Optimize ExecuTorch fallback with CMSIS-NN on Cortex-M
Introduction
Understand and prepare the environment
Lower a model to the Cortex-M backend
Add Ethos-U delegation and CPU fallback
Inspect the generated CMSIS Pack selection
Understand the reusable CMSIS project
Build and run the mixed application
Next Steps
Optimize ExecuTorch fallback with CMSIS-NN on Cortex-M
Review frequently asked questions
These FAQs were drafted with an approved AI-assisted workflow and reviewed by Arm contributors before publication. Human technical review remains part of the process so the final page reflects engineering rigor, accuracy, and Arm editorial standards.
Check that the exported graph has no
executorch_call_delegate node. You’ll see supported patterns lowered to cortex_m.* operators, while aten.topk remains portable. In this model, the quantized convolutions use CMSIS-NN, but other Cortex-M operators can use different implementations.Inspect the exported operator list. You’ll find the stride-1 convolution in an Ethos-U55 delegate and the stride-4 convolution on Cortex-M55 through CMSIS-NN. The remaining
topk operation uses a portable ExecuTorch CPU kernel.Inspect
cmsis-executorch-u55.cbuild-pack.yml for the resolved ExecuTorch and CMSIS-NN Pack versions, then inspect ai_layer/ai_layer.clayer.yml for the components selected from those Packs. You’ll find the Ethos-U backend, Cortex-M operator wrappers, CMSIS-NN library, and portable topk component.Inspect
cmsis-executorch-u55.csolution.yml to see the SSE-300-U55 target and its Board and AI layer references. The generated AI layer selects the runtime components of the model. The project generates model_pte.c from model.pte, so the bare-metal application can load the program without a filesystem.compile_commands.json confirms that the build compiled the expected ExecuTorch, Ethos-U, CMSIS-NN, and portable topk sources. On the FVP, Test_result: PASS tells you that inference returned a nonempty output list. Check that the output also reports ten values. You need a reference comparison to assess numerical accuracy.