Convolution Neural Network (CNN) has achieved great successes and has been widely applied in various visual applications such as object detections and image classifications. However, large amount of data and complicated computing process lower the inference performance if all data are collected and sent to a central process node for computation. Edge AI, which directly process the convolution on the edge device itself, has attracted huge attentions. Nevertheless, limited computation power and memory makes it impractical to process entire CNN computation. To overcome the challenge, light-weight neural networks have been proposed to reduce computation complexity with the accuracy drops as a cost. Recently, two light-weight neural networks, MobileNet and ShuffleNet, are widely discussed with acceptable accuracy degradation. However, most of the state-of-the-art DNN accelerators are not suitable for MobileNet and ShuffleNet. Therefore, in this paper, we propose a novel CNN accelerator which can support both depthwise separable convolution and channel shuffle for these light-weight neural networks. Experimental results show that our design can successfully compute MobileNet and ShuffleNet. Moreover, compare with previous works, we can achieve at most 20% area reduction while maintaining performance.