898 Systolic Binary Neural Network Accelerator
898 : Systolic Binary Neural Network Accelerator
- Author: Dean Foulds
- Description: V2 systolic 16-neuron BNN accelerator. XNOR dot product engine, hardware feature expansion, signed bias, balanced popcount tree. 8-cycle systolic compute reuses hardware for smaller silicon footprint.
- GitHub repository
- Open in 3D viewer
- Clock: 10000000 Hz
Systolic Binary Neural Network Accelerator
What is this chip?
This chip implements a 16-neuron Binary Neural Network (BNN) inference accelerator in silicon. It classifies an 8-bit input vector using 16 independently programmable neurons, each with learnable binary weights and a signed bias. The design is inspired by the systolic array architecture invented by H.T. Kung at Carnegie Mellon University in 1978, and implements the XNOR-popcount computation that is the standard in modern BNN research.
How it works
Each neuron computes a dot product between its weight vector and a set of derived input features, then fires if the result exceeds a learned bias:
feat = feature_expand(ui_in)
S[n] = popcount( XNOR( weights[n], feat ) )
y[n] = 1 if S[n] + bias[n] >= 0
y[n] = 0 otherwise
The compute engine processes one bit per clock cycle over 8 cycles (systolic), reusing hardware rather than duplicating it 16 times. This reduces silicon area significantly compared to a fully parallel design.
See the detailed documentation:
Pin mapping
| Pin | Direction | Function |
|---|---|---|
clk |
in | System clock |
rst_n |
in | Active-low reset |
ui_in[7:0] |
in | Input features (infer) or load data (load) |
uio_in[0] |
in | Mode: 0=load, 1=infer |
uio_in[1] |
in | Target: 0=weights, 1=bias |
uio_in[5:2] |
in | Neuron select 0–15 |
uo_out[7:0] |
out | Fire signals neurons 0–7 |
uio_out[7:0] |
out | Fire signals neurons 8–15 |
How to test
Load weights for neuron n
- Set
uio_in[0]=0,uio_in[1]=0,uio_in[5:2]=n - Set
ui_in[7:0]= 8-bit weight pattern - Pulse clock
Load bias for neuron n
- Set
uio_in[0]=0,uio_in[1]=1,uio_in[5:2]=n - Set
ui_in[3:0]= bias magnitude,ui_in[4]= sign (1=negative) - Pulse clock
Run inference
- Set
uio_in[0]=1 - Set
ui_in[7:0]= input feature vector - After 8 clock cycles read
uo_outanduio_out
Train weights 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, bias, _ = train_perceptron(X, y, epochs=100)
generate_load_instructions(weights, bias)
External hardware
No external hardware required. A Raspberry Pi or Arduino can load trained weights and run inference via the ui_in and uio_in pins. See info_v2.md for full Raspberry Pi wiring and Python code.
End
IO
| # | Input | Output | Bidirectional |
|---|---|---|---|
| 0 | x0 - input feature bit 0 | fire0 - neuron 0 output | mode - 0=load 1=infer |
| 1 | x1 - input feature bit 1 | fire1 - neuron 1 output | target - 0=weights 1=thresholds |
| 2 | x2 - input feature bit 2 | fire2 - neuron 2 output | sel0 - neuron select bit 0 |
| 3 | x3 - input feature bit 3 | fire3 - neuron 3 output | sel1 - neuron select bit 1 |
| 4 | x4 - input feature bit 4 | fire4 - neuron 4 output | sel2 - neuron select bit 2 |
| 5 | x5 - input feature bit 5 | fire5 - neuron 5 output | sel3 - neuron select bit 3 |
| 6 | x6 - input feature bit 6 | fire6 - neuron 6 output | fire8 - neuron 8 output |
| 7 | x7 - input feature bit 7 | fire7 - neuron 7 output | fire9 - neuron 9 output |