The Photonic Computing Landscape

This module has assembled a worldview in five steps: computing is throttled by power and data movement; light is a phenomenal carrier with five knobs to encode on; the photon's very virtues make it a hopeless switch; the energy bill of optical signalling is all negotiable engineering; and history punishes anyone who asks light to imitate a transistor. One task remains: turn the worldview into a map. "Photonic computing" is not one field — it is at least half a dozen, with different physics, different competitors and wildly different maturity. Confusing them is the single most common analytical error in the area (and in its press coverage). This lesson draws the taxonomy along three axes and then lays the rest of the course onto it.

Three axes

Axis 1 — analog vs digital. Does the system compute with continuous physical quantities (field amplitudes, phases), or with restored, discrete logic levels? Analog photonics inherits interference's effortless linear algebra but pays in noise and precision — recall the 3-dB-per-bit law from the encoding lesson. Digital optical logic must clear Keyes' bar of gain, cascadability and restoration — the bar it historically failed.

Axis 2 — compute vs interconnect. Does light transform the data, or transport it? Interconnect asks nothing of light that light doesn't love doing — which is why it is the one photonic product already shipping in enormous volume.

Axis 3 — classical vs quantum. Classical photonics computes with fields — many photons, wave physics, no fragile quantum states. Quantum photonics encodes information in individual photons or squeezed states, and computes with interfering probability amplitudes — a different machine for different problems, sharing the same chips and couplers.

The map

FamilyQuadrantWhat light doesModuleStatus
MZI meshes / optical NNsanalog · computematrix–vector products by interference4–5startups, lab-to-market
Fourier / 4f processorsanalog · computetransforms and convolutions via lenses6venerable, niche
Ising machines, reservoirsanalog · computephysical dynamics as optimisation/inference6research
All-optical logicdigital · computenonlinear switching of light by light7largely abandoned
WDM links, co-packaged opticsdigital · interconnecttransporting bits between chips and racks8shipping in volume
Photonic quantum computingquantum · computeinterfering single-photon amplitudes9heavily funded research
Neuromorphic / emerginganalog · computespiking dynamics, phase-change weights, combs10research

Two readings of the map are worth memorising. Reading one: maturity increases toward the interconnect row — the less you ask light to decide, the sooner you can buy the product. Reading two: every viable compute entry is linear-algebra-shaped — meshes, Fourier engines, even quantum interference are all "propagate and superpose", never "switch mid-flight". The map is this module's one-sentence physics, drawn as geography.

The accelerator's fine print: Amdahl

One more instrument belongs in the toolkit before the course descends into hardware. Every photonic compute system on the map is an accelerator bolted to an electronic host — and accelerators are governed by Amdahl's law. If light speeds up only the fraction f of the workload (say, the matrix products), the overall gain is capped no matter how fast the optics:

S \;=\; \frac{1}{(1-f) + f/s} \;\xrightarrow{\;s\to\infty\;}\; \frac{1}{1-f}.

A workload that is 90% matrix multiplication caps at 10×, even with infinitely fast photonics — and the data must still cross the electronic–photonic border, at conversion costs the energy lesson priced. This single formula disciplines every glossy claim in Modules 5, 6 and 11:

function speedup(f: number, s: number): number { return 1 / ((1 - f) + f / s); } for (const f of [0.5, 0.9, 0.99]) { const line = [4, 100, 1e6].map((s) => "s=" + s + " → " + speedup(f, s).toFixed(2) + "×"); console.log((f * 100) + "% accelerable: " + line.join(" ") + " cap " + (1 / (1 - f)).toFixed(0) + "×"); }

Note the shape: pushing s from 100 to a million barely moves the answer. For an accelerator, the leverage is in f and in the border crossing — which is why Module 11 devotes a whole lesson to the electronic–photonic interface.

No machine on this map — not one — is a computer in the sense of your laptop. Every real photonic system is a photonic organ inside an electronic body: electronics holds the memory (light cannot be parked, as the photon's zero rest mass decreed), runs the control flow, makes the nonlinear decisions, and converts data across the border in both directions. This is not a temporary embarrassment to be engineered away; it follows from the physics of Module 1 and the history of the 1990s. So when a headline says a photonic chip "runs a neural network", the precise claim is that it performs the linear layers — and Amdahl's law plus the border tax decide how much that is worth. The practitioners' term of art is hybrid, and the sooner "photonic computing" is read everywhere as "electronic–photonic computing", the fewer disappointed investors the field will produce. The course's systems module returns to this with benchmarking rules for honest comparisons.

Follow the money and the map lights up unevenly. Optical transceivers for datacenters — the digital-interconnect corner — are a market worth tens of billions of dollars a year, growing ferociously with AI clusters, and every hyperscaler buys them by the container-load. That corner funds the fabs, the packaging lines, the design tools and the trained engineers that every other quadrant borrows. The analog-compute startups have raised billions in venture capital but ship in small volumes; quantum photonics runs on strategic and venture funding with revenues far in the future. The historical rhyme is exact: as telecom incubated the components that enabled the 2010s return, today's interconnect boom is incubating the platform — cheap lasers, dense modulators, co-packaging — on which the compute quadrants will stand or fall. In photonics, the boring quadrant has always been the patron of the exciting ones.

The road through the map

Module 1's work is done: you can argue from carrier physics, particle properties, energy ledgers and history to a defensible map of the field. The course now walks the map bottom-up. Module 2 puts light on a chip — waveguides, fibre, the silicon platform. Module 3 builds the component zoo: the Mach–Zehnder interferometer, modulators, detectors, on-chip lasers. Modules 4–6 occupy the analog-compute quadrant, from beamsplitter matrices to Ising machines; Module 7 performs the digital post-mortem; Module 8 tours the quadrant that ships; Module 9 climbs the quantum axis to boson sampling and beyond; Modules 10–11 cover the emerging fringe and close with a full-system capstone. First stop: how do you trap a beam of light inside a wire of glass?