# Benchmark tab: scope and reproducibility

Added after Conversion. Access date 2026-10-23. All values retain their stated
metric boundaries; no peak TOPS conversion or hardware scaling was used.

## AiOnIc

Use board reports under `/home/aionic/work/vlm/`:
- v124 README/METHODOLOGY and retained saved-frame complete reports.
- v125 README saved-frame phase-switch comparison.
- v139 README and fresh-camera timing/ONNX reports.
- v143 VOICE_AND_PERFORMANCE history-reuse control (v144 voice extension).
- v145 `reports/paris_voice_filter_validation.json`, Osaka equivalent,
  `reports/speech_optimization/fresh_vs_reuse.json`, `reports/cold_speech/`,
  `reports/live_japanese/`, and each corresponding `runs/voice_*/chat.json`.

Prompt/generation length and loading boundaries differ across model releases.
Retained observations are not presented as medians. Question-end estimates use
an amplitude endpoint; USB ALSA completion is not an acoustic microphone probe.
Real speech errors and synthesized controls remain separate.

## External primary/firsthand sources

1. https://github.com/hybridgroup/yzma/blob/main/benchmarks/linux.md
   Raw b11146 / 24 September 2026 rows: SmolLM-135M base Q2_K. Four-thread,
   five-run medians. Public overview has different/newer numbers, so the guide
   uses the stated raw version consistently. Inference tokens/s includes prompt
   tokens. It is not equivalent to AiOnIc output-only decode throughput.
   Metric source: https://github.com/hybridgroup/yzma/blob/main/pkg/llama/benchmark_test.go
   The multimodal metric also adds image/text input token counts:
   https://github.com/hybridgroup/yzma/blob/main/pkg/mtmd/benchmark_test.go
   Therefore no published aggregate VLM token/s value is graphed as decode.
2. https://docs.radxa.com/e/e54c/app-development/artificial-intelligence/rkllm-smolvlm2
   ROCK5B+ RK3588, W8A8 language network and FP16 vision at 384x384.
   The 62.81 tok/s number is generated-token decoding, not camera-to-caption.
   No matching revision/ONNX token-agreement evidence is supplied for our frames.
3. https://itsfoss.com/smollm-raspberry-pi/
   Firsthand Pi5 test is SmolLM **1.7B**, not135M. It appears only as context;
   quantization is not specified in that article.
4. https://developer.nvidia.com/blog/nvidia-jetpack-6-2-brings-super-mode-to-nvidia-jetson-orin-nano-and-jetson-orin-nx-modules/
   NVIDIA JetPack6.2 LLM tables4/6. SmolLM2-1.7B is a different generation and
   model size. It does not stand in for our135M/150M models.
5. https://www.raspberrypi.com/news/seeing-understanding-and-responding-low-power-cnn-vlm-and-slm-workloads-on-raspberry-pi-5/
   Pi5 with DEEPX DX-M1M and DX-M1: Qwen3-1.7B,96 prefill tokens. Context only.

No verified matching result was found for our LLM-jp model on Pi4/Pi5/Jetson,
legacy Jetson Nano, AGX Orin or Hailo. Missing values stay unreported. We have not
run those platforms. Precision/prompt differences prevent a fair hardware rank.
To extend the table, measure a pinned checkpoint and the same inputs on each
runtime; retain EOS/output count, power mode, versions, accuracy and repetitions.