Profile with hpx

Measure the module on the board rather than estimating it from the conversion log. The two configurations select different engines and retain separate results. Record engine, runtime and build differences when comparing them.

Field	Value
Model	kws_ref, MLPerf Tiny keyword spotting, int8
Target	apollo510_evb
Shows	The two-config comparison recipe and how to read it
CI	Manual. hpx builds firmware, flashes a board over a debug probe and runs the model on it; no runner here has one.

The model is audio/mlperf-tiny/kws_ref/model.tflite from helia-model-zoo: int8, 53,936 bytes, SHA-256 aeea4368…bd0ae, fetched by run.sh and checked against the fixed SHA-256 recorded in the checked-in run script. Model licensing is not established by these hashes; see Model provenance.

Prerequisites

Use a repository checkout with Bash and the environment from Running examples: uv sync --frozen --group ci, then activate .venv. Fetched models require network access, curl and sha256sum or shasum. Run the commands below from this example directory. Unlike the conversion-only examples, this script builds, flashes and runs firmware on the connected board.

Run it

hpx --version
hpx doctor
./run.sh

Install and provision heliaPROFILER using its own version-matched setup instructions before these commands. This AOT recipe does not pin or qualify a current profiler release, probe or board setup. Record hpx --version, review the configuration against that version, and reserve your board before running run.sh, which builds and flashes firmware.

hpx doctor checks installed host tools and Python dependencies, not a connected probe or board. Install the profiler’s aot extra for this recipe; a successful default doctor check alone does not establish AOT readiness. The script then runs both configs and the compare:

hpx profile --config hpx_rt.yml
hpx profile --config hpx_aot.yml
hpx compare results/comparison_rt results/comparison_aot --output-dir results/rt-vs-aot

The two configs

hpx_rt.yml and hpx_aot.yml name the same model, the same board and the same counters. They differ in the engine block and the result destination (output.dir):

engine:
  type: helia-rt
  config:
    variant: release-with-logs

engine:
  type: helia-aot
  config:
    prefix: hpx
    module_name: hpx_model

How to read the comparison

The script requests compare_summary.json under results/rt-vs-aot. Inspect that summary’s comparability decisions before quoting a difference. Model identity, validity, topology and power scope can qualify different metric families separately; a produced report is not blanket approval of all metrics. Exact rules belong to the profiler version you ran.

Keep model inputs and intended board settings fixed. Runtime libraries, compiler flags, logging and placement can still differ between engines; record these as confounders. A profiling-image size is not runtime-only support size, and source or flash bytes do not measure inference cycles.

Quote any measurement with the board, the two engine versions and the date it was taken. A figure without them is not reproducible and does not survive the next release of either engine.

Next

• Measure a conversion for the canonical AOT measurement workflow and interpretation.
• MLPerf Tiny benchmark for a comparison already run, with its labels.