Skip to content

Engine Comparison#

Goal: run the same model through heliaRT and heliaAOT to compare binary size, cycles, and memory usage.

Setup#

One config per engine, same model and board:

hpx_rt.yml
model:
  path: my_model.tflite
  arena_size: 131072

engine:
  type: helia-rt
  config:
    variant: release-with-logs
    dist_path: path/to/helia_rt_v1_7_0

target:
  board: apollo510_evb

profiling:
  pmu_counters:
    cpu: all
  per_layer: true
  iterations: 5
  warmup: 2

output:
  format: csv
  dir: ./results/comparison_rt
  detailed: true
hpx_aot.yml
model:
  path: my_model.tflite
  arena_size: 131072

engine:
  type: helia-aot
  config:
    prefix: hpx
    module_name: hpx_model
    cmsis_nn_path: path/to/ns-cmsis-nn

target:
  board: apollo510_evb

profiling:
  pmu_counters:
    cpu: all
  per_layer: true
  iterations: 5
  warmup: 2

output:
  format: csv
  dir: ./results/comparison_aot
  detailed: true

Run#

hpx profile --config hpx_rt.yml
hpx profile --config hpx_aot.yml

What you get#

import json

rt = json.load(open("results/comparison_rt/summary.json"))
aot = json.load(open("results/comparison_aot/summary.json"))

print(f"{'Metric':<20} {'heliaRT':>12} {'heliaAOT':>12}")
print(f"{'Total cycles':<20} {rt['total_cycles']:>12,} {aot['total_cycles']:>12,}")
print(f"{'Binary (total)':<20} {rt['binary']['total']:>12,} {aot['binary']['total']:>12,}")

Metric                    heliaRT     heliaAOT
Total cycles            2,016,376    2,010,842
Binary (total)             752,436       96,100
Metric What to expect
Binary size heliaAOT is typically much smaller — no interpreter, dispatch tables, or op resolver
Total cycles Similar for models using the same CMSIS-NN kernels underneath
Arena usage heliaAOT's per-tensor placement can pack the arena tighter than a single runtime arena

Isolate the engine effect

Keep toolchain, board, and PMU counters identical across both configs so the only variable is the engine. See Toolchain Comparison if you also want to vary the compiler.

Where to go deeper#