Tiny Sentinel is a one-tile hardware sensor-integrity watchdog for Tiny Tapeout. It monitors an 8-bit sensor stream, maintains an adaptive baseline, and flags spike, stuck-at, and slow-drift anomalies in real time using deterministic shift/add logic, small counters, and saturating arithmetic.
The project demonstrates a practical embedded ASIC block: catching sensor failures directly in silicon before software has to poll, filter, or diagnose the stream.
Tiny Sentinel watches ui_in[7:0] as a digital sensor sample. After calibration, it compares each sample against the previous sample and the learned baseline:
- Sudden jumps raise
spike_detected. - Repeated identical samples raise
stuck_detected, including repeated zero. - Sustained deviation from the baseline raises
drift_detected. - A 3-bit score rises on anomaly events and decays on clean samples.
- A latch records that an anomaly happened until cleared.
| Pin | Direction | Meaning |
|---|---|---|
ui_in[7:0] |
input | Current 8-bit sensor sample |
uo_out[0] |
output | Any anomaly |
uo_out[1] |
output | Spike / sudden jump detected |
uo_out[2] |
output | Stuck-at sample detected |
uo_out[3] |
output | Slow drift detected |
uo_out[6:4] |
output | Saturating anomaly score |
uo_out[7] |
output | Latched alert |
uio_in[0] |
input | Learn/adapt baseline enable |
uio_in[1] |
input | Strict mode |
uio_in[2] |
input | Freeze baseline |
uio_in[3] |
input | Clear latched alert |
uio_out[7:4] |
output | Baseline high nibble |
uio_out[3:0] |
output | Tied low |
uio_oe is fixed to 8'b11110000, so uio[3:0] are controls and uio[7:4] are debug outputs.
The baseline tracker initializes on the first learned sample. After that, when learning is enabled and freeze is low, it moves toward the current sample by a power-of-two fraction of the difference. If the shifted difference is zero but the sample differs, the baseline still moves by one count.
The detector uses two distances:
abs_diff = |sample - baseline|
jump_diff = |sample - previous_sample|
Normal mode thresholds:
| Detector | Threshold |
|---|---|
| Spike | jump_diff >= 24 |
| Stuck | same sample for 8 repeats |
| Drift | abs_diff >= 16 for 4 non-spiking cycles |
Strict mode thresholds:
| Detector | Threshold |
|---|---|
| Spike | jump_diff >= 12 |
| Stuck | same sample for 4 repeats |
| Drift | abs_diff >= 8 for 2 non-spiking cycles |
During a demo, use learn mode for calibration. For detection, deassert learn mode or assert freeze mode. If learn remains enabled, the detector still works, but slow-drift detection becomes less sensitive because the baseline tracks the signal.
The cocotb testbench covers:
- Reset and initialization.
- Baseline learning and debug output.
- Normal mild variation.
- Spike detection.
- Stuck detection, including repeated
8'd0. - Drift detection with exact threshold-cycle timing.
- Strict-mode thresholds.
- Frozen baseline behavior.
- Alert latch clear priority.
- Score decay.
- Fixed
uio_oemapping.
Run:
cd test
make -BThe passing simulation writes test/tb.fst for waveform inspection.
The design is fully synchronous to clk with active-low reset through rst_n. It uses synthesizable Verilog only: no delays, no initial blocks in source, no implicit wires, and all outputs are assigned. It is designed to fit within a standard 1x1 TinyTapeout tile using simple counters, comparators, shift-based averaging, and saturating arithmetic.
Passing RTL waveform:
GDS render from the TinyTapeout GDS workflow:
The deployed TinyTapeout viewer is available at taz33m.github.io/tiny-sentinel.
Tiny Sentinel expects an 8-bit digital sample stream and does not digitize analog signals. The baseline is intentionally simple, so it is easy to synthesize and explain, but it is not a statistical model. Thresholds are fixed in hardware except for normal versus strict mode.