[Detector Support]: Frigate 0.17.2 – recommended setup for RTX 5070 Ti / i7-14700KF?

[ThorstenUm]

Describe the problem you are having

I am migrating my Frigate installation from an older system using two Coral USB TPUs to a new system:
Debian 13
Intel Core i7-14700KF
NVIDIA GeForce RTX 5070 Ti 16 GB (Blackwell)
32 GB RAM
Frigate 0.17.2
Docker image: ghcr.io/blakeblackshear/frigate:0.17.2-tensorrt
NVIDIA driver 610.57.04
NVIDIA Container Toolkit
about 20–25 cameras
ONNX YOLOv9-S 320x320 (yolo-generic)

CUDA is working and Frigate uses the RTX 5070 Ti for both FFmpeg hardware acceleration and object detection. Detection inference is currently around 10.5 ms.

However, with the default CUDA Graph configuration the ONNX detector repeatedly became stuck and Frigate restarted the detector (Detection appears to be stuck). For testing I changed enable_cuda_graph from True to False in detection_runners.py. Since then the detector appears to run normally and the system statistics page also works correctly.

My questions:
Is disabling CUDA Graphs the recommended workaround for RTX 50-series/Blackwell with Frigate 0.17.2?
Is there a supported/configurable way to disable CUDA Graphs permanently without modifying files inside the container?
For this hardware (RTX 5070 Ti 16 GB + i7-14700KF), what detector/model configuration would you recommend for roughly 20–25 cameras?
Is YOLOv9-S 320 currently a good choice, or would you recommend another model/input size for this GPU?
Should video decoding and object detection both stay on the RTX 5070 Ti, or would you recommend a different distribution of work between CPU and GPU?
Are there any specific recommendations for Blackwell GPUs regarding CUDA, TensorRT, FP16, model size, or Frigate 0.17/0.18?

What is the recommended installation method for my hardware configuration when using Docker?
Do I have to run the model myself, as the AI ​​suggests? The AI ​​recommends like this:
https://docs.frigate.video/configuration/object_detectors/#downloading-yolo-models
Or is there a much simpler way to do this—like a standard setup—similar to how I used Coral Sticks on my old system? Is there a difference between yolov9-t-320.onnx and yolov9-s-320.onnx?
Could the aforementioned problem with CUDA Graphs be resolved by using a different model?

