[HW Accel Support]: Move from Intel iGPU to dual Quadro P200

[kavejo]

Describe the problem you are having

I am in the process of changing the hardware I am using for Frigate as I'd love to pair it with Double Take.

Currently I have a Lenovo M900 with i5-6500T having a built-in Intel HD Graphics 530 which with 11 Full HD streams (10 are h.246 and 1 is h.265) is practically maxed out. Detection is done by 2 Coral USB.

I've got hold of a Lenovo P520 with a Xeon CPU, more RAM, and 2 Nvidia Quadro P2000 built-in which I am planning to move Frigate to. I'd love to use both GPU for h/w decoding and for object detection, this last task paired with the Coral USBs.

I've read around and found https://docs.frigate.video/configuration/hardware_acceleration/ which suggest I need to modify the docker compose file and add the "deploy" portion, but how would I pass 2 Quadro? Would I need to increase the "count" or add more "device_ids"?

As I had issues with h.265 streams with the current hardware, would I be able to use h.265 streams (that are 4K instead of 1080p) in the new hardware?

Would I still need to use go2rtc? Or would it be unnecessary at that point?

Also, is there a way to simplify the config file as I feel there might be parts that are redundant but I am not entorely sure.

Thanks!

Version

0.14.1-f4f3cfa

Frigate config file

# yaml-language-server: $schema=http://10.1.4.2:5000/api/config/schema.json
# Stream 0: 2688x1520 h.265
# Stream 1: 704x576 mpeg
# Stream 2: 1920x1080 h.264

logger:
  default: warning

mqtt:
  host: x.x.x.x
  port: xxxx
  user: xxxxx
  password: xxxx

detectors:
  coral-right:
    type: edgetpu
    device: usb:0
  coral-left:
    type: edgetpu
    device: usb:1

objects:
  track:
    - person
    - car
    - motorcycle
  filters:
    person:
      min_score: 0.6
      threshold: 0.8
    car:
      min_score: 0.6
      threshold: 0.7
    motorcycle:
      min_score: 0.5
      threshold: 0.7

review:
  alerts:
    labels:
      - car
      - person
      - motorcycle
  detections:
    labels:
      - car
      - person
      - motorcycle

record:
  enabled: true
  retain:
    days: 10
    mode: all
  events:
    pre_capture: 5
    post_capture: 5
    retain:
      default: 30
      mode: active_objects

snapshots:
  enabled: true
  clean_copy: false
  timestamp: false
  bounding_box: true
  crop: false
  retain:
    default: 14

birdseye:
  enabled: false

live:
  height: 1080
  quality: 8

ui:
  timezone: Europe/Rome

detect:
  width: 1920
  height: 1080
  fps: 5
  enabled: true

audio:
  enabled: false
  listen:
    - bark
    - chatter
    - dog
    - explosion
    - scream
    - speech
    - yell

go2rtc:
  streams:
    01:
      - rtsp://USER:PASS@IP:PORT/cam/realmonitor?channel=1&subtype=2
      - ffmpeg:01#video=h264#hardware
    02:
      - rtsp://USER:PASS@IP:PORT/cam/realmonitor?channel=2&subtype=2
      - ffmpeg:02#video=h264#hardware
    03:
      - rtsp://USER:PASS@IP:PORT/cam/realmonitor?channel=3&subtype=2
      - ffmpeg:03#video=h264#hardware
    04:
      - rtsp://USER:PASS@IP:PORT/cam/realmonitor?channel=4&subtype=2
      - ffmpeg:04#video=h264#hardware
    05:
      - rtsp://USER:PASS@IP:PORT/cam/realmonitor?channel=5&subtype=2
      - ffmpeg:05#video=h264#hardware
    06:
      - rtsp://USER:PASS@IP:PORT/cam/realmonitor?channel=6&subtype=2
      - ffmpeg:06#video=h264#hardware
    07:
      - rtsp://USER:PASS@IP:PORT/cam/realmonitor?channel=7&subtype=2
      - ffmpeg:07#video=h264#hardware
    08:
      - rtsp://USER:PASS@IP:PORT/cam/realmonitor?channel=8&subtype=2
      - ffmpeg:08#video=h264#hardware
    09:
      - rtsp://USER:PASS@IP:PORT/cam/realmonitor?channel=9&subtype=2
      - ffmpeg:09#video=h264#hardware
    10:
      - rtsp://USER:PASS@IP:PORT/cam/realmonitor?channel=10&subtype=2
      - ffmpeg:10#video=h264#hardware
    11:
      - rtsp://USER:PASS@IP:PORT/cam/realmonitor?channel=11&subtype=1
      - ffmpeg:11#video=h264#hardware
  webrtc:
    candidates:
      - x.x.x.x:8555

cameras:

  01:
    ffmpeg:
      inputs:
        - path: rtsp://LOCAL:PORT/01
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
      output_args:
        record: preset-record-generic
    detect:
      width: 1920
      height: 1080
      fps: 5
    record:
      enabled: true
    mqtt:
      enabled: true
      timestamp: false
      bounding_box: false
      crop: false

  02:
    ffmpeg:
      inputs:
        - path: rtsp://LOCAL:PORT/02
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
      output_args:
        record: preset-record-generic
    detect:
      width: 1920
      height: 1080
      fps: 5
    record:
      enabled: true
    mqtt:
      enabled: true
      timestamp: false
      bounding_box: false
      crop: false

  03:
    ffmpeg:
      inputs:
        - path: rtsp://LOCAL:PORT/03
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
      output_args:
        record: preset-record-generic
    detect:
      width: 1920
      height: 1080
      fps: 5
    record:
      enabled: true
    mqtt:
      enabled: true
      timestamp: false
      bounding_box: false
      crop: false

