Every command from dataset to live NPU inference: YOLOv8n fine-tuning, QAIRT A16W8 quantization, and a resident C++ daemon on the IQ-8275 EVK. 9 steps, no shortcuts.
No QAIRT toolchain, no C++. Train on the platform, download one binary, get 2 ms NPU inference on the IQ-8275 EVK. Under two hours.
From first photo to live NPU inference on the IQ-8275 EVK: every design decision, failure, fix, and benchmark. CNN from scratch vs YOLOv8n fine-tuning vs Edge Impulse, all paths converge at 2 ms.
Quantize YOLO depth with AI Hub, compile to QNN DLC/context, run it from a persistent C++ QNN app, and display a live USB-camera depth overlay with Python/OpenCV.
Take Pi0.5, a 3-billion-parameter vision-language-action model, from download to robot motion entirely on the Hexagon NPUs of an IQ-9075 EVK — 1.1 s action chunks, 4.3× real time, no CPU fallback.
Quantize Nemotron Nano with AIMET W4A16, compile it in Qualcomm AI Hub, and run it with Genie/QnnHtp on the IQ-9075 EVK.
Build two OSTree images, flash once, then move a live IQ-8275 EVK to the new version over the air, with rollback safety and an 8 MB delta.
Bring up a TurtleBot3 Burger on a Dragonwing SBC, build a map with Cartographer SLAM, and drive it autonomously with Nav2, all on ROS 2 Jazzy.
Get live, programmatic NPU utilization data on Qualcomm Ubuntu using libqcperf, including Q6, HVX, HMX, and clock metrics over FastRPC.
Build LiteRT-LM from source and run Google’s Gemma-4 E2B on the Hexagon NPU of a Dragonwing IQ-8275, from an Ubuntu prototype to a Qualcomm Linux production image.