Player not loading? Watch on YouTube
This tutorial adds an RTX 3060 with 12 GB of VRAM to a Ryzen 7 2700 running Proxmox VE 9. The card goes to a Debian 13 VM with Frigate 0.17. The speaker explains GPU object detection and video decoding, then configures semantic search and camera event summaries through Ollama.
The setup covers BIOS settings, VFIO binding, PCIe passthrough, the Nvidia driver and NVIDIA Container Toolkit. Frigate uses its TensorRT image with a YOLOv9 model converted to ONNX. The speaker starts with the small model at 320 input resolution and warns that the first TensorRT engine build takes 2 to 10 minutes. On this motherboard, passing through the HDMI audio function caused a reset failure, so the VM receives only the graphics function.
Frigate and Ollama share the card through separate Docker containers in the same VM. The walkthrough raises VM memory to 16 GB and uses Qwen 3.5's 4 billion parameter vision model. It keeps the model loaded and disables thinking for summaries. Semantic search indexes event thumbnails; older events need a reindex step.
Review summaries use up to 20 frames after an event ends. The speaker cautions that captions can misidentify people or shadows, and generation briefly competes with detection for GPU time. He reports single-digit millisecond detection and says this self-hosted setup keeps frames on the network and works offline.
