Vanilla TFLM baseline

Goal

Produce the baseline that gives the heliaRT and heliaAOT numbers their meaning. You will learn what the tflm engine is and is not, how to pick its kernel backend, and how its binary compares.

Hardware and wiring

An Apollo510 EVB on its J-Link USB port; no other connections.

Software prerequisites

hpx, the GNU Arm toolchain, CMake, Ninja and J-Link software (Install). No extra: the tflm engine resolves the nsx-tflite-micro module from the registry.

Config file

examples/quickstart/hpx_tflm_baseline.yml:

model:
  path: examples/quickstart/kws_model.tflite
  arena_size: 131072
engine:
  type: tflm
  backend: cmsis_nn    # reference or cmsis_nn
target:
  board: apollo510_evb
  toolchain: arm-none-eabi-gcc
profiling:
  pmu_counters:
    cpu: default
    memory: default
  per_layer: true
  iterations: 3
  warmup: 1
power:
  enabled: false
output:
  format: csv
  dir: ./results/results_tflm_baseline

engine.backend: cmsis_nn links upstream Arm CMSIS-NN; reference selects TFLM’s portable C kernels. Neither enables Ambiq’s HELIA kernels; that is the point of the baseline.

Run

From the repository root (the model path is repository-relative):

hpx profile --config examples/quickstart/hpx_tflm_baseline.yml

Expected output

From the nightly case named under Validated on (the arena progress bar is omitted):

  Engine              tflm
  Board               apollo510_evb
  Layers              13
  Total cycles        2,118,339
  Clean E2E cycles    2,109,120  (-0.4% vs per-layer sum)
  Arena      27,300 / 32,768 bytes  83%
  Model      53,744 bytes
  Binary Sections
  Section          Size
  text          507,932
  data          154,796
  bss            89,284
  reserved      362,380
  total       1,114,392
  Validity: VALID

How to read it

• 2,109,120 clean cycles against heliaRT’s 2,051,232 and heliaAOT’s 2,044,512 on the same board, model and toolchain: for this model the arithmetic runs in CMSIS-NN either way, so the engines separate by a few percent. Models with operators that heliaRT fuses show a larger gap; the baseline is what tells you how large.
• .text is 507,932 bytes, 1.8 times heliaRT’s 278,596, because the upstream interpreter links its full operator set and resolver.
• The per-layer table has the same columns as every other engine, so a row-by-row comparison with hpx compare works (Compare engines).
• The result describes the vanilla port; it is not an Ambiq-optimised number and should not be quoted as one.

Variations

• Reference kernels: copy the config, set backend: reference, keep everything else, and compare the two TFLM runs to see what CMSIS-NN buys on this core.
• Another model: change model.path and model.arena_size; the arena usage line tells you how much you needed.
• Cortex-M4 boards (apollo3p_evb, apollo4l_blue_evb) run the same engine with cycle counts only.

Validated on

Board	apollo510_evb, lp clock tier
Run	nightly hardware validation run 35078330526, 2026-09-16, hpx 0.1.6 at 0714893
Case	apollo510_evb-kws-tflm-arm-arm-none-eabi-gcc-rtt-auto (arm is the CMSIS-NN provider), verdict valid
nsx modules	nsx-tflite-micro 7afcf2b, arm-cmsis-nn 6d21a6f, nsx-ambiq-sdk a9f4ec2, nsx-pmu-armv8m 5725c06 (bundle nsx.lock)
Toolchain	arm-none-eabi-gcc 15.2.1, CMake 4.1.2
Differences from the config above	arena_size: 32768, pmu_counters: {cpu: default} in the nightly case
reference backend	not yet validated: the nightly suite runs the CMSIS-NN backend only
Bundle	examples/results/hardware-validation-2026-09-16/apollo510_evb-kws-tflm-arm-arm-none-eabi-gcc-rtt-auto/