Skip to main content

Neural Technology Target Profiles

Overview​

This package provides the MLIA target plugin for Arm Neural Technology analysis. It adds target-specific configuration, standardized-result collection, and backend wiring for estimator and measured-profiling flows. Core MLIA owns output post-processing and rendering.

Bundled profiles​

The package ships built-in profiles under src/mlia/resources/target_profiles/:

  • neural-technology
  • NX-peak-12SC-8NX-600MHz
  • NX-sustained-12SC-8NX-350MHz

These profiles are the starting point for MLIA analysis and can be referenced by name from CLI commands.

Supported input formats​

This package participates in workflows for several model formats:

  • LiteRT / TensorFlow Lite (.tflite).
  • TOSA (.tosa, .tosamlir).
  • VGF (.vgf).
  • PyTorch export (.pt2).
  • ExecuTorch PTE (.pte).
  • ExecuTorch ETRecord (.etrecord, .bin), when paired with ETDump for measured performance analysis.

Not every format is consumed directly by the NX Performance Estimator. Some are handled through conversion integrations in this package or through installed converter plugins.

Compatible converter plugins​

Use these converter plugins when the input is not already in a directly usable format for the Neural Technology flow:

  • mlia-converters-litert for LiteRT / TensorFlow Lite (.tflite) inputs.
  • mlia-converters-pytorch for PyTorch export (.pt2) and ExecuTorch PTE (.pte) inputs.

If you are documenting or demonstrating the Neural Technology plugin on its own, a .tosa model is usually the clearest default example because it avoids assuming those extra converter packages are installed.

Typical usage​

Compatibility check:

mlia check my_model.tosa --target-profile neural-technology \
  --compatibility --backend nx-performance-estimator

Performance analysis:

mlia check my_model.tosa --target-profile neural-technology \
  --performance --backend nx-performance-estimator

PyTorch export flow:

mlia check my_model.pt2 --target-profile neural-technology --performance

In the PyTorch case, MLIA can automatically invoke the PyTorch-to-TOSA conversion path before the performance estimator runs, provided mlia-converters-pytorch is installed.

LiteRT / TensorFlow Lite .tflite inputs can also flow through the Neural Technology path, but they depend on mlia-converters-litert to prepare the estimator input.

Measured profiling data​

Structured captures from VK_LAYER_LGL_neural_statistics can be analyzed without a model when one dispatch is selected. A VGF model can instead correlate one or more graph segments with captured pipelines. Profiling-data runs support performance analysis only and use the built-in neural-technology-profiling-data backend automatically.

See CLI for the accepted capture layout and dispatch-selection rules.

An ETRecord and matching ETDump use this same measured backend, with the optional ExecuTorch dependency. Target-profile selection does not choose the ETDump decoder's statistics mode or device. See ETRecord and ETDump for supported inputs and the current CLI decoder defaults.

Configuration concepts​

Neural Technology flows commonly depend on:

  • Target profile selection.
  • System configuration files.
  • Compiler configuration files.
  • Conversion steps between model formats and estimator inputs.

This package provides both the target-side logic and much of the backend plumbing needed to keep that workflow together.

Outputs​

The target and selected analysis mode contribute standardized results containing:

  • Model-level cycles, throughput, utilization, memory, and availability metrics.
  • Per-chain and per-cascade breakdowns with explicit aggregation policies.
  • Segment, source-operator, module, and source-code provenance entities.
  • Measured-mode metadata when structured profiling captures are analyzed.
  • Intermediate conversion or estimator artifacts used for diagnostics.

Core MLIA validates, post-processes, and renders these results. See Outputs and metrics for the detailed structure.