2298503_en-US

Yes

iMX95 neutron eIQ converter for 6.12.3_1.0.0

Hello,

I am using an iMX95 with the BSP: 6.12.3_1.0.0.

I can't move from that BSP, and I am trying to run tensorflow-lite examples, CPU only works, but when using NPU I get the following:

# python3 label_image.py -e /usr/lib64/libneutron_delegate.so -m /tmp/converted.tflite
Loading external delegate from /usr/lib64/libneutron_delegate.so with args: {}
INFO: NeutronDelegate delegate: 1 nodes delegated out of 4 nodes with 1 partitions.

Error in cpuinfo: prctl(PR_SVE_GET_VL) failed
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
Warning: microcode version mismatch! 0xcebb80a (expected 0x24d93dcf)
Warm-up time: 0.8 ms

Inference time: 0.4 ms

0.000000: toilet tissue
0.000000: starfish
0.000000: zebra
0.000000: sorrel
0.000000: guinea pig

I have been trying multiple neutron converters, but I always get the firmware mismatch. I can't really find the one for 6.12.3, can you help me by pointing me which eIQ version has it?

Thanks,

Roberto

Re: iMX95 neutron eIQ converter for 6.12.3_1.0.0

Hi @Chavira ,

Ok I found the real issue, without using any custom model following your guidance, I get everything on 0 but no firmware mismatch:

# python3 label_image.py -e /usr/lib64/libneutron_delegate.so
Loading external delegate from /usr/lib64/libneutron_delegate.so with args: {}
INFO: NeutronDelegate delegate: 29 nodes delegated out of 31 nodes with 1 partitions.

Error in cpuinfo: prctl(PR_SVE_GET_VL) failed
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
Warm-up time: 56.6 ms

Inference time: 0.3 ms

0.000000: toilet tissue
0.000000: starfish
0.000000: zebra
0.000000: sorrel
0.000000: guinea pig

After upgrading only the neutron firmware (and tf neutron delegate) to 6.12.20, the error is gone, so probably there was something wrong with the older firmware. Also it took a lot of time loading everything before inference. Anyway, I think I am just able to upgrade this section of the BSP and move on. Thank you.

Re: iMX95 neutron eIQ converter for 6.12.3_1.0.0

Hi @robertogs21,

Thank you for contacting NXP Support!

Our reference examples do not require model conversion specifically for the NPU.

If you are using a different model, you must first quantize the model to INT8 before performing the conversion using eIQ. This step is required for compatibility with the NPU delegate.

If you are only following our provided demos, no additional conversion step is needed. You can run the demo directly using the following command:

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

This command enables execution using the Neutron delegate.

For more details about this demo and model requirements, please refer to the Machine Learning User Guide, which provides a complete explanation of the workflow and supported configurations.

Best regards,
Chavira