A 596M-parameter ternary model on a $299 Digilent Arty A7-100T. All twenty-eight transformer blocks run on the FPGA, reading 110 MB of {−1, 0, +1} weights out of the board's own DDR3. The host tokenizes, looks up the embedding, and turns the last hidden state back into logits — nothing else.
This student was distilled on SST-2, so binary sentiment is the job it was trained for and this is it doing that job. One forward pass per token of the input; the answer is the gap between two entries of the logit vector.
Roughly 2.9 s a token at the start of a context and 4.7 s at 512,
because attention walks the KV cache and the rest of a block does not.
The board's arithmetic is about a tenth of a per cent of that; the rest is
a soft CPU doing elementwise work and handing the array bytes one 32-bit
register write at a time. That is the finding, not the speed.
On the generate tab the output repeats one word. That is the model, not
the machine — a sentiment classifier asked to write prose. What it
demonstrates is that the board matches a float64 reference token for token.
github.com/Ternarycore/ternarycore
· CERN-OHL-S v2