🌲

Woodland Wizard

Photonic Computing

Light has six independently addressable degrees of freedom: wavelength, amplitude, phase, polarization, orbital angular momentum, and spatial mode. An electrical wire carries one bit per clock cycle. A single optical channel carries a vector across all six dimensions simultaneously. This is the basis for photonic computing.

Current State (2025-2026)

Optical interconnects are shipping now. TSMC's COUPE platform co-packages photonic and electronic dies. Lightmatter's Passage M1000 delivers 114 Tbps optical bandwidth. NVIDIA adopted it for Spectrum-X. Specialized photonic AI acceleration works: Lightmatter published in Nature (April 2025) running ResNet, BERT, and deep RL on a photonic processor at near-32-bit precision without model modification.

General-purpose optical computing does not exist. Every working system is hybrid: photonic compute, electronic memory and control flow.

How Photonic Computation Works

A Mach-Zehnder interferometer (MZI) splits a beam, phase-shifts one arm, and recombines. The phase difference determines the output ratio. A mesh of MZIs implements any unitary matrix transformation. Light propagating through the mesh performs matrix-vector multiplication in a single pass at the speed of light. No clock cycles per multiply. No carry chains. The mesh configuration (phase shifter settings) is the program.

Wavelength multiplexing scales this without additional hardware. A microcomb generates 100+ wavelengths. All enter the same mesh and compute independently. 100 wavelengths = 100 parallel matrix operations in one pass. The Shanghai/NTU group demonstrated this in 2025 (eLight): 100-frequency channel parallel computing with >0.9 matrix consistency across all channels.

Addition is beam combining (constructive interference). Negation is a pi phase shift. Multiplication is attenuation. These are physical processes, not logic gate sequences.

Four Unsolved Problems

Optical memory. Photons move at c. They do not sit in registers. Phase-change optical memories (GST on waveguides) fail after ~100K write cycles. Electronic memory handles quadrillions. Every optical computer converts to electronics for storage. A ferroelectric Pockels memory (2025) achieved 6 non-volatile states per cell at femtojoule energy. A 209-state photonic memory broke density records. Neither is mature.

Photon-photon interaction. Photons pass through each other. Electronic transistors are cheap because one electron stream easily controls another. Conventional optical nonlinearities are extremely weak. The Purdue single-photon switch (Nov 2025, Nature Nanotechnology) changes this: one photon triggers avalanche multiplication creating a refractive index change 15 orders of magnitude larger than silicon's conventional optical nonlinearity. Room temperature, CMOS-compatible, GHz speeds. Lab-stage only.

Cascading and fan-out. Optical signals lose power at every split. After a few logic stages the signal is too weak. Most optical neural networks delegate nonlinear activation to electronics because optical nonlinearities cannot cascade reliably.

Miniaturization floor. Light wavelength: ~1 micrometer. Transistor feature size: ~2 nanometers. Three orders of magnitude. Photonic chips compensate with parallelism (wavelength, mode, polarization multiplexing) rather than density.

Biological Approaches to Optical Memory

Bacteriorhodopsin (from Halobacterium salinarum) has two stable conformational states with different absorption spectra. Green light writes (bR to Q state). Blue light erases (Q back to bR). Low-intensity red light reads without altering state. Billions of write/erase cycles demonstrated. Q state stable for years at room temperature. Robert Birge's group built 3D optical memory cubes using two-photon absorption for volumetric addressing. The protein self-repairs in living systems, giving theoretically unlimited write endurance.

Photoswitchable fluorescent proteins (Dronpa, rsEGFP, Padron) switch between fluorescent and non-fluorescent states with specific wavelengths. Different variants respond to different colors. Multiple proteins in the same location enable wavelength-multiplexed storage: bits addressed by color, natively compatible with photonic computing's wavelength channels.

Phytochromes (plant photoreceptors) are bistable optical flip-flops. Red light sets one state, far-red light sets the other. Holds state indefinitely in darkness.

