C/C++, with support for integrating C++ libraries.

Currently supports TensorFlow framework. Models under PyTorch need to be converted to TensorFlow first.

All software libraries used are open-source: TensorFlow Lite Micro, CMSIS-NN, Ethos-U65 Driver, and CMSIS-DSP Math

Vela Compiler: Used to convert models into int8 data types recognizable by the NPU.
Development Toolchain: Such as processor Compiler for code compilation.
Neural Network Development Tools: Such as processor NN inference engines to accelerate ML inference tasks.
Model Conversion Tools: Convert trained models into formats suitable for NPU execution.

Model Quantization: Convert floating-point numbers to fixed-point numbers (e.g., 8-bit or 16-bit integers) to reduce memory and computational requirements.
Model Pruning: Remove less important parameters from the model to decrease computation and storage demands.
Weight Sharing: Reduce the number of model parameters through weight sharing.
Using Official Toolchains: Utilize relevant tools provided by the processor for model conversion and optimization to ensure model compatibility.

The NPU in ZPC1Pro operates at 800Hz and can utilize 1MB SRAM + 512KB ITCM + 128KB DTCM of RAM space, along with external PSRAM space (typically 8MB).

The NPU in ZPC1Pro supports Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), Recurrent Neural Networks (RNN), and more, capable of handling complex workloads. It has been validated on the chip for common AI models such as keyword recognition, image classification, audio classification, voice wake-up, visual wake-up, and audio noise reduction.

The NPU delivers up to 1 TOPS computing power at 1GHz.

It is suitable for emerging edge AI fields including smart cameras, environmental sensors, industrial automation, mobile devices, smart homes, retail, and industry. It can be used for real-time image classification, object recognition, voice assistants, and other applications.