Optical Link Budget

An optical link budget is the running ledger that decides whether a photonic interconnect actually works before a single bit ever gets sent: start with the optical power a laser can deliver, subtract every loss the light accumulates crossing couplers, modulators, fiber, and connectors on the way to the receiver, and check whether what is left still clears the minimum power a photodetector needs to distinguish a one from a zero with an acceptable error rate.

The decibel is the unit that makes this entire exercise tractable, because it converts a chain of multiplicative power losses into a simple additive subtraction. Optical power loss at each stage — coupling light from a laser into a waveguide, modulating that light to carry data, propagating through fiber, passing through connectors — is naturally multiplicative: each stage transmits some fraction of the power it received. Expressed in decibels, each stage's loss becomes a number that simply subtracts from a running total, so a link with five lossy stages is analyzed by adding five negative numbers rather than multiplying five fractions together — the same reason decibels are the standard unit throughout RF and optical engineering rather than linear power ratios.

Each stage in a typical silicon-photonic link contributes a specific, physically distinct loss mechanism, and a link budget has to account for every one of them individually rather than using one average number. Coupling loss occurs getting light from a laser source into an on-chip waveguide, where mode mismatch between a laser's beam shape and the waveguide's much smaller cross-section wastes a predictable fraction of the power. Modulator insertion loss comes from whatever optical structure encodes data onto the light — a ring resonator or Mach-Zehnder interferometer, both covered elsewhere on this site — since encoding data onto an optical carrier is itself not a lossless process. Fiber attenuation and connector loss accumulate with physical distance and at every mechanical joint the signal passes through. None of these losses can be wished away by a better receiver; they are physically fixed once the link's architecture and component choices are set, which is why the budget has to be built component by component rather than assumed from a system-level target.

Receiver sensitivity — the minimum optical power a photodetector needs to reliably tell a one from a zero at a target bit-error rate — is the other half of the budget, and it is not a single fixed number but one that moves with data rate. Higher symbol rates compress each bit into a shorter time window, giving the detector less integrated signal energy to work with per bit, which pushes the minimum required power up as speed increases. This is why pushing an optical link to a higher data rate is never simply a matter of driving the laser harder; it simultaneously demands more sensitive receiver design, lower-loss components throughout the chain, or both, since the margin between delivered power and required sensitivity has to clear a bar that is itself rising with speed.

Co-packaged optics — mentioned elsewhere on this site as the technique moving the electrical-to-optical conversion physically inside the package, closer to the compute silicon — stakes its entire value proposition on precisely this budget closing more comfortably than it would over a longer electrical or conventional optical path. Every centimeter of distance a signal travels as light rather than as copper changes its position in the budget: shorter on-package optical paths mean less fiber attenuation and fewer connector transitions to subtract, which directly widens the margin available for everything else in the chain, including aging and manufacturing variation this series has already covered as unavoidable sources of degradation over a product's lifetime. The entire argument for moving optics inside the package, rather than keeping it as a separate pluggable module at the rack's edge, is a link-budget argument before it is anything else.

An optical link budget is a chip's own version of accounting — every component on the path either adds power or spends it, and the link only works if what survives the spending still clears the bar the receiver sets. It is the reason an optical interconnect's architecture — how many components the light passes through, how far it travels, and at what data rate — is decided by arithmetic on a spreadsheet long before any photonic circuit is actually fabricated.