SensorForge Optimizer · we shrink models to fit edge devices

Measured on real models, across sensors

Your int8 file as given against the optimized int8 file, within a 1-point accuracy limit. SRAM is measured under TensorFlow Lite Micro built for a 32-bit RISC-V core.

Motion

20×

less memory, more accurate than plain int8

Gesture recognition, GRUPyTorch model as ONNX

SRAM	103.8 → 5.9 KB
Flash	264.4 → 12.1 KB
Accuracy	92.0% → 93.4%

368.2 KB → 18.0 KB of memory

Audio

4.4×

less memory on a public Arm model

Keyword spotting, DS-CNN-MArm ML-Zoo, int8 .tflite

SRAM	67.4 → 19.4 KB
Flash	181.9 → 36.7 KB
Accuracy	85.6% → 89.5%

249.4 KB → 56.1 KB of memory

Images

Fits

a camera model that did not fit now runs on the chip

Person detection, MobileNetV2MCU board, 188 KB of app RAM

Needed	207.5 KB of RAM
On the chip	85.8 KB used
Outputs	same as on the PC, 496/496 photos

Lossless mode: no data, no retraining

Model	SRAM	Flash	Accuracy	Less memory
AudioKeyword spotting, CNN-LArm ML-Zoo, int8 .tflite	23.6 → 12.8 KB	473.2 → 106.7 KB	84.4% → 87.8% (+3.4 pts)	4.2×
AudioWake word, GRUKeras, int8 .tflite	48.2 → 10.2 KB	157.7 → 63.5 KB	89.7% → 89.8% (unchanged)	2.8×
AudioWake word, DS-CNN-SArm ML-Zoo, int8 .tflite	22.6 → 13.0 KB	46.5 → 12.6 KB	84.8% → 84.7% (within 0.1 pts)	2.7×
AudioKeyword spotting, DS-CNN-LArm ML-Zoo, int8 .tflite	111.4 → 105.3 KB	492.0 → 127.9 KB	86.0% → 92.7% (+6.7 pts)	2.6×
ImagesPerson detection, MobileNetV1public model zoo, int8 .tflite	98.7 → 69.9 KB	325.5 → 145.8 KB	80.9% → 80.9% (unchanged)	2.0×

Not just the best cases. On 14 public models, SRAM came down on every one (median 44%, from 3% to 89%) and flash by a median of 66%. A quantization-only tool run on the same 14 models changed neither.

How it works

1Upload

Your trained model and a sample of its training data. No code changes, no new runtime.

2Pick the chip

RISC-V, Arm Cortex-M or an NPU. Minimize SRAM, flash or both, within your accuracy limit.

3It searches

Removes channels and picks bit widths layer by layer, re-plans memory, and retrains against your model.

4Download

A .tflite for stock TensorFlow Lite Micro with your file's inputs and outputs, in the same order and under the same names (the report lists any change of scale), its C sources, and a report of every measurement. About 15 minutes.

Formats

.tflite (int8, int16, float) and ONNX (float, QDQ, QOperator); Keras and PyTorch models export to either

Models

CNNs and depthwise CNNs, GRU and LSTM (streaming too), small transformer classifiers, multi-output models and detectors

Checks

Your file and the optimized file both run under TensorFlow Lite Micro on every test sample, and the accuracy limit is judged there, so a file that runs but computes differently on the chip cannot pass

Try it on your own model