mmWave MIMO demapping goes block-wise: up to 17.6× lower BER over the air, 34.3× faster on RFSoC
Synopsis
Soft demapping in mmWave MIMO receivers no longer treats each symbol on its own. Prior learned demappers used multilayer perceptrons on single symbols, while classical demappers rely on analytical decision rules and a Gaussian noise assumption, so neither exploits residual distortion structure across neighboring symbols. This work applies a one-dimensional convolutional network to whole blocks of equalized symbols; on over-the-air data from a 60 GHz, four-stream, 1.76 GHz-per-stream link, the single-stream variant reaches up to 2.6× BER gain over the conventional soft demapper for 16QAM and up to 17.6× for 64QAM, and a fixed-point RFSoC implementation runs about 34.3× faster.
Interpretation
The unit of demapping changes from a single symbol to a whole block of equalized symbols, while the rest of the receiver chain stays unchanged. Earlier learned demappers either processed symbols one at a time with multilayer perceptrons or used convolutional self-attention for joint processing at higher computational cost, and classical demappers likewise worked symbol by symbol on analytical decision rules. On an IEEE 802.11ay single-carrier waveform at 60 GHz with four spatial streams and 1.76 GHz per stream, tested on over-the-air data collected across multiple days, the single-stream variant reaches a maximum BER gain of 2.6 and an average of 1.8 for 16QAM, and a maximum of 17.6 and an average of 8.7 for 64QAM over the conventional soft demapper.
A one-dimensional convolution extracts local features along time, bringing neighboring-symbol context into soft-bit estimation. Symbol-wise methods cannot use correlation between adjacent symbols, whereas measured residual errors show a weak short-range dependency in phase, mainly at lag one. When the convolution kernel covers only a single symbol, the learned demapper's BER is nearly identical to the conventional demapper across most of the SINR range; widening the receptive field so each symbol is processed jointly with its neighbors makes it consistently outperform the conventional demapper across the whole range. Randomly shuffling symbol order within each block pushes the learned demapper's BER below the conventional demapper, showing the gain comes from temporal relationships.
The trained single-stream model keeps comparable BER in new environments without retraining. Earlier evaluation drew data from different days but within the same laboratory environment, without exposure to markedly different propagation scenarios. In an indoor server room and an outdoor open space, neither present in training, the single-stream demapper's BER curves for 16QAM and 64QAM closely match its behavior in the original laboratory.
The fixed-point demapper becomes a reusable convolution engine that streams into the receiver chain on an RFSoC. A processor-only implementation needs 17.5 ms per block, and a single-precision floating-point accelerator brings that to 0.95 ms. With a fixed-point format and two accelerators in parallel, execution time falls to 0.51 ms per block, about 34.3× faster than the processor implementation, while DSP usage drops 71.4% relative to the floating-point design.
Perspective
The demonstrated scope is an IEEE 802.11ay single-carrier mmWave MIMO link at 60 GHz with four spatial streams and 1.76 GHz per stream, measured indoors and outdoors under line-of-sight conditions across multiple days and transmit gains, with training and test data separated by day. The intended readers are researchers and engineers working on wireless physical-layer receivers and hardware acceleration. The authors propose future work on higher-order MIMO configurations and on mobility and rapidly time-varying channels.
The multi-stream variant's BER flattens around 19 dB SINR for 64QAM while the single-stream variant keeps falling across the range, which the authors attribute to residual distortion after equalization being largely stream-local. Adding channel frequency response or guard interval as extra inputs yields only limited gains, and the guard interval more than doubles trainable parameters. Fixed-point implementation raises BER from 0.0101 to 0.0108 on a representative dataset.