Version

0.17.2

Frigate config file

mqtt:
  enabled: true
  host: example-host-01.local
  port: 1883
  user: "xxxx"
  password: "xxxx"
auth:
  session_length: 8640000
tls:
  enabled: false
ui:
  time_format: 24hour
  date_style: short
  time_style: medium
birdseye:
  enabled: true
  mode: objects
  inactivity_threshold: 15
  layout:
    max_cameras: 12
detectors:
  onnx:
    type: onnx
model:
  model_type: yolo-generic
  width: 320
  height: 320
  input_tensor: nchw
  input_dtype: float
  path: /config/model_cache/yolov9-s-320.onnx
  labelmap_path: /labelmap/coco-80.txt
snapshots:
  enabled: true
  required_zones: []
  retain:
    default: 20
    objects:
      person: 15
  quality: 70
objects:
  track:
    - person
    - car
    - bicycle
    - bus
    - motorcycle
    - dog
review:
  alerts:
    labels:
      - person
  detections:
    labels:
      - bus
      - car
      - bicycle
      - motorcycle
      - dog
semantic_search:
  enabled: true
  reindex: false
  model_size: small
record:
  enabled: true
  continuous:
    days: 4
  motion:
    days: 8
  alerts:
    retain:
      days: 30
      mode: all
  detections:
    retain:
      days: 30
      mode: all
go2rtc:
  streams:
    camera_02_Live:
      - rtsp://xxxx:xxxx@example-host-02.local/Streaming/Channels/1
    camera_03_Live:
      - rtsp://xxxx:xxxx@example-host-03.local/Streaming/Channels/1
    camera_04_Live:
      - rtsp://xxxx:xxxx@example-host-04.local/Streaming/Channels/1
    camera_05_Live:
      - rtsp://xxxx:xxxx@example-host-05.local/Streaming/Channels/1
    camera_06_Live:
      - rtsp://xxxx:xxxx@example-host-06.local/Streaming/Channels/1
    camera_07_Live:
      - rtsp://xxxx:xxxx@example-host-07.local/Streaming/Channels/1
    camera_08_Live:
      - rtsp://xxxx:xxxx@example-host-08.local/Streaming/Channels/1
    camera_09_Live:
      - rtsp://xxxx:xxxx@example-host-09.local/Streaming/Channels/1
    camera_01_Live:
      - rtsp://xxxx:xxxx@example-host-10.local/Streaming/Channels/1
    camera_10_Live:
      - rtsp://xxxx:xxxx@example-host-11.local/Streaming/Channels/1
    camera_11_Live:
      - rtsp://xxxx:xxxx@example-host-12.local/Streaming/Channels/1
    camera_12_Live:
      - rtsp://xxxx:xxxx@example-host-13.local/Streaming/Channels/1
    camera_13_Live:
      - rtsp://xxxx:xxxx@example-host-14.local/Streaming/Channels/1
    camera_14_Live:
      - rtsp://xxxx:xxxx@example-host-15.local/Streaming/Channels/1
    camera_15_Live:
      - rtsp://xxxx:xxxx@example-host-16.local:10554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
    camera_16_Live:
      - rtsp://xxxx:xxxx@example-host-16.local:11554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
    camera_17_Live:
      - rtsp://xxxx:xxxx@example-host-16.local:12554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
    camera_18_Live:
      - rtsp://xxxx:xxxx@example-host-16.local:13554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
    camera_19_Live:
      - rtsp://xxxx:xxxx@example-host-16.local:14554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
    camera_20_Live:
      - rtsp://xxxx:xxxx@example-host-16.local:15554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
    camera_21_Live:
      - rtsp://xxxx:xxxx@example-host-16.local:16554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
    camera_22_Live:
      - rtsp://xxxx:xxxx@example-host-16.local:17554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
    camera_23_Live:
      - rtsp://xxxx:xxxx@example-host-16.local:18554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
    camera_24_Live:
      - rtsp://xxxx:xxxx@example-host-16.local:19554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
cameras:
  camera_01:
    enabled: true
    detect:
      enabled: false
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-10.local/Streaming/Channels/2
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-10.local/Streaming/Channels/1
          roles:
            - record
    live:
      streams:
        camera_01_Live: camera_01_Live
    motion:
      enabled: true
      threshold: 60
      contour_area: 20
      improve_contrast: true
      mask: 0,0.94,1,0.937,1,1,0,1
    objects:
      filters:
        person: {}
        car: {}
    snapshots:
      bounding_box: true
      enabled: true
      timestamp: true
      retain:
        default: 20
    zones:
      zone_01:
        coordinates: 0,0,1,0,1,0.929,0,0.937
        inertia: 3
        loitering_time: 0
        objects: person
    review: {}
    onvif:
      host: example-host-10.local
      port: 80
      user: "xxxx"
      password: "xxxx"
      autotracking:
        enabled: false
        calibrate_on_startup: false
        zooming: disabled
        zoom_factor: 0.3
        track:
          - person
        required_zones:
          - zone_02
        return_preset: Preset 1
        timeout: 10
  camera_02:
    enabled: true
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-02.local/Streaming/Channels/2
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-02.local/Streaming/Channels/1
          roles:
            - record
    live:
      streams:
        camera_02_Live: camera_02_Live
    motion:
      enabled: true
      threshold: 60
      contour_area: 20
