Parallelism-flexible Convolution Core for Sparse Convolutional Neural Networks on FPGA

This paper proposes a convolution core for sparse CNN that is capable of flexibly alternating the parallelism schemes and degree exploiting intraand inter-output parallelism of the convolutional layer, and leveraging weight sparsity using a compressed sparse model in the compressed sparse column format and output-stationary dataflow. The experimental results show that the performance is improved by 3.9 times even in the deeper layer where the conventional accelerator could not fully exploit the parallelism due to the small layer size. The proposed architecture could also exploit the weight sparsity. Then, by combining both the multi-parallelism and the weight sparsity, the proposed architecture achieved 5.2 times better performance than the conventional accelerator.