 Optical computing can be used to improve the speed and energy efficiency of convolutional neural networks, CNNs. A new design was developed which uses 3-2 by two correlated real-valued kernels made up of two multi-mode interference cells and four phase shifters to perform parallel convolution operations. This design allows for linear scalability with respect to computational size, making it suitable for large-scale integration. Experimental results show that this design can achieve 10-class classification of handwritten digits from the MNIST database. This article was authored by Xiang Yanming, Guo Jiejiong, Nuan Yuanxia, and others.