---
license: cc-by-4.0
task_categories:
- image-classification
tags:
- benchmark
- edge
- mobile
- latency
- quantization
pretty_name: TinyEdge edge-inference benchmarks
---

# TinyEdge edge-inference benchmarks

Independently **measured** latency and accuracy for well-known vision models on
real edge devices (phones, tablets — fleet growing), produced by
[TinyEdge](https://tinyedge.ai), a device cloud for edge-AI benchmarking.
Nothing here is taken from papers or spec sheets: every row is a job executed
on the physical device through TinyEdge's production agent, with accuracy
measured on a fixed 500-image stratified sample of
[ImageNet-V2](https://github.com/modestyachts/ImageNetV2) (matched-frequency)
using a standardized preprocessing spec, batch 1, 224x224.

ImageNet-V2 top-1 runs systematically below classic ImageNet val for every
model (known distribution shift) — treat `accuracy_value` as a *relative*
capacity indicator; rankings and deltas are the signal.

## Headline findings (snapshot 2026.06)

- **fp16 is a placebo on these NPUs** (onnxruntime-android + NNAPI): identical
  latency and accuracy to fp32 on every model tested.
- **Static int8 (QDQ) is the real lever: 2.6–2.8x faster at bit-identical
  accuracy** for MobileNetV2 — when the architecture tolerates it.
- **Quantization safety is architecture-dependent:** naive post-training int8
  collapsed MobileNetV3-Small (−57 pts) and EfficientNet-B0 (−32 pts);
  MobileNetV3-Large lost 4 pts; MobileNetV2 lost nothing.
- **Dynamic ("data-free") int8 does not run at all** on Android ONNX Runtime
  (`ConvInteger` not implemented).
- **Published accuracy can invert on-device rankings:** EfficientNet-B2
  (published +2.9 over B0) measured *no better* than B0 while 51% slower.

## LLM benchmarks (llm-benchmarks.csv)

GGUF quantization ladders measured on-device via llama.cpp (CPU): decode and
prefill tok/s at the measured-optimal thread count (big cores only — the
llama.cpp default of all-cores is 28-39% slower on big.LITTLE SoCs), with
WikiText-2 perplexity (fixed 32-chunk subset) as the device-independent
quality column. First snapshot: Qwen2.5-0.5B-Instruct, Q8_0→Q3_K_M, on a
Snapdragon 662 phone — 14-19 tok/s decode (conversational) at every level;
Q6_K dominates Q8_0; prefill (~30-40 tok/s) is the practical bottleneck for
long prompts.

## Schema

One row = one (model, device, precision/quant) measurement.
Key columns: `latency_ms_p50`, `throughput_per_s`, `accuracy_value`
(`accuracy_metric` = imagenetv2_top1), `quant_method`, `runtime` (includes the
execution provider that actually ran), `sweep` (provenance).
The `task` column is `image-classification` for all current rows; LLM rows
(decode tok/s, TTFT, RAM) will use the same table with `task=text-generation`.

## Method & provenance

Sweep tooling and raw per-run JSON live in the TinyEdge repos; results are
snapshot-versioned via the `snapshot` column. Jobs run on battery-policy-aware
agents; each result records the exact runtime + execution provider. Eval images
are NOT redistributed here — only measurements.

Questions / want your model or device measured: https://tinyedge.ai/contact