FPGA Implementation of Radio Frequency Neural Networks

Recent advances in Neural Network (NN) models for the Radio Frequency (RF) domain have made them a dominant force in realizing robust architectures that generalize well to novel operating conditions. While the performance of NN models when running on a Graphics Processing Unit (GPU) are generally very good, many applications require lower latency and higher throughput to be edge deployable. We have recently developed physics-driven NN models to perform Digital Signal Processing (DSP) functions for a Long Term Evolution (LTE) receiver application, demonstrating equal or better performance than their DSP equivalents. This paper discusses moving some of these NN models to Field Programmable Gate Array (FPGA) to tackle the latency and throughput goals and evaluate the performance at different quantization levels. We compare the FPGA performance results at different quantization levels with their GPU performance counterpart and discuss the path forward towards an RF edge solution.