Towards Resource Efficient and Low Latency CNN Accelerator for FPGAs: Review and Evaluation

Authors:

Abstract
Convolutional Neural Networks (CNNs) have seen widespread adoption in various domains, including speech recognition, image classification, and segmentation. However, their computational demands necessitate dedicated and reconfigurable hardware architectures, such as FieldProgrammable Gate Arrays (FPGAs). Despite the numerous advantages of CNN accelerators including enhanced efficiency within diverse application contexts, their deployment on FPGAs must address challenges related to energy consumption, computational demands, and resource optimization. Both academia and industries are actively engaged in research and development aimed at designing resource-efficient and lowlatency CNN accelerators for FPGAs. However, there has been relatively little attention paid to the dynamic comprehensive analysis and evaluation in which CNN accelerators equipped with FPGAs operate. This paper aims to bridge this gap by conducting a comprehensive survey and evaluation of CNN accelerator techniques for FPGAs, focusing on resource utilization, latency, and power consumption. This study investigates the performance of CNN accelerators for FPGAs, considering a range of hardware architectures and optimization techniques, from multiplier designs to complex heterogeneous pipelining. Furthermore, the paper highlights the challenges faced by FPGA-based CNN accelerators, particularly the tradeoffs between resource utilization and latency, based on the characteristics of the data and devices involved. The insights gleaned from this review can offer valuable guidance and inspiration to researchers, engineers, and students working in these domains, facilitating further advancements in FPGAbased CNN accelerator technology.

This publication has 51 references indexed in Scilit: