 Gbps symmetric, current rollout.
- NG-PON2 (G.989): 40 Gbps via four wavelengths.
- 25G-PON, 50G-PON: emerging.
The two wavelengths used in GPON: 1490 nm downstream from the optical line terminal (OLT) to the optical network terminal (ONT), 1310 nm upstream from ONT to OLT. The same fiber carries both, separated by WDM filters at each end.
Verizon FiOS, AT&T Fiber, Google Fiber, Comcast (Xfinity 10G), and most European carriers use GPON or XGS-PON to deliver internet to tens of millions of homes. The "G" in 5G is sometimes celebrated, but the unsung 10 Gbps of XGS-PON at the curb is what actually delivers most of those gigabits to your modem.
11.6 Submarine cables
The internet's geographic backbone is a network of submarine fiber cables on the seafloor. About 550 active submarine cables span 1.4 million km of seabed, carrying 99% of all intercontinental data traffic.
Anatomy of a typical submarine cable:
Cable cross-section (deep sea, light armor):
┌─────────────────────────────┐
│ polyethylene sheath │
│ ┌───────────────────────┐ │
│ │ steel armor wires │ │
│ │ ┌─────────────────┐ │ │
│ │ │ copper conductor│ │ │ <- HVDC for repeater power
│ │ │ ┌────────────┐ │ │ │
│ │ │ │ fiber │ │ │ │ <- 4-24 fiber pairs in plastic tube
│ │ │ │ bundle │ │ │ │
│ │ │ └────────────┘ │ │ │
│ │ └─────────────────┘ │ │
│ └───────────────────────┘ │
└─────────────────────────────┘A modern cable like 2Africa or MAREA carries 16 to 24 fiber pairs, each running 8-16 Tbps with coherent DWDM transponders, for 200-300 Tbps total. Spread across the cable, that is enough capacity to support millions of HD video streams in parallel.
The sheath protects the fiber bundle from saltwater. The armor protects against fishing trawls and ship anchors. The copper conductor at center carries up to 18 kV DC from shore stations to power the EDFA repeaters (which are passive optically but need electrical power for the pump lasers).
Cable laying ships pay out cable from a ship's stern at 8 km/hour, with a plow that buries the cable under the seabed in shallow regions. The cable is engineered for a 25-year working life. Repairs are done by fishing the broken cable up to a repair ship, splicing in a new section, and dropping it back. Repair operations take days to weeks and cost millions of dollars per incident.
The internet backbone geography is dictated by where these cables make landfall. Clusters of cables come ashore at landing stations in places like Virginia Beach, Marseille, Singapore, and Mumbai. The lightning-fast routes between continents are a matter of where cables happen to lie on the ocean floor, which makes the geography of the modern internet a literal map of where ships have laid glass.