Electronics Guide

Articles in This Category

Fundamental Principles

Optical computing exploits the wave nature of light. Light waves naturally exhibit interference, diffraction, and, in suitable materials, nonlinear interactions, and each of these can be mapped onto a useful mathematical operation. The result is a machine in which the arithmetic is performed by propagation rather than by a sequence of clocked logic states, which sets both the appeal of the approach and the boundary of what it does well.

Linear Operations in Passive Optics

Passing a set of optical amplitudes through a network of beam splitters and phase shifters realizes a linear transformation of those amplitudes. A lossless network of this kind implements a unitary matrix, and adding attenuation or gain extends the reach to general linear maps. Two operations follow directly and account for most of the practical interest in the field.

The first is the Fourier transform. A converging lens performs it in the spatial domain: with an object placed in the front focal plane, the field in the back focal plane is proportional to the two-dimensional Fourier transform of the input, computed in the time light takes to cross the optic. Correlators, matched filters, and convolution engines all follow from that single property.

The second is matrix-vector multiplication, the dominant operation in neural-network inference. A coherent mesh of interferometers or a bank of wavelength-selective resonators performs it in a single pass, with the latency set by the optical transit time rather than by a loop count. Because these primitives are realized by physics rather than by repeated multiply-accumulate instructions, they reach high throughput at low latency, which is why photonic hardware is usually positioned as an accelerator for specific kernels rather than as a replacement for the general-purpose processor.

The Missing Nonlinearity

Linear optics alone cannot compute. Neural networks require an activation function between layers, and Boolean logic requires a thresholding element; both demand that one light field change the medium seen by another. Optical nonlinearities in transparent materials are weak, so obtaining useful switching contrast generally requires high optical intensity, a long interaction length, a resonant structure that builds up field strength, or a material with strong resonant absorption. Each remedy carries a cost in power, area, bandwidth, or fabrication tolerance. In practice, most working systems sidestep the problem: they detect the optical result, apply the nonlinearity in the electronic domain, and modulate the outcome back onto light for the next stage.

Encoding, Precision, and Noise

Photonic processors are analog machines. A value is encoded in optical amplitude, in intensity, in phase, or in some combination, and the achievable precision is bounded by shot noise, laser relative-intensity noise, thermal drift of the waveguides, fabrication mismatch between nominally identical components, and the resolution of the digital-to-analog and analog-to-digital converters at the boundary. Reported effective precision for integrated photonic multiply-accumulate hardware has commonly fallen in the range of roughly four to eight bits, which suits inference on quantized neural networks but not scientific computing in double precision. Raising precision usually means slowing the system down, averaging over more photons, or spending more energy on calibration and stabilization.

Hardware Platforms

Several hardware families compete within the field, and they differ less in the mathematics they perform than in how they scale, what they cost to build, and where their errors come from.

Integrated Interferometer Meshes

A coherent mesh encodes a matrix in the phase settings of a network of waveguide couplers, most often Mach-Zehnder interferometers arranged in a triangular or rectangular lattice. Any unitary transformation on N optical modes can be decomposed into a mesh of N(N − 1)/2 two-mode interferometers, so a 128-mode processor requires 8,128 of them, each with its own phase shifter, driver, and calibration state. That quadratic growth is the central scaling problem of the approach: control complexity, thermal crosstalk, and cumulative insertion loss all rise with mesh size. The same programmable-mesh hardware underlies programmable photonics, where one fabric is reconfigured for filtering, switching, or sensing rather than for arithmetic alone.

Wavelength-Multiplexed Resonator Banks

An alternative encodes each input on a separate wavelength and weights it with a tuned microring resonator. Because the rings on a shared bus are addressed independently by wavelength, a single waveguide carries many products, and a photodetector at the output sums them by simply integrating the total incident power. Chip-scale optical frequency combs, including soliton microcombs, supply the dozens of equally spaced carriers such a scheme needs from one pump laser. Microrings are compact and naturally parallel, but their resonances are narrow: a shift of a fraction of a nanometer in radius, or a few degrees of temperature change, detunes a ring, so each one needs a local heater and a feedback loop.

