FPGA implementation of CNN for modulation format identification

modulation format identification (MFI) has become an increasingly attractive avenue for research. Especially as the need for optical layer performance monitoring increases. Deep learning methods provide an effective recognition method with high accuracy. At the same time, the need to implement deep learning models on embedded devices is growing. Field-Programmable Gate Array (FPGA) are the best choice due to their parallelism, flexibility, and energy efficiency. In this paper, an FPGA-based CNN network design scheme for modulation format identification is proposed. Our proposed design explores and provides a suitable solution for better performance while maintaining accuracy. We generated modulation format datasets for MFI and Optical Signal-to-Noise Ratio (OSNR) in simulation environment, training and fine-tuning the network parameters base on this dataset, and then changing some suitable components in the CNN networks to fit for the deployment on FPGA. Our experiments in the Xilinx XC7A200 development kit show that despite model compression, the accuracy of our implementation is comparable to the original model running on the CPU and GPU. In terms of speed, the FPGA implementation of the XC7A200 kit is inferior to the high-end GPU RTX 2080 Ti but superior to the Intel® Core™ i7-7700HQ CPU. In terms of power consumption, the FPGA implementation is significantly lower than the Intel® Core™ i7-7700HQ and the RTX 2080 Ti, being approximately about 13–14 times lower than them. In terms of energy efficiency, FPGA are completely superior to GPU, being 5–6 times more efficient than the RTX 2080 Ti and 29 times more efficient than the Intel® Core™ i7-7700HQ.