Programmable Photonics

Digital electronics offers two ways to get a chip: design an ASIC — years, millions of dollars, one function forever — or buy an FPGA, whose sea of LUTs and programmable routing becomes your circuit an hour after you describe it. Photonics, so far, has only had ASICs. Every chip in this course — the modulator, the weight bank, the Clements mesh — was drawn as fixed geometry and fabricated for one job; change your mind and you are back in the foundry queue for months. Programmable photonics is the field's bid for its own FPGA: a general-purpose chip covered in a regular lattice of identical tunable elements, meaning nothing in particular when it ships, becoming a filter, a delay line, a switch matrix or an interferometer — or several at once — when software writes phase settings into it. Same fabric today, different instrument tomorrow, no foundry involved.

The fabric: a lattice of tunable couplers

The building block is the tunable basic unit (TBU): a Mach–Zehnder interferometer with a phase shifter in each arm, acting as a 2×2 coupler whose splitting ratio and phase are both set electrically. Three settings matter. Bar: light stays in its own waveguide — the TBU is a wire. Cross: light swaps waveguides — the TBU is a crossing. Partial: anything in between — the TBU is a beamsplitter with a programmable ratio. Tile TBUs edge-to-edge into a lattice of closed cells — hexagonal in the most-studied designs — and you get a photonic fabric in which every edge is a programmable coupler:

The crucial structural difference from the Reck and Clements meshes of Module 4 is that this lattice contains loops. A feed-forward triangle of MZIs can only implement transformations where light marches from inputs to outputs; it can never make light circulate. The hexagonal fabric can — and the moment a fabric supports feedback paths, a whole class of circuits that were impossible before falls out of it: resonators, cavities, and the sharp spectral filters they enable.

Synthesis: three instruments from one fabric

Suppose each TBU is 300\ \mu\mathrm{m} of waveguide with group index n_g = 4.2, and let's build things purely by choosing states.

A delay line. Set twenty TBUs along a snaking path to bar state: L = 20 \times 300\ \mu\mathrm{m} = 6\ \mathrm{mm} of guided path, a delay of

\tau \;=\; \frac{n_g L}{c} \;=\; \frac{4.2 \times 6\ \mathrm{mm}}{3 \times 10^{8}\ \mathrm{m/s}} \;\approx\; 84\ \mathrm{ps},

programmable in ~4 ps steps by lengthening or shortening the snake — the tapped, adjustable delay that microwave photonics and beamforming systems pay dearly for as custom chips.

A ring resonator. Close the six TBUs around one hexagonal cell into a loop: perimeter 6 \times 300\ \mu\mathrm{m} = 1.8\ \mathrm{mm}, giving a free spectral range

\mathrm{FSR} \;=\; \frac{c}{n_g L} \;=\; \frac{3\times10^{8}}{4.2 \times 1.8\times10^{-3}} \;\approx\; 40\ \mathrm{GHz},

with the coupling into the loop — which sets the filter's sharpness — dialled by the partial state of one boundary TBU. Merge two cells into a twelve-TBU loop and the FSR halves: resonator size is a software parameter.

An interferometer. Split light along two different mesh paths and recombine it, and you have synthesized an unbalanced MZI whose path imbalance — hence filter period — is chosen in units of whole TBUs. The published demonstrations (Pérez et al., 2017, on a seven-cell hexagonal mesh) ran through over twenty distinct circuits on one chip: FIR and IIR filters, delay lines, splitters, and simultaneous independent functions on different corners of the same fabric.

Software-defined photonics

Once the hardware is generic, the intelligence moves into software, and the field borrows its vocabulary wholesale from electronics: a netlist (the circuit you want), a place-and-route step (graph algorithms mapping it onto healthy, unoccupied TBUs), a bitstream (the vector of phase settings driven onto the chip). The fabric even supports the photonic version of self-test: built-in monitor photodiodes let control loops discover each TBU's actual calibration, route around defective units, and hold circuits on target as the chip drifts thermally — David Miller's self-configuring mesh ideas put to work. This is why the field calls itself software-defined photonics: the deliverable is not a chip but a chip plus a compiler.

Frontier accounting, as ever. Demonstrated: university and startup fabrics of tens of cells (the field's first commercial general-purpose processor, iPronics' SmartLight, shipped from the Valencia group's work), reprogrammed in milliseconds among dozens of RF-photonic functions. Speculative: the thousand-cell fabrics that would make "FPGA of light" literally true — blocked today by accumulated loss, by the electrical I/O of driving thousands of phase shifters, and by the control software's scaling.

A mesh cell could be triangular, square or hexagonal — the three regular tilings of the plane — and the choice was a genuine research question. Triangles give the shortest possible loops (three TBUs), hence the largest synthesizable FSR, but their sharp 60° bends are unkind to waveguides and their routing options tangle quickly. Squares are tidy but support fewer distinct path lengths per area. The Valencia group's analysis (Pérez, Capmany et al.) found the hexagonal lattice the sweet spot: gentle 120° junctions, the richest menu of synthesizable delays, loops and couplings per unit of chip area, and graceful scaling as cells are added. There is a pleasing rhyme here — biology's honeycombs win on material per unit of enclosed area, and photonics' honeycombs win on circuits per unit of silicon. Hexagons, it seems, are simply good engineering, whoever the engineer is.

The FPGA analogy is irresistible and this course has leaned on it — but know exactly where it breaks. A LUT does not merely route its input; it regenerates it. Every stage of an FPGA snaps voltages back to clean digital rails, so a signal can traverse ten thousand LUTs and emerge perfect. The photonic mesh has no such mercy: it is a passive, analog fabric. Each real TBU costs ~0.3–0.5 dB of loss, so a twenty-TBU circuit loses a couple of dB and a thousand-TBU circuit would lose hundreds — there is no optical "level restorer" to call on, only amplifiers that add noise. Phase errors likewise accumulate rather than reset, which is why calibration is a control loop and not a one-off. The flip side of passivity is a genuine superpower the FPGA lacks: the fabric is transparent — it does not know or care what waveform passes through, so a single configuration carries 100 Gb/s data, analog RF and anything else inside its bandwidth untouched. Analog transparency and digital restoration are a trade, not a ranking — but never claim the mesh scales like the FPGA does. It does not, and loss is the reason.

Where this goes next

A programmable fabric routes and filters whatever light you feed it — but feeding it remains the expensive part: every wavelength channel wants its own laser. The next lesson conjures hundreds of perfectly spaced, phase-locked channels out of a single ring and a single pump — the frequency comb — and turns them into the multi-channel engine that photonic tensor processors run on.