      improve_contrast: true
      mask: 0,0.94,1,0.944,1,1,0,1
    objects:
      filters:
        person: {}
        car:
          mask:
            - 0.002,0.624,0.188,0.586,0.19,0.294,0.154,0.196,0.032,0.184,0,0.195
            - 0.704,0,0.696,0.328,1,0.498,1,0
    snapshots:
      bounding_box: true
      enabled: true
      timestamp: true
      retain:
        default: 20
    zones: {}
    review: {}
    onvif:
      host: example-host-02.local
      port: 80
      user: "xxxx"
      password: "xxxx"
      autotracking:
        enabled: false
        calibrate_on_startup: false
        zooming: disabled
        zoom_factor: 0.3
        track:
          - person
        required_zones:
          - zone_03
        return_preset: Preset 1
        timeout: 10
  camera_03:
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    enabled: true
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-03.local/Streaming/Channels/2
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-03.local/Streaming/Channels/1
          roles:
            - record
    live:
      streams:
        camera_03_Live: camera_03_Live
    motion:
      enabled: true
      threshold: 60
      contour_area: 20
      improve_contrast: true
      mask:
        - 0,0.296,0.137,0.233,0.268,0.211,0.474,0.191,0.614,0.207,0.7,0.162,0.73,0.167,0.767,0.166,0.772,0,0,0
        - 0,0.937,1,0.938,1,1,0,1
    objects:
      filters:
        person: {}
        car:
          mask:
            - 0.755,0,0.74,0.418,0.703,0.761,0.667,1,1,1,1,0
            - 0,0.174,0.628,0.148,0.628,0,0,0
    snapshots:
      bounding_box: true
      enabled: true
      timestamp: true
      retain:
        default: 20
    zones: {}
    review: {}
  camera_04:
    enabled: true
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-04.local/Streaming/Channels/2
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-04.local/Streaming/Channels/1
          roles:
            - record
    live:
      streams:
        camera_04_Live: camera_04_Live
    motion:
      enabled: true
      threshold: 60
      contour_area: 20
      improve_contrast: true
      mask: 0,0.938,1,0.938,1,1,0,1
    objects:
      filters:
        person: {}
        car:
          mask: 0.313,0.105,0.402,0.119,0.405,0.18,0.345,0.18,0.308,0.17
      mask: 0.481,0.515,0.733,0.516,0.722,0.383,0.205,0.367,0.211,0.5
    snapshots:
      bounding_box: true
      enabled: true
      timestamp: true
      retain:
        default: 20
    zones: {}
    review: {}
    onvif:
      host: example-host-04.local
      port: 80
      user: "xxxx"
      password: "xxxx"
      autotracking:
        enabled: false
        calibrate_on_startup: false
        zooming: disabled
        zoom_factor: 0.3
        track:
          - person
        required_zones:
          - zone_04
        return_preset: Preset 1
        timeout: 10
  camera_05:
    enabled: true
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-05.local/Streaming/Channels/2
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-05.local/Streaming/Channels/1
          roles:
            - record
    live:
      streams:
        camera_05_Live: camera_05_Live
    motion:
      enabled: true
      threshold: 15
      contour_area: 8
      improve_contrast: true
      mask: 0.002,0.94,1,0.934,1,1,0,1
    objects:
      filters:
        person: {}
        car: {}
      mask:
        - 0.637,0.13,0.874,0.214,0.806,0.051,0.64,0.024
        - 0.195,0.208,0.4,0.218,0.379,0.333,0.098,0.266,0.132,0.198
    snapshots:
      bounding_box: true
      enabled: true
      timestamp: true
      retain:
        default: 20
    zones:
      zone_02:
        coordinates:
          0,0.075,0.124,0.013,0.395,0.019,0.649,0.031,0.849,0.065,1,0.091,1,0.193,1,0.452,1,0.929,0,0.937
        loitering_time: 0
        objects: person
        inertia: 3
    review: {}
    onvif:
      host: example-host-05.local
      port: 80
      user: "xxxx"
      password: "xxxx"
      autotracking:
        enabled: false
        calibrate_on_startup: false
        zooming: disabled
        zoom_factor: 0.3
        track:
          - person
        required_zones:
          - zone_02
        return_preset: Preset 1
        timeout: 10
  camera_06:
    enabled: true
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-06.local/Streaming/Channels/2
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-06.local/Streaming/Channels/1
          roles:
            - record
    live:
      streams:
        camera_06_Live: camera_06_Live
    motion:
      enabled: true
      threshold: 60
      contour_area: 20
      improve_contrast: true
      mask: 0,0.938,1,0.94,1,1,0,1
    objects:
      filters:
        person: {}
        car:
          mask:
            - 0,0.457,0,1,0.749,1,1,1,1,0,0.641,0,0,0
            - 0.731,0.211,0.712,0.648,0.793,0.878,0.904,0.869,0.979,0.423
            - 0.731,0.211,0.712,0.648,0.793,0.878,0.904,0.869,0.979,0.423
      mask: 0.731,0.211,0.712,0.648,0.793,0.878,0.904,0.869,0.979,0.423
    snapshots:
      bounding_box: true
      enabled: true
      timestamp: true
      retain:
        default: 20
    zones:
      zone_05:
        coordinates:
          0.231,0.023,0.579,0.034,0.628,0.426,0.715,0.508,0.833,0.928,0,0.932,0,0.488,0,0.091
