NVIDIA CMP 50HX for object detection

[belovkb]

Describe the problem you are having

Good afternoon. At the moment, I use a GTX 1060 with 6 gigabytes of memory to recognize objects in video streams. It does its job to process 4 streams from the cameras, but they offer me a CMP 50HX 11Gb at a very tempting price. I know that these cards are intended for mining. At the moment, I am interested in the question of whether this card can improve the performance of my system at the moment, and how much will the performance benefit be relative to the GTX 1060. I plan to connect several more cameras to the video surveillance system.

Version

0.16.0-c2f8de9

Frigate config file

not need

docker-compose file or Docker CLI command

not need

Relevant Frigate log output

not need

Install method

Docker Compose

Object Detector

TensorRT

Screenshots of the Frigate UI's System metrics pages

Any other information that may be helpful

No response

[inkeep-ai-assistant[bot]]

Frigate supports object detection on Nvidia GPUs using TensorRT, as long as the GPU supports CUDA 12.x and has a compute capability of 5.0 or greater(1). The CMP 50HX is based on the same architecture as the RTX 2080 (Turing, compute capability 7.5), so it meets the technical requirements for TensorRT use in Frigate(1).

You can expect performance to be better than the GTX 1060, since the CMP 50HX has more CUDA cores and significantly more memory. Actual inference speed depends on the model you use and your overall system configuration, but Turing-class cards generally outperform Pascal-class ones like the GTX 1060(1).

To use the CMP 50HX with Frigate and TensorRT:

1. Ensure you are running the Frigate Docker image with the -tensorrt tag(2).
2. Make sure your host system has the NVIDIA driver (>=545) and nvidia-container-runtime installed(1).
3. In your config, specify the TensorRT detector:

   detectors:
     tensorrt:
       type: tensorrt
       device: 0 # Selects the first GPU
   model:
     path: /config/model_cache/tensorrt/yolov7-320.trt
     labelmap_path: /labelmap/coco-80.txt
     input_tensor: nchw
     input_pixel_format: rgb
     width: 320
     height: 320

4. Models are generated automatically at startup if not present, or you can specify them with the YOLO_MODELS environment variable(2)(3)(4).

Frigate does not provide specific performance benchmarks for the CMP 50HX, but based on the architecture, you should see a substantial improvement over the GTX 1060, especially for more streams or higher resolution models(1).

📚 Sources:

1. Recommended hardware
2. Supported Hardware
3. Supported Hardware
4. How to use Nvidia GTX 1080 card for Object detection and also off loading to the GPU

---

Have another question? Just tag @inkeep.

[NickM-27]

I am not aware of anyone that has run Frigate with that hardware

[belovkb]

here are the approximate characteristics of this card. the only thing that bothers me is that the PCI-Express bus is version 1.0*4

[belovkb]

this 1060