Optical parallel computing of convolution using a Dammann grating

As artificial intelligence consumes more and more power than traditional electronic computers, it becomes urgent to find a new optical computing system to calculate the enormous amount of matrix convolution in a parallel and energy-saving way. The naturally high parallelism of photons offers a potential solution for dramatically increasing computer computing power. This paper proposes a parallel optical computing architecture based on the optical shadow-casting method using a Dammann grating. Firstly, using the unique beam-splitting ability of Dammann grating, the convolution kernel is diffracted to the targeted places on the convolution matrix at different angles, which can realize the flexible sliding of the convolution kernel on the convolution matrix. Secondly, the image conjugation of the 4𝑓 system can significantly increase the alignment accuracy, and the convolution kernel of different diffraction orders is located in the specific position of the convolution matrix. Finally, matrix multiplication and addition are further realized by loading matrix information through the light modulator and lens convergence. The system has significant advantages in energy consumption ratio and computing power improvement, and parallel optical computing might be attractive in competition with the traditional electronic computers in some large-scale matrix computing fields.