Efficient Segment-wise Pruning for DCNN Inference Accelerators

This paper presents a structure pruning with design space exploration for DCNN accelerators. The design space exploration tool, called NNArch, generates optimized design scheduling for the row-stationary DCNN accelerators with fast and accurate analytical performance/energy models. Based on the NNArch, the proposed filter-segment pruning can efficiently compress the DCNNs with a simple filter index table, optimizing model accuracy, accelerator performance, and energy consumption. The experiment result shows that the proposed pruning scheme can achieve 2 times speedup and improve the energy-delay-product by 3.42 times on ResNet50 with the accuracy drop of 0.2%, on the accelerator of 168 PEs.