A 22 nm 10 TOPS Mixed-Precision Neural Network SoC for Image Processing with Energy-Efficient Dilated Convolution Support

Dilated convolutions have become popular in recent Deep Neural Networks (DNNs). However, they introduce additional zeros within the weights. To efficiently compute these layers within image processing DNNs, we implemented a 10 TOPS neural network SoC in 22 nm that supports dilated convolutions with reduced energy consumption. By exploiting this advantage, the energy can be reduced by a factor of 4.46 even for DNNs such as DeepLabV3+ with a majority of standard convolutional layers. Keywords: accelerator, DNN, dilated convolution, image pipeline