Earlier projects here fitted a small streaming keyword-spotting engine into Tiny Tapeout tiles, but stopped the model sweep at 64 hidden units, so every tile from a 1x2 up just got "the biggest model I tried". research/accel-ttbig removes that limit: 64 Yosys/SKY130 synthesis runs at 96, 128, 192 and 256 hidden units, with the RTL checked against an integer reference.
What came out:
- Area is nearly linear, about 250 µm² per hidden unit.
- The best model that fits a 1x2 tile (36,347 µm²) scores 0.737 on the test set. The next one up, 128 units, scores 0.752 and needs a 2x2 tile (75,603 µm²): +0.015 accuracy for 2.08 times the area.
- At 256 units with 8-bit weights, a random model overflows the 20-bit accumulator the small designs use. A 24-bit one costs 0.1 to 2.1% more area.
What is not claimed: accuracy above 64 units was not retrained (the dataset host returned 403 in the container; the 128-unit figures are from an earlier project), trained models were not checked against the accumulator limit, and utilisation and the clock are assumptions. Details and the numbers: research/accel-ttbig.