        inertia: 3
        loitering_time: 0
        objects: person
    review:
      alerts: {}
      detections: {}
    onvif:
      host: example-host-06.local
      port: 80
      user: "xxxx"
      password: "xxxx"
      autotracking:
        enabled: false
        calibrate_on_startup: false
        zooming: disabled
        zoom_factor: 0.3
        track:
          - person
        required_zones:
          - zone_05
        return_preset: Preset 1
        timeout: 10
  camera_07:
    enabled: true
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-07.local/Streaming/Channels/2
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-07.local/Streaming/Channels/1
          roles:
            - record
    live:
      streams:
        camera_07_Live: camera_07_Live
    motion:
      enabled: true
      threshold: 60
      contour_area: 20
      improve_contrast: true
      mask: 0,0.938,1,0.938,1,1,0,1
    objects:
      filters:
        person: {}
        car:
          mask:
            0.446,0.732,0.463,0.799,0.65,0.788,0.709,0.885,0.906,1,1,1,1,0.761,0.855,0.666,0.709,0.594,0.447,0.67
    snapshots:
      bounding_box: true
      enabled: true
      timestamp: true
      retain:
        default: 20
    zones: {}
    review: {}
    onvif:
      host: example-host-07.local
      port: 80
      user: "xxxx"
      password: "xxxx"
      autotracking:
        enabled: false
        calibrate_on_startup: false
        zooming: disabled
        zoom_factor: 0.3
        track:
          - person
        required_zones:
          - zone_04
        return_preset: Preset 1
        timeout: 10
  camera_08:
    enabled: true
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-08.local/Streaming/Channels/2
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-08.local/Streaming/Channels/1
          roles:
            - record
    live:
      streams:
        camera_08_Live: camera_08_Live
    motion:
      enabled: true
      threshold: 15
      contour_area: 8
      improve_contrast: true
      mask: 0,0.941,1,0.94,1,1,0,1
    objects:
      filters:
        person: {}
        car: {}
      mask: 0,0.163,0.176,0.165,0.185,0.101,0.31,0.126,0.339,0,0.003,0.002
    snapshots:
      bounding_box: true
      enabled: true
      timestamp: true
      retain:
        default: 20
    record:
      enabled: true
    zones: {}
    review:
      alerts: {}
      detections: {}
    onvif:
      host: example-host-08.local
      port: 80
      user: "xxxx"
      password: "xxxx"
      autotracking:
        enabled: false
        calibrate_on_startup: false
        zooming: disabled
        zoom_factor: 0.3
        track:
          - person
        required_zones:
          - zone_06
        return_preset: Preset 1
        timeout: 10
  camera_09:
    enabled: true
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-09.local/Streaming/Channels/2
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-09.local/Streaming/Channels/1
          roles:
            - record
    live:
      streams:
        camera_09_Live: camera_09_Live
    motion:
      enabled: true
      threshold: 60
      contour_area: 20
      improve_contrast: true
      mask:
        - 0.276,1,0.275,0.894,0.128,0.898,0.156,0.651,0.149,0.475,0.195,0.295,0.297,0.177,0.35,0.137,0.409,0.103,0.425,0,0,0,0,1
        - 0.613,0,0.61,0.169,0.753,0.224,0.792,0.14,0.867,0.15,0.894,0.286,0.87,0.438,0.779,0.525,0.767,0.678,0.644,0.697,0.621,0.747,0.73,0.894,0.682,1,1,1,1,0
    objects:
      filters:
        person:
          mask:
            - 0.581,0.198,0.584,0.248,0.599,0.249,0.602,0.197
            - 0.713,0.262,0.735,0.264,0.736,0.198,0.713,0.195
        car:
          mask: 0,0,0,1,1,1,1,0
    snapshots:
      bounding_box: true
      enabled: true
      timestamp: true
      retain:
        default: 20
    zones: {}
    review: {}
    onvif:
      host: example-host-09.local
      port: 80
      user: "xxxx"
      password: "xxxx"
      autotracking:
        enabled: false
        calibrate_on_startup: false
        zooming: disabled
        zoom_factor: 0.3
        track:
          - person
        required_zones:
          - zone_07
        return_preset: Preset 1
        timeout: 10
  camera_10:
    enabled: true
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-11.local/Streaming/Channels/2
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-11.local/Streaming/Channels/1
          roles:
            - record
    live:
      streams:
        camera_10_Live: camera_10_Live
    motion:
      enabled: true
      threshold: 25
      contour_area: 10
      improve_contrast: true
      mask:
        - 0,0.94,1,0.937,1,1,0,1
        - 0,0.157,0.049,0.143,0.086,0.127,0.129,0.101,0.285,0.079,0.565,0.052,0.679,0,0,0
    objects:
      filters:
        person: {}
        car:
          mask: 0.93,0.125,0.909,0.17,0.967,0.205,0.969,0.149
    snapshots:
      bounding_box: true
      enabled: true
      timestamp: true
      retain:
        default: 20
    zones:
      zone_02:
        coordinates:
          0,0.276,0,0.156,0.108,0.111,0.292,0.075,0.563,0.06,0.68,0,1,0,1,0.452,1,0.929,0,0.937
        loitering_time: 0
        objects: person
        inertia: 3
    review: {}
