Three new references this month.

x86 RTX GPU inference

x86-rtx runs GPU inference on Intel x86-64 with a discrete NVIDIA RTX GPU, two ways selectable at provision time: host-native TensorRT-RTX that compiles an ONNX model into an engine on-device, and a containerized path that runs an nvidia/cuda image on the GPU through the NVIDIA Container Toolkit via CDI. Both ride the same driver stack, supplied by the board BSP chosen with --target-board (Neousys Nuvo-9000 or LattePanda Sigma). The CUDA image is pre-seeded into /var at build time, so the device runs offline with no registry pull.

RZ/V2N DRP-AI3 object detection

rzv2n-drpai-yolo runs YOLOv3 object detection on the on-chip DRP-AI3 accelerator of the SolidRun RZ/V2N HummingBoard SoM, using Renesas's TVM-compiled model from the RZ/V2N AI SDK. It pulls frames from a looping video file, draws bounding boxes per frame, and renders the annotated feed full-screen on Wayland/Weston via GStreamer.

Custom signed U-Boot for i.MX 8M Plus EVK

uboot-custom-imx8mp-evk cross-compiles a custom imx-boot bundle (TF-A BL31 + U-Boot SPL/proper + DDR firmware) from source in the SDK container and replaces the upstream bootloader on the NXP i.MX 8M Plus EVK. It builds HAB-ready (so the image can be signed against the SoC's SRK fuses) with FIT signature verification, and an insert-fit-pubkey.sh step injects your RSA public key so you can rotate FIT signing keys without rebuilding U-Boot.