Optical Computing and Processing
Optical computing and processing encompasses technologies that use the properties of light to perform computation, signal manipulation, and data processing. Unlike electronic systems, which move charge through semiconductor circuits, optical systems carry information on photons. This enables massively parallel operations, very high bandwidths, and propagation delays limited only by the speed of light in the medium.
The field spans a wide range of maturity. At one end, all-optical signal processing is already embedded in telecommunications networks, where signals are amplified, switched, and reshaped without conversion to the electrical domain. At the other end, emerging photonic computing architectures aim to accelerate specific workloads such as neural network inference and combinatorial optimization. As electronic devices approach fundamental limits in clock speed and energy per operation, optical approaches offer one pathway toward continued performance scaling for the tasks that suit them.
This category covers the technologies, techniques, and applications that harness light for computation and signal manipulation, from established optical network functions to research-stage photonic processors. The subcategories below organize the topic from fully optical computation through hybrid optical-electronic systems, optical-domain signal processing, and light-based machine learning. A companion category, photonic and optical computing, approaches the same physics from the architecture side and describes the machines assembled from these devices; the pages here stay with the devices, the signal-domain techniques, and the processing functions themselves.
Articles in This Category
Fundamental Concepts
Advantages of Optical Processing
Light offers several intrinsic advantages for information processing. In a linear medium, photons do not interact with one another, so signals at different wavelengths can share the same component without crosstalk. This property underlies wavelength division multiplexing, in which dozens of channels travel down a single fiber on a standardized grid. ITU-T Recommendation G.694.1 anchors the dense WDM grid at 193.1 THz and defines fixed channel spacings from 12.5 GHz up to 100 GHz and integer multiples of it, along with a flexible grid whose spectral slots are sized in 12.5 GHz increments. The same wavelength dimension that carries independent traffic channels in a network can carry independent data streams into a photonic processor, giving optics a form of parallelism that has no direct electronic analogue.
Propagation occurs at the speed of light in the medium, and signals can travel through free space or guided waveguides, enabling three-dimensional interconnection topologies that planar electronic circuits cannot match. Optical links also avoid the resistive and capacitive charging losses that dominate electrical interconnect energy, so the cost of moving a bit becomes largely independent of distance beyond a few centimeters. Some mathematical operations map naturally onto optics as well: a single lens performs a two-dimensional spatial Fourier transform of the field in its front focal plane, producing the transform in its back focal plane without any active computation, and a passive interferometer mesh applies a matrix to a vector of optical amplitudes as the light simply propagates through it.
Nonlinear Optical Effects
Linear optics handles routing and filtering, but active signal processing and computation require nonlinear optical effects, in which one optical field changes the medium's response to another. Third-order effects dominate in glasses and centrosymmetric semiconductors: four-wave mixing, cross-phase modulation, self-phase modulation, and stimulated Brillouin and Raman scattering allow signals to interact and influence each other, enabling wavelength conversion, signal regeneration, all-optical switching, and logic operations. Second-order effects, available in non-centrosymmetric crystals such as lithium niobate and periodically poled lithium niobate, give efficient frequency conversion and parametric amplification at lower power.
The central engineering challenge is achieving useful nonlinear interaction at practical power levels. Because most materials have weak nonlinear coefficients, designers raise the effective interaction strength by concentrating optical power in a small mode area and holding it there for a long interaction length. Highly nonlinear fibers, semiconductor optical amplifiers, resonant structures such as microring and photonic-crystal cavities, and tightly confined integrated waveguides made from silicon, silicon nitride, chalcogenide glass, or thin-film lithium niobate all serve this purpose. Each choice trades nonlinear efficiency against loss, optical bandwidth, and unwanted side effects such as two-photon absorption and free-carrier absorption, which limit silicon waveguides at telecom wavelengths.
Optical-Electronic Integration
Most practical systems combine optical and electronic elements to exploit the strengths of each. Optical components excel at high-bandwidth transport and parallel linear operations, while electronics provide flexible control, dense storage, nonlinear thresholding, and complex decision-making. The absence of a low-power, cascadable optical equivalent of the transistor is the main reason general-purpose all-optical computers remain a research goal rather than a product. An electronic logic gate restores signal levels, provides gain, and drives many downstream gates from one output; no optical device does all three at comparable energy.
