Foundry workbench
A working 64–32–10 binary neural net in NAND and LATCH. TapeOut cells take the dense XNOR array. Behemoth cells take the popcount trees and weight RAM.
Probe station
64-bit vector
Argmax
—
Draw a digit or load a glyph, then run the pipeline. Scores are 2·feature XNOR + template XNOR (max 128).
Hybrid die
Dense XNOR synapses sit in TapeOut NAND. Wide popcount trees and weight RAM use Behemoth cells at 6× drive so the critical path stays inside one clock at 12 MHz.
Clocked pipeline
cycle 0
Seven stages from pad latch through winner-take-all. Step or run the clock.
Traveler / BOM
Gate probe
pop 5 / thr 3 · 1
Each shown cell is a 5-NAND XNOR against a weight-1 mask bit. The Behemoth popcount tree sums all 8 hits and the threshold latch writes hidden[8].
Use case
————
Utility meters, elevator floors, and kiosk PINs are four bright glyphs on a dark field. A 64–32–10 binary net built only from NAND and LATCH classifies each cell at microwatts, and the same netlist can be re-simulated bit-for-bit — the inference is the circuit.
·
·
·
·