Striping input feature map cache for reducing off-chip memory traffic in CNN accelerators

Data movement between the Convolutional Neural Network (CNN) accelerators and off-chip memory is critical concerning the overall power consumption. Minimizing power consumption is particularly important for low power embedded applications. Specific CNN computes patterns offer a possibility of significant data reuse, leading to the idea of using specialized on-chip cache memories which enable a significant improvement in power consumption. However, due to the unique caching pattern present within CNNs, standard cache memories would not be efficient. In this paper, a novel on-chip cache memory architecture, based on the idea of input feature map striping, is proposed, which requires significantly less on-chip memory resources compared to previously proposed solutions. Experiment results show that the proposed cache architecture can reduce on-chip memory size by a factor of 16 or more, while increasing power consumption no more than 15%, compared to some of the previously proposed solutions.