Adaptive Inference for FPGA-Based 5G Automatic Modulation Classification

Automatic Modulation Classification (AMC) is key to the efficient use of the radio frequency spectrum in modern applications, like 5G-based IoT. Optimizing AMC is crucial to achieving the latency, throughput, and energy levels expected by the final user. State-of-the-art solutions to the AMC problem are based on Deep Learning methods (e.g., Deep Neural Networks - DNNs). However, these methods require heavy processing and high energy consumption up to the point that accelerators (e.g., FPGA) are used to carry out such computations. Based on the observation that the classification becomes computationally harder or easier depending on the amount of noise the signal is subjected (i.e., Signal-to-Noise Ratio - SNR), this work proposes a fully adaptive FPGA-based inference system that selects the most appropriate DNN according to the current signal quality (SNR level). Compared to the state-of-the-art static approach, the framework reduces energy consumption by up to 43% while delivering 8.9x more inferences per second.