Computational Efficiency of Deep Learning-Based Super Resolution Methods for 5G-NR Channel Estimation

The increasing demand for high-quality and efficient Channel Estimation (CE) in 5G New Radio (5G-NR) systems has prompted the exploration of advanced Deep Learning (DL) techniques. While traditional methods, such as Linear Interpolation (LI) and Least Squares (LS), provide reasonable accuracy and are practical for real-time physical layer processing, recent DL-based CE approaches have primarily focused on accuracy, often without evidence of real-time capabilities. In this paper, we present a comprehensive evaluation of DL-based Super-resolution (SR) methods for CE, comparing models like Super Resolution Convolutional Neural Network (SRCNN), ChannelNet, and Enhanced Deep Super-Resolution (EDSR) in both 1D and 2D convolutional architectures. We optimize these models using NVIDIA TensorRT to reduce computational complexity and latency. Our results show that the optimized 1D-EDSR model achieves the best performance with a Mean Squared Error (MSE) of 0.0126, outperforming all other models in terms of accuracy. However, the optimized 1D-EDSR model fails to meet real-time constraints due to additional computational overhead (0.6798 ms/sample). In contrast, the 1D-SRCNN model offers a balanced trade-off between MSE (0.01738) and inference time (0.0866ms/sample), achieving 40% higher accuracy than LS (0.0288) while maintaining the best energy efficiency (1.48 mJ/sample).