Configurable Multi-directional Systolic Array Architecture for Convolutional Neural Networks

The systolic array architecture is one of the most popular choices for convolutional neural network hardware accelerators. biggest advantage its simple and efficient design principle. Without complicated control dataflow, accelerators with can calculate traditional convolution very efficiently. However, this also brings new challenges to array. When computing special types convolution, such as small-scale or depthwise processing element (PE) utilization rate decreases sharply. main reason that limits flexibility In article, we a configurable multi-directional (CMSA) address these issues. First, added data path It allows users split through configuration speed up calculation convolution. Second, redesigned PE unit so has multiple transmission modes dataflow strategies. This switch addition, unlike other works, only make few changes modifications existing architecture. avoids additional overheads be easily deployed in application scenarios require small arrays mobile terminals. Based on our evaluation, CMSA increase by 1.6 times compared typical when running last layers ResNet-18. MobileNet, 14.8 times. At same time, are similar area energy consumption.