STM32N6: why can a fast NPU still produce a slow vision application?

Measure capture, preprocessing, transfers, inference and postprocessing separately. Check the conversion report for unsupported or differently mapped operations. A theoretical compute rate describes neither the whole model nor the complete application, especially when memory traffic or CPU-side work dominates.

Use the same image dimensions, numeric format and processing steps as the deployed application. For example, a 640 × 480 RGB frame contains 921,600 bytes before additional buffers. Copying and converting several frames can become significant even when model execution is short.

Profile the integrated workload with logging, display and communication enabled, and report observation age as well as frame rate. Keep a reproducible input set and tool versions with the results. Optimize the measured bottleneck, then recheck accuracy after model or preprocessing changes.