Lab 11.32 — Two-board coarse-CFO acquisition in the FPGA fabric

Objective

Demonstrate that the shipped qpsk_coarse_cfo.v block acquires a live QPSK frame inside the PL across a controlled 0–55 kHz inter-board carrier offset. The comparison keeps the RF path, frame, attempt budget and timing-phase sweep identical and changes only gp_ctrl[13]: coarse correction enabled versus Costas-only reception.

This is an acquisition-range experiment, not a long-duration BER claim. A zero-error frame at one timing phase proves that the carrier loop can acquire that CFO. Repeatability is reported separately as clean-attempt rate, lock rate and aggregate BER over full frames.

Conducted setup

board A: vendor image, cyclic DMA TX at 915 MHz + Δ
    TX1 ── 30 dB attenuator ── SMA cable ── RX1
board B: course bitstream, fabric RX + optional coarse-CFO + Costas + BER counter

Parameter	Value
Carrier	915 MHz
Symbol/sample rate	480 ksym/s / 3.84 Msps
Frame	140 QPSK symbols, 280 scored bits
TX gain	−30 dB
RX gain	50 dB
Start offsets	0…7
Attempts	3 per offset and mode, 24 per CFO and mode
CFO sweep	0…55 kHz in 5 kHz steps

The transmitter replays a bit-exact frame reconstructed from bpsk_frame_bits.mem and the course QPSK mapper. The receiver and BER counter run in the course bitstream. Both boards are forced to −89.75 dB before setup and in finally, including after an exception.

Metric contract

Metric	Meaning
reached_zero / best_ber	At least one full 280-bit frame was decoded without errors; evidence of acquisition.
clean_attempt_rate	Zero-error full frames divided by every attempt, including no-lock attempts.
lock_rate	Attempts that returned all 140 symbols divided by every attempt.
aggregate_ber	BER over full frames only; no-lock attempts are not silently counted as zero errors.
Wilson interval	95% interval for lock and clean-attempt proportions.

The JSON retains every attempt, so the headline can be regenerated without selecting rows by hand.

Measured result

Measured 2026-07-20 over the conducted 30 dB path. Coarse correction acquired at BER=0 at 12/12 CFO points and produced 75/288 clean attempts (26.0%). Costas-only produced 0/216 clean attempts at the nine discriminating points from 15 to 55 kHz.

CFO, kHz	Coarse clean	Coarse lock	Coarse aggregate BER	Costas best BER	Costas clean	Costas lock
0	7/24	50.0%	0.175	0	8/24	66.7%
5	6/24	70.8%	0.267	0.400	0/24	70.8%
10	6/24	79.2%	0.302	0.464	0/24	83.3%
15	5/24	66.7%	0.245	0.493	0/24	25.0%
20	9/24	79.2%	0.203	no lock	0/24	0%
25	7/24	83.3%	0.227	no lock	0/24	0%
30	7/24	70.8%	0.232	0.439	0/24	12.5%
35	8/24	87.5%	0.276	0.468	0/24	25.0%
40	2/24	58.3%	0.395	0.443	0/24	16.7%
45	3/24	58.3%	0.361	0.514	0/24	4.2%
50	10/24	70.8%	0.136	0.454	0/24	4.2%
55	5/24	66.7%	0.171	no lock	0/24	0%

The result validates the CFO acquisition range: the in-fabric estimator finds a clean frame at every tested CFO while Costas alone does not at 15–55 kHz. It does not yet validate a robust continuous link. The 26% clean-attempt rate and high conditional aggregate BER show the remaining sample-clock/timing-phase problem clearly; the next receiver improvement is continuous timing recovery or a disciplined burst-phase search.

Reproduce

First check the bit-exact host waveform without radio:

python python/lab_11_32_two_board_fabric_coarse_cfo.py --self-test

Then run the measured sweep from the repository root:

python blocks/block_11_integrated_sdr_project/python/lab_11_32_two_board_fabric_coarse_cfo.py \
  --host-a 192.168.40.1 --host-b 192.168.20.1 \
  --cfo-start 0 --cfo-stop 55000 --cfo-step 5000 \
  --retries-per-offset 3

The canonical machine-readable result is docs/assets/lab1132_two_board_fabric_coarse_cfo.json. The command exits non-zero unless coarse correction reaches BER=0 at every CFO point and Costas-only reaches none of the points with |CFO| ≥ 15 kHz.

RF safety

Use only the contained cabled path with the 30 dB attenuator. Do not attach antennas with these settings. Do not raise TX power to compensate for a weak link; inspect the cable, attenuator and RX gain instead. On interruption, verify both boards are back at −89.75 dB before disconnecting them.