The rapid development of artificial intelligence has prompted the convolutional neural network (CNN) to process huge amount of data, which has caused a great burden on convolution operations. Therefore, according to the characteristics of the systolic array architecture, the acceleration structure of CNN is constructed by fusing it with CNN. Besides, it is optimized in practical application, and its effectiveness is verified. The experimental results show that in the broadcast architecture, the time required by the CNN acceleration architecture is at least 0.005, while the maximum throughput is 16.83, which is far higher than the acceleration architecture under the systolic array architecture. In the case of small change in the maximum frequency, the error rate is the same as that of the systolic array, which is about 3.62%. In the comparison of various methods proposed on the systolic array, the accuracy rate of CNN acceleration architecture is 94.7%, and the utilization rate is 81.95%. The correctness and effectiveness of the algorithm are proved. To sum up, the improved CNN acceleration structure based on pulse array optimization reduces the response time and meets the requirements of terminal calculation force, which is of high significance in practical application.