摘要
This paper presents the tile-based DCNN accelerator architecture. The hierarchical interconnection networks enable distributed data delivery to maximize the data bandwidth both for the conventional and depthwise convolution layers. An exploration tool is also developed to optimize architectural parameters under the trade-off among specific performance, power, and area. The case study shows that our accelerator can outperform the state-of-the-art by up to 36% faster and up to 25% lower energy on different modern DCNNs. The experiment also justifies the scalability of our approach.