On the Performance of Quantized Neural Networks based Digital Predistortion for PA linearization in OFDM systems

Neural networks (NNs) based digital predistortion (DPD) have been shown to be a very promising technique to enhance power amplifier linearity. However, studies consider high level of quantization NNs (32-bits) whose hardware implementation is not feasible in practice. This paper addresses the challenge of low level quantization of NNs based DPD for PA linearization in OFDM communication systems. The goal is to improve the overall inference time, resources and energy efficiency. To perform quantization, we first operate a post training quantization to assess the impact of quantization on the NN. Second, we perform a quantization aware training to cope with the quantization noise. Thus, we truly believe that the proposed quantization approach puts forward an efficient transmission chain using NN DPD for OFDM based wireless communication systems. Indeed, numerical simulations show that using 4-bits quantization, resource usage is reduced by 62% compared to the 32-bits quantization. Moreover, the error vector magnitude is still lower than −35dB which is a slight degradation compared to 32-bits quantization.