Estimate neural accelerator (NX) performance with ML Inference Advisor
Introduction
Understand MLIA and the NX Performance Estimator
Set up your MLIA environment
Download a public baseline model
Check baseline operator compatibility
Estimate performance and inspect configurations
Compare a smaller model under the same target profile
(Optional) Run and profile the model on an NX device
Next Steps
Estimate neural accelerator (NX) performance with ML Inference Advisor
Introduction
Understand MLIA and the NX Performance Estimator
Set up your MLIA environment
Download a public baseline model
Check baseline operator compatibility
Estimate performance and inspect configurations
Compare a smaller model under the same target profile
(Optional) Run and profile the model on an NX device
Next Steps
Check the baseline operators
Pass models/baseline.tflite to MLIA. It converts the model automatically and selects the backend for the neural-technology target profile:
On your first check, MLIA offers to install the bundled NX Performance Estimator. Confirm installation when prompted, then read its Arm license and continue only if you agree. If the backend is already installed, MLIA skips this setup.
mlia check models/baseline.tflite --target-profile neural-technology --compatibility --output-dir runs/baseline-compatibility
Each check writes to its own directory under runs, created during the model download.
The standardized report is written to runs/baseline-compatibility/mlia-output/mlia-output.json. Read the compatibility result and validate its status:
python - <<'PY'
import json
from pathlib import Path
report = json.loads(Path("runs/baseline-compatibility/mlia-output/mlia-output.json").read_text())
result = next(result for result in report["results"] if result["kind"] == "compatibility")
failed_checks = [check for check in result.get("checks", []) if check["status"] != "pass"]
print(f"Compatibility status: {result['status']}")
print(f"Non-passing checks: {len(failed_checks)}")
assert result["status"] == "ok", "Review the compatibility report before estimating performance."
assert not failed_checks, "Review the non-passing operator checks."
PY
The expected output for the pinned packages and baseline is:
Compatibility status: ok
Non-passing checks: 0
Interpret compatibility correctly
The compatibility result includes checks, source-operator entities, and metrics. For this model, accelerator_operator_percentage is 100.0. MLIA reports ml-sdk-model-converter as the compatibility producer because operator support is checked during conversion.
A passing result means that this graph can proceed through the supported conversion flow. It doesn’t establish classification accuracy, guarantee a deployable application, or measure device performance.
If a check fails, inspect the report’s non-passing checks and linked entities. Verify that you’re analyzing models/baseline.tflite and that both plugins from the environment setup are installed.
What you’ve accomplished and what’s next
You’ve checked the baseline’s operator compatibility without a manual conversion step. Next, you’ll collect performance estimates and inspect their breakdowns.