    onvif:
      host: example-host-11.local
      port: 80
      user: "xxxx"
      password: "xxxx"
      autotracking:
        enabled: false
        calibrate_on_startup: false
        zooming: disabled
        zoom_factor: 0.3
        track:
          - person
        required_zones:
          - zone_02
        return_preset: Preset 1
        timeout: 10
  camera_11:
    enabled: false
    birdseye:
      enabled: false
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-12.local/Streaming/Channels/2
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-12.local/Streaming/Channels/1
          roles:
            - record
    live:
      streams:
        camera_11_Live: camera_11_Live
    motion:
      enabled: true
      threshold: 25
      contour_area: 10
      improve_contrast: true
      mask: 0,0.926,1,0.927,1,1,0,1
    objects:
      filters:
        person: {}
        car: {}
    snapshots:
      bounding_box: true
      enabled: true
      timestamp: true
      retain:
        default: 20
    zones: {}
    review: {}
    onvif:
      host: example-host-12.local
      port: 80
      user: "xxxx"
      password: "xxxx"
      autotracking:
        enabled: false
        calibrate_on_startup: false
        zooming: disabled
        zoom_factor: 0.3
        track:
          - person
        required_zones:
          - zone_02
        return_preset: Preset 1
        timeout: 10
  camera_12:
    enabled: false
    birdseye:
      enabled: false
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-13.local/Streaming/Channels/2
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-13.local/Streaming/Channels/1
          roles:
            - record
    live:
      streams:
        camera_12_Live: camera_12_Live
    motion:
      enabled: true
      threshold: 25
      contour_area: 10
      improve_contrast: true
      mask: 0,0.926,1,0.927,1,1,0,1
    objects:
      filters:
        person: {}
        car: {}
    snapshots:
      bounding_box: true
      enabled: true
      timestamp: true
      retain:
        default: 20
    zones: {}
    review: {}
    onvif:
      host: example-host-13.local
      port: 80
      user: "xxxx"
      password: "xxxx"
      autotracking:
        enabled: false
        calibrate_on_startup: false
        zooming: disabled
        zoom_factor: 0.3
        track:
          - person
        required_zones:
          - zone_02
        return_preset: Preset 1
        timeout: 10
  camera_13:
    enabled: false
    birdseye:
      enabled: false
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-14.local/Streaming/Channels/2
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-14.local/Streaming/Channels/1
          roles:
            - record
    live:
      streams:
        camera_13_Live: camera_13_Live
    motion:
      enabled: true
      threshold: 25
      contour_area: 10
      improve_contrast: true
      mask: 0,0.926,1,0.927,1,1,0,1
    objects:
      filters:
        person: {}
        car: {}
    snapshots:
      bounding_box: true
      enabled: true
      timestamp: true
      retain:
        default: 20
    zones: {}
    review: {}
    onvif:
      host: example-host-14.local
      port: 80
      user: "xxxx"
      password: "xxxx"
      autotracking:
        enabled: false
        calibrate_on_startup: false
        zooming: disabled
        zoom_factor: 0.3
        track:
          - person
        required_zones:
          - zone_02
        return_preset: Preset 1
        timeout: 10
  camera_14:
    enabled: false
    birdseye:
      enabled: false
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-15.local/Streaming/Channels/2
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-15.local/Streaming/Channels/1
          roles:
            - record
    live:
      streams:
        camera_14_Live: camera_14_Live
    motion:
      enabled: true
      threshold: 25
      contour_area: 10
      improve_contrast: true
      mask: 0,0.926,1,0.927,1,1,0,1
    objects:
      filters:
        person: {}
        car: {}
    snapshots:
      bounding_box: true
      enabled: true
      timestamp: true
      retain:
        default: 20
    zones: {}
    review: {}
    onvif:
      host: example-host-15.local
      port: 80
      user: "xxxx"
      password: "xxxx"
      autotracking:
        enabled: false
        calibrate_on_startup: false
        zooming: disabled
        zoom_factor: 0.3
        track:
          - person
        required_zones:
          - zone_02
        return_preset: Preset 1
        timeout: 10
  camera_15:
    enabled: true
    birdseye:
      enabled: false
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:10554/axis-media/media.amp?videocodec=h264&resolution=640x360&fps=5
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:10554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
          roles:
            - record
    live:
      streams:
        camera_15_Live: camera_15_Live
    motion:
      mask: 0,0.824,0.539,0.033,0.549,0.003,0,0
    objects:
      filters:
        car:
          mask:
            0.487,0.924,1,0.909,1,0.375,0.835,0.188,0.754,0.142,0.731,0.128,0.622,0.051,0.546,0.07
  camera_16:
    enabled: true
    birdseye:
      enabled: false
