An Efficient Frequency Encoding Scheme for Optical Convolution Accelerator

In today’s era where the demand for computational resources by large models is increasingly high, optical computing offers an alternative physical platform for computation. With its high parallelism and the maturation of integrated photonic technologies, optical computing is expected to further support the computational resources required by large models. For one-dimensional optical convolution accelerators, existing methods can fully utilize the working bandwidth of electro-optic modulators and the frequency resources of light sources. However, most convolution computations currently require the execution of two-dimensional matrix convolutions, and existing encoding schemes suffer a drop in terms of effective computations per second when performing two-dimensional matrix convolutions. In response to this, we propose a new encoding scheme that can fully utilize the computational resources of optical convolution accelerators. For convolution operations with a kernel size of M×M, compared to existing encoding schemes, it can achieve an M-fold increase in effective computation rate. This implies that under the constraints of essentially the same hardware physical performance, adopting our newly proposed encoding scheme can achieve a significant improvement in computational performance. We construct an optical convolution acceleration system to demonstrate the principles of the new encoding.