Computing with Light
Optical & Photonic Computing · Frontiers track
7 min read
In this module
Photons vs. Electrons
Every computer we've looked at so far computes with electrons — charges shoved through transistors, along copper, across silicon gates. Electrons are the workhorse, and the entire semiconductor industry is built around bullying them around. They have hard limits, though: they dissipate heat in resistive materials, they leak electromagnetic fields into their neighbours, and they amble along through wires at well below the speed of light.
Photons play by completely different rules. They travel at light speed (about 3 x 10^8 m/s in vacuum, around 2 x 10^8 m/s in optical fibre). Under normal conditions, they ignore each other entirely — two beams can cross and walk away unaffected. They carry massive bandwidth via wavelength-division multiplexing (multiple colours sharing one fibre). And many linear operations, matrix multiplication chief among them, happen as light passes through an optical component — effectively for free, at the speed of light.
Optical and photonic computing chases one question: can we use these properties to build computers that are faster, more energy-efficient, or simply more capable than their electronic cousins?
A Brief History of Optical Computing
Computing with light isn't a new idea. In the 1960s and 70s, researchers built optical processors that did Fourier transforms, correlations, and pattern matching using lenses and spatial light modulators. The clever observation: a plain lens performs a Fourier transform on an image — O(N log N) work digitally, instant as light passes through the glass.
These early analogue optical machines were table-sized, finicky to align, hard to reprogram, and they couldn't keep up with how fast digital electronics were improving. By the 1990s, optical computing had drifted out of the mainstream conversation.
What flipped the script was the rise of silicon photonics — building photonic devices on silicon wafers using standard semiconductor processes. Suddenly photonic components were small, cheap, and could share a chip with electronics. Optical computing got pulled back off the shelf.
Fundamental Properties of Photons for Computing
Speed of Light Propagation
Light in a waveguide moves at roughly 2 x 10^8 m/s — about two-thirds of vacuum c, throttled by the material's refractive index. An electrical signal in copper hits 0.5–0.7 c depending on the wire. Over short on-chip distances, photons aren't dramatically faster than electrons.
The lead grows over distance. Optical fibre carries signals over kilometres with negligible loss, which is why the internet's backbone is optical. Inside data centres, optical interconnects are eating electrical ones for chip-to-chip and board-to-board links — more bandwidth, less power, and zero electromagnetic interference.
No Electromagnetic Interference
Electrons in wires throw off electromagnetic fields that couple into their neighbours. The result is crosstalk, and as features shrink it becomes the gating constraint on how many signals you can run side by side.
Photons in waveguides don't care. Two waveguides can sit right next to each other and remain electromagnetically oblivious, provided they're not within evanescent-coupling range (a few hundred nanometres). Different wavelengths can also share a single waveguide with no interaction at all.
Wavelength-Division Multiplexing (WDM)
Perhaps the most powerful property of photons for computing: different wavelengths (colours) of light do not interact and can travel simultaneously through the same waveguide. This is wavelength-division multiplexing - the same principle that allows a single optical fibre to carry hundreds of independent communication channels.
For computing, WDM means that a single photonic circuit can process multiple independent data streams simultaneously, one per wavelength. This provides a parallelism dimension that has no direct analogue in electronic computing.
Single waveguide carrying multiple wavelengths:
λ1 (red) ─╲λ2 (green) ──→ Single Waveguide → ──→ λ1, λ2, λ3, λ4 (all independent)λ3 (blue) ─╱λ4 (violet) ╱Wavelength-division multiplexing hands photonic computing a free dimension of parallelism. A single waveguide carrying 64 wavelengths runs 64 independent computations side by side. Scale up by adding wavelengths — easy with tunable lasers — not by duplicating physical circuits the way electronics demands.
Linear Operations Are "Free"
When light passes through an optical element — a lens, a beam splitter, a waveguide coupler — it goes through a linear transformation. Amplitude and phase get modified by the element's properties, and the transformation happens as a side effect of propagation. No clock, no extra energy beyond keeping the element there.
Matrix multiplication, the core operation in neural nets and a heap of scientific computing, is linear. Configure your optical elements just right and a photonic circuit performs a matrix-vector multiply in one pass, at light speed.
That's the punchline behind optical neural networks: the most expensive operation in deep learning is essentially free in the optical domain.
Basic Optical Logic
Can light do AND, OR, and NOT? Yes, but it's a lot more painful than in electronics.
Linear Optical Logic
Light hitting a beam splitter interferes with itself. Get the phases of the input beams right and constructive interference (beams add) versus destructive interference (beams cancel) gives you logic:
- Optical NOT: interfere the input with a reference beam of opposite phase
- Optical AND/OR: use nonlinear optical effects or threshold detectors
The catch: pure optical logic needs nonlinearities to be general, and optical nonlinearities are feeble. Most demand a lot of optical power before they kick in.
Electro-Optic Hybrid Logic
In practice, optical systems split the work. Light handles the linear stuff (data transport, matrix multiplication); electronics handle the nonlinear stuff (activations, comparisons, control flow). Data hops between domains at the boundaries:
- Modulators turn electrical signals into optical ones (encoding data onto light)
- Photodetectors go the other way (measuring light intensity)
Modern silicon photonic modulators run at 50–100 GHz, and photodetectors are even quicker.
Optical Interconnects: Where Photonics Already Dominates
Even before optical computing matures, photons have already won one battle outright: data communication. The internet's backbone is optical fibre, pushing terabits per second across thousands of kilometres.
Inside data centres, optical interconnects are crowding out copper for rack-to-rack and increasingly chip-to-chip links. Why:
- Higher bandwidth: a single fibre can carry terabits per second under WDM
- Lower power per bit: optical signalling beats electrical past a few centimetres
- No EMI: optical links shrug off the RF mess of a dense server rack
- Lighter and smaller: a fibre cable beats an equivalent-bandwidth copper bundle on every dimension
Broadcom, Cisco, and Intel are all shipping silicon photonic transceivers that hop between electrical signals on the chip and optical signals heading off it. The next leap is co-packaged optics — putting photonic transceivers right on the processor or switch package, killing off the last few centimetres of electrical signalling.
Industry consensus right now: photons own data movement and specific linear operations like matrix multiplication; electrons own general-purpose compute, storage, and anything nonlinear. The practical photonic systems are therefore hybrid — photonics where it wins, electronics everywhere else. A fully optical general-purpose computer remains a long way off.
The Promise of Photonic Computing
The potential payoff is substantial. Speed: matrix multiplication at the speed of light, capped only by how long the photons take to cross the chip — picoseconds. Energy efficiency: the linear computation doesn't generate resistive heat, so the energy bill lives in the modulators and photodetectors, not in the multiplication itself. Bandwidth: WDM gives photonics a free dimension of parallelism, letting one chip chew through dozens of independent data streams in the same pass. And scalability: bigger matrices mean a bigger photonic circuit, but the wall-clock time stays constant — light propagation doesn't care about matrix size, so long as you stay inside the coherence length.
In the next lesson, we will examine how these properties are realised in photonic integrated circuits built on silicon wafers.