Vanilla TFLM Baseline#

Goal: profile the vanilla TFLM port as a baseline for comparison with heliaRT and heliaAOT.

The tflm engine uses the nsx-tflite-micro NSX module. It is intentionally separate from heliaRT: the baseline does not enable Ambiq's HELIA kernels. Choose either the reference kernels or upstream CMSIS-NN with engine.backend.

Setup#

Start with examples/quickstart/hpx_tflm_baseline.yml:

engine:
  type: tflm
  backend: cmsis_nn    # reference or cmsis_nn

The bundled KWS model is used by default. Replace model.path with another .tflite model and adjust model.arena_size if the first run reports a larger arena requirement.

Run#

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

To compare the two TFLM kernel choices, make a copy of the config and change backend to reference. Keep the model, placement, board, toolchain, and profiling settings identical when comparing results.

What you get#

The output contains the same cycle, PMU, memory, and per-layer reports as the other profiler engines. The result is a baseline measurement for the vanilla TFLM port; it should not be interpreted as an Ambiq-optimized runtime result.

Where to go deeper#

• Inference Engines
• Configuration
• Engine Comparison