A survey of architectures of neural network accelerators

Nowadays, with the growth in data demand and the improvement in hardware computing power, artificial intelligence (AI) can be applied to a wide range of applications. Among them, neural network algorithms have successfully solved some practical problems, such as face recognition and autonomous driving. Although these algorithms perform well, their computing performance on traditional hardware platforms is still inefficient. To address the issue, some customized accelerators are proposed. In this survey, we introduce some architecture designs of typical accelerators, including the computing unit, data flow, the characteristics of different neural networks to be accelerated, and design considerations on emerging platforms. Finally, we also provide our insights of future trend of neural network accelerators.