Frequency Combs for Computing

Every WDM computing scheme in this course — the crossbars, the weight banks, last lesson's programmable fabrics — quietly assumed a supporting cast nobody priced: one laser per wavelength channel. Sixty-four channels means sixty-four lasers, each with its own driver, its own temperature servo holding it on grid, its own slow drift pulling it off again — a rack of equipment propping up a chip. This lesson replaces the rack with a single ring of glass a hundred microns across. Pump one microresonator with one continuous-wave laser and, through the Kerr nonlinearity, it spontaneously erupts into a frequency comb: hundreds of new wavelengths, spaced with metronomic exactness, phase-locked to each other, all streaming out of one waveguide. One knob to rule every channel. For photonic computing this is the difference between a laboratory demo and a light engine you could actually ship — which is why the headline photonic-AI experiments of the 2020s all have a comb hiding in their block diagram.

A Fourier series made of glass

You already own the mathematics of combs — it is the theory of complex Fourier series wearing safety glasses. A signal that is periodic in time with period T has a spectrum that is discrete: delta-like lines at exact multiples of 1/T,

E(t) \;=\; \sum_{n} c_n\, e^{\,2\pi i\,(f_0 + n f_{\mathrm{rep}})\,t}, \qquad f_{\mathrm{rep}} = \frac{1}{T}.

Inside a pumped ring resonator, the stable pattern that forms is a dissipative Kerr soliton — a single short pulse of light circulating the ring forever, held together by a double bargain: the Kerr nonlinearity cancels the dispersion that would spread the pulse, and parametric gain from the pump replaces the light lost each lap. Every round trip, time T_R, the pulse passes the output coupler and a copy leaks out: a perfectly periodic pulse train. Its spectrum is therefore forced to be a comb with line spacing f_{\mathrm{rep}} = 1/T_R — the ring's free spectral range. The comb's exquisite regularity is not precision engineering; it is the Fourier theorem acting as physical law. And because a short pulse needs many Fourier components, the comb is automatically broad: a soliton a few tens of femtoseconds long spans terahertz of spectrum, with the smooth \mathrm{sech}^2 envelope that is the soliton's spectral signature.

Comb arithmetic

Two numbers specify a comb — the line spacing and the span — and both come from geometry you can compute. The spacing is the ring's FSR. A silicon-nitride ring of radius 100\ \mu\mathrm{m} has circumference L = 2\pi R \approx 628\ \mu\mathrm{m}; with group index n_g \approx 2.1,

f_{\mathrm{rep}} \;=\; \frac{c}{n_g L} \;=\; \frac{3\times10^{8}}{2.1 \times 628\times10^{-6}} \;\approx\; 227\ \mathrm{GHz}.

Want the denser 100 GHz telecom grid instead? The spacing is inversely proportional to the circumference, so grow the ring to L \approx 1.43\ \mathrm{mm}. Bigger ring, finer comb — the trade is that simple. The channel count follows by division: a comb spanning 8\ \mathrm{THz} of usable bandwidth at 100\ \mathrm{GHz} spacing supplies 8000/100 = 80 channels; halve the spacing and you double the count. Each line sits at f_n = f_0 + n f_{\mathrm{rep}} — two knobs (offset and spacing) pin down hundreds of frequencies at once, which is exactly why a comb-fed weight bank never needs per-channel wavelength servos.

Now bolt the comb onto the machinery of Module 5 and count operations. Comb lines become the input vector: N lines, each modulated with one activation. Fan the ensemble across M weight rows — microrings or last lesson's phase-change cells — with a photodetector summing each row. Every modulation interval, B times per second, the crossbar completes N \times M multiply–accumulates:

\text{throughput} \;=\; 2\,N\,M\,B \ \ \text{ops/s} \qquad\text{e.g.}\quad 2 \times 64 \times 16 \times 10^{10} \;\approx\; 2\times10^{13} \ \text{ops/s},

twenty tera-ops from one pump laser and one ring. This is not hypothetical bookkeeping: in 2021 two landmark experiments — a Münster–Oxford–Exeter phase-change crossbar and a Swinburne group's time-interleaved convolver — used exactly this comb-driven architecture to demonstrate convolutional processing at multi-TOPS rates, the fastest photonic computation shown to that date. Demonstrated: comb-driven matvec and convolution engines at tera-op scale, turnkey soliton combs packaged with their pump. Speculative: full accelerators where combs feed thousands-of-channel meshes, and the co-integration of comb, modulators and detectors on one chip at a competitive energy per op.

Frequency combs were not invented for computing — they were invented for measuring. In the late 1990s, John Hall and Theodor Hänsch realised that a mode-locked laser's comb is a ruler drawn on the frequency axis: to measure an unknown optical frequency (~10¹⁴ Hz, far too fast for any electronic counter), you beat it against the nearest comb line and read off the low-frequency difference — reducing the hardest measurement in physics to counting with a ruler whose ticks are known to fifteen digits. Optical atomic clocks, tests of whether fundamental constants drift, and the hunt for exoplanets via centimetre-per-second stellar wobbles all run on comb rulers; Hall and Hänsch shared the 2005 Nobel prize. The computing connection arrived through miniaturisation: in 2007 Tobias Kippenberg's group showed a millimetre glass microresonator could generate a comb via the Kerr effect alone — no mode-locked laser, no optical table. Chip-scale combs went looking for jobs beyond metrology, and photonic computing — hungry for exactly "many perfect wavelengths, cheap" — was standing first in line.

"Hundreds of channels from one laser" invites the fantasy of hundreds of lasers' worth of power. Run the energy audit. A bright-soliton microcomb converts pump light into comb lines with single-digit-percent efficiency, and that trickle is split across every line: tens of milliwatts of pump can become mere microwatts per channel. Downstream, each weak channel must still clear the detector's noise floor for your target precision — the analog-precision arguments of Module 5 apply per line, and they often force optical amplifiers back into the very system the comb was meant to simplify. Second myth: the comb does not eliminate per-channel electronics. Every line still needs its own modulator and, ultimately, its own detection path — the comb replaces the laser array, not the I/O. What it buys is precious but specific: one pump, rigid phase-locked spacing with no per-channel servos, and phase coherence across all lines. What it does not buy is watts. Darker-soliton and pulse-pumped designs are clawing the efficiency up — but check the per-line power in any comb-computing paper before believing its throughput headline.

Where this goes next

Spiking lasers, atomic-scale memory, software-defined fabrics, and now a light engine of hundreds of channels — this module has stayed on the waveguide. The final emerging lesson abandons even that: no waveguides, no ring, no chip — a neural network 3-D printed as a stack of frosted plastic layers, computing on light in free space, with no power supply at all.