The EN683 NPU provides 2.7 TOPS of compute. It supports a limited set of operators, and any model built solely from these supported operators can run on the device. The NPU SDK includes a compiler and a simulator, so customers can verify model accuracy and performance before deployment. As long as the customer provides a labeled dataset, a model can be trained, quantized, compiled, and deployed on the EN683.

NPU SDK contents
Environment: Ubuntu 20 / 22 / 24, Python 3.10

NPU Compiler
- Input model format: ONNX (recommend), tflite
- Compute precision: INT8
- Quantization scheme: per-channel symmetric (weights) / per-tensor asymmetric (activations)
- PTQ only (QAT not supported)

NPU Simulator
- Distributed as a Python package; runs on a local PC

EYENIX MODEL ZOO
EYENIX MODEL ZOO URL : https://github.com/EyenixG/eyenix_model_zoo

Supported Model
- mobilenetv2
- resnet
- efficientnet-lite
- osnet
- yolov3 to v9
- yolov8-seg
- and more