Efficient Quantization and Hardware Implementation of AlexNet on Resource-limited FPGAs

As Deep Neural Networks (DNNs) have been growing to more and more complex models, their computation and storage requirements have increased dramatically. Recent research works on quantized neural networks (QNNs) have shown their great advantages for real-time machine learning on resource-limited embedded devices. FPGAs have emerged as a popular and efficient hardware platform to accelerate QNNs and DNNs when considering both performance and energy efficiencies. On the other hand, with the increasing complexity of hardware design, high-level synthesis tools are becoming popular to facilitate fast implementation. In this paper, we present our quantization and hardware implementation of the AlexNet using the Xilinx FINN framework. Extensive experiments with QNNs of various bits on a low-cost and resource-limited FPGA board demonstrate their good performance with much smaller memory footprint and lower computation complexity.