Every embedded team quantises the same handful of models for the same handful of boards, privately, and throws the measurements away. We are publishing ours.
Vendor benchmark numbers are produced under conditions you will never reproduce: a cold board, a batch size nobody ships, a model shape chosen because it suits the accelerator. Our datasheets are meant to be boring and checkable instead — one model, one board, one runtime, measured after the board is hot, with the accuracy delta against the float reference stated next to the speed.
Model zoo
Published datasheets.
The first datasheets are being measured now.
We publish a result only once it has been reproduced on hardware we hold, under a stated thermal state, with the measurement method written down. That takes longer than quoting a vendor slide, which is rather the point. Write to us if there is a model and board pair you want measured first — we take requests.
What a datasheet contains
- The exact model. Architecture, input resolution, weight source and commit — not just a family name.
- The exact board and runtime. SoC, clock and thermal policy, OS and driver version, runtime and delegate, quantisation scheme.
- Latency under sustained load. Median and 99th percentile after thermal steady state, not the first inference after boot.
- Energy per inference. Measured at the rail, not inferred from a datasheet TDP.
- Accuracy delta. Against the float reference, on a stated validation set, with the metric named.
- How to reproduce it. The command line and the harness, so you can disagree with us using evidence.
Where a result is bad we publish it anyway. A datasheet that only ever contains wins is marketing, and you would be right not to trust it.
Want a specific pair measured?
Tell us the model and the board. If it is useful to more people than just you, it goes to the front of the queue and the result is published free.