MR-Transformer model sets

RML2016.10a, no-convolution MR-Transformer for the Arty S7-50 FPGA (Vitis HLS + Vivado 2026.1).

Summary

Accuracy at +18 dB
93.36 %6 enc
92.59 %3 enc
Parameters
1,196,8116 enc
601,9953 enc
Weights at 16-bit (BRAM is 337.5 KB)
2,338 KB6 enc
1,176 KB3 enc
Metric6 encoders3 encodersChange
Parameters1,196,811601,995−50 %
Accuracy +18 dB93.36 %92.59 %−0.77
Mean accuracy SNR ≥ 0 dB92.60 %92.06 %−0.54
Accuracy, all SNR62.76 %62.60 %−0.16
Weights 16-bit2,338 KB1,176 KB−50 %
Weights 8-bit1,169 KB588 KB−50 %
QPSK test signal 199018, P(QPSK)0.92190.9265+0.0046
Front-end C-sim error (S1–S7)1.79e-051.88e-05*—

* Predicted with a bit-accurate emulation; confirm by running C-sim on S7.

Accuracy by class at +18 dB

Class6 encoders3 encodersChange
8PSK98.5 %97.0 %−1.5
AM-DSB, CPFSK, GFSK100.0 %100.0 %0.0
AM-SSB96.5 %96.5 %0.0
BPSK98.5 %98.5 %0.0
PAM499.0 %99.5 %+0.5
QPSK99.0 %99.0 %0.0
QAM1697.0 %93.5 %−3.5
QAM6494.5 %90.0 %−4.5
WBFM44.0 %44.5 %+0.5

Files

6 encoders

Original model
MRT_FrontEnd_GuideArch.zip7-step FPGA front end (S1–S7), ROM data, roadmap 624 KBDownload
Complete_Flow_6Enc_results_verification_reference.zipReference outputs, verification, test results, function sheet 637 KBDownload
RML2016_Transformer_Complete_Flow.zipFull flow: training, weight CSVs, outputs, results 13.6 MBTo be added

3 encoders

Recommended for FPGA
MRT_FrontEnd_GuideArch_3Encoders.zip7-step FPGA front end, ROM data for the 3-encoder model 632 KBDownload
RML2016_Transformer_Complete_Flow_3Encoders.zipFull flow: training, weight CSVs (Wq/Wk/Wv split), outputs, results 8.3 MBTo be added