Interconnect is where hybrid integration has advanced furthest. Front-panel pluggable transceivers consume on the order of tens of picojoules per bit once their electrical retiming and drive circuitry are counted, and the electrical channel from the switch package to the front panel accounts for a large share of that. Co-packaged optics moves the optical engines onto the same substrate as the switch or processor die, shortening the electrical reach to millimeters; commercial co-packaged optics switch platforms began reaching the market in the mid-2020s. Monolithic and hybrid photonic-electronic integration on common substrates promises tighter coupling still, with the photonic layer fabricated in a silicon process and the driver and receiver electronics stacked or bonded directly above it. Choosing between all-optical and hybrid approaches, and deciding where to place the optical-electrical boundary, is central to system design: every crossing costs energy and latency, so a useful photonic accelerator must perform enough work between conversions to earn them back.
Encoding, Precision, and Noise
A photonic processor is an analog machine, so how numbers are represented determines what it can compute and how accurately. Values may be encoded in optical amplitude or intensity, in phase, in wavelength, or in time slots, and practical systems mix these dimensions. Intensity encoding is simple to detect but confined to non-negative values, which forces signed weights to be represented as differences between pairs of channels or as balanced photodetector outputs. Coherent encoding in field amplitude and phase handles signed and complex values naturally but demands stable optical path lengths, since a fraction of a wavelength of drift changes the result.
Accuracy is bounded by shot noise, detector and amplifier noise, crosstalk between channels, and drift in the programmable elements. Thermo-optic phase shifters, the most common tuning mechanism in silicon photonics, respond in microseconds and consume steady power, and they couple thermally to their neighbors; electro-optic shifters in lithium niobate or on carrier-depletion junctions switch far faster but usually cost more loss or area. Effective numerical precision in reported analog photonic processors typically falls well short of the 16-bit and 32-bit formats of digital hardware, which is why the workloads targeted first, such as neural network inference, tolerate low precision. Periodic calibration, on-chip monitor photodiodes, dithering, and digital error correction in the electronic layer all help hold accuracy over temperature and time.
Key Applications
Telecommunications Signal Processing
Optical networks perform a growing share of signal processing in the optical domain to avoid electronic bottlenecks. Erbium-doped fiber amplifiers boost every channel in the conventional band, roughly 1530 to 1565 nm, in a single device without per-channel conversion, and L-band variants and distributed Raman amplification extend the reach and the usable spectrum. Wavelength-selective switches route channels by wavelength inside reconfigurable optical add-drop multiplexers, allowing an operator to add, drop, or redirect a wavelength at a node under software control rather than by rewiring.
Optical dispersion compensation, using dispersion-compensating fiber or chirped fiber Bragg gratings, preserves pulse shape over long distances, and all-optical regeneration can restore signal quality without detection and retransmission. The balance here has shifted: since coherent transceivers with digital signal processing became standard on long-haul routes, chromatic dispersion and polarization-mode dispersion are usually corrected electronically in the receiver, and inline optical compensation is now most valuable on direct-detection links and in reach extension. Optical-domain functions that remain firmly in place are those electronics cannot replicate economically, above all broadband amplification and wavelength-granular switching, which together keep modern long-haul and metropolitan networks scalable.
Photonic Neural Networks
Optical implementations of neural networks exploit the parallelism and speed of photonics for machine learning inference. Matrix-vector multiplication, the dominant operation in neural networks, maps onto optics through interference and wavelength multiplexing: a mesh of Mach-Zehnder interferometers can apply a programmed weight matrix to an optical input vector, and microring-resonator banks can do the same by routing weighted contributions onto a shared waveguide for summation at a photodetector. Free-space and diffractive designs take a third route, using spatial light modulators or fabricated phase masks as the weight layer. The linear, analog multiply-accumulate happens essentially at the speed of light, while electronics supply the inputs, the nonlinear activation, and the output readout.