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:11554/axis-media/media.amp?videocodec=h264&resolution=640x360&fps=5
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:11554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
          roles:
            - record
    live:
      streams:
        camera_16_Live: camera_16_Live
    motion:
      mask:
        0.171,0,0.174,0.047,0.396,0.06,0.443,0.079,0.487,0.104,0.514,0.116,0.59,0.123,0.651,0.136,1,0.32,1,0
    objects:
      filters:
        car:
          mask: 0.133,0.201,0.342,0.881,0.705,0.604,0.305,0.097
  camera_17:
    enabled: true
    birdseye:
      enabled: false
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:12554/axis-media/media.amp?videocodec=h264&resolution=640x360&fps=5
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:12554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
          roles:
            - record
    live:
      streams:
        camera_17_Live: camera_17_Live
    motion:
      mask:
        - 0.001,0.097,0.159,0.066,0.311,0.057,0.332,0,0,0
        - 0.588,0.065,0.586,0.143,0.64,0.175,0.647,0.07,0.736,0.095,0.735,0.347,0.937,0.517,1,0.584,1,0,0.41,0,0.409,0.038
    objects:
      filters:
        car: {}
      mask: 0.496,0.265,0.818,0.57,0.85,0.6,1,0.755,1,0,0.409,0,0.402,0.188
  camera_18:
    enabled: false
    birdseye:
      enabled: false
    detect:
      enabled: true
      width: 200
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:13554/axis-media/media.amp?videocodec=h264&resolution=640x360&fps=5
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:13554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
          roles:
            - record
    live:
      streams:
        camera_17_Live: camera_18_Live
    motion:
      mask:
        - 0.775,0.203,0.757,0.706,0.996,0.932,1,0.132
        - 0.473,0,0.534,0.09,0.87,0.082,1,0.064,1,0
    objects:
      filters:
        car: {}
      mask:
        - 1,0.139,0.775,0.202,0.762,0.715,1,0.936
        - 0.251,0.265,0,0.488,0,0,0.698,0,0.662,0.157
  camera_19:
    enabled: false
    birdseye:
      enabled: false
    detect:
      enabled: true
      width: 200
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:14554/axis-media/media.amp?videocodec=h264&resolution=640x360&fps=5
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:14554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
          roles:
            - record
    live:
      streams:
        camera_17_Live: camera_19_Live
    motion:
      mask: 0,0.749,0.105,0.345,0.148,0.216,0.238,0.087,0.317,0,0,0
    objects:
      filters:
        car: {}
      mask:
        - -0.001,0.737,0.08,0.427,0.152,0.208,0.236,0.082,0.333,0,0,0
        - 0.495,0.092,0.947,0.73,1,0.866,1,0,0.617,0,0.588,0.075
  camera_20:
    enabled: false
    birdseye:
      enabled: false
    detect:
      enabled: true
      width: 200
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:15554/axis-media/media.amp?videocodec=h264&resolution=640x360&fps=5
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:15554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
          roles:
            - record
    live:
      streams:
        camera_17_Live: camera_20_Live
    motion:
      mask: 0.725,0.134,0.807,0.358,0.93,1,1,1,1,0,0.725,0
    objects:
      filters:
        car: {}
      mask:
        - 0.693,0.123,0.776,0.346,0.936,1,1,1,1,0,0.716,0
        - 0.008,0.63,0.578,0.132,0.573,0,0,0
  camera_21:
    enabled: true
    birdseye:
      enabled: false
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:16554/axis-media/media.amp?videocodec=h264&resolution=640x360&fps=5
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:16554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
          roles:
            - record
    live:
      streams:
        camera_17_Live: camera_21_Live
    motion:
      mask:
        - 0.494,0.048,0.422,0.196,0.282,0.486,0.158,0.857,0.1,1,0,1,0,0,0.828,0,0.817,0.047
        - 0.691,0.24,0.789,0.729,1,0.734,1,0.505,1,0.248
    objects:
      filters:
        car:
          mask:
            - 0.691,0.231,0.786,0.729,1,0.738,1,0.244
            - 0.835,0.047,0.498,0.051,0.39,0.281,0.283,0.503,0.165,0.858,0.1,1,0,1,0,0,0.844,0
            - 0,0,1,0,1,0.141,0.582,0.141,0.463,1,0,1
      mask:
        - 0.691,0.231,0.786,0.729,1,0.738,1,0.244
        - 0.835,0.047,0.498,0.051,0.39,0.281,0.283,0.503,0.165,0.858,0.1,1,0,1,0,0,0.844,0
  camera_22:
    enabled: true
    birdseye:
      enabled: false
    detect:
      enabled: true
      width: 200
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:17554/axis-media/media.amp?videocodec=h264&resolution=640x360&fps=5
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:17554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
          roles:
            - record
    live:
      streams:
        camera_17_Live: camera_22_Live
    motion:
      mask:
        0,0.635,0.053,0.635,0.141,0.415,0.32,0.122,0.342,0.08,0.456,0.064,0.47,0,0,0
    objects:
      filters:
        car: {}
      mask:
        - 0,0,0.478,0,0.456,0.061,0.363,0.093,0.324,0.173,0.208,0.515,0.129,0.751,0,0.75
        - 0.293,0.154,0.497,0.146,0.714,0.276,0.905,0.346,0.988,0.357,1,0.147,0.492,0.064,0.324,0.084
