A 5.76 Gb/s 79.7 pJ/b 128×32 Massive Deep-Learning Uplink MIMO Detector in 28nm CMOS Technology

This work presents the first fabricated massive deep-learning (DL) multiple-input multiple-output (MIMO) detector for a $128 \times 32$ antenna configuration and 4 to 256QAM modulations, offering near-optimal detection across diverse channel conditions. Our design employs the fully parallel initialization procedure followed by the fully unfolded DL operations, enhancing the data rate by 13 times compared to the baseline design. We also propose a novel approximate projection to relax the processing cost per layer and a gradient reusing to eliminate all the matrix-vector multiplications (MVMs) in the last few layers, saving the overall detector complexity by 38%. Targeting the 256QAM system, our prototype in a 28 nm CMOS process offers a throughput of $5.76 \mathrm{~Gb} / \mathrm{s}$ and an energy efficiency of $79.7 \mathrm{pJ} / \mathrm{b}$.