Evaluation on User Equipment Chip for Deep Learning based Channel Estimation in 5G Advanced System

Nowadays, the integration of conventional 5G communication systems and artificial intelligence (AI) has become one of the important trends for the evolution of future advanced systems. With the employment of AI, many works have verified that significant performance gains can be achieved for varieties of tasks. However, these works mainly focus on the performance improvement and ignore the practical feasibility. In this paper, we investigate the performance of a convolutional neural network (CNN) based channel estimation scheme to run on two powerful mobile terminal chips with different quantization types. In particular, the accuracy of channel estimation and the computation capability of terminal chips for supporting model inference are evaluated. Simulation results show that the accuracy losses caused by both quantization types are small in the mobile scenario of speed at 120 km/h. However, in the case of speed at 350 km/h, the INT8 quantization type leads to a performance degradation while the $\Gamma$P16 quantization type can still maintain a satisfying performance. In addition, the results also show that further optimization for AI based computing is essential for guaranteeing a promising pratical deployment more reliably. Finally, the achievable demodulation performance gain is much smaller than the channel estimation gain, which should be explored fully in the future works.