Exploration and Generation of Efficient FPGA-based Deep Neural Network Accelerators

Abstract

Convolutional Neural Networks (CNNs) have emerged as an answer to next-generation applications such as complex image recognition and object detection. Embedding such compute-intensive and memory-hungry algorithms on edge systems will lead to smarter high-value applications. However, the algorithmic innovations in the CNN field leave the hardware accelerators one step behind. Reconfigurable hardware (e.g. FPGAs) allows designing custom accelerators adapted to new algorithms. Furthermore, new design approaches such as high-level synthesis (HLS) enable to generate RTL code based on high-level function descriptions. This paper presents a high-level CNN accelerator generation framework for FPGAs. A first phase of the framework characterizes CNN descriptions using hardware-aware metrics. These metrics then drive a hardware generation phase which builds the proper C source code implementation for each layer of the network. Finally, an HLS tool outputs the synthesizable RTL code of the accelerator. This approach aims at reducing the gap between the evolving applications based on artificial intelligence and hardware accelerators, thus reducing time-to-market of new systems.