Free-Space and Diffractive Optics

Free-space systems process two-dimensional optical fields with lenses, mirrors, spatial light modulators, and diffractive elements, operating on millions of points in parallel across a beam. Diffractive networks push this further by training a stack of passive phase masks offline and then fabricating them, after which the assembly classifies an input as light simply passes through it, with no electrical power consumed in the optical layers themselves. The parallelism is unmatched, but the systems occupy volume, demand precise alignment and mechanical stability, and are reprogrammed slowly, since spatial light modulators refresh at video rates rather than at gigahertz.

Fiber and Delay-Line Architectures

Fiber-based systems exploit low loss and long delays rather than dense integration. A single nonlinear node coupled to a long fiber delay line implements time-multiplexed reservoir computing, in which a fixed recurrent dynamical system is driven by the input and only a linear readout is trained. Related time-multiplexed architectures underpin coherent Ising machines, which relax combinatorial optimization problems onto the phases of a train of optical pulses. These machines trade the parallelism of integrated meshes for very large effective node counts at modest hardware cost.

The Silicon Photonics Process Platform

Most integrated work builds on silicon-on-insulator wafers with a 220-nanometer-thick silicon device layer, patterned on the same deep-ultraviolet lithography tools used for electronics. The platform supplies single-mode strip waveguides, directional couplers, grating and edge couplers, thermo-optic and carrier-depletion phase shifters, and germanium photodetectors monolithically integrated for the 1310- and 1550-nanometer bands. Routine single-mode routing loses on the order of a decibel per centimeter, while wide, shallow-etched geometries optimized to reduce sidewall scattering have measured below 0.1 decibel per centimeter in commercial foundry runs. Silicon has no useful direct bandgap, so lasers are attached by flip-chip assembly, micro-transfer printing, or wafer bonding of III-V material rather than grown in place, and that laser integration step remains one of the harder yield and cost problems. Design flows depend on specialized photonic design automation tools, because layout, electromagnetic simulation, and circuit-level modeling interact far more tightly than in digital electronics.

Demonstrations and Commercial Progress

The field has moved from proof-of-principle experiments to systems that run recognizable workloads, though the published record is easier to interpret when the specific conditions are kept in view.

Research Milestones

In 2017 a team centered at the Massachusetts Institute of Technology reported a programmable nanophotonic processor built from a cascaded array of fifty-six Mach-Zehnder interferometers on a silicon chip, and used it to run a small neural network for vowel recognition. The experiment established that a fabricated interferometer mesh could be programmed to hold trained weights, and it seeded much of the commercial activity that followed.

A 2021 result published in Nature demonstrated an integrated photonic tensor core that combined phase-change-material memory cells, which hold a weight without continuous power, with a chip-based soliton frequency comb for wavelength parallelism. The reported hardware reached about two trillion multiply-accumulate operations per second at a compute density near 555 billion multiply-accumulate operations per second per square millimeter, at roughly five bits of precision, and performed convolutional image processing directly in the optical domain.

In 2025 a photonic processor described in Nature integrated four 128 × 128 interferometer-based tensor cores, stacked vertically with 12-nanometer CMOS control dies, and executed ResNet and BERT models along with a deep reinforcement-learning agent, reaching accuracy close to that of 32-bit floating-point digital hardware. The system delivered about 65.5 trillion operations per second at roughly 78 watts of electronic power. That result is significant less for its raw efficiency, which does not yet lead the field, than for showing that a photonic core can hold enough effective precision to run standard models rather than curated demonstrations.

Optics in the Interconnect

Optics is already entrenched one level below computation. Long-haul, metropolitan, and intra-data-center links have been optical for years, and the pressure now falls on the last few centimeters. Optical interconnects replace copper where electrical channel loss, equalization power, and reach limits bite hardest, and co-packaged optics moves the optical engine into the switch or accelerator package instead of a faceplate module. Broadcom's Bailly platform places 6.4-terabit-per-second silicon-photonic engines inside a 51.2-terabit-per-second switch package, and NVIDIA introduced co-packaged Quantum-X Photonics InfiniBand and Spectrum-X Photonics Ethernet switches in 2025, with the InfiniBand product offering 144 ports of 800 gigabits per second and liquid cooling for the photonic engines. Vendors report several-fold improvements in energy per bit relative to pluggable transceivers. This is the commercially decisive use of photonics in computing today, and it also builds the manufacturing base, the packaging techniques, and the volume that a photonic compute engine would eventually need.

