Profiling on i.MX95

[EOF]

Hi all,

Is there a tool available to profile the runtime performance of the NPU and CPU (separately) on the i.MX95?

So far, I found a related discussion pointing to Section 7.1.2 “Profiling on hardware accelerators” in the ML User Guide, but this appears to apply only to the i.MX8 platform.

I also found that in Section 7.2.10 of the ML User Guide there is a functionality described specifically for the i.MX93.

1) Is there an equivalent vela-prof profiling tool or workflow available for the i.MX95?

2) I also want to please know which command did you use to get the benchmark numbers (see photo) gathered from nxp-demo 

e. g. my guess is something like: ./benchmark_model --graph=label.tflite --enable_op_profiling=true --external_delegate_path=/usr/lib/libneutron_delegate.so

3) Which Yocto BSD version was used for the benchmark shown in the photo?

I would appreciate if you can please reply my three questions, thank you in advance!

Best regards,
EOF

[yipingwang · Accepted solution]

The Neutron Profiling feature will be supported in the new release of eIQ AI toolkit (new version of eIQ toolkit). 

The work flow will be:

1. Run the model on the i.MX95 target, and the profiling data will be collected.

2. Export the profiling data and copy it to the PC where eIQ AI toolkit is installed, and use eIQ AI toolkit to visualize the profiling data.

The eIQ AI toolkit will be released soon (should be ready by the end of 2026 March), you can try the feature after the release.