Power capture with a Joulescope

Goal

Capture current, charge and energy for the clean inference window of a profile run. You will learn what the config adds, what the run does after the profile pass, and which health signals decide whether the energy number can be quoted.

Hardware and wiring

An Apollo510 EVB, a J-Link, and a Joulescope JS110, JS220 or JS320 in series with the board’s supply, with its GPIO input on the board’s sync pin and, on the JS220 and JS320, its Vref pin wired. The wiring, the pin table and the instrument differences are on Joulescope; read it before plugging anything in.

Software prerequisites

hpx with the aot extra (pyjoulescope_driver ships with hpx), the toolchain and J-Link software, and on Linux the Joulescope udev rule. hpx power-on is the smoke test that the instrument is seen.

Config file

examples/quickstart/hpx_aot_power.yml:

model:
  path: kws_model.tflite
  arena_size: 131072
engine:
  type: helia-aot
  config:
    prefix: hpx
    module_name: hpx_model
target:
  board: apollo510_evb
  toolchain: arm-none-eabi-gcc
profiling:
  pmu_counters:
    cpu: default
  per_layer: true
  iterations: 3
  warmup: 1
power:
  enabled: true
  driver: joulescope
  mode: external
  duration_s: 20
  io_voltage: 1.8
  sync_gpio_pin: 29
output:
  format: csv
  dir: ./results/results_aot_power

power.enabled turns the five power stages on; driver, mode, duration_s, io_voltage and sync_gpio_pin are explained on Power measurement and the generated configuration reference.

Run

cd examples/quickstart
hpx power-on --driver joulescope
hpx profile --config hpx_aot_power.yml

The run profiles as usual, then builds and flashes a dedicated power binary, arms the Joulescope, resets the board and captures the gated clean window.

Expected output

Not yet captured on this bench. A completed run adds a power block to summary.json with measurement_scope, energy_per_inference_j, capture_duration_s, the gated-window statistics and the health signals; the fields are listed on Output files and the terminal panel is described on Verifying a capture.

How to read it

• measurement_scope must read gpio_gated_clean_window for the energy per inference to mean one inference; the other scopes are explained on Windows and lock-step.
• The power.window_* and gated_window_duration_suspect codes in issues are the health checks; a degraded verdict with one of them means the number is a caveat, not a measurement.
• Compare two power runs only through hpx compare, which withholds the power deltas when the scope, mode, firmware, probe or clean workload differ (Analysis and comparison).

Variations

• Reading the block the run wrote: Read a power summary; holding two captures against each other: Compare two power runs.
• heliaRT instead of heliaAOT: copy the power block into hpx_rt.yml.
• INA228 on the board instead of a Joulescope: power.driver: ina228, a different measurement scope, on INA228.
• Longer captures for tighter statistics: raise duration_s; the fixed inference count is derived from the profile pass’s clean timing.

Validated on

Not yet validated. The nightly hardware validation runs with power: off, and no captured power bundle is shipped with the documentation. The configuration above is the checked-in quickstart file; the bench facts on the linked measurement pages are cited there.