Challenges and Limitations

The same properties that make light attractive also make general-purpose optical computing difficult, and the obstacles are structural rather than incidental.

Logic Devices Are Hard to Build

A device that is to serve as an optical equivalent of the transistor must satisfy several criteria at once, set out clearly by David Miller in a 2010 commentary in Nature Photonics: it must restore logic levels so that degradation does not accumulate, provide fan-out so that one output can drive several inputs, isolate its input from its output, remain cascadable so that one stage drives the next directly, and avoid critical biasing that would demand impractical tolerances. Miller observed that most devices then promoted as optical transistors failed most of these tests, and no candidate has since become the general-purpose optical switching element the field once anticipated.

Analog Error Accumulates

Encoding values in optical amplitude and phase exposes computation to amplitude noise, phase noise, loss, and component drift, and these errors compound as operations are cascaded. Depth is therefore limited in a way that has no counterpart in digital electronics, where each gate regenerates its output. Large meshes also require calibration: phase shifters interact thermally, fabrication leaves every coupler slightly different, and characterizing a mesh with thousands of tunable elements is itself a substantial computational task that must be repeated as the chip ages and its environment changes.

The System Budget Rules

Conversions between the electronic and optical domains cost energy and time. Modulators, drivers, transimpedance amplifiers, and data converters sit on both sides of the optical core, and their combined power frequently exceeds that of the optical computation. Laser wall-plug efficiency, thermal tuning of resonators, and the electrical power needed to keep phase shifters at their set points add further overhead that scales with the number of components rather than with the useful work performed. A hybrid system wins only when the optical work it accelerates is large enough to amortize all of that, which is why honest comparison against mature electronics must account for the whole system rather than the optical core alone. Memory is a further asymmetry: light cannot be held at rest, so optical storage means either recirculating delay lines or a return to an electronic or material state, and no optical memory approaches the density and access economics of DRAM or SRAM.

A Moving Target

Electronic accelerators are not standing still. Aggressive quantization to eight-bit and four-bit formats, structured sparsity, and advanced packaging with high-bandwidth memory have captured part of the efficiency margin that photonic architectures once claimed. Any photonic proposal must be measured against the electronics that will exist when it ships, not against the electronics it was conceived to beat.

Outlook

The near-term trajectory is hybrid. Systems that keep linear, high-throughput work in the optical domain while handling control, memory, and nonlinear activation electronically avoid the unsolved problems and exploit the solved ones, and they inherit a supply chain that co-packaged optics is already building. Related work in neuromorphic photonics pursues the same hybrid instinct from the direction of neuromorphic computing, using optical fan-in and wavelength parallelism for the synaptic layer while leaving spiking dynamics and learning to electronics or to resonant devices.

Longer-term progress depends on device physics: more efficient optical nonlinearities for switching and activation, nonvolatile weight cells that hold their state without power, denser and more uniform integration, and laser sources efficient enough that the optical power budget stops dominating. The convergence of optics with quantum information adds a separate dimension. Photons carry quantum states well at room temperature, travel over fiber without a cryostat, and interfere naturally, which makes them strong candidates for quantum communication and for measurement-based and boson-sampling approaches to quantum computation, at the cost of probabilistic gates and demanding requirements on single-photon sources and detectors.

Whether photonic computing remains a set of specialized accelerators or grows into a broader computing platform, it is best understood as a complement to electronics. Its record so far supports that reading: light has won decisively where it moves information, and it earns its place in computation exactly where a large linear operation must be performed quickly and repeatedly enough to pay for the trip in and out of the optical domain. The articles above examine the devices, materials, and architectures on which that case rests.

Related Topics