Skip to main content

Validate accuracy and performance

After you compile a model, validate accuracy and measure performance before you deploy it.

Validate accuracy​

Run the compiled model through the execute path. Compare its outputs against the original float model on representative inputs.

For a worked execute accuracy check, see the validation step in Compile Your First Model.

Use representative inputs from your target workload. Compare the compiled model against a trusted floating-point reference. Acceptable tolerances depend on the model and quantization scheme.

Start with model-level metrics such as classification accuracy or detection mAP. If a metric regresses, inspect per-output differences.

Measure performance​

To measure on-device latency, throughput, power, and energy, use the benchmark walkthrough in the beginner tutorials:

  • Benchmark Your Model — run a compiled model with deterministic synthetic tensors and print the headline latency, throughput, power, and energy.

Use benchmark results to confirm that the selected batch size, quantization scheme, and tessellation settings meet your application requirements. If latency or throughput misses the target, revisit compilation options before you change the runtime application.