Progress has moved from single-layer proofs of concept to integrated processors that run recognizable models. Research demonstrations published in the mid-2020s combined photonic tensor cores with CMOS control chips through advanced packaging and executed standard networks such as convolutional image classifiers and transformer language models end to end, with the electronic layer handling activations and data movement. For suitable workloads, these accelerators promise low latency and potentially lower energy per operation than electronic hardware, and interest centers on data center and edge inference where batch sizes are small and latency matters more than throughput. Practical deployment still contends with analog noise, limited numerical precision, optical loss that grows with mesh depth, the energy cost of digital-to-analog and analog-to-digital conversion at the boundaries, and thermal sensitivity of the phase shifters. Weight reconfiguration speed is a further constraint: a mesh that takes microseconds to reprogram suits fixed inference weights far better than training.
Optical Sensing and Measurement
Optical processing enables high-speed measurement and characterization that electronics alone cannot reach. Optical sampling gates a signal with a train of ultrashort pulses, so the timing resolution is set by the pulse width rather than by the bandwidth of an electronic sampler, and effective rates far exceed those of electronic analog-to-digital converters. Optical spectrum analyzers resolve signal spectra with fine resolution across an entire communication band at once.
Time-lens systems apply a quadratic phase to stretch or compress signals in time. Slowing an ultrafast event by a large factor makes it slow enough for a conventional digitizer, which is the basis of time-stretch imaging, time-stretch spectroscopy, and single-shot capture of rare transient events that a triggered oscilloscope would miss. Optical frequency combs supply the stable, broadband reference these techniques depend on and have become instruments in their own right for precision spectroscopy and frequency metrology. Together these capabilities are essential for characterizing high-symbol-rate communication links and for scientific instrumentation that observes phenomena on picosecond and femtosecond timescales.
Specialized Computing Applications
Optics shows particular promise for specific computational problems, including combinatorial optimization, pattern recognition, and quantum simulation. Coherent Ising machines encode an optimization problem as the couplings of an Ising Hamiltonian and let a network of optical oscillators, commonly degenerate optical parametric oscillators circulating as pulses in a fiber ring, settle toward a low-energy state that represents a good solution. Couplings are applied by measuring each pulse and feeding back a computed drive, which yields all-to-all connectivity without physical wiring between every pair. A machine of this type built by NTT and the National Institute of Informatics and reported in 2021 held roughly 100,000 spin pulses, and unlike superconducting quantum annealers it operates at room temperature. Such systems produce good approximate solutions quickly rather than certified optima, so they compete with heuristics such as simulated annealing rather than with exact solvers.
Reservoir computing takes a different tack, using the rich transient dynamics of an optical system, such as a semiconductor laser with delayed feedback, as a fixed nonlinear feature map and training only a simple linear readout. Because the reservoir itself is never trained, the hardware may be fast, physically simple, and imperfect, which suits analog photonics well. Other specialized uses include optical correlators for pattern matching and photonic simulators for physics problems that share the mathematics of wave propagation. As these approaches mature, they are likely to complement, rather than replace, electronic computing, serving workloads that reward massive parallelism or ultrafast response.
Challenges and Outlook
Optical computing faces real obstacles that temper its advantages. Cascading optical logic without intermediate electronic regeneration is difficult, because passive optics provides no gain and each stage adds loss and noise. Optical memory remains limited compared with electronic storage: light must be kept moving, so holding a bit means circulating it in a delay line or trapping it in a bistable cavity, neither of which approaches the density or retention of a DRAM array. Analog photonic processors must further manage shot noise, calibration drift, and the energy cost of repeated conversions between the optical and electrical domains, and honest comparisons with electronic accelerators must count the converters, lasers, and thermal control, not only the optical core.
Manufacturing maturity is improving as photonic integrated circuits adopt silicon-based fabrication in established foundries with published process design kits, yet packaging, fiber coupling, laser integration, and thermal stabilization still dominate cost and remain far less standardized than electronic assembly. Test is likewise harder, since a photonic die cannot be probed as freely as an electronic one. Meanwhile the electronic baseline keeps moving: advanced digital accelerators continue to improve in energy per operation, so any photonic approach must beat a target that does not stand still.
The near-term trajectory therefore favors targeted accelerators and signal-processing functions, where light's parallelism and bandwidth deliver clear benefits, and above all optical interconnect, which is already displacing electrical links inside systems. General-purpose optical computing remains a longer-term research aim, contingent on a practical, low-energy, cascadable optical nonlinearity that does not yet exist.