Issue with npu on imx95 Linux 6.12.20_2.0.0

[eirgut]

Hi

i am trying to run the label_image.py example on the imx95-a1-19x19-lpddr5-evk npu. But none of the nodes are delegated to the npu:

I have tried to convert it using the eIQ neutron-converter:

\nxp\eIQ_Toolkit_v1.16.0\bin\neutron-converter\MCU_SDK_25.06.00+Linux_6.12.20_2.0.0>neutron-converter

When running the converted model i only get random and incorrect predictions:

Am i doing something wrong or is the 6.12 neutron-converter corrupting the model?

I tried the same apporach on Linux 6.6.36 and it worked perfect

[Chavira]

Hi @eirgut!
Thank you for contacting NXP Support!

According to our Machine Learning User Guide, it looks like the command you're using is incorrect.

To run the model accelerated on the NPU hardware using the Neutron Delegate on the i.MX 95, please use the following command:

$ python3 label_image.py --ext_delegate /usr/lib/libneutron_delegate.so

The output should be as follows:

Loading external delegate from /usr/lib/libneutron_delegate.so with args: {}
INFO: NeutronDelegate delegate: 29 nodes delegated out of 31 nodes with 1
partitions.
Warm-up time: 1.9 ms
Inference time: 1.7 ms
0.850980: military uniform
0.058824: Windsor tie
0.011765: bulletproof vest
0.007843: bow tie
0.007843: mortarboard

Best Regards!

Chavira

[eirgut]

Hi @Chavira 

Thank you for your reply!

According to the python script, i should be able to use this command:

$ python3 label_image.py --ext_delegate /usr/lib/libneutron_delegate.so

I have tried with both scripts, but none of them are able to delegate nodes to the NPU

Also the one i converted my self using the neutron-converter (\nxp\eIQ_Toolkit_v1.16.0\bin\neutron-converter\MCU_SDK_25.06.00+Linux_6.12.20_2.0.0>neutron-converter)

runs on the npu, but seems like it just gives random variables:

(When running the preconverted model on the cpu it predicts correctly)

It seems that the MCU_SDK_25.06.00+Linux_6.12.20_2.0.0 converter corrupts the model.

[eirgut]

I mean this command:
"python3 label_image.py -e /usr/lib/libneutron_delegate.so"

[Chavira]

HI @eirgut!

After discussing with the internal team, we’ve identified some issues in the latest BSP release.

Due to ongoing silicon changes in the i.MX95, certain features may not be compatible with earlier silicon revisions.

We appreciate your patience and recommend keeping an eye on upcoming BSP updates and silicon version announcements.

Best regards,
Chavira