Hi, I’m the Toradex AI Assistant.
I might make mistakes or occasionally get you wrong. Our support team will step in if needed.
I’d be glad to help troubleshoot this Neutron delegate issue on your Verdin i.MX95! To provide you with the most accurate guidance, could you let me know which OS and version you are running on your device? You can use the tdx-info command to get detailed information about your setup - more details at https://developer.toradex.com/software/development-resources/tdx-info/.
Initial Analysis
Based on your detailed testing, you’ve correctly identified that the issue is specific to the Neutron delegate on i.MX95, while the VX delegate works properly on i.MX8M Plus (1).
Known Issues with i.MX95 Neutron Conversion
The community has reported similar issues with Neutron delegate functionality on i.MX95 (2). Specifically, there have been reports of internal errors during model conversion for fully connected layers in classifiers like InceptionV3, ResNet50, and MobileNetV3Small (2).
Neutron Delegate Architecture
The Neutron delegate is designed for i.MX95 platforms with the Neutron-S NPU (1). According to the machine learning documentation, the Neutron delegate should capture operators and aggregate them as neutron graph nodes for acceleration by the Neutron-S NPU.
Troubleshooting Steps
Once you provide your OS version information, I can give you more specific guidance. In the meantime, here are some general approaches:
- Verify Model Compatibility: Ensure your model meets the Neutron delegate’s operator constraints
- Check Library Versions: Confirm you have the correct TensorFlow Lite and Neutron delegate versions for your BSP
- Test with Simpler Models: Try with minimal models to isolate the issue
Alternative Approaches
For i.MX95, you might also consider using the GPU delegate as an alternative acceleration method, which leverages the Arm Mali G310 GPU (1).
Please share your OS version details so I can provide more targeted assistance for your specific configuration.
Best regards,
Toradex AI Assistant.
If I answered your question, please tick the “Solution” checkbox.
For follow-up questions to me, please use @ToradexAI in your post.