Why Optical Logic Lost to CMOS

In 1985 — with optical bistability on magazine covers and "the optical computer" drawing serious funding — Robert Keyes published a quantitative comparison of optical and electronic logic and concluded, calmly, that electronics would win and keep winning. The field's peak was still five years away; the verdict has held for forty. This lesson reconstructs that argument with modern numbers, because the reasons optical digital logic lost are not accidents of 1980s engineering. They flow from constants of nature — the photon's energy, the photon's size, and the photon's refusal to stand still — colliding with the criteria for logic, while the competition rode Moore's law and Dennard scaling down an exponential for six decades. Treat what follows as a post-mortem conducted with respect: the corpse was a serious idea, examined honestly.

Count one: the photon is an expensive particle

A telecom photon at \lambda = 1.55\ \mathrm{\mu m} carries E = hc/\lambda \approx 0.8\ \mathrm{eV} = 1.28\times 10^{-19}\ \mathrm{J} — about thirty times the room-temperature thermal energy kT \approx 0.026\ \mathrm{eV}, and (not coincidentally) the same neighbourhood as a semiconductor band gap. And a logic level can never be one photon: light arrives in Poissonian clumps, so a level read against shot noise with error rates worthy of digital logic (10^{-15}-ish, remembering the 10^{15} operations per second) needs thousands of photons per decision once real detectors, contrast ratios and margins are counted. The floor arithmetic:

const h = 6.626e-34, c = 3e8, q = 1.602e-19; const eV = (lambdaNm: number): number => (h * c) / (lambdaNm * 1e-9) / q; console.log("photon at 1550 nm: " + eV(1550).toFixed(2) + " eV"); console.log("photon at 1310 nm: " + eV(1310).toFixed(2) + " eV"); // An optical logic level: N photons at 1550 nm. for (const N of [1e3, 1e4]) { console.log(N + " photons = " + (N * eV(1550) * q * 1e15).toFixed(2) + " fJ"); } // A CMOS gate: E = C V^2 (charge and discharge), tiny C, sub-volt V. const C = 0.1e-15, V = 0.75; console.log("CMOS node, C=0.1 fF at 0.75 V: " + (C * V * V * 1e15).toFixed(3) + " fJ");

Read the punchline twice: a thousand-photon optical level costs about 0.13\ \mathrm{fJ} — roughly the same as switching a modern CMOS node. Optics' floor and electronics' practice already coincide, before the optical gate has paid for its nonlinearity (last lesson: tens of femtojoules to picojoules), its laser's wall-plug efficiency (a divide-by-3 to divide-by-10), or its detection and re-modulation. Worse, the two costs scale differently. Dennard scaling spent decades shrinking CV^2 by shrinking both C and V; the photon's hc/\lambda is a constant of nature, and moving to shorter \lambda makes each photon more expensive, not less. Electronics was an energy curve pointed downward; optics was a floor.

Count two: the photon is a large particle

Light cannot be confined much below half its wavelength — the diffraction limit. At 1.55\ \mathrm{\mu m} that is \lambda/2 \approx 775\ \mathrm{nm} in air, or about 220\ \mathrm{nm} inside silicon (n \approx 3.5) — and a practical photonic component is far bigger than the bare limit: single-micron waveguide pitches, five-micron bend radii, hundred-micron interferometers. A 2020s transistor has a gate pitch near 50\ \mathrm{nm} and packs >10^{8} devices per square millimetre; photonic "gates" manage 10^{2}– 10^{3} in the same area. That is four to six orders of magnitude of density — which is also four to six orders of magnitude of wiring length, capacitance, latency and cost. The chart puts the two length scales on one axis, on a log scale, across the decades:

The crossing happened around 1990 — poetically, the very moment digital optical computing peaked. Every year since, the transistor has pulled further below anything a guided photon can match, and the gap is now roughly the gap between a house and a doormat. Note also the shape of the escape routes: plasmonics can squeeze light below the diffraction limit by riding it on electron oscillations in metal — at the price of resistive loss so severe that the signal dies in microns, which is to say, by becoming electronics.

