Optical computing
Performing computation with light rather than only transmitting data with it. In its most developed form, matrix-vector multiplication carried out by interference in a mesh of interferometers or a microring array, where the result appears in the propagation itself. Distinct from optical interconnect, which moves data between electronic processors.
Optical computing means doing arithmetic in the optical domain, as opposed to optical interconnect, which uses light only to carry data between electronic chips. The distinction matters because the two have had opposite fortunes, and they are constantly conflated.
The central primitive is linear: a mesh of Mach-Zehnder interferometers, or an array of microring weights, performs a matrix-vector multiplication as light passes through it. The input vector is encoded on the amplitudes of several waveguides, the matrix is set by the phase shifters in the mesh, and the interference at the output produces the product. The multiplication happens at the speed of propagation, in nominally constant latency and with no switching energy spent on the arithmetic itself, which is the entire appeal: matrix multiplication is the dominant operation in neural-network inference, and here it appears to come for free.
The difficulty is that the arithmetic is only part of the cost. The vectors must be converted from electronic to optical form at the input and back at the output, and those data conversions, along with the digital-to-analog and analog-to-digital steps around them, dominate the energy budget and erase much of the apparent free multiplication. Analog optical computation carries limited effective precision, on the order of a handful of bits, against the higher precision digital hardware supplies on demand. The phase shifters that hold the matrix drift with temperature and must be trimmed and reloaded. And the nonlinear activation between layers, along with the memory, generally remains electronic, so a practical system converts in and out of the optical domain repeatedly. The result is that for general-purpose inference, digital electronics still wins the system-level comparison once conversion and precision are counted honestly, and optical computation is confined for now to niches: fixed linear transforms, and specific low-precision workloads where the conversion overhead can be amortized.
The commercial history is the clearest evidence. Several companies that set out to build optical processors have redirected their shipping products toward optical interconnect, where light has a decisive and uncontested advantage, while continuing optical compute as research; others pursue analog photonic processors on thin-film lithium niobate with native nonlinear functions as the current serious attempt at compute by light. The one-line summary that survives scrutiny: light has won the moving of data and has not yet won the doing of arithmetic. The full argument is in Why light moves data and electrons still compute.
References: Shen et al., "Deep learning with coherent nanophotonic circuits," Nature Photonics 11, 441 (2017); Hamerly et al., analyses of optical versus electronic matrix multiplication energy (2019 onward). See also co-packaged optics for the interconnect side that succeeded.