Visualize Ethos-U NPU performance with ExecuTorch on Arm FVPs
Introduction
Overview
Understand the ExecuTorch workflow
Set up your ExecuTorch environment
Set up the Corstone-320 Fixed Virtual Platform
Deploy and run Mobilenet V2 on the Corstone-320 FVP
Enable GUI and deploy a model on Corstone-320 FVP
Evaluate Ethos-U Performance
Next Steps

  Log an issue

  Fork and edit

  Discuss on Discord

Deploy Mobilenet V2 with ExecuTorch

With your environment and FVP now set up, you’re ready to deploy and run a real TinyML model using ExecuTorch.

This example deploys the MobileNet V2 computer vision model. The model is a convolutional neural network (CNN) that extracts visual features from an image. It is used for image classification and object detection.

The Python code for the MobileNet V2 model is in your local executorch repo: executorch/examples/models/mobilenet_v2/model.py . You can deploy it using run.sh , just like you did in the previous step, with some extra parameters:

Tip

On macOS, make sure Docker is running. FVPs execute inside a Docker container.

    

        
        
./examples/arm/run.sh \
--aot_arm_compiler_flags="--delegate --quantize --intermediates mv2_u85/ --debug --evaluate" \
--output=mv2_u85 \
--target=ethos-u85-128 \
--model_name=mv2

The --model_name=mv2 flag tells run.sh to use the Mobilenet V2 model defined in examples/models/mobilenet_v2/model.py.

Explanation of run.sh Parameters

run.sh ParameterMeaning / Context
–aot_arm_compiler_flagsPasses a string of compiler options to the ExecuTorch Ahead-of-Time (AOT) compiler
–delegateEnables backend delegation
–quantizeConverts the floating-point model to int8 quantized format using post-training quantization
Essential for running on NPUs
–intermediates mv2_u85/Directory where intermediate files (e.g., TOSA, YAMLs, debug graphs) will be saved
Useful output files for manual debugging
–debugVerbose debugging logging
–evaluateValidates model output, provides timing estimates

What to expect

ExecuTorch will:

  • Compile the PyTorch model to .pte format
  • Generate intermediate files (YAMLs, graphs, etc.)
  • Run the compiled model on the FVP
  • Output execution timing, operator delegation, and performance stats

You should see output like:

    

        
        
Batch Inference time     4.94 ms,  202.34 inferences/s
Total delegated subgraphs: 1
Number of delegated nodes: 419

A high number of delegated nodes means the majority of model execution was successfully offloaded to the Ethos-U NPU for acceleration. This confirms that the model was successfully compiled, deployed, and run with NPU acceleration.

Next steps

If you’d like to visualize instruction counts and performance using the GUI, continue to the next (optional) section.

Back
Next