Understand MLIA and the NX Performance Estimator

Performance estimation with MLIA

Arm ML Inference Advisor (MLIA) helps you evaluate a model for an Arm target before you build a deployment pipeline or tune an application on a device. You provide a model and a target profile, then inspect compatibility checks, performance metrics, and advice.

You’ll use the mlia-neural-technology package to evaluate two pretrained, quantized MobileNet V1 image classifiers from the TensorFlow public model catalog. You’ll compare a baseline with a smaller-input variant under the same Neural Technology configuration. You don’t need a model stored in this repository or a device to run the estimator.

The tested example reduces estimated total cycles from 136,962 to 79,824. These are compiler estimates, not device measurements or an accuracy result. You’ll learn about the model and configuration assumptions behind this comparison in Compare a smaller model under the same target profile .

If you have a neural accelerator (NX)-capable Android device and a compatible native Vulkan driver, you can execute the model on the device. You can then bring a measured capture back to MLIA. For more information, see (Optional) Run and profile the model on an NX device .

How the NX and Ethos-U workflows are connected

The Analyze Ethos-U models with MLIA Learning Path applies the same workflow to Arm Ethos-U: select a target profile, check compatibility, inspect performance, and use the advice to guide your next change. It uses Vela for compiler estimates and Corstone for whole-model NPU performance counters. In this Learning Path, you’ll use the Neural Technology plugin and NX Performance Estimator, with optional measurements from an Android device.

The report-reading approach carries across targets, but their backends, supported operators, and performance metrics differ. Don’t compare cycle counts across targets assuming that they describe the same hardware configuration.

Understand the conversion and analysis flow

The NX Performance Estimator backend evaluates Arm Neural Technology targets. You’ll download the models and pass their .tflite files directly to MLIA. With the LiteRT converter plugin installed, MLIA handles the conversion and analysis in the following order:

1. Converts the LiteRT model to Tensor Operator Set Architecture (TOSA) multi-level intermediate representation (MLIR) and checks operator compatibility.
2. Compiles the graph to a Vulkan Graph Format (VGF) artifact and runs the NX Performance Estimator for a performance check.
3. Writes standardized JSON results and the intermediate model artifacts to your output directory.

You don’t need to run a separate converter or manage intermediate model files.

Select a target profile

The Neural Technology plugin supplies neural-technology, NX-peak-12SC-8NX-600MHz, and NX-sustained-12SC-8NX-350MHz profiles. Each profile selects system and compiler configuration files. Select one profile and use the same profile and configuration when comparing model variants. Changing both the model and the target makes the result harder to interpret.

What you’ve learned and what’s next

You’ve learned how model conversion, compatibility analysis, and performance estimation fit together.

Next, you’ll install the public packages and verify that the Neural Technology profiles are available.