12. Performance Benchmark

This section reports inference performance - latency and memory - for the reference AM62Dx EdgeAI models. The same methodology (perfsim estimate + on-target measured) applies to any model you compile through the workflow. It is not an accuracy or audio-quality benchmark: model quality (classification accuracy, PESQ / STOI / SI-SDR) is measured by the evaluate step in Model Compile and Evaluate.

12.1. Methodology

Two independent figures are reported for each model, and they must not be confused:

Figure	How it is obtained
perfsim estimate	Produced by the TIDL performance simulator during compile. The compiler writes a result.yaml per model under its work directory; the perfsim latency, GMAC count, and estimated DDR transfer are read directly from it. This is a static estimate for the C7™ NPU subgraph - it does not include Arm-side pre/post-processing or runtime overhead.
on-target measured	Wall-clock latency timed on the AM62D target while running the Model Zoo inference scripts. For the streaming speech-enhancement models it is also expressed as a real-time factor (RTF) = processing time / audio duration; RTF < 1 means the model keeps up with real-time audio.

Only models compiled for the C7™ NPU (TVM-RT + TIDL) have a perfsim estimate.

Note

The perfsim values below are read from result.yaml and are labeled estimated (perfsim). On-target measured latency and RTF are marked TBD until measured on the AM62D EVM.

12.2. Per-model results

Model	Model ID	Precision	perfsim latency (estimated)	GMACs	DDR transfer (estimated)
VGGish11	ac-0101	8-bit	5.235 ms	1.1515	10.73 MB
YAMNet	ac-0201	8-bit	1.990 ms	0.0692	3.55 MB
GCRN	ase-0201	16-bit	TBD	TBD	TBD

VGGish11 and YAMNet are compiled and have result.yaml perfsim values (shown above). GCRN is compiled at 16-bit but its result.yaml is not yet captured, so its perfsim figures are TBD.

On-target measured latency and RTF for all four models are TBD - to be filled in from the Model Zoo inference scripts on the AM62D EVM:

Model	On target	Measured latency	RTF
VGGish11	C7™ NPU	TBD	(clip classifier)
YAMNet	C7™ NPU	TBD	(clip classifier)
GCRN	C7™ NPU	TBD	TBD

RTF is reported only for the streaming speech-enhancement models (GCRN); VGGish11 and YAMNet classify a whole clip, so a real-time factor is not meaningful for them.

12.3. Scope & Limitations

This section does not assess model quality. Classification accuracy and speech-enhancement quality metrics are produced by tidlrunner-cli evaluate and documented in Model Compile and Evaluate.