Paper Reading: DNPU: An 8.1TOPS/W Reconfigurable CNN-RNN Processor for Deep Neural Networks

1. Introduction

        This paper is published in ISSCC in 2017. Recently, deep learning with convolutional neural networks (CNNs) and recurrent neural networks (RNNs) has become universal in all-around applications. But there has been no work on a combined CNN-RNN processor. Hence, the researchs present a reconfigurable CNN-RNN processor.

2. Innovation point

        The computational requirements in CNNs are quite different from those of RNNs, as shown in the following figure.

         While convolution layers require a massive amount of computation with a relatively small number of filter weights, fully-connected layers and RNN-LSTM layers require a relatively small amount of computation with a huge number of filter weights. Thus, a combined CNN-RNN processor with high energy-efficiency is implemented. The following is the overall architecture of DNPU.

 3. Summary

        The chip photograph and specifications are shown below.