Skip to main content
Be Bitwise

Photonic Integrated Circuits

Optical & Photonic Computing · Frontiers track

9 min read

In this module

Shrinking Optics onto a Chip

Early optical computing meant lab benches stacked with lenses, mirrors, and beam splitters — precise, bulky, fragile. Photonic integrated circuits (PICs) do for optics what ICs did for electronics: shrink the components onto a chip and unlock mass production, compact form factors, and tight coupling to electronics.

A PIC packs waveguides, modulators, detectors, couplers, and filters onto a single substrate. Light flows through waveguides (the optical analogue of wires), gets reshaped by modulators, split and recombined by couplers, and finally read out by photodetectors — all in a chip measured in millimetres.

Silicon Photonics: The Key Enabler

Silicon photonics is what makes photonic computing practical. It uses the same silicon wafers and CMOS-compatible processes that fab electronic chips, so photonic devices can be manufactured at scale in existing foundries.

Why Silicon?

Silicon is transparent at telecom wavelengths (1310 nm and 1550 nm). Its refractive index is high — around 3.5 — so silicon waveguides confine light tightly. Waveguide cross-sections can shrink to 500 nm × 220 nm, in the same ballpark as electronic interconnects.

The big index contrast between silicon (n=3.5) and its silicon-dioxide cladding (n=1.45) lets you bend light hard with very little loss. Photonic circuits stay compact — square millimetres to square centimetres, depending on what they're doing.

The deciding factor, though, is that silicon photonics rides on top of the multi-trillion-dollar CMOS manufacturing stack. Photonic chips fab alongside electronic ones, sharing lithography tools, etch processes, and clean rooms. That's an enormous cost cut versus exotic optical materials.

CMOS Compatibility Is Everything

The single property that matters most about silicon photonics is CMOS compatibility. Photonic chips fab in the same foundries that make CPUs and GPUs, using the same proven, high-yield processes. It also opens the door to putting photonic and electronic components on the same die or in the same package — the tight electro-optic integration that real photonic computing demands.

Core Photonic Components

Waveguides

Optical waveguides are the wires of a photonic circuit. A silicon waveguide is a strip of silicon — usually 500 nm wide and 220 nm tall — clad in silicon dioxide. Light stays inside the silicon core via total internal reflection, the same trick that runs an optical fibre.

Waveguides can ferry light across a chip (centimetres) with modest loss — 1–3 dB per centimetre on state-of-the-art silicon. They route, bend (radii down to about 5 µm), and cross each other with minimal loss or crosstalk.

Modulators

Electro-optic modulators turn electrical signals into optical ones by tweaking the light passing through a waveguide. The workhorse in silicon photonics is the carrier-depletion modulator: drive a voltage onto a p-n junction in the waveguide, free-carrier concentration shifts, refractive index shifts, the phase of the light shifts.

Drop that modulator into an interferometer (a Mach-Zehnder, say) and the phase shift becomes an intensity shift — your digital or analogue signal is now encoded on the beam.

Modern silicon modulators run at 50–100 Gbaud (50–100 billion symbols per second), with research pushing towards 200 Gbaud and beyond.

Photodetectors

Photodetectors convert optical signals back to electrical signals. In silicon photonics, germanium is typically used for detection (silicon itself is transparent at telecom wavelengths, but germanium absorbs at 1550 nm). Germanium photodetectors are integrated directly on the silicon chip and can operate at speeds exceeding 100 GHz.

Couplers and Splitters

Directional couplers split light between two waveguides by placing them close together so that their evanescent fields overlap. By controlling the coupling length and gap, any splitting ratio (from 100:0 to 50:50) can be achieved.

Multimode interferometers (MMIs) are another common splitting element, using a wider waveguide where multiple modes interfere to produce desired splitting ratios. MMIs are more tolerant of fabrication variations than directional couplers.

Microring Resonators

Microring resonators are tiny circular waveguides (radii of 5-50 micrometers) that resonate at specific wavelengths. When the light's wavelength matches the ring's resonance, light couples into the ring and circulates. This makes rings excellent as:

  • Wavelength filters - selecting a specific wavelength from a WDM signal
  • Modulators - tuning the resonance with a voltage to switch a wavelength channel on or off
  • Weight elements - controlling how much light of a given wavelength passes through
text
Waveguide Ring Resonator
──────────────╮ ╭─────────╮
╰───│ │
│ ○ Ring │
╭───│ │
──────────────╯ ╰─────────╯
Light at the resonant wavelength couples into the ring.
Light at other wavelengths passes through undisturbed.
Microring Resonators as Programmable Weights

In photonic neural networks, microring resonators play the role of programmable weights. Tune a ring's resonance electrically — by heating it or injecting carriers — and you control how much light couples through. That's a multiplication of the optical signal by a weight value. An array of rings, each set to a different weight, implements a row of a weight matrix across multiple wavelengths in parallel.

