Dear NXP Team,
We are trying to run an ONNX model on the Neutron NPU of an i.MX95 device.
We found the following NXP documentation:
https://docs.nxp.com/bundle/UG10166/page/topics/onnx_model_test.html
In the section “Running a CNN Model”, the documentation states:
“ONNX Runtime now supports executing CNN models on the Neutron NPU as an experimental feature. To enable this, the model must be quantized and converted with the onnx2neutron tool (see README of the onnx2neutron tool from the eIQ Toolkit).”
However, we are not able to find the onnx2neutron tool in the eIQ Toolkit or in the documentation. Without this conversion step, the ONNX model appears to run only on the CPU instead of the NPU.
Our questions are:
1. Does the onnx2neutron tool currently exist and is it publicly available?
2. If yes, where can we download or find it?
3. Is ONNX Runtime execution on the Neutron NPU currently supported on i.MX95, or is the TFLite path the only supported route at the moment?
We have successfully tested the TFLite example flow. However, for our own custom model, the TFLite route does not work correctly: it appears to run, but silently produces incorrect results. Therefore, we would like to investigate whether the ONNX path is a viable alternative.
Any guidance on the status of onnx2neutron and ONNX support for the Neutron NPU would be appreciated.
Best regards,
Peter
iMX95