Light can process information as well as carry it. Photonic computing encodes numbers in properties such as intensity or phase. Engineered optical components transform those signals through transmission and interference. Suitable arrangements implement multiplication and addition operations that occur repeatedly in neural networks.
The attraction for AI accelerators is parallel optical processing and the possibility of reducing some electrical data movement. Practical systems commonly retain electronic memory, control and processing. Data must enter the optical circuit, and its output must be measured and often converted back into electrical form.
Conversion, calibration and the light source all consume resources. A fast optical component therefore does not establish the efficiency of a complete computer. Accelerator designs remain an active research and engineering field with different levels of maturity. Classical optical computing also differs from photonic quantum computing: using light does not imply using qubits. Photonic crystals and metamaterials contribute ways to control that light.