The Limits of Analog Optical Computing

This module has been a museum of near-miracles. A lens that computes a megapixel Fourier transform in a nanosecond. A photograph that multiplies complex matrices. Crystals holding terabits, fibre loops that learn, lasers that optimize. Every one of them real, demonstrated, published — and every one of them, from the Fourier engine to the Ising machine, stuck at the same four walls. This closing lesson names the walls precisely: dynamic range, drift, cascading error, and the I/O bottleneck. They are not engineering embarrassments awaiting a cleverer graduate student; they follow from what "analog" means. Understanding them turns you from a consumer of photonic press releases into someone who can predict — usually in one envelope-back calculation — which optical computing proposals can work and which are structurally doomed.

Wall 1: precision is a photon budget

A digital number carries as many bits as you write down. An analog optical value is a physical intensity, and its precision is its signal-to-noise ratio — a resource you must buy, photon by photon. Even with perfect components, light arrives in quanta with Poisson statistics: detecting N_{\text{ph}} photons gives shot-noise-limited \text{SNR} = \sqrt{N_{\text{ph}}}, so the effective resolution is

b \;\approx\; \log_2 \sqrt{N_{\text{ph}}} \;=\; \tfrac{1}{2}\log_2 N_{\text{ph}} \qquad\text{bits per readout.}

The logarithm is the brutal part. A million photons per result — a generous budget at speed — buys \text{SNR} \approx 10^3, about 10 bits. Each additional bit costs the photons (and 4× the energy) forever; 16 bits wants N_{\text{ph}} \approx 4\times10^9 per value, at which point the energy argument for optics has quietly evaporated. Real systems, with detector noise, crosstalk and component mismatch stacked on top of shot noise, typically demonstrate 4–8 effective bits. Now recall what floating-point representation hands you for free: a float32 carries a 24-bit significand — sixteen orders of magnitude beyond an 8-bit analog pipe in relative precision — and a float64 carries 53. Analog optics does not compete on precision. It cannot; the scaling law is against it.

Wall 2: errors cascade — and analog has no eraser

Digital logic's deepest virtue is not speed but restoration: every gate snaps its output back to a clean 0 or 1, discarding the noise of the previous stage. A digital error below half a level simply vanishes. Analog optics has no such eraser: every stage's noise, mismatch and distortion is faithfully amplified and passed on. With k cascaded stages of independent noise, error grows as \sqrt{k} — costing \tfrac12\log_2 k bits of precision — and correlated errors (a systematic gain miscalibration) grow linearly, worse. Watch the two philosophies diverge:

// A value passes through 30 processing stages, each adding a little noise. // Analog keeps the noise; digital quantizes (restores) after every stage. let seed = 11; const rand = (): number => { seed = (seed * 1664525 + 1013904223) % 4294967296; return seed / 4294967296; }; const noise = (sigma: number): number => sigma * Math.sqrt(-2 * Math.log(rand() + 1e-12)) * Math.cos(2 * Math.PI * rand()); const SIGMA = 0.002; // per-stage noise const LSB = 1 / 64; // a 6-bit digital grid const TRIALS = 400, STAGES = 30; const rms: { analog: number[]; digital: number[] } = { analog: [], digital: [] }; for (let k = 0; k < STAGES; k++) { rms.analog.push(0); rms.digital.push(0); } for (let t = 0; t < TRIALS; t++) { let a = 0.5, d = 0.5; for (let k = 0; k < STAGES; k++) { a += noise(SIGMA); // analog: noise accumulates d += noise(SIGMA); d = Math.round(d / LSB) * LSB; // digital: restored to the nearest level rms.analog[k] += (a - 0.5) ** 2; rms.digital[k] += (d - 0.5) ** 2; } } for (const k of [0, 4, 9, 29]) { const ra = Math.sqrt(rms.analog[k] / TRIALS); const rd = Math.sqrt(rms.digital[k] / TRIALS); console.log("after stage " + (k + 1) + ": analog RMS error " + ra.toFixed(4) + " digital RMS error " + rd.toFixed(4)); } console.log("analog error grows like sqrt(k); digital noise below LSB/2 is simply erased.");

This is why every successful analog optical processor is shallow: one transform, one mask, one readout. The 4f correlator is two lenses deep. Proposals that cascade tens of analog layers without intermediate restoration are asking \sqrt{k} to be kind; it never is. (Restoring optically at each stage is precisely the "optical transistor" problem — Module 7's story, and its graveyard.)

