Convolutional Neural Networks Inference Accelerator Design using Selective Convolutional Layer

Convolutional Neural Networks (CNNs) often require a huge amount of multiplication. The current approach of multiplication reduction requires data preprocessing, which is power-hungry and time-consuming. The paper proposed an Image-wised Selective Processing Engine (SPE-I) for accelerating CNN processing by eliminating unessential operations through algorithm-hardware co-designs. The SPE-I compares the similarity of two input images and identifies any redundant calculations that can be skipped. A modified LeNet-5 network, LeNet3x3 was designed to validate the performance improvement of SPE-I using the MNIST dataset. LeNet3x3 with and without SPE-I were implemented in TSMC 90-nm CMOS technology at 87.5 MHz operating frequency. Compared to the network without SPE-I, the network with SPE-I only has 0.12% - 1.79% accuracy drop, achieving 43.1% power saving due to 73% - 81% multiplication reduction. Regarding timing, SPE-I takes 20% of total clock cycles to provide convolutional data compared to the convolutional layer using preprocessing.