Scenarios, not client stories: where a vision model meets a fixed board, and what we measure there.
One edge box at a crossroads analyses several camera feeds. The model sets how many cameras each box serves, and so how many boxes a city buys.
Who brings us inBoard distributors and camera makers.
Jetson Orin NX, YOLOX-Nano: 16 cameras per box at most instead of 11, at 30 fps assumed per camera, after one change in the application (no retraining).
Demonstration on public data (COCO val2017). An upper bound computed from the GPU latency, for this model alone, one image at a time; host-device copies, image decoding and NMS excluded.
A classifier sorts good parts from defects on a conveyor. It is accurate in the lab, too slow on the line’s board, and a bigger board at every station is not an option.
Who brings us inVision software vendors and integrators.
Jetson Orin NX, ResNet-18 classifier: 0.850 → 0.449 ms.
See it block by blockDemonstration on public data (ImageNet). Latency measured on weights not yet retrained; accuracy known after retraining.
A robot, a drone or a smart sensor carries its own chip. The chip is set by weight and power, not by the model.
Who brings us inChip and board makers.
STM32N6 microcontroller, person detection: 3.5 ms for our model against 9.4 ms for the chipmaker’s own model of the same size, at the same accuracy (44.2 vs 43.8 AP50).
Verify it yourselfST Edge AI 4.0.1 · AP50 simulated on PC on 1,345 COCO val2017 person images · compared after evaluation.