Measure every model on every chip.
Throughput, latency, power, and accuracy for every model and every chip, updated with every build.
The datasheet number is not your number.
Know the chip before you buy it
Set a power budget and a frame-rate floor, and see which boards actually hit it with your model.
Track a model across builds
Watch throughput and latency move as you iterate, chip by chip.
Accuracy, not just speed
Every run reports accuracy measured on the device alongside throughput and power, so a quantization regression shows up as a number.
Measure everything that matters.
Throughput
Mean FPS over a steady-state window.
Latency
p99 in milliseconds, end to end, postprocessing included.
Power
Watts sampled at the rail.
Accuracy
mAP 50:95 on the COCO validation set for detection models.
Efficiency
FPS per watt, for the fleets that run on batteries.
Build over build
Track every metric across builds and watch a model get faster over time.
From build to fleet and back.
The fleet reports back from the field, and that data makes the next build better.
Measure it. Then ship it.
Run the matrix against your constraints and see which chips clear them.