Reconfigurable MAC Systolic Array Architecture Design for Three-Dimensional Convolution Neural Network

Nowadays, Convolutional Neural Network (CNN) is widely used in many applications. Plenty of MAC operations may cause long latency for CNN. To solve this problem, 3D stacking of CNN (3D CNN) is applied. However, 3D CNN may lead to the thermal problem. In this work, we propose a reconfigurable MAC systolic array architecture (RMACA) for 3D CNN. If the hotspot in the 3D CNN occurs, partial MACs can be turned off by using RMACA. Thus, the heat accumulation can be reduced and the hotspot can be removed. RMACA is implemented by TSMC 90-nm CMOS technology, 21,670 gates is required to realize 4x4 CNN with RMACA.