  camera_23:
    enabled: false
    birdseye:
      enabled: false
    detect:
      enabled: false
      width: 200
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:18554/axis-media/media.amp?videocodec=h264&resolution=640x360&fps=5
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:18554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
          roles:
            - record
    live:
      streams:
        camera_17_Live: camera_23_Live
    motion: {}
    objects:
      filters:
        car: {}
  camera_24:
    enabled: true
    birdseye:
      enabled: false
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:19554/axis-media/media.amp?videocodec=h264&resolution=640x360&fps=5
          roles:
            - detect
        - path:
            rtsp://xxxx:xxxx@example-host-16.local:19554/axis-media/media.amp?videocodec=h264&resolution=1920x1080&fps=15
          roles:
            - record
    live:
      streams:
        camera_17_Live: camera_24_Live
    motion:
      mask:
        0,0,0.215,0,0.225,0.067,0.29,0.058,0.452,0.506,0.63,0.492,0.936,1,0.374,1,0,1
    objects:
      filters:
        car: {}
      mask:
        0,0,0.208,0,0.221,0.081,0.288,0.071,0.452,0.495,0.627,0.477,0.953,1,0,1
  camera_25:
    enabled: false
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path: /config/restream_test_15.mp4
          input_args: -re -stream_loop -1 -fflags +genpts
          roles:
            - detect
    motion:
      enabled: true
      threshold: 15
      contour_area: 8
      improve_contrast: true
      mask: 0.002,0.94,1,0.934,1,1,0,1
    objects:
      filters:
        person: {}
        car: {}
    snapshots:
      bounding_box: true
      enabled: true
      timestamp: true
      retain:
        default: 20
    zones:
      zone_02:
        coordinates:
          0,0.075,0.124,0.013,0.395,0.019,0.649,0.031,0.849,0.065,1,0.091,1,0.193,1,0.452,1,0.929,0,0.937
        loitering_time: 0
        objects: person
        inertia: 3
    review: {}
  camera_26:
    enabled: false
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path: /config/restream_test_18.mp4
          input_args: -re -stream_loop -1 -fflags +genpts
          roles:
            - detect
    motion:
      enabled: true
      threshold: 15
      contour_area: 8
      improve_contrast: true
      mask: 0,0.941,1,0.94,1,1,0,1
    objects:
      filters:
        person: {}
        car: {}
      mask: 0,0.163,0.176,0.165,0.185,0.101,0.31,0.126,0.339,0,0.003,0.002
    snapshots:
      bounding_box: true
      enabled: true
      timestamp: true
      retain:
        default: 20
    zones: {}
    review:
      alerts: {}
      detections: {}
  camera_27:
    enabled: false
    birdseye:
      enabled: false
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path: /config/restream_test.mp4
          input_args: -re -stream_loop -1 -fflags +genpts
          roles:
            - detect
    motion:
      enabled: true
      threshold: 60
      contour_area: 20
      improve_contrast: true
      mask:
        - 0.329,1,0.304,0.657,0.28,0.467,0.331,0.377,0.402,0.253,0.444,0.187,0.474,0.136,0.52,0.108,0.545,0.092,0.545,0,0,0,0,1
        - 0.907,0,0.88,0.16,0.894,0.286,0.87,0.438,0.793,0.585,0.752,0.601,0.734,0.672,0.756,0.79,0.731,0.864,0.734,1,1,1,1,0
    objects:
      filters:
        person: {}
        car:
          mask: 0,0,0,1,1,1,1,0
    zones: {}
    review: {}
  camera_28:
    enabled: false
    detect:
      enabled: true
      width: 640
      height: 360
      fps: 6
    ffmpeg:
      hwaccel_args: preset-nvidia
      inputs:
        - path: /config/restream_test_25.mp4
          input_args: -re -stream_loop -1 -fflags +genpts
          roles:
            - detect
    motion:
      enabled: true
      threshold: 25
      contour_area: 10
      improve_contrast: true
      mask:
        - 0,0.94,1,0.937,1,1,0,1
        - 0,0.157,0.049,0.143,0.086,0.127,0.129,0.101,0.285,0.079,0.565,0.052,0.679,0,0,0
    objects:
      filters:
        person: {}
        car: {}
    snapshots:
      bounding_box: true
      enabled: true
      timestamp: true
      retain:
        default: 20
    zones:
      zone_02:
        coordinates:
          0,0.276,0,0.156,0.108,0.111,0.292,0.075,0.563,0.06,0.68,0,1,0,1,0.452,1,0.929,0,0.937
        loitering_time: 0
        objects: person
        inertia: 3
    review: {}
camera_groups:
  group_01:
    order: 2
    icon: LuActivity
    cameras:
      - camera_02
      - camera_03
      - camera_04
      - camera_05
      - camera_06
      - camera_08
      - camera_03
      - camera_09
      - camera_01
      - camera_10
      - camera_07
      - camera_12
  group_02:
    order: 3
    icon: LuBadgeAlert
    cameras:
      - camera_15
      - camera_16
      - camera_17
      - camera_18
      - camera_19
      - camera_20
      - camera_21
      - camera_22
      - camera_23
      - camera_24
  group_03:
    order: 1
    icon: LuCat
    cameras: birdseye
  group_04:
    order: 4
    icon: LuDribbble
    cameras:
      - camera_01
      - camera_02
      - camera_03
      - camera_04
      - camera_05
      - camera_06
      - camera_07
      - camera_08
      - camera_09
      - camera_10
version: 0.17-2
detect:
  enabled: true

