Examples

Choose a recipe for the result you need. Each example includes its configuration, commands, expected output and validation status. Hardware recipes require the listed board and tools; saved-result recipes can run without a connected board.

Start here

First results→

Profile a model, every counter per layer, the TFLM baseline

One valid bundle from a board, then the full counter catalogue and the upstream baseline the Ambiq engines are measured against.

hpx compare

Comparisons→

Engines and toolchains on one model

Separate captures of the same model across engines or toolchains, with comparison reports and the rules that withhold a delta.

Joulescope

Power→

Energy per inference beside the profile

The power capture recipe, how to read the power block a capture writes, and the rules that decide whether two captures compare.

Experimental

NPU→

Ethos-U85 on the Atomiq110 FPGA

A Vela-compiled model on the NPU with both counter families, under the experimental label its sources carry.

Python and batch

Automation and analysis→

Sessions, hpx validate, saved bundles, overlays

Drive hpx from Python, run a matrix in one command, re-analyse bundles without a board, and put per-layer cost on the model graph.

Find an example

Search by model, engine, board or measurement, or narrow the list by task and setup.

13 of 13 examples

Search examples

Example	Description
Atomiq110 NPU profilingNPU	Profile Ethos-U85 on the experimental Atomiq110 FPGA with heliaRT or heliaAOT. Recipe not yet validated.
Batch runs from YAMLAutomate	Run a model matrix with hpx validate. Includes a historical nightly run across four boards.
Capture per-layer countersProfile	Capture CPU, memory and MVE counters with heliaAOT on Apollo510 EVB. Recipe not yet hardware-validated.
Compare enginesCompare	Compare heliaRT, heliaAOT and LiteRT on one model and board. Includes captured bundles and a host comparison.
Compare toolchainsCompare	Compare GNU Arm and ATfE builds using recorded runs. Arm Compiler 6 remains unvalidated in this recipe.
Compare two power runsPower	Check power comparability before interpreting energy deltas. No captured power bundle.
Interactive Python sessionAutomate	Explore Session configuration, profiling and comparison from Python. Host steps checked; hardware steps unvalidated.
Per-layer cost on the model graphAnalyze	View per-layer counters on a model graph. Output format documented; no captured overlay bundle.
Power capture with a JoulescopePower	Add Joulescope power capture to a heliaAOT run. Configuration provided; no captured power bundle.
Profile a modelProfile	Capture timing and layer counters with heliaRT on Apollo510 EVB. Includes a historical hardware bundle.
Re-analyse saved resultsAnalyze	Verify, load and compare the supplied result bundles without a board. Host steps checked on those bundles.
Read a power summaryPower	Read energy, measurement scope and capture health from a saved power summary. No captured power bundle.
Vanilla TFLM baselineProfile	Profile the keyword-spotting model with LiteRT for Microcontrollers and CMSIS-NN. Includes a captured baseline.

What the validation status means

“Validated on” names a hardware run record: board, nsx module versions (from the bundle’s nsx.lock), date and source commit. Nine bundles from the nightly hardware validation run of 2026-09-16 ship under examples/results/hardware-validation-2026-09-16/, published with probe serials redacted and manifest digests refreshed, so load_result_manifest(..., verify=True) passes on each. A page whose recipe has no such record says “not yet validated”, and a compile-only run is never called profiled. Host-only steps (loaders, hpx compare, Session.resolve()) record the date and hpx version they were executed with.

What next

Getting started if hpx is not installed yet; the User guide for the mechanics behind every recipe; the Reference for every command, key and field.