CNNX: A Low Cost, CNN Accelerator for Embedded System in Vision at Edge

Convolutional neural networks (CNNs) have been widely deployed in artificial intelligence, including computer vision and pattern recognition. In these applications, CNN is the most computationally intensive part. Recently, many researchers have used depthwise convolution to decrease the computational load in the execution of CNNs; on the other hand, today, CNNs have become larger and larger. Consequently, they need more computational budget for their executions. The problem is more serious when this application is run in an embedded system, especially in the edge devices, as the embedded processor can hardly handle these heavy computational loads. This paper proposes a lightweight, low-power, and efficient CNN hardware accelerator for edge computing devices. This accelerator is explicitly designed for depthwise CNN. The proposed accelerator can be configured and programmed to run any lightweight CNN of a wide range of AI networks such as MobileNet, Xception, and shuffleNet. Our experimental results show that our accelerator can run MobileNet 70 times per second in a remote sensing AI application with a 224 × 224 pixel image from the ImageNet dataset.