[Detector Support]: Unable to use GTX 1070 for Object Detection #21803
Replies: 2 comments 3 replies
-
|
Your GTX 1070 is experiencing CUDA errors when using the TensorRT image with ONNX models. The error "CUDA failure 700: an illegal memory access was encountered" indicates a compatibility issue between your GPU architecture and the CUDA/cuDNN versions in the current Frigate TensorRT image.(1) Recommended Solution: Switch to the Default Frigate Image For your GTX 1070 (Pascal architecture), use the default Frigate image instead of the TensorRT image. The default image will automatically detect and use OpenVINO for ONNX models, which should work better with your GPU.(1) Update your docker-compose.yml: services:
frigate:
container_name: frigate
restart: unless-stopped
stop_grace_period: 30s
image: ghcr.io/blakeblackshear/frigate:stable
shm_size: "4gb"Remove the TensorRT-specific configuration:
For Hardware Accelerated Video Decoding Keep the NVIDIA preset for hardware-accelerated video decoding in your config.yml: ffmpeg:
hwaccel_args: preset-nvidiaYou'll still need to pass through the GPU for video decoding. Add this to your docker-compose.yml: services:
frigate:
devices:
- /dev/dri:/dev/driWhy This Works The default Frigate image will automatically detect and use OpenVINO with your ONNX models for object detection, while still allowing hardware-accelerated video decoding via NVIDIA.(1) The TensorRT image uses CUDA 12.3, which may have compatibility issues with older Pascal GPUs like the GTX 1070.(3) 📚 Sources: Have another question? Just tag @inkeep. |
Beta Was this translation helpful? Give feedback.
-
|
I will give this a shot I guess. |
Beta Was this translation helpful? Give feedback.
-
|
We did have a user confirm that Maxwell architecture GPUs (900 series) were no longer functional in 0.17 due to the upgrades to onnxruntime. Not sure about Pascal cards, though. I think both architectures were recently moved to the "legacy" driver branch by nvidia. |
Beta Was this translation helpful? Give feedback.
-
|
Thank you. From what I can tell it should work. especially since I am on 0.16. |
Beta Was this translation helpful? Give feedback.
-
I'm using a 1080Ti successfully with 0.16.4. |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Describe the problem you are having
I have been trying to test out Frigate as a docker container on my Ubuntu system that has a GTX 1070 GPU. I want to use the GPU for Object detection and decoding if possible.
I have downloaded multiple models (yolov9, yolov7, yolovx) and I always get this error:
I am not really to sure what I am needing to do to get this to work. Does the TensorRT image not support Pascal GPUs anymore? Do I need to use an older version? I am not sure if its just an issue with the age of the GPU or a version issue.
Here is the nvidia-smi output:
The config file and the docker compose file I also have added here. Please let me know if there is anything that I can do to make this work.
Version
0.16.3-9034454
Frigate config file
docker-compose file or Docker CLI command
Relevant Frigate log output
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
Beta Was this translation helpful? Give feedback.
All reactions