  04:
    ffmpeg:
      inputs:
        - path: rtsp://LOCAL:PORT/04
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
      output_args:
        record: preset-record-generic
    detect:
      width: 1920
      height: 1080
      fps: 5
    record:
      enabled: true
    mqtt:
      enabled: true
      timestamp: false
      bounding_box: false
      crop: false

  05:
    ffmpeg:
      inputs:
        - path: rtsp://LOCAL:PORT/05
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
      output_args:
        record: preset-record-generic
    detect:
      width: 1920
      height: 1080
      fps: 5
    record:
      enabled: true
    mqtt:
      enabled: true
      timestamp: false
      bounding_box: false
      crop: false

  06:
    ffmpeg:
      inputs:
        - path: rtsp://LOCAL:PORT/06
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
      output_args:
        record: preset-record-generic
    detect:
      width: 1920
      height: 1080
      fps: 5
    record:
      enabled: true
    mqtt:
      enabled: true
      timestamp: false
      bounding_box: false
      crop: false

  07:
    ffmpeg:
      inputs:
        - path: rtsp://LOCAL:PORT/07
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
      output_args:
        record: preset-record-generic
    detect:
      width: 1920
      height: 1080
      fps: 5
    record:
      enabled: true
    mqtt:
      enabled: true
      timestamp: false
      bounding_box: false
      crop: false

  08:
    ffmpeg:
      inputs:
        - path: rtsp://LOCAL:PORT/08
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
      output_args:
        record: preset-record-generic
    detect:
      width: 1920
      height: 1080
      fps: 5
    record:
      enabled: true
    mqtt:
      enabled: true
      timestamp: false
      bounding_box: false
      crop: false

  09:
    ffmpeg:
      inputs:
        - path: rtsp://LOCAL:PORT/09
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
      output_args:
        record: preset-record-generic
    detect:
      width: 1920
      height: 1080
      fps: 5
    record:
      enabled: true
    mqtt:
      enabled: true
      timestamp: false
      bounding_box: false
      crop: false

  10:
    ffmpeg:
      inputs:
        - path: rtsp://LOCAL:PORT/10
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
      output_args:
        record: preset-record-generic
    detect:
      width: 1920
      height: 1080
      fps: 5
    record:
      enabled: true
    mqtt:
      enabled: true
      timestamp: false
      bounding_box: false
      crop: false

  11:
    ffmpeg:
      inputs:
        - path: rtsp://LOCAL:PORT/11
          input_args: preset-rtsp-restream
          roles:
            - detect
            - record
      output_args:
        record: preset-record-generic
    detect:
      width: 1920
      height: 1080
      fps: 5
    record:
      enabled: true
    mqtt:
      enabled: true
      timestamp: false
      bounding_box: false
      crop: false

telemetry:
  network_interfaces:
    - eth
    - enp
    - eno
    - ens
  stats:
    intel_gpu_stats: true
    network_bandwidth: true
  version_check: true
version: 0.14

docker-compose file or Docker CLI command

version: "3.9"
services:

  frigate:
    container_name: frigate
    privileged: true
    restart: always
    image: ghcr.io/blakeblackshear/frigate:stable
    shm_size: "1024mb"

    devices:
      - /dev/bus/usb:/dev/bus/usb
      - /dev/dri/renderD128

    volumes:
      - /etc/localtime:/etc/localtime:ro
      - /mnt/CONFIG/Frigate/config.yml:/config/config.yml
      - /mnt/USB/Friguate/config:/config
      - /mnt/USB/Frigate/media:/media/frigate
      - type: tmpfs
        target: /tmp/cache
        tmpfs:
          size: 1000000000

    ports:
      - "5000:5000" # Web UI
      - "1935:1935" # RTMP
      - "8554:8554" # RTSP
      - "8555:8555/tcp" # WebRTC over TCP
      - "8555:8555/udp" # WebRTC over UDP

Relevant Frigate log output

N/A

Relevant go2rtc log output

N/A

FFprobe output from your camera

N/A

Install method

Docker Compose

Object Detector

Coral

Network connection

Wired

Camera make and model

Dahua

Screenshots of the Frigate UI's System metrics pages

N/A

Any other information that may be helpful

No response

[NickM-27]

1. You can't run corals and Tensortt for detection at the same time
2. Yes just set the count to 2
3. This change wouldn't affect anything regarding h265 compatibility, that is client dependent
4. Go2rtc is always necessary for the best live view experience

[kavejo]

Thanks @NickM-27.

1. In such case would I be best running Coral or Tensor? Is it just a matter to test and verify which one has a higher inference speed and lower false positives?

2. Ok, so something like the below? Is there any other capability that need to be added?

    deploy:    # <------------- Add this section
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids: ['0','0'] # this is only needed when using multiple GPUs
              count: 2 # number of GPUs
              capabilities: [gpu]

3. Understood - So would I be better using h.265 for record/detect as the stream has a higher resolution and decode it to h.264 for the viewing on the clients (as for example notifications through HA were not viewable in iOS when the notification came with a h.265 feed)?

4. How would I best configure go2rtc for this? Would I just need to change the string "ffmpeg:01#video=h264#hardware"? Is it then best to do detect/record on the go2rtc stream or the source stream?

THanks!

[NickM-27]

1. That depends on you, the coral would be more power efficient but can't run large models like the GPU can
2. You should comment out the device_ids part
3. Not sure I know where you would be decoding the stream
4. I don't think it makes sense to transcode to h264 personally, provided the client supports h265 which most do