These are slow (microsecond-scale conformational changes) but the self-repair property is unique. No engineered material matches it.

The Brain's Architectural Lesson

The visual system is a working optical-to-persistent-storage pipeline. Rhodopsin (same protein family as bacteriorhodopsin) detects single photons. A G-protein cascade amplifies: 1 photon triggers ~1,000,000 molecular events. Ion channel modulation creates electrical signals. 130 million photoreceptors compress to 1 million optic nerve fibers. Synaptic weight changes (long-term potentiation and depression) store the result.

The key insight: in the brain, there is no separation between memory and computation. A synapse stores a weight and multiplies the input signal by that weight. Reading memory is computing with it. There is no bus. There is no von Neumann bottleneck.

This maps directly onto photonic computing. Phase-change material phase shifters in an MZI mesh store weights non-volatilely. Light propagating through the mesh reads those weights by computing with them. The mesh configuration is both the program and the memory. Intermediate results are photons in transit between stages. A deep enough mesh computes input-to-output in one propagation pass with no intermediate storage.

Neuromorphic photonic systems implementing this exist in labs: photonic synapses (phase-change-on-waveguide), photonic neurons (microring resonators with excitable dynamics), and spike-timing dependent plasticity demonstrated in photonic circuits.

Holographic Sphere as Bulk Storage

Volume holographic storage records interference patterns as refractive index changes in a photosensitive crystal. Different reference beam angles store different pages in the same volume (angular multiplexing). Different wavelengths at the same angle store additional independent pages (wavelength multiplexing).

A sphere is geometrically optimal for this. A flat slab accepts reference beams from limited angles (~1 steradian). A sphere accepts beams from any direction (full 4pi steradians, ~12.6 steradians). This provides roughly 12x more angular addresses. The IEEE demonstrated spherical holographic storage with angle multiplexing (paper 906895).

Address space of a 10cm sphere: ~10,000-100,000 angular addresses x 3-100 wavelengths x 2 polarization states. At Microsoft's demonstrated density of 9.6 GB/cm3 (Project HSD, Dec 2024, iron-doped lithium niobate), a 524 cm3 sphere holds ~5 TB before accounting for the angular advantage.

The data format is natively compatible with photonic computing. The same wavelength channels used for parallel computation address holographic memory pages. Write: computation output beams record interference patterns. Read: reference beam at the stored angle and wavelength reconstructs the data page. No optoelectronic conversion. No serialization.

Current status: Microsoft's Project HSD acknowledges 1-2 orders of magnitude improvement in energy efficiency needed before competing with electronic storage. HoloMem targets 200 TB per cartridge (ribbon format, not sphere) for 2027 mass production. SPhotonix demonstrated 360 TB per disc in nanostructured glass with femtosecond lasers. No one has built the sphere commercially. Fabricating optically perfect lithium niobate spheres and omnidirectional beam steering remain engineering challenges.

Realistic Architecture

A photonic computer is not a faster electronic computer. It is a different computational model: a dataflow machine where physics performs linear algebra and the architecture steers light.

Electronic host CPU (control flow, branching, memory management)
  |
  +-- Photonic accelerator (matrix math, neural inference)
  |     wavelength-multiplexed MZI mesh, teraMAC throughput
  |
  +-- Photonic interconnect fabric
  |     chip-to-chip, full wavelength-encoded payload over fiber
  |
  +-- Memory hierarchy:
        Tier 0: photons in flight (picoseconds, no storage)
        Tier 1: photonic SRAM latches (nanoseconds, volatile, small)
        Tier 2: phase-change waveguide memory (microseconds, non-volatile, moderate)
        Tier 3: holographic volume storage (milliseconds, massive capacity)
        Tier 4: bio-optical archive (seconds, self-repairing, unlimited endurance)

Optical interconnects are here now. Specialized photonic accelerators in 2-5 years. General-purpose optical computing depends on solving optical memory and is 2030s at the earliest.

Key Players