Hardware-oriented Groupwise Topology for CNN Accelerator

For edge AI inference applications, one of the major challenges is to accomplish huge computation load using limited hardware resources and lowest power consumption. A groupwise architecture is proposed in this paper for accelerating neural network workload via hardware group parallelism and optimized processing element utilization. To achieve better efficiency and higher accuracy, the hardware is optimized by using customized groupwise networks and configurable processing element. The experiment results show that it could improve the accuracy by 2%, using CIFAR-10 dataset and MobileNet as the baseline model, at the expense of little growth in the number of parameters. Moreover, the groupwise topology also enhances hardware inference performance, in terms of frames per second, without additional hardware resource for running YOLO-Face network.