Abstract
Hardware-efficient quantization strategies are imperative for massive MIMO detectors. However, conventional methods such as unified quantization and manual tuning respectively suffer from bitwidth redundancy and prohibitive exploration overhead. To address these challenges, this letter proposes a deep learning-driven sensitivity-aware quantization (DeepSAQ) framework, which includes: 1) A gradient-driven framework with a Bayesian optimized loss function to minimize fractional bitwidths; 2) Sensitivity-aware tuning to compress the fractional bitwidth redundancy; 3) Range-extended integral bitwidths to further improve the quantization efficiency. Applying to the unfolded MMSE-wNSA detection, the proposed framework delivers floating-point comparable performance but reduces 50.04% average bitwidths and 50.47% computational complexity compared to the unified quantization. Benchmarking with the recent art, the proposed framework succeeds in squeezing up to 18.32% bitwidths and 21.23% computational complexity in 8 $\times $ 128 64-QAM MIMO systems.
Funding Information
  • National Natural Science Foundation of China (62331009)
  • National Natural Science Foundation of China (62401642)
  • Natural Science Foundation of Jiangsu Province (BG2024004)
  • Natural Science Foundation of Jiangsu Province (BM2023016)
  • Fundamental Research Funds for the Central Universities

This publication has 14 references indexed in Scilit: