Enabling hardware acceleration does not take effect

[zjl767864079zjl]

I need to configure 8 cameras, but when using the CPU, I can only configure up to 4. If there are more than 4, all camera decoding will fail. I have enabled NVIDIA GPU acceleration, and with Nvidia smi, I am able to check usage status. However, I can only configure up to 4 cameras, and if there are more than 4, all cameras will fail to decode.

[inkeep-ai-assistant[bot]]

Based on your nvidia-smi output showing ffmpeg processes using the GPU, your hardware acceleration appears to be working. However, there are several configuration aspects to verify for NVIDIA GPU acceleration in Frigate.

First, ensure you're using the correct Frigate Docker image for NVIDIA GPUs. For NVIDIA hardware acceleration, you need the -tensorrt tagged image (1).

For NVIDIA GPU hardware acceleration, configure your ffmpeg settings with:

ffmpeg:
  hwaccel_args: preset-nvidia

(1)

You should also verify that hardware decoding is working correctly by running nvidia-smi, which should show ffmpeg processes (1). Your screenshot confirms this is happening.

However, the issue might be related to your Docker configuration. When using NVIDIA GPUs, you need to pass through the GPU to the Docker container using the same methods described in the hardware acceleration documentation (1).

The fact that you can only configure 4 cameras before all fail suggests a resource limitation. This could be:

1. Shared memory (shm-size) - Ensure your Docker container has sufficient shared memory allocated. The default may be too small for 8 cameras (2).

2. GPU memory limits - Your Tesla V100 shows 16GB memory, but each camera stream consumes GPU memory for decoding.

3. Container resource limits - Check if your Docker container has CPU or memory limits that could cause failures.

For troubleshooting, check the docker logs for the container and look for decoding errors when you add the 5th camera (1).

The community has successfully run many cameras with similar setups. One user reported running up to 32 FHD video streams on an Intel N100 with Coral (3), suggesting the issue may be configuration-related rather than a fundamental limitation.

📚 Sources:

1. Video Decoding
2. Getting Started
3. [HW Accel Support]: Unsure if hardware acceleration actually working, or working as it should be

---

Have another question? Just tag @inkeep.

[zjl767864079zjl]

Thank you very much, the detector is now work with GPUs. However, the current situation is not very ideal. Although both the detector and video streaming process can be detected in the GPU, the CPU usage is still high and the GPU usage is low. Is there any other parameter that allows it to use GPU instead of CPU as much as possible?

Comparison of GPU and CPU usage rates:

CPU usage:

[hawkeye217]

What is your new config now?

[zjl767864079zjl]

config1024.yaml

[hawkeye217]

Multiple detectors cannot be mixed for object detection. You have 4 different ONNX detectors defined, but Frigate only supports one detector type for object detection. Docs: https://docs.frigate.video/configuration/object_detectors#onnx

Your model configuration is incorrectly nested - The model section should be at the global level, not under each detector.

[zjl767864079zjl]

ok,thank you.