About this module
In the first three modules, you deployed models through Arm CPU backends. In this module, you move the main neural-network compute to an Arm Ethos-U Neural Processing Unit (NPU). You will prepare and quantize MobileNetV2, compile it for an Ethos-U85, and run the resulting ExecuTorch program on a Corstone-320 Fixed Virtual Platform (FVP).
The module compares three model variants and two execution stages:
| Model/runtime path | What it shows |
|---|---|
| FP32 MobileNetV2 with Ethos-U partitioning | Why quantization is necessary for Ethos-U deployment |
| INT8 MobileNetV2 with Ethos-U | The main quantized deployment path for Ethos-U. |
| INT8 MobileNetV2 with inserted LRN | How a model change can split delegation and introduce an unsupported operator. |
| INT8 MobileNetV2 on Corstone-320 FVP | Whether the compiled .pte artifact executes in a simulation of the target. |
| INT8 MobileNetV2 classification app | How to add model-specific output handling to a generic runner. |
Optional sections cover Model Explorer and the Arm ML Inference Advisor (MLIA).
Course complete
Nice work
You completed Optimize and deploy AI inference with ExecuTorch on Arm.