Counts three and four: nothing stands still, nothing pushes back

Two final counts, quicker but just as fatal. Storage: computing is not only logic but state — latches, registers, caches. Electronics stores state effortlessly because charge stays put: a DRAM cell is one transistor and a capacitor; an SRAM cell, six transistors holding each other's hands. A photon, by contrast, travels at c/n always; it can be delayed (a kilometre of fibre buys 5 µs) or absorbed (destroying it), but there is no optical capacitor — no way to park a photon and read it later. Every "optical memory" is a loop that must be refreshed, a cavity that leaks in nanoseconds, or a conversion into matter — at which point it is electronic (or atomic) memory with optical I/O. Gain and restoration: the transistor is a power amplifier with built-in isolation, manufactured by the billion for nothing; optical gain means a pumped medium — a laser amplifier per gate — with spontaneous-emission noise and no inherent isolation. The scorecard, then, on the criteria that the last lesson established actually decide the contest:

Criterion	CMOS transistor	All-optical gate
Switching energy	~0.05–1 fJ, still falling	fJ floor; fJ–pJ in practice
Device pitch	~50 nm	~1–100 µm (λ-bound)
Density	10⁸/mm²	10²–10³/mm²
Gain + isolation	intrinsic, free	pumped amplifier per gate
Level restoration	every gate, automatic	rare; needs engineered sigmoid
State storage	capacitor: 1T + C	none — light won't stand still
Manufacturing	10¹⁰ devices/chip, matched	critically-biased resonators

Not one row favours photons. This is why the loss was not a photonics failure to try harder: on the terms by which digital logic is scored, the electron is simply the better particle — small, cheap, storable, and blessed with the transistor.

The era got its moon-landing moment in January 1990, when Alan Huang's group at Bell Labs unveiled the first digital optical processor — racks of lasers and lenslet arrays clocking data through arrays of S-SEED devices, quantum-well switches invented at the same labs by David Miller. It made the evening news as the dawn of light-speed computing. Read today, it is a beautifully honest experiment that demonstrated the opposite of its billing: the processor ran at kilohertz-to- megahertz rates, dissipated more than the electronics it emulated, and its S-SEEDs were not even all-optical — they were photodiodes and quantum-well modulators, electronic devices wearing optical I/O. The deepest irony is who drew the lesson. Miller, whose SEED was the best optical logic device ever built, spent the following decades as the field's clearest sceptic — his criteria paper "Are optical transistors the logical next step?" (2010) is the definitive "no" — while redirecting optics toward what it demonstrably does win: interconnects. The 1990 machine's true legacy was teaching a generation, at industrial expense, exactly where the boundary between optics and electronics belongs — the subject of the next lesson.

This lesson's conclusion is routinely over-read, in both directions. What died was a specific proposition: replacing the transistor with an optical gate in general-purpose digital logic — Boolean operations, deep cascades, dense state. What did not die is everything this course spent six modules building, precisely because none of it fights the criteria: optical interconnects carry bits without needing gain or restoration at every stage (one amplifier per kilometre, not per gate); Mach–Zehnder meshes compute matrix products with linear optics, no nonlinearity per operation, with electronics handling thresholds and memory; analog Fourier processors, Ising machines and boson sampling all exploit propagation physics rather than cascaded switching. When you hear "optical computing failed", the honest translation is "optical digital logic failed its physical exam in the 1980s, and the field grew up and specialised". Conversely, when a startup deck promises "all-optical AND gates will replace CPUs", every number in this lesson still applies — the photon has not become smaller, cheaper or more stationary since 1985.

Where this goes next

Keyes' argument aged well because it was built on invariants, and it is worth carrying beyond photonics: the same audit — energy per operation against the Landauer floor in numbers, density, gain, restoration, storage — is how you should greet every candidate logic technology for the rest of your career. But a post-mortem is not an obituary for the whole family. Limited, criteria-compliant slivers of optical signal processing did survive in production systems, doing jobs electronics could not reach — and the final lesson of this module tours those survivors and extracts the partitioning principle that governs every photonic system that actually ships.