266 Frozen ternary backbone + loadable head
266 : Frozen ternary backbone + loadable head
- Author: Rahul Mascarenhas
- Description: 8x8 image classifier. Frozen ternary feature layer compiled into logic; 6-class head reloaded over SPI.
- GitHub repository
- Open in 3D viewer
- Open in Tiny Tapeout Explorer
- Clock: 10000000 Hz
How it works
A 12x12 binary image, classified in two stages.
Frozen backbone, 144 -> 56. 3,584 ternary weights compiled into the
netlist at synthesis: +1 is a wire, -1 an inverter, 0 deletes the input
and its branch of the adder tree. About half are zero. One pixel enters per
cycle and eight accumulator lanes sweep the image seven times; the sign of
each accumulator is a feature. These weights cannot be changed.
Loadable head, 56 -> 6. 336 ternary weights and six 8-bit biases in a 720-bit shift register, reloaded as 90 bytes. One accumulator walks them and emits six signed scores, then picks the winner on-die.
1,344 cycles per inference, 134 us at 10 MHz.
One inference gives you three things: the six raw scores, the winning class, and the 56 frozen features. Loading a different 90 bytes retargets the chip to six different characters; reading the features instead lets a host run its own classifier.
Trained on EMNIST byclass. A fresh head on six unseen characters loses about five points against a fully trained network.
How to test
| Pin | Function |
|---|---|
ui[0] / ui[1] |
image bit / shift |
ui[2] |
start |
ui[3] / ui[4] |
head bit / shift |
ui[5] |
advance feature readout |
uo[7:0] |
signed score while busy; winning class in uo[2:0] once uio[1] is high |
uio[0] / uio[1] / uio[2] |
score valid / done / busy |
uio[5:3] |
which class the current score belongs to |
uio[6] / uio[7] |
feature bit / last feature |
Reset. Shift 720 head bits in on ui[3], MSB first. Shift 144 pixels in on
ui[0], MSB first. Pulse ui[2]. Read uo on each uio[0] pulse for the
six scores, then read it once more after uio[1] goes high for the winner.
The features latch when the backbone finishes and hold until the next start,
so ui[5] can be pulsed 56 times to read them on uio[6] at any point after
that, including while the head is still running.
Input must be normalised the way EMNIST is or accuracy drops sharply: crop to the ink, scale so the longer side fills 20/28 of the frame, centre by centre of mass, then average-pool to 12x12 and threshold at 0.15. Do the scaling at full resolution, not on the 12x12 grid.
A head bitstream is only valid for the preprocessing it was fitted against.
External hardware
None. The chip reports its own answer, so a host is optional.
IO
| # | Input | Output | Bidirectional |
|---|---|---|---|
| 0 | img_bit | score[0] / class[0] | score_valid |
| 1 | img_shift | score[1] / class[1] | done |
| 2 | go | score[2] / class[2] | busy |
| 3 | head_bit | score[3] | class[0] |
| 4 | head_shift | score[4] | class[1] |
| 5 | feat_shift | score[5] | class[2] |
| 6 | unused | score[6] | feat_bit |
| 7 | unused | score[7] | feat_wrap |