A Comparison Between CPU and GPU Computing in DNN-Based DoA Estimation

Deep Learning algorithms have recently become accessible to modern Telecommunication systems due to the advancement in both hardware and software technology. The massively parallel computational tasks of Artificial Intelligence (AI) training are now directly performed in Graphic Processing Units (GPUs), which contain many processing cores. This paper tests the improvement in computing time for Deep Neural Network (DNN) architectures that solve the Direction of Arrival (DoA) estimation problem as a classification task by utilizing the Nvidia CUDA framework. The training and testing data are generated by receiving signal information from a simulated antenna array. The input of the DNN is a 64 × 64 image with two channels, whereas the problem is modelled as a multi-label classification task. The purpose of this research is to observe the acceleration that the GPU provides in the DoA estimation classification problem. The project includes benchmarks from a wide range of computing systems and the results demonstrate the potential of DNN-based DoA estimators as a modern technology that can be deployed on available commercial hardware.