Skip to main content

Ethos-U Target Profiles

Overview​

This package provides the MLIA target plugin for Arm Ethos-U NPUs. It is the part of the MLIA ecosystem responsible for Ethos-U-specific target profiles, operator analysis, result-owned advice, standardized output, and the backend wiring needed for performance and compatibility flows.

The plugin is intended for LiteRT / TensorFlow Lite .tflite models that will be analysed against Ethos-U55, Ethos-U65, or Ethos-U85 configurations. It also participates in supported Corstone ExecuTorch flows for .pt2 and .pte inputs.

Bundled profiles​

The package ships built-in TOML target profiles under src/mlia/resources/target_profiles/.

Bundled profiles include:

  • ethos-u55-128
  • ethos-u55-256
  • ethos-u65-256
  • ethos-u65-512
  • ethos-u85-128
  • ethos-u85-256
  • ethos-u85-512
  • ethos-u85-1024
  • ethos-u85-2048

These profiles let MLIA select the correct accelerator configuration, memory mode, and backend assumptions without requiring a custom profile for common hardware targets.

Supported input formats​

The most common input format in this package is:

  • LiteRT / TensorFlow Lite (.tflite).

Additional formats can also participate in Ethos-U workflows:

  • TOSA (.tosa, .tosamlir).
  • ExecuTorch exported program (.pte).
  • PyTorch exported program (.pt2) when mlia-converters-pytorch is installed.

Quantized LiteRT / TensorFlow Lite .tflite models are the standard input for Vela and most Ethos-U analysis. PyTorch .pt2 inputs also support Vela performance estimation, which is the only performance route for .pt2 on Ethos-U65. ExecuTorch .pte performance is supported only on Ethos-U55 with corstone-300 and Ethos-U85 with corstone-320; it is unsupported on Ethos-U65. See the support matrix.

Typical usage​

Check compatibility:

mlia check my_model.tflite --target-profile ethos-u65-512 --compatibility

Estimate performance with Vela:

mlia check my_model.tflite --target-profile ethos-u55-256 --performance --backend vela

Run a more detailed performance flow with a Corstone backend:

mlia check my_model.tflite --target-profile ethos-u65-512 --performance --backend corstone-310

Run a supported ExecuTorch AOT flow through Corstone:

mlia check my_model.pt2 --target-profile ethos-u55-256 --performance --backend corstone-300

You can also point the Corstone path at a prepared .pte artifact directly:

mlia check my_model.pte --target-profile ethos-u85-256 --performance --backend corstone-320

Configuration concepts​

Ethos-U target profiles in MLIA typically encode:

  • Accelerator or MAC configuration.
  • Memory mode assumptions.
  • System configuration names used by backend tooling.
  • Paths to bundled or custom backend configuration assets.

When the built-in profiles are not enough, custom profiles can be layered on top of this package's bundled backend assets.

Performance and compatibility outputs​

This package collects target-specific standardized results and attaches advice to the result that produced it. Core MLIA then post-processes and renders that output. The plugin contributes:

  • Operator compatibility information.
  • Model-level performance estimates.
  • Per-operator and per-layer breakdowns linked to standardized entities.
  • Target-aware advice for compatibility and performance findings.

The exact metrics depend on which backend is used. See backends.md for the split between Vela and Corstone responsibilities.