What has actually been measured?

Query the normalized edge AI benchmark database by hardware, model family, precision, runtime and task. Every row keeps its evidence class — C measured and published by the source named on the row, D EdgeAIStack-derived with the derivation stated — plus its source line. Nothing is scored; you get rows, an expected range per precision, and a cross-hardware comparison.

01 · Define query

Hardware & model

Hardware

Loading benchmark catalog…

Model family

Any

Precision

Any

FP32

FP16

FP8

INT8

INT4

Task

Any

Evidence & sort

Evidence

Measured + derived

Measured only

Sort

Throughput ↓

Latency ↑

Evidence class

Newest

Advanced

Runtime

Resolution

Power mode

Limit

Compare across hardware

Off

All hardware for this model

Select a hardware, model family or task to continue

03 · Submit a measurement

Ran the measurement protocol on your own hardware? Submit the result below. It is reviewed against the protocol before it appears as a class C row with your attribution — nothing is published automatically.

Submit a measurement

Hardware id *

Model family *

Variant *

Precision *

Runtime *

Task

Resolution

FPS

Latency (ms)

Tokens / s

JetPack (if applicable)

Power mode

Exact benchmark command *

Samples

Warm-up run

Attribution (name/handle) *

Contact (optional)

Evidence URL (optional)

Notes (optional)

04 · How this works

Every row keeps its evidence class

Class C means the number was measured and published by the source named on the row — a vendor benchmark page, a paper, or a community submission reviewed against the measurement protocol. Class D means EdgeAIStack derived it — scaled or interpolated from measured rows on related hardware — and the row states the derivation method. Rows are deduplicated per hardware × model × runtime × precision × input × power mode, keeping the best evidence.

The expected range is not a score

For a given hardware, model and variant, the expected-range card is the min / median / max of every matching row per precision — not a single "recommended" number. When a precision has no row at all, the estimator's benchmark hierarchy (measured on this hardware → close sibling → cross-platform scale) is consulted and its tier chain is reported, so a gap in the data is explicit rather than silently filled in.

Comparison is the best row per hardware, not a ranking score

The cross-hardware comparison takes the fastest matching row for each hardware for the same model and precision and shows it relative to the fastest — it does not weigh cost, power or ecosystem the way Hardware Selector does.

Full formulas, source list and what invalidates a result: methodology. Static tables: /benchmarks/. Raw dataset: datasets/benchmarks-v2.json. Measurement protocol: how to submit a comparable number.

05 · FAQ

What is the difference between class C and class D rows?

Class C rows are measured and published by the source named on the row — a vendor benchmark page, a paper, or a community measurement reviewed against the measurement protocol. Class D rows are EdgeAIStack-derived: scaled or interpolated from measured rows on related hardware, with the derivation method stated on the row.

What does the expected range mean when it says fallback?

When no row matches the exact hardware / model / precision, the expected-range card falls back to the same benchmark hierarchy the estimator engines use (measured on this hardware, then a close sibling, then a cross-platform scale) and reports the chain it walked so the gap is explicit rather than silent.

Are these numbers full-pipeline or inference-only?

Unless a row's scope says end_to_end, every number is inference-only at batch 1 — decode, pre/post-processing and tracking are not included. Use Camera Stream Capacity for the full per-stream budget.

Can I contribute a measurement?

Yes — run the published measurement protocol on your hardware and submit the result from the Submit a measurement section above. It is reviewed against the protocol before it appears as a class C row with your attribution.