MCX N94x NPU benchmarks: what must match before comparing inference speed?

Record the exact model, input size, numeric format, compiler and device. Check which operations are mapped to the accelerator and which remain elsewhere. A speedup measured for one supported graph cannot be applied to an unrelated network by multiplying its CPU execution time.

Separate model execution from input acquisition, preprocessing and output handling. For a streaming application, measure sustained throughput and the age of each result. Include memory traffic and simultaneous peripheral activity in the test rather than benchmarking an isolated call only.

Keep an accuracy evaluation alongside the timing results, especially after quantization or graph changes. Compare identical inputs against a trusted reference and inspect difficult cases. Faster inference is useful only if the deployed pipeline still meets the application's error and response requirements.