NVIDIA GPU detection with GTX 1650 in Frigate 0.17

[igortau]

Hello,

I am running Frigate 0.17.0 in Docker on Debian.

Hardware:
GPU: NVIDIA GTX 1650 (4 GB)
Driver: 535
CUDA: 12.2

System:
x86_64
Docker

Current configuration:

• 5 cameras
• TP-Link cameras using two streams (stream1 for recording, stream2 for detection)
• D-Link cameras using two streams (live1.sdp for recording, live2.sdp for detection)
• NVDEC hardware acceleration for video decoding is working correctly

Object detection is currently running on CPU.

Frigate works stable, but CPU usage is around 70-80%.

Question:

What is the recommended way to run object detection on NVIDIA GPU in Frigate 0.17?

Should I use:

• the stable-tensorrt Docker image
• detectors: onnx
• another configuration?

Is GTX 1650 suitable for GPU detection in Frigate or is CPU detection preferable for this GPU?

Thank you.

[inkeep-ai-assistant[bot]]

For Frigate 0.17 with your GTX 1650, use the ONNX detector with the -tensorrt Docker image.(1)

Nvidia GPUs automatically detect and use ONNX when you configure a supported ONNX model in the -tensorrt Frigate image.(1)

Recommended Configuration:

detectors:
  onnx:
    type: onnx

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

(1)

Key Points:

• Use the ghcr.io/blakeblackshear/frigate:stable-tensorrt image(1)
• The ONNX detector supports YOLOv9, RF-DETR, YOLO-NAS, YOLOX, and D-FINE models(1)
• YOLOv9 and RF-DETR support CUDA Graphs for optimal Nvidia performance(1)
• Multiple ONNX detectors can be defined if GPU resources are available(1)

Your GTX 1650 is suitable - a GTX 1660 Super achieved 7ms inference time at 7% GPU usage with 3 cameras(2), so your GTX 1650 should handle 5 cameras efficiently and reduce your CPU usage significantly.

The GPU will be automatically detected when using the -tensorrt image with ONNX configuration.(1)

📚 Sources:

1. Supported Hardware
2. [Detector Support] RTX 1660 Super | AttributeError: 'TensorRtDetector' object has no attribute 'outputs'

---

Have another question? Just tag @inkeep.

[igortau]

Thank you.

I previously tried switching my working setup and it caused instability, so I want to avoid changing the production container directly.

To confirm the recommended path for my specific setup:

• Frigate 0.17.0
• Debian + Docker
• NVIDIA GTX 1650 4GB
• 5 cameras
• NVDEC decode already working
• detection currently on CPU

Is the correct and supported migration path really:

1. switch to ghcr.io/blakeblackshear/frigate:stable-tensorrt
2. configure detectors: onnx
3. explicitly define the ONNX model: section

Also, is GTX 1650 known to work well enough for this in practice?

I can test it in a separate container, but I would like to confirm the recommended configuration first.

[NickM-27]

yes that is what you would do and expected to work