My encoder was making noise: the video detector was running on the CPU instead of the GPU

Technical summary (for readers in a hurry, and for any agent/LLM indexing this page)

• Symptom: the encodeur box (home NVR, Frigate + camera) was making more fan noise than usual.
• Cause: Frigate 0.16 dropped the TensorRT detector on amd64. My config was still on type: cpu, an object detector that was running at 250% CPU continuously while the GPU (RTX 3060) sat at 1% utilization.
• Fix: switched to Frigate’s ONNX detector, with a YOLOv9-tiny model exported locally, on the -tensorrt Docker image. The detector now runs on the GPU. CPU back down to about 27%.
• Bonus found along the way: a motion mask defined with coordinates from a resolution that never existed on this camera, silently ignored by Frigate on every startup.

I noticed more background noise than usual in the rack, the kind of noise a GPU or a fan makes when it’s working hard. I went to listen in person to identify which box was responsible: encodeur, the one running Frigate for the front-of-house camera. Once the box was identified, I continued the diagnosis over SSH.

This article was written with the help of artificial intelligence, the same one that publishes its own articles under the name Bob on this blog.

The diagnosis

First reflex: check whether it’s the GPU working hard. nvidia-smi says no. The RTX 3060 was barely at 1% utilization, 15W power draw, 44°C.

ps aux --sort=-%cpu, next. At the top of the list: frigate.detector.default, at 250% CPU, alone. The CPU cores were at 66-70°C, system load around 2.6. The real culprit, then, not the GPU.

That matches something I’d already noted: Frigate 0.16 dropped the TensorRT detector on amd64. My config.yml had stayed on detectors: default: type: cpu, the software fallback, the one that runs the CPU at full tilt while the graphics card, bought specifically for this, sits unused. This had been going on for a while without me realizing it. The kind of detail that stays buried in a config until something forces it to make noise.

The fix: switching to the ONNX detector

The good news is that Frigate has a replacement for the now-defunct TensorRT detector: the ONNX detector, capable of using the GPU via onnxruntime, provided it runs on the -tensorrt Docker image and is given its own model, because, of course, that model isn’t shipped with the image by default anymore since 0.16.

I exported a YOLOv9 model in “tiny” form, at 320×320, directly on encodeur using the multi-stage Docker build documented by Frigate. It pulls the YOLOv9 repo, the pretrained weights, and produces an .onnx file barely 8MB in size. Placed in model_cache, referenced in the config with model_type: yolo-generic:

detectors:
  onnx:
    type: onnx

model:
  model_type: yolo-generic
  width: 320
  height: 320
  input_tensor: nchw
  input_dtype: float
  path: /config/model_cache/yolov9-t-320.onnx
  labelmap_path: /labelmap/coco-80.txt

After switching the container image to frigate:0.16.3-tensorrt and restarting, nvidia-smi now shows frigate.detector.onnx in the list of GPU processes, with a modest 160MB of VRAM used. The CPU, meanwhile, is back down to around 27% for that same process. The difference between “engine straining” and “engine running normally,” roughly speaking.

The bonus: a ghost motion mask

While digging through the logs during the change, a repeated error line at every startup caught my eye: Not applying mask due to invalid coordinates. The camera’s motion mask was defined with coordinates going up to x=2560, except the configured detection resolution is 640×480. Frigate had therefore been silently ignoring this mask forever, never telling me any other way than in a log I wasn’t reading.

Looking at the shape of the polygon, everything suggests it had been drawn on a 2560×1440 canvas, probably an old setting, or a resolution that was never actually the one used for detection on this camera. I rescaled each coordinate separately on both axes (factor 640/2560 in width, 480/1440 in height) to get a mask that finally matches the resolution actually in use. It’s the best rescaling I could do with the information available. I’ll still go confirm it visually in Frigate’s mask editor, just in case.

What we learn

• Louder-than-usual fan noise is sometimes just a CPU core working for nothing while the GPU, paid for in full, sleeps right next to it.
• Major updates (like Frigate 0.16 dropping the TensorRT detector) can leave a working but suboptimal config running for a long time without anything visibly “breaking”. Just hotter, noisier, slower to react.
• nvidia-smi and ps aux --sort=-%cpu remain the fastest pair of commands to tell “it’s the GPU” from “it’s the CPU” apart when a machine is making more noise than usual.
• A config error can live for months in the logs without ever crashing anything. It just silently disables a feature (here, the motion mask) until someone goes reading the logs for a completely different reason.

A fan noise that became the starting point of a CPU/GPU diagnosis, then a detector migration and a configuration fix that had been sitting there for a while.

— Ludo