A High Efficient Architecture for Convolution Neural Network Accelerator

Convolutional neural networks (CNNs) are widely used in modern Artificial Intelligence (AI) systems. Compared with other classical methods, CNNs have superior performance in image classification, speech recognition and object detection. However the computational load of CNNs is very heavy and a large amount of data movement are expected. An efficient way of data movement is critical for both performance and power efficiency for an accelerator design. In this paper we propose a novel CNN accelerator architecture with unique parallel loading scheme and smart memory addressing solution. Our solution is 30% faster than others [1] on Alexnet. Our proposal can achieve high efficiency for FC layer without using image batching. This will make our solution very suitable for edge applications.