docker-compose file or Docker CLI command

services:
  frigate:
    container_name: frigate0172stable
    privileged: true
    restart: unless-stopped
    image: ghcr.io/blakeblackshear/frigate:0.17.2-tensorrt
    shm_size: "512mb"
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    volumes:
      - /etc/localtime:/etc/localtime:ro
      - /DOCKER/frigate/config_v0.17.2:/config
      - /datastore/frigate_media:/media/frigate
      - type: tmpfs
        target: /tmp/cache
        tmpfs:
          size: 1000000000
    ports:
      - "8971:8971"
      - "5000:5000"
      - "8554:8554"
      - "8555:8555/tcp"
      - "8555:8555/udp"
    environment:
      FRIGATE_MQTT_USER: "xxxx"
      FRIGATE_MQTT_PASSWORD: "xxxx"
      FRIGATE_RTSP_USER: "xxxx"
      FRIGATE_RTSP_PASSWORD: "xxxx"
      FRIGATE_RTSP_P1USER: "xxxx"
      FRIGATE_RTSP_P1PASSWORD: "xxxx"

Relevant Frigate log output

CUDA Graph configuration the ONNX detector repeatedly became stuck

Install method

Docker Compose

Object Detector

TensorRT

Screenshots of the Frigate UI's System metrics pages

no screens

Any other information that may be helpful

No response

[NickM-27]

Frigate 0.17 technically has no support for blackwell. 0.18 should work fully with it.

That said, the 14700K iGPU should be able to fully handle all of Frigates needs, just need to setup multiple OpenVINO detectors.

Personally, I would use that and leave the 5070Ti to run an LLM to be used with the Frigate AI chat and review summaries.

[ThorstenUm]

Thanks a lot for the clarification and for taking the time to help!

Just one small correction regarding my hardware: I am using an Intel i7-14700KF, not the 14700K. So unfortunately this CPU does not have an Intel iGPU, which means the OpenVINO/iGPU solution you suggested isn't available on this particular system.

#21559
Regarding the RTX 5070 Ti: I had understood from an earlier discussion (#21559), before buying the new hardware, that NVIDIA 40 and 50 series GPUs would generally be suitable for Frigate. That's one of the reasons I chose the 5070 Ti. I now understand from your explanation that Blackwell is not technically supported by Frigate 0.17 and that proper support is coming with 0.18. No problem at all – I just wanted to explain why I ended up with this particular hardware combination.

Interestingly, I may have found a usable workaround for 0.17.2:
I'm currently running YOLOv9-S 320x320 via ONNX/CUDA on the RTX 5070 Ti. CUDA itself works and frigate.detector:onnx is running on the GPU. Inference time is around 10 ms.

With the default configuration, however, the detector repeatedly became stuck and Frigate restarted the detection process (Detection appears to be stuck). After looking through the 0.17.2 code, I found that yolo-generic uses the CudaGraphRunner and enables CUDA Graphs.
As a test I changed:
enable_cuda_graph: True
to:
enable_cuda_graph: False
in detection_runners.py.

Since doing that, the detector appears to be running normally so far. CUDA is still being used for detection, so this only seems to bypass the CUDA Graph path.

Would you consider disabling CUDA Graphs a reasonable temporary workaround for running a Blackwell GPU on 0.17.2 until 0.18 is stable? Or are there other known Blackwell incompatibilities in 0.17 that would make you advise against using this setup in production even if it proves stable?

I still have the two Coral USB TPUs from my previous Frigate system, so another option would be to use those for detection temporarily. I could also simply continue running the old Frigate machine and keep this new system as a backup until 0.18 is ready.
I would also really appreciate your opinion on the rest of the setup. For this machine (i7-14700KF + RTX 5070 Ti 16 GB, around 20–25 cameras), does the following generally look sensible to you for 0.17.2?
+0.17.2-tensorrt image
+NVIDIA hardware decoding (preset-nvidia)
+ONNX detector using CUDA
+YOLOv9-S, 320x320, yolo-generic
+CUDA Graphs disabled as described above
Or would you change any of these choices while waiting for 0.18?

And finally, a huge thank you to you and everyone working on Frigate. I really appreciate the amount of work that goes into a project like this, especially supporting so many different hardware platforms, accelerators and camera setups. It's easy for us as users to ask questions and report problems, but we shouldn't forget that we can be very grateful that you guys develop and maintain all of this in the first place. Thank you!

[NickM-27]

It's fine if that's what works for you, CUDA Graphs do make the GPU run faster and more consistent especially with lots of cameras. 0.18 is going to RC soon so generally I'd just recommend jumping onto beta3

[ThorstenUm]

Thanks, that's exactly the information I needed.
I'll keep the 0.17.2 setup with CUDA Graphs disabled for a little while to verify that it remains stable, but given that 0.18 is already approaching RC, I'll also start testing beta3 on this machine rather than spending too much time working around Blackwell limitations in 0.17.
Thanks again for the quick help and clarification – very much appreciated!