Python · SystemVerilog · ngspice · Verilator · Yosys/nextpnr (ECP5)
A modelled implant sensing chain that decodes finger velocity within 5% of a conventional design at ~40× less power (0.48 µW/channel).
- Modelled a full implant sensing chain (electrode → amplifier → ADC → features → decoder) under one power model; the best design uses 0.48 µW/channel and decodes finger velocity within 5% of a 19.6 µW conventional chain (~40× less power). All results are simulated; nothing has run on hardware yet.
- Swept amplifier current, ADC resolution, sample rate and feature type on public Neuralink, Nason and NoMAD recordings to map decode accuracy against µW, arriving at a 250 nA amplifier, 8-bit ADC at 2 kS/s and spiking-band power (amplifier = 93% of power).
- Modelled how stimulation artifacts spread and saturate the front end, and built a fixed-point template canceller that keeps 86% of samples (vs 60% for blanking) and holds decoding at R² 0.707 (vs 0.718 without stimulation).
- Simulated 39 weeks of electrode degradation; a running z-score stabiliser recovers decoding from R² 0.587 to 0.627 for 0.00034 µW/channel.
- Implemented the feature path in SystemVerilog, verified bit-exact against the Python reference over 348M samples (0 mismatches); synthesises for a Lattice ECP5 in 489 LUTs for 96 channels.
- Next: an ADS1299-based recording board with a saline phantom and stimulator to test the model on real signals.