Hello NXP Team,
I am currently using LF_v6.6.52-2.2.2_images_IMX95 and I'm trying to identify the correct version of the Neutron Converter SDK to convert a TensorFlow Lite INT8 quantized model for the NPU.
I have tried multiple versions of the Neutron Converter SDK, but every attempt results in the following error:
INFO: NeutronDelegate delegate: 1 nodes delegated out of 27 nodes with 1 partitions.
INFO: Created TensorFlow Lite XNNPACK delegate for CPU.
Warning: microcode version mismatch! 0x359f358d (expected 0xa186aaf2)
Warning: microcode version mismatch! 0x359f358d (expected 0xa186aaf2)
Inference poll timeout
Error: component='Neutron Driver', category='timeout', code=754
Traceback (most recent call last):
File "/home/object_detc/main.py", line 80, in
interpreter.invoke()
File "/usr/lib/python3.12/site-packages/tflite_runtime/interpreter.py", line 941, in invoke
self._interpreter.Invoke()
RuntimeError: /usr/src/debug/tensorflow-lite-neutron-delegate/2.16.2/neutron_delegate.cc:355 neutronRC != ENONE (193099 != 0)Node number 27 (NeutronD.
Could you clarify why support for LF_v6.6.52-2.2.2_images_IMX95 was removed? Is this because the BSP is still considered an alpha release?
Could you also please advise on the following:
Thank you for your help. I look forward to your guidance.
Hi @boopathi123,
Thank you for contacting NXP Support!
It appears that you are using a very early silicon revision. In that case, I recommend using the Neutron Converter included with the eIQ Toolkit. However, please note that this BSP version is not officially supported for the i.MX95.
The i.MX95 was officially released with the B0 silicon revision and BSP 6.12.34. Earlier silicon revisions were intended for evaluation and pre production purposes, and therefore may not provide the same level of functionality, stability, compatibility, or performance as the currently supported devices.
For this reason, I strongly recommend migrating to a supported combination of:
Using a supported software and hardware configuration will ensure that you benefit from the latest fixes, optimizations, and NPU software support available for the i.MX95 platform.
The behavior you are observing may be related to limitations or known issues present in the early silicon revisions rather than the model itself.
Best regards,
Chavira