NPU Microcode Version Mismatch on iMX95 (Toradex SMARC V1.2A) - TensorFlow Lite Neutron Delegate

Technical Support
, , , , , , ,
1

Hi everyone,

I am encountering an NPU initialization issue while trying to run a TensorFlow Lite benchmark model using the Neutron delegate on our new Toradex SMARC i.MX95 board (silicon version V1.2A).

When executing the benchmark tool, the Neutron firmware boots successfully, but it fails to apply the delegate due to a microcode version mismatch.

EIQ tool versions used: 3.0.1 and 3.2.2

Reference link used: https://developer.toradex.com/software/ai-computer-vision-machine-learning/ai-and-machine-learning-on-imx95/#configure-and-build-the-image

Error Log:

INFO: External delegate path: [/usr/lib/libneutron_delegate.so]
...
[  680.343224] remoteproc remoteproc0: Booting fw image NeutronFirmware.elf, size 41304
[  680.352344] remoteproc remoteproc0: remote processor neutron-rproc is now up
...
INFO: NeutronDelegate delegate: 1 nodes delegated out of 3 nodes with 1 partitions.
...
Error: microcode version mismatch! 0xf157aeb4 (expected 0x5bfe0804)
ERROR: /usr/src/debug/tensorflow-lite-neutron-delegate/2.16.2/neutron_delegate.cc:233 neutronRC != ENONE (15635 != 0)
ERROR: Node number 3 (NeutronDelegate) failed to prepare.
ERROR: Failed to apply the default TensorFlow Lite delegate indexed at 0.

System Information (tdx-info output):

  • HW model: Toradex SMARC iMX95 on Toradex SMARC Development Board

  • Toradex version: 0096 V1.2A

  • Kernel version: 6.6.147-7.8.0-devel #1 SMP PREEMPT Thu Jul 30 14:24:25 UTC 2026

  • Distro version: VERSION_ID=7.8.0-devel-20260824133657-build.0

Has anyone encountered this specific mismatch 0xf157aeb4 (expected 0x5bfe0804) with the i.MX95? Do I need to update the libneutron_delegate.so, the NPU firmware (NeutronFirmware.elf), or recompile the model with a newer Vela/Neutron compiler version to match this specific silicon?

Thanks in advance for the help!

image (20)
image (20)1920×514 111 KB

Screenshot from 2026-08-27 12-48-56
Screenshot from 2026-08-27 12-48-56916×396 44.8 KB

2

Hi, I’m the Toradex AI Assistant.

I might make mistakes or occasionally get you wrong. Our support team will step in if needed.

The microcode version mismatch you’re encountering indicates an incompatibility between the model’s compiled microcode and the Neutron runtime environment on your system.

Root Cause

The error microcode version mismatch! 0xf157aeb4 (expected 0x5bfe0804) suggests that your TensorFlow Lite model was converted with a different version of the eIQ Neutron converter than what your runtime Neutron delegate and firmware expect(1)(2).

Solution Steps

1. Verify Your Build Environment

Your system is running a development build (VERSION_ID=7.8.0-devel-20260824133657-build.0), which may have a different Neutron software stack version than the eIQ toolkit versions you used (3.0.1 and 3.2.2).

2. Reconvert Your Model

You need to convert your model using the eIQ Neutron converter that matches your runtime environment. According to the documentation, the conversion process should be:

$ ./bin/neutron-converter --input mobilenet_v1_1.0_224_quant.tflite --output mobilenet_v1_1.0_224_quant_converted.tflite --target imx95

(1)

However, ensure you’re using the eIQ toolkit version that corresponds to your BSP release.

3. Check Component Versions

The Neutron software stack consists of multiple components that must be aligned(1)(2):

  • Neutron delegate (libneutron_delegate.so)
  • Neutron firmware (NeutronFirmware.elf)
  • Neutron converter (from eIQ toolkit)

All three must be from compatible versions.

4. Verify Silicon Support

Your board uses silicon version V1.2A. According to the documentation, NPU acceleration support is available for i.MX95 devices based on B0 silicon(1). Confirm that V1.2A corresponds to B0 silicon and is fully supported by your BSP version.

Recommendation

Since you’re using a development build (7.8.0-devel), I recommend checking which eIQ toolkit version is compatible with this specific BSP release. The mismatch strongly suggests you need to either:

  1. Update to a matching eIQ toolkit version and reconvert your model, or
  2. Use a stable Torizon OS release where component versions are guaranteed to be aligned

For further assistance with version compatibility, you may want to visit the Toradex Community or Toradex GitHub.

Best regards,
Toradex AI Assistant.

Did you like my answer? :+1: :-1:

If I answered your question, please tick the “Solution” checkbox.

For follow-up questions to me, please use @ToradexAI in your post.