Time delay The basic idea is to delay a signal in order to perform useful computations. Of interest would be to solve
NP-complete problems as those are difficult problems for conventional computers. Two basic properties of light are used in this approach: • Light can be delayed by passing it through an optical fiber. • Light can be split into multiple rays. This property allows multiple solutions to be evaluated concurrently. Solving a problem with time-delays involves the following steps: • Create a graph-like structure made from optical cables and splitters. Each graph has a start node and a destination node. • Light enters through the start node and traverses the graph until it reaches the destination. It is delayed when passing through arcs and divided inside nodes. • Light is marked when passing through an arc or through a node to identify that fact at the destination node. • The destination node waits for a signal (fluctuation in the intensity of the signal) which arrives at a particular moment in time. If no signal arrives at that moment, it means no solution was found. Otherwise the problem has a solution. Fluctuations can be read with a
photodetector and an
oscilloscope. The first problem attacked in this way was the
Hamiltonian path problem. An optical device solving an instance with four numbers {
a1, a2, a3, a4} is depicted below: The light enters Start node where it divides into two rays of smaller intensity. These two rays arrive at the second node at moments
a1 and 0. Each is further divided into two rays that arrive at the third node at moments 0,
a1,
a2 and
a1 + a2. These represent all subsets of set {
a1, a2}. Intensity fluctuations occur at no more than four moments. The destination node expects fluctuations at no more than 16 different moments (subsets of the initial). A fluctuation at the target moment
B means that a solution has arisen, otherwise no subset sums to
B. Zero-length cables are not possible, thus all cables are lengthened by a small (fixed for all) value
k. In this case the solution is expected at moment
B+n×k.
Photonic tensor operations With increasing demands on GPU-based accelerator technologies, the 2010s experienced emphasis on on-chip integrated optics. The emergence of deep learning neural networks based on phase modulation, and more recently amplitude modulation using photonic memories has created photonic technologies assisting
neuromorphic computing. Evolving technology had allowed these parallel operations to be performed on-chip on an integrated photonic tensor core. In a 2025 paper titled "Direct tensor processing with coherent light," researchers demonstrated "single-shot" tensor computing through an algorithm titled "parallel optical matrix–matrix multiplication (POMMM)." POMMM allows for
tensor operations such as multiplication to be performed in a single shot of light at high speeds. POMMM has the potential to replace GPUs for tasks such as
convolutions and
attention layers.
Wavelength-based computing Wavelength-based computing can be used to solve the
3-SAT problem with
n variables,
m clauses and with no more than three variables per clause. Each wavelength, contained in a light ray, is considered as possible value-assignments to
n variables. The optical device contains prisms and mirrors that discriminate wavelengths which satisfy the formula.
Computing by xeroxing on transparencies This approach uses a photocopier and transparent sheets for performing computations. The
k-SAT problem with
n variables,
m clauses and at most
k variables per clause has been solved in three steps: • All 2n possible assignments of
n variables are generated by performing
n photocopies. • Using at most 2
k copies of the truth table, each clause is evaluated at every row of the truth table simultaneously. • The solution is obtained by making a single copy operation of the overlapped transparencies of all
m clauses.
Masking optical beams The
travelling salesman problem was solved by Shaked
et al. (2007) via an optical approach. All possible TSP paths were generated and stored in a binary matrix that was multiplied with another gray-scale vector containing the distances between cities. The multiplication is performed optically by using an optical correlator.
Optical Fourier co-processors Many computations, particularly in scientific applications, require frequent use of the 2D
discrete Fourier transform (DFT) – for example in solving
differential equations describing wave propagation of waves or heat transfer. Though
GPU technologies typically enable high-speed computation of large 2D DFTs, other techniques can perform continuous Fourier transform optically by utilising the natural
Fourier transforming property of lenses. The input is encoded using a
liquid crystal spatial light modulator and the result is measured using a conventional
CMOS or
CCD image sensor. Such optical architectures can offer superior scaling of computational complexity due to the inherently highly interconnected nature of optical propagation, and have been used to solve 2D heat equations.
Ising machines Ising machines are computers whose design was inspired by the theoretical
Ising model.
Yoshihisa Yamamoto's lab at
Stanford pioneered building Ising machines using photons. Initially Yamamoto and his colleagues built an Ising machine using lasers, mirrors, and other optical components. Later a team at
Hewlett Packard Labs developed
photonic chip design tools and used them to build a single chip Ising machine, integrating 1,052 optical components. ==Industry==