An Energy-Efficient Systolic Pipeline Architecture for Binary Convolutional Neural Network

Memory and computation cost limit the applications of Convolutional Neural Networks (CNN) on mobile devices. Binary CNN (BCNN) is a quantized neural network, which can reduce memory requirement and achieve multiplication-free computation. This paper focuses on BCNN. First, we proposed a hardware friendly CNN model to decrease the inference accuracy loss by eliminating the decimal part of each floating number. Second, we presented a fully pipelined on-chip BCNN architecture. The architecture has systolic data flow and an inter-layer pipeline, which ensures weights reuse and high throughputs. The results show that we achieve an inference accuracy of 99.04% for the MNIST dataset on the Pytorch platform and 98.91% in the hardware architecture, which means the inference accuracy loss is only 0.13%; the inference accuracy loss without this model is 0.41%. Besides, this architecture can achieve 23.08k: qfps and 353.5 GOP/s/W at 120MHz with small resource use while processing the BCNN.