AMB82-mini edge AI capability

Knowledge AI
1

Hi,

We are currently developing a project that utilizes edge AI to implement a smart camera application. We would like to learn the specific capabilities of the AMB82-mini platform’s edge AI including:

  1. Which AI models does the AMB82-mini’s built-in NPU currently support?

  2. In what specific scenarios can the currently supported AI models be applied? Are model examples provided for our basic evaluation, or even further optimization?

  3. Are there any available architectures for the currently supported AI models that we can leverage our own trained models(e.g. we have our own Yolo8 trained model already)? We understand that the AMB82-mini cannot directly run big models, but believe Realtek has solved the problem of converting big models into tiny models to run on the AMB82-mini while maintaining the model’s efficiency and accuracy. We would like to learn more about the detailed design of the AMB82-mini’s edge AI architecture to better achieve our current project’s goals in true edge AI

your great assitance will be appreciated!

Reagrds,

Allen

2

Hi @allenyang ,

Great to hear about your smart camera project! You may refer to the responses below:

  1. Here’s the list of models supported on AMB82-mini here. Additionally, mobilenetv2 and YOLOv9 is supported on our latest FreeRTOS SDK too.
  2. The SDK provides ready-to-run examples for evaluation, this includes object detection, face recognition, etc.:
    1. If you are using FreeRTOS SDK, you may compile and run the VIPNN examples.
    2. If you are using Arduino SDK, here are the NN examples, POC examples
    3. Pre-built binary model files are available in the SDK, NN Model Prep
  3. Yes, as long as the model is currently supported on our SDK. Please take a look at this NN deployment guide on AmebaPro2.

Feel free to drop in your questions, if there is anything unclear after reviewing the documentation.

Thank you.