How to use NPU on G720 without the NDA access

NeuroPilot NeuroPilot - Analytical AI
,
1

Just like the title, I am tring to use the NPU on G720 EVK.

But I can’t apply the NDA account for some reason.

For I know, if I want to use NPU I must use the NeuroPilot 8 Public to converted the model first.

When I finish the model converted , I want to publish it on the board and use NPU to inference.

For the Public I can’t use Neuro SDK , I only can use Embedded TFLite.

But Embedded TFLite is not 100% sure the model run on the NPU.

Is there anyway I can use NPU without NDA?

2

Thank you for this inquiry.

MediaTek currently does not have plans to release a standalone Public version of the NeuroPilot SDK (which remains restricted to NDA customers for the full bundle). However, developers using public resources can still achieve NPU acceleration on Genio platforms—including Genio-720—by using the LiteRT (TensorFlow Lite) Stable Delegate.

NPU Acceleration via Stable Delegate

The Stable Delegate acts as the bridge between the LiteRT runtime and the MediaTek NPU (MDLA). It allows the system to partition the model graph and offload supported operations to the NPU without requiring the full proprietary Neuron SDK suite.

To utilize the NPU on Genio-720 Yocto (NP8 generation):

  1. Model Conversion: Use the NP8 Converter (available to public developers) to convert your pretrained models to the TFLite format.
  2. Inference: Execute the model using the LiteRT runtime with the stable_delegate enabled. This ensures that compatible operators are scheduled on the NPU.

Benchmark and Implementation Example

For a practical demonstration of how to execute models on the NPU using the Stable Delegate, refer to the IoT AI Hub documentation. It provides detailed steps for using the benchmark tool to verify NPU performance: Execute on NPU via Neuron Stable Delegate (IoT AI Hub)

3

Thanks for the reply it really helpful !