16-Neuron Binary Neural Network

A Binary Neural Network (BNN) inference layer implemented directly in silicon on a Tiny Tapeout tile. 16 independent binary perceptron neurons each classify the same 8-bit input vector simultaneously, producing a 16-bit output in a single clock cycle — with no CPU, no software, no operating system.

Silicon Layout

The image below shows the actual physical layout of the 16-neuron BNN as it will be etched into silicon — generated by the OpenLane ASIC toolchain for Tiny Tapeout.

The Mathematical Model

Each neuron n computes:

y[n] = 1   if   Σ( w[n][i] AND x[i] ) >= θ[n]
y[n] = 0   otherwise

Because all values are binary, multiplication reduces to a logical AND — orders of magnitude cheaper in silicon area and power than floating-point arithmetic.

Inside Each Neuron

Stage 1 — AND Array: Each input bit is ANDed with its corresponding weight bit in parallel across 8 gates.

Stage 2 — Adder Tree (Popcount): The 8 product bits are summed using a tree of half-adders and full-adders, producing a 4-bit count S[n] between 0 and 8.

Stage 3 — Threshold Comparator: If S[n] >= θ[n], the neuron fires. A single bit — yes or no.

Why 16 Neurons in Parallel?

All 16 neurons share the same 8-bit input bus but each has its own independent weight register (8 bits) and threshold register (4 bits). Each can be programmed to recognise a different pattern. The result is a 16-bit output vector answering 16 different yes/no questions about the input simultaneously.

Why this is genuinely AI

This implements the McCulloch-Pitts neuron (1943) — the mathematical model that founded neural networks. Every modern AI system is built from billions of this computation. BNNs are an active research area for ultra-low-power AI inference at the edge, where AND+popcount replaces expensive multiply-accumulate operations.

Weights can be trained in Python

from perceptron_trainer import train_perceptron, generate_load_instructions
import numpy as np

X = np.random.randint(0, 2, (200, 8))
y = (X.sum(axis=1) > 4).astype(int)

weights, threshold, _ = train_perceptron(X, y, epochs=100)
generate_load_instructions(weights, threshold)

Pin Mapping

View on GitHub

Tiny Tapeout