Lab 11.29 — Cold-boot BER reliability campaign¶

Objective¶

Turn "the modem decodes" into a number with a confidence interval. A single 10/10 says nothing about the tails; this reboots the board many times, reloads the bitstream from a clean state each time, and counts frames through the deterministic PL fabric loopback so that a run of zero errors becomes a bounded BER rather than an unsupported "BER = 0".

What it measures¶

Each run:

cold-boots the board (soft reboot by default: Linux + AD9361 re-init + overlay reload from stock; or a true power cycle, see below);
reloads our bitstream and checks the gpreg core id 0x4250534B — a load-failure detector;
sprays --frames frames through the fabric loopback (gp_ctrl[6]: TX looped into RX inside the PL, no AD9361, no RF, bit-deterministic) for BPSK (281 bits/frame) and QPSK (280 bits/frame), counting received frames, clean frames, and bit errors.

Fabric loopback is the PL-correctness metric: it exercises the whole synthesized mapper → upsampler → RRC → sampler → decision → frame-sync → BER-counter through the real gpreg/CDC plane. Because it is error-free, every clean frame adds to a large deterministic bit count.

Statistics (no third-party dependency)¶

Run rates (loaded, all-clean) — Wilson score interval, plus Clopper–Pearson when scipy is importable.
BER with zero errors in m bits — one-sided upper bound 1 − 0.05^(1/m) ≈ 3/m (rule of three). Reported as BER < X, never BER = 0.
FER (frames with ≥1 error / returned frames) — Wilson interval.

The JSON summary is rewritten after every run, so a stopped or killed campaign still leaves a complete, honest partial result.

Run¶

python blocks/block_11_integrated_sdr_project/python/lab_11_29_cold_boot_ber_campaign.py \
  --runs 50 --frames 200

--runs 30–100 gives a useful interval width. The default --bit is tmp/bridge_txrx_mux.qpsk.wordswap.bit.bin. The board is rebooted to stock at the end.

True power cycle¶

The default "cold boot" is a soft reboot — what can be scripted without switchable power. With a controllable supply, pass a command that cycles the board's power and the campaign runs it in place of the soft reboot, then waits for the board to return:

python .../lab_11_29_cold_boot_ber_campaign.py --runs 50 \
  --power-cycle-cmd "your-smart-plug-cli --off && sleep 3 && your-smart-plug-cli --on"

Result (2026-07-11, 50 soft cold boots, 200 frames/mode)¶

	loaded	all-clean	bits	BER (95% upper)
BPSK fabric	—	—	2,810,000	< 1.07 × 10⁻⁶
QPSK fabric	—	—	2,800,000	< 1.07 × 10⁻⁶
combined	50/50	50/50	5,610,000	< 5.34 × 10⁻⁷

Run-clean rate 50/50, Wilson 95% [0.929, 1.0]; zero load/boot failures.
FER 0 on both modulations, Wilson 95% [0, 3.84 × 10⁻⁴]; error histogram {0: 10000} each.

So the synthesized PL modem comes up decoding on every cold boot, with a sub-microsecond BER bound over 5.6 M deterministic bits.

Timing-closure robustness¶

Companion tool tools/timing_directive_sweep.tcl checks whether the shipped timing closure is a property of the design or of one lucky placement. Vivado 2021.1 has no numeric placement seed, so it sweeps the placement directive (a stronger test — it swaps the algorithm):

/g/Xilinx/Vivado/2021.1/bin/vivado.bat -mode batch -notrace -source tools/timing_directive_sweep.tcl

Result — ROBUST, all five strategies met timing:

directive	WNS (ns)	TNS
Default	+0.087	0
Explore (shipped)	+0.123	0
WLDrivenBlockPlacement	+0.017	0
ExtraNetDelay_high	+0.171	0
ExtraPostPlacementOpt	+0.122	0

Worst across genuinely different placements is +0.017 ns (WLDriven) — positive, and the margin to watch as the RTL grows. The sweep leaves impl_1 on the last directive; re-run the normal build to restore the shipped Explore placement.

RF safety¶

Fabric loopback never configures or raises TX — the board stays at the stock −89.75 dB throughout, and the campaign reboots to stock at the end. Nothing radiates.