Introduction

• This project presents a hardware-software co-design implementation of a neural network-based solution for applying digital predistortion (DPD) on RF signals to offset nonlinearities effects in power amplifiers, inspired by the approach detailed in this MathWorks article.
• you can obtain the code in that article by running this command in matlab: openExample('deeplearning_shared/NeuralNetworkDigitalPredistortionOfflineTrainingExample')
• The power amplifier (PA) used here is a simulated version of the NXP™ Airfast LDMOS Doherty PA that you can access from the same matlab script

• dnn.ipynb: Contains the neural network model (same structure as in the presceding article) written in Python using TensorFlow.

• hls_fpga_accelerator.cpp: This file contains a hardware accelerator implemented in High-Level Synthesis (HLS) C++ and deployed on the Zynq UltraScale+ MPSoC SOM for inference (achieved 47.2% reduction in inference time compared to the SoC CPU).

The corresponding driver and PYNQ deployment code will be released soon. For early access, please reach out at [email protected].

System Setup

Note: the upper part is the training setup and the other is for inference

FPGA Block Diagram