Bypassing the AVX2 Tax (Part 4): AMD on the Line and the Windows Release

The piece reports a Windows deployment test of Fruit Inspector, a real-time fruit inspection pipeline, on a compact Firebat R3 mini PC with a Ryzen 5 7430U and integrated Radeon graphics. Each frame is captured as CPU-side NV12 video, processed by YOLOX-nano at 640×640 for detection and tracking, then cropped to a 256×256 region around the tracked item. A LightGBM classifier running through ONNX Runtime grades the fruit from 44 features. The intended factory setup favors a self-contained folder of an executable, DLLs, and models, with no Python or CUDA installation. The comparison focuses on CPU, OpenVINO, and DirectML execution providers, selectable through settings without rebuilding the application.

On this AMD machine, CPU execution is fastest: YOLO inference takes about 6.8-7.5 ms, producing roughly 77-88 frames per second for the pipeline’s YOLO passes. OpenVINO’s automatic selection performs similarly, but its bundled GPU support targets Intel hardware, not the Radeon iGPU. DirectML takes about 11 ms, as GPU and D3D12 overhead outweigh the benefit for this small network; the recommendation is to benchmark it before use. The public release therefore includes only CPU files, while optional providers remain in the development tree to avoid a larger archive and extra setup. The comparison also places the Ryzen between an i5-11400 at 124-132 fps and an older, non-AVX2 i3-2120 at 11-15 fps, arguing that newer SIMD support and an iGPU do not automatically make inference faster.