SYMBEX-1

Inference. Reduced to one bit.

A fully binary neural network. Compiled to C. Checked against its reference on the chip.

Try the bitwise lab See measured results ↗

Count / thresholdToy example · 8 bits

INPUT

10110010

XOR each position

WEIGHT

11010100

XOR

01100110

One small primitive. Not a full neural layer.

Scroll to see the whole system

1-bitWeights & activationsXORFind the differencesPOPCNTCount the onesTInteger thresholds.hGenerated C headerS3ESP32-S3 PIE SIMD100%Bit-exact on chip1-bitWeights & activationsXORFind the differencesPOPCNTCount the onesTInteger thresholds.hGenerated C headerS3ESP32-S3 PIE SIMD100%Bit-exact on chip

24.3KB of ResNet-8 weights40.7ms on one coreCIFARImage classificationFOMOPerson detectionHALHardware layerORResidual shortcutsAGPLOpen source24.3KB of ResNet-8 weights40.7ms on one coreCIFARImage classificationFOMOPerson detectionHALHardware layerORResidual shortcutsAGPLOpen source

By the numbers

Small by design.
Measured on silicon.

Every figure below links to the repository result it came from.

01bit

Weights and activations, every layer.

How it works ›

24.3KB

ResNet-8 model weights · CIFAR-10.

Repository-reported ›

40.7ms

ESP32-S3 · 240 MHz · 1 core.

1.28× faster than INT8 ›

71.28%

CIFAR-10 accuracy · 87.0% INT8.

Accuracy is part of the result ›

The constraint is the opportunity

When space is scarce, every bit matters.

A microcontroller has a small memory and a finite compute budget. What if a network could work with just two values?

RAM

Pack intermediate activations into bits.

FLASH

Store binary weights in compact words.

COMPUTE

Count bits. Compare an integer threshold.

Representation · 8 elementsIllustrative

-2416103-865-721492

INT8 values · 8 bytes

Binary representation · 1 byte

01101011

A conceptual representation change, not a direct model conversion. Training, thresholds, alignment and metadata also matter.

From numbers to bits ↗

Less abstract. More interactive.

Don’t just read it. Flip a bit.

Change an input. Watch XOR find the differences, popcount count the ones, and an integer threshold decide the output.

Enter the digital lab

Real calculations. Toy example.
No hardware measurements in this demo.

XOR / POPCOUNT / THRESHOLD8-BIT EXPERIMENT

INPUT

WEIGHT

XOR01100110

THRESHOLD4

POPCOUNT

4 / 8

4 ≥ 4

OUTPUT
BIT1

Tap any input or weight bit.

Illustrative operation · 0/1 bits · XOR counts differences.
Example rule: output = 1 when count ≥ threshold.

From an idea to an inference

Follow the bit.
All the way to silicon.

A complete path from model and data to generated C, verified execution, and measurements you can inspect.

Input & constraints

CONFIG

Train & prepare

WORKSTATION

Integer thresholds

COUNT ≥ T

Export to C

C HEADER

PC reference check

PC CHECK

Execute on chip

MICROCONTROLLER

Verify & benchmark

REAL RESULTS

The engine under the experience

Small operations.
A complete toolchain.

Inside the architecture ↗

Prepare

The CDT

Train, fold, export and validate. Take a model and its constraints to a generated C header.

tools/cdt.py ↗Execute

lib_symbex

A portable C engine that replaces much of the arithmetic with bit counting and integer thresholds.

C inference + hardware layer ↗Verify

Real chip. Real checks.

Compare with the PC reference. Measure device latency and memory against documented INT8 baselines.

chip_test.py + benchmarks/ ↗

Evidence, with the tradeoffs

Measured on hardware.
Open to inspection.

Explore every comparison ↗

ESP32-S3 · CIFAR-10

Image classification

ResNet-8 · official MLPerf Tiny INT8 baseline

40.7 msper binary inference

SYMBEX · BINARY40.7 ms

TFLITE MICRO · INT852.0 ms

Accuracy: 71.28 ± 0.81% binary · 87.0% INT8.

View result & methodology ↗

ESP32-S3 · Pascal VOC persons

Person-center detection

FOMO · same resized architecture and training recipe

126.3 msper binary inference

SYMBEX · BINARY126.3 ms

TFLITE MICRO · INT8256.9 ms

F1: 31.2 ± 0.7% binary · 30.1 ± 0.2% INT8.

View result & methodology ↗

Repository-reported results · 240 MHz · one core. Validation-split fixes and affected re-measurements are ongoing; FOMO sizing used test-set selection. See all conditions and caveats.

A workbench for your curiosity

Make the invisible
computation visible.

Five experiments. Real deterministic math. Every step is yours to change.

Experiment 01

Your next inference starts in the source.

Build the board example. Run the library tests. Train and export a model, then verify it on your hardware.

Read the documentation Explore the source ↗

START WITH THE LIBRARY · REPOSITORY ROOT

# Bit-exact C library tests (gcc required)
python Symbex1_PIO/lib/lib_symbex/tests/run_tests.py

# Build & upload the configured S3 example
cd Symbex1_PIO
pio run -t upload
pio device monitor -b 115200

Open source. Open questions. Open possibilities.

Built to be inspected.

Explore the engine. Reproduce a benchmark. Help make it better.

Find your way in View on GitHub ↗

Inference. Reduced to one bit.

Small by design.
Measured on silicon.

When space is scarce, every bit matters.

Don’t just read it. Flip a bit.

Follow the bit.
All the way to silicon.

Input & constraints

Train & prepare

Integer thresholds

Export to C

PC reference check

Execute on chip

Verify & benchmark

Small operations.
A complete toolchain.

The CDT

lib_symbex

Real chip. Real checks.

Measured on hardware.
Open to inspection.

Image classification

Person-center detection

Make the invisible
computation visible.

Bitwise inference

Memory packing

Thermometer encoding

A tiny binary network

Hardware tradeoffs

Your next inference starts in the source.

Built to be inspected.

Inference. Reduced to one bit.

Small by design.Measured on silicon.

When space is scarce, every bit matters.

Don’t just read it. Flip a bit.

Follow the bit.All the way to silicon.

Input & constraints

Train & prepare

Integer thresholds

Export to C

PC reference check

Execute on chip

Verify & benchmark

Small operations.A complete toolchain.

The CDT

lib_symbex

Real chip. Real checks.

Measured on hardware.Open to inspection.

Image classification

Person-center detection

Make the invisiblecomputation visible.

Bitwise inference

Memory packing

Thermometer encoding

A tiny binary network

Hardware tradeoffs

Your next inference starts in the source.

Built to be inspected.

Small by design.
Measured on silicon.

Follow the bit.
All the way to silicon.

Small operations.
A complete toolchain.

Measured on hardware.
Open to inspection.

Make the invisible
computation visible.