This paper presented a roadmap of optical computing, particularly, for optical implementation of convolution, in comparison with supercomputer. Optical computing power presented in this roadmap is predicted based on a simple mathematic equation of convolution where optics can do matrix convolution in full parallelism with single pass of light, which could be 500 times faster than supercomputer( Fugaku), while energy consumption of optical computing might be 1000 times lower than supercomputer (Fugaku), supposing 8K Spatial Light Modulator with extremely high updating rate <22000 frame per second might be obtained in the future. This roadmap indicates a bright future of optical computing for matrix convolution, which might take up 80% of total computing power required in deep-learning neural network.
|