Walls 3 and 4: drift, and the toll booth at the border

Drift. Analog optical state is stored in physical quantities — path lengths, refractive indices, alignments — that move with temperature, age and vibration. A silicon waveguide's phase shifts measurably with millikelvin changes; a fringe pattern that stores your matrix detunes as the crystal breathes. Digital hardware drifts too, but restoration hides it; analog systems must instead measure and correct continuously, spending duty cycle, control electronics and engineering budget on standing still. Every machine in this module paid this tax: the hologram's Bragg condition, the reservoir's retrained readout, the Ising machine's interferometric feedback, the mesh calibration loops of Module 11.

The I/O bottleneck. The transform is free; the border crossing is not. Getting a number into the optical domain costs a DAC conversion and a modulator drive; getting it out costs photodetection and an ADC conversion — each of order a picojoule, a figure that improves at glacial pace compared with logic. The arithmetic of amortization decides everything. An N-sample transform with N^2-MAC equivalent work costs 2N conversions: per MAC, the I/O toll is 2/N picojoules — negligible for N = 10^3, which is why optical matrix engines quote femtojoules per MAC. But run a pointwise operation (N MACs for the same 2N conversions) and the toll is 2 pJ per MAC — worse than just doing it digitally. The rule of thumb of the whole field: analog optics wins exactly when the operations-per-conversion ratio is enormous, and every architecture in this module is an attempt to make that ratio big.

The verdict, in one table

Property Analog optics Digital electronics
Precision 4–8 effective bits; each extra bit costs 4× the photons Arbitrary — float32/float64 by convention
Cascade depth Shallow — error compounds as √k or worse Unlimited — every gate restores
Latency (fixed transform) Nanoseconds — time of flight Micro- to milliseconds
Parallelism Free over space and wavelength Paid for in cores, watts and wires
Programmability Limited, slow, drift-prone Total and instantaneous
Energy ~0 for the transform; I/O and laser dominate Per-operation, still falling

Read as a design rule rather than a scorecard: analog optics wins when the workload is a large, fixed, reused transform at low precision over massively parallel data — a first CNN layer, a Fourier preprocessor, a beam-forming front end, an Ising relaxation — ideally with the data already in optical form (a fibre link, an image) so one border crossing disappears entirely. It structurally cannot win where the workload needs high precision (the photon budget explodes), deep dependent chains (no restoration), frequent reprogramming (drift plus slow modulator updates), or control flow — branching on intermediate values forces a conversion at every branch, which hands the game to the ADC. There is no shame in this list; it is a job description. GPUs, too, are machines that lose at almost everything except the thing they are enormous at.

Why obsess over converters? Because they are the slowest-improving mainstream component in electronics. Surveys of thirty years of published ADCs (Walden's, then Murmann's) show the energy-per-conversion figure of merit improving by roughly a factor of two every few years — respectable, but nothing like logic's ride down the process-node curve — and pinned near hard physics: an ADC's comparators must resolve exponentially finer voltage differences per added bit against thermal noise kT/C, so cost rises steeply with both resolution and speed. This is not optics' problem to fix — it inherits it. Any analog accelerator, optical or electronic, that must digitise its results at high rate and high resolution is capped by converter physics, no matter how magical the middle. The systems that escape do so by not converting: keeping precision low (1-bit correlator outputs — a peak is there or it isn't), keeping results sparse (read only the argmax), or keeping the signal analog end-to-end (the reservoir equalizer feeding an analog decision circuit). When you evaluate a photonic-computing paper, find the ADC in the block diagram first. Its spec sheet, not the optics, usually sets the headline number.

The classic sleight of hand in analog-computing papers — optical and electronic alike — is the subsystem energy figure: "0.1 fJ per MAC" counting only the photons that traverse the modulator, while the laser's wall-plug inefficiency (multiply by 3–10×), the DAC and ADC conversions, the SLM refresh, the thermal stabilisation and the calibration duty cycle all live off the books. System-level accounting routinely moves headline numbers by two to three orders of magnitude, which is the difference between "revolutionary" and "roughly a GPU". The checklist for reading any claim in this field: (1) does the energy figure include I/O conversions and the laser at wall-plug? (2) is the precision at which the throughput is quoted the same precision at which the accuracy was measured? (3) is the workload's operations-per-conversion ratio one the target application actually has? (4) is the digital baseline contemporary hardware, tuned, at matched accuracy — or a strawman? Papers that survive all four exist, and they are the ones worth reading twice. The habit generalises far beyond photonics; it is how you read benchmarks, forever.

Where this goes next

The deepest wall was restoration: analog optics cannot clean up its own signals, so it cannot go deep. The obvious response is to ask whether optics can be made digital — discrete levels, restoring gates, an optical transistor. People have tried for fifty years, the attempt produced beautiful physics and one of computing's most instructive failures, and that story — nonlinear optics, optical logic gates, and why the all-optical computer lost to CMOS — is Module 7.