The Mach-Zehnder Interferometer (MZI)

The Mach-Zehnder interferometer is the load-bearing building block of programmable photonic circuits. It has three pieces:

  1. An input coupler (beam splitter) that splits incoming light into two paths
  2. Two waveguide arms, one of which has a tunable phase shifter
  3. An output coupler that recombines the light
text
Input light → ─┬──── Phase φ ────┬── → Output 1
│ │
└─────────────────┘── → Output 2

By adjusting the phase shift φ, the MZI controls how the light from the two arms interferes at the output:

  • φ = 0: the arms interfere constructively at one output port, so all the light leaves there (call it Output 1)
  • φ = π: they now interfere destructively at that port, so all the light switches to Output 2
  • Intermediate φ: Output 1 gets a cos²(φ/2) share of the power and Output 2 gets sin²(φ/2)

Which physical port plays "Output 1" depends on the couplers. With two 50:50 directional couplers, for instance, φ = 0 sends the light to the crossed port, the one diagonally opposite the input.

That makes the MZI a universal tunable element: pair a phase shift with beam splitting and you get any 2x2 unitary matrix. A mesh of MZIs can implement any NxN unitary, as proved by Reck et al. in 1994 and later refined into a more compact layout by Clements et al.

Programmable Photonic Processors

A programmable photonic processor is an array of MZIs (and sometimes microring resonators) whose phase shifts can be individually controlled by an electronic controller. The array can be reconfigured to implement different matrix transformations by adjusting the phase settings.

MZI Mesh Architectures

Two main flavours of NxN unitary mesh exist:

Reck architecture (triangular): N(N-1)/2 MZIs in a triangular layout. Any NxN unitary decomposes into this form, but the triangle means different input-output paths cross different numbers of MZIs, which introduces asymmetry.

Clements architecture (rectangular): same MZI count, more symmetric rectangular layout. Optical paths are better balanced and the design tolerates fabrication imperfections better.

Here are both meshes for N = 4, six MZIs each. Read each box as one MZI and its column, and each arrow as a waveguide carrying light on to the next MZI.

Reck needs five columns (2N - 3) and leaves mode 4 touching a single MZI. Clements packs the same six into four columns (N).

Count the columns: that's the optical depth, the most MZIs light can cross. Fewer columns means less loss on the worst path, and in Clements every path sees roughly the same number, so the losses stay balanced.

Either way, an NxN matrix wants O(N²) MZIs and O(N²) individually controlled phase shifters. The matrix-vector multiplication itself runs in O(1) time — the time for light to cross the mesh, usually picoseconds — regardless of N.

Photonic Chip Examples

Intel Silicon Photonics

Intel is the volume leader in silicon photonics, churning out millions of photonic transceivers a year for data-centre interconnects. Their stack includes integrated lasers, high-speed modulators, and germanium photodetectors on a single silicon chip.

The current products focus on communication, but the same platform underpins photonic computing. Intel has shown photonic circuits with hundreds of components on a chip.

Lightmatter Envise

Lightmatter's Envise is a photonic processor purpose-built for AI inference. An MZI mesh handles the matrix multiplication optically; electronics take care of nonlinear activations, data conversion, and control. The chip uses a CMOS-compatible process and slots into standard electronic systems.

Xanadu X-Series

Xanadu builds photonic processors that span both quantum and classical workloads. The X-series uses squeezed light states for quantum computation and can also crunch classical linear algebra through its photonic circuits.

Integration with Electronics

No photonic processor stands alone. Every real photonic computing system needs electronics close at hand for:

  • Data input: turning electronic data into optical signals via modulators
  • Data output: turning optical results back into electrons via photodetectors
  • Nonlinear operations: activation functions (ReLU, sigmoid) applied electronically
  • Control: programming phase shifters and orchestrating the computation
  • Memory: storing weights, inputs, and intermediate results — practical photonic memory just doesn't exist

The integration approaches break down roughly like this:

Monolithic integration: photonic and electronic components on the same silicon die. Tightest possible coupling, but you compromise both fabrication processes to get there.

Heterogeneous integration: photonic and electronic dies fabbed separately, then bonded together in one package. Each die gets optimised on its own, at the cost of packaging complexity.

Co-packaged optics: photonic chips dropped onto the same interposer as the electronic processor, connected over short electrical links. This is where the data-centre industry is currently heading.

The level of integration directly drives the energy story. Send data the long way round between optical and electrical domains and the conversion overhead eats your optical savings whole.

In the next lesson, we will examine how photonic circuits are used specifically for neural network inference - optical neural networks that perform the core matrix multiplications of deep learning at the speed of light.