FPGA based neural network accelerators

Machine learning (ML) and artificial intelligence (AI) technology are revolutionizing many different fields of study in computer science as well as a wide range of industry sectors such as information technology, mobile communication, automotive, and manufacturing. As more people are using the technology in their everyday life, the demand for new hardware that enables faster and more energy-efficient AI processing is ever increasing. Over the last few years, traditional hardware makers such as Intel and Nvidia as well as start-up companies such as Graphcore and Habana Labs were trying to offer the best computing platform for complex AI workloads. Although GPU still remains the most preferred platform due to its generic programming interface, it is certainly not suitable for mobile/edge applications due to its low hardware utilization and huge power consumption. On the other hand, FPGA is a promising hardware platform for accelerating deep neural networks (DNNs) thanks to its re-programmability and power efficiency. In this chapter, we review essential computations in latest DNN models and their algorithmic optimizations. We then investigate various accelerator architectures based on FPGAs and design automation frameworks. Finally, we discuss the device's strengths and weaknesses over other types of hardware platforms and conclude with future research directions.