[HW Accel Support]: Power Draw of new GPU too high due to FFMPEG.

[bananaboy1990]

Describe the problem you are having

Power Draw of new GPU too high due to FFMPEG.

Greetings Everyone.

Im finally switched from my old GTX1070 to a RTX 3090.
After the switch i realized that the new card is drawing much more power than the old one for the same operations.
130Watts of power while only decoding 2 rstp streams seems to much. My old GTX1070 used only round about 30 Watts for the same operation(as far as i remember). I've tried to stop my other docker containers using this GPU but with no luck. Finally i saw that stopping the GPU decoding reduced the power draw to the expected 30 Watts. This doesn't make sense for me at all.

Imma pretty experienced linux administrator but not experienced at all about that FFMPEG thing so i need ur help to reduce my monthy power bill at least a little bit :-)

Version

0.16.0-f141b58

Frigate config file

mqtt:
  enabled: true
  host: x
  port: 1883
  user: x
  password: x
  topic_prefix: x
  client_id: fx
  stats_interval: 20
face_recognition:
  enabled: true
  model_size: large
ffmpeg:
  hwaccel_args: preset-nvidia

cameras:
  cam_2:
    enabled: true
    ffmpeg:
      inputs:
        - path: rtsp://x.x.x.x:554/stream1  # WQHD Stream
          roles:
            - record

    zones:
      Private_area:
        coordinates: 1,1,1,0,0.563,0.028,0.003,0.524,0.001,1
        loitering_time: 0
  cam_1:
    enabled: true
    ffmpeg:
      inputs:
        - path: rtsp://x.x.x.x:554/stream1  # WQHD Stream
          roles:
            - record
        - path: rtsp://x.x.x.x:554/stream2  # Lowres Stream
          roles:
            - detect

    detect:
      enabled: true #
      width: 704
      height: 576
      fps: 5
    objects:
      track:
        - person
        - face
    snapshots:
      enabled: true


version: 0.16-0
detectors:
  tensorrt:
    type: tensorrt
    device: 0 
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
detect:
  enabled: true
semantic_search:
  enabled: false
  model_size: small

docker-compose file or Docker CLI command

frigate:
    container_name: frigate
    restart: unless-stopped
    stop_grace_period: 30s
    image: ghcr.io/blakeblackshear/frigate:0.16.0-beta3-tensorrt
    shm_size: "512mb"
    volumes:
      - /etc/localtime:/etc/localtime:ro
      - ./frigate-data/config/:/config
      - ./frigate-data/storage/:/media/frigate
      - type: tmpfs 
        target: /tmp/cache
        tmpfs:
          size: 1000000000
    ports:
      - "5000:5000"
      - "8971:8971"
      - "8554:8554" # RTSP feeds
    networks:
      - net-x
    environment:
      - YOLO_MODELS=yolov7-320,yolov7x-640
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1 
              capabilities: [gpu]

Relevant Frigate log output

No errors

Relevant go2rtc log output

No errors

FFprobe output from your camera

No ffprobe installed inside container and this Container is in a special hardened environment with no internet.

Install method

Docker Compose

Object Detector

TensorRT

Network connection

Wired

Camera make and model

reolink

Screenshots of the Frigate UI's System metrics pages

Any other information that may be helpful

No response

[inkeep-ai-assistant[bot]]

High power draw when using an NVIDIA GPU with Frigate for decoding is typically related to how the GPU is being used by FFmpeg and the underlying driver/runtime configuration. Using the preset-nvidia hardware acceleration argument is the recommended method for NVIDIA GPUs, as it allows FFmpeg to select the appropriate decoding profile and should provide efficient offloading for H.264 and H.265 streams(1).

However, some NVIDIA cards—especially high-end models like the RTX 3090—may not downclock as aggressively when running FFmpeg hardware decode, even at low utilization. This is a hardware/driver behavior and not something Frigate itself can control. It's also important to make sure you are using a supported and compatible NVIDIA driver version, as mismatches can result in suboptimal power/performance(1).

You can review NVIDIA's GPU support matrix and driver compatibility here:
NVIDIA Video Encode and Decode GPU Support Matrix
FFmpeg NVIDIA Codec Headers Info

Additionally, the Frigate documentation suggests using the NVIDIA Container Toolkit and specifying the GPU for Docker, as you have, which is the correct approach(1).

If you want to experiment further, try the following:

• Reduce the number of streams or the resolution/framerate, which may allow the GPU to enter lower power states.
• Monitor the nvidia-smi output while Frigate is running to see if decoder utilization is low but power usage remains high—this confirms it's a hardware/driver power management issue(1).
• If you have other NVIDIA tools available, check for any processes that may be keeping the GPU in a high-power state.

Ultimately, some high-end GPUs like the RTX 3090 may not be as power efficient for light video decode workloads as older or smaller GPUs.

Frigate is using the correct configuration for NVIDIA hardware acceleration:

ffmpeg:
  hwaccel_args: preset-nvidia

(1)

For more details: Video Decoding

📚 Sources:

1. Video Decoding

---

Have another question? Just tag @inkeep.

[NickM-27]

This is normal, just because they're doing the same operation doesn't mean the power draw will be the same. The 3090 has entirely different architecture, much more RAM, and other components that require powering.

[NickM-27]

I personally manually downclocked my 3070 GPU and memory to drop the power state to P3 and reduce power usage a bit

[bananaboy1990]

Thanks alot for all the feedback. I wil try to mess with the driver.

[bananaboy1990]

I've found this:
https://github.com/sasha0552/nvidia-pstated
The deamon manually sets the P States according to GPU Load.
Got it down to 38Watts with a Power Target of 200W and this attributes in systemd script:

--performance-state low --utilization-threshold 10 --iterations-before-switch 1

According to my understanding it should switch from P8 State not before more than 10 % GPU Load and it will do the switch within the first loop.

[NickM-27]

Nice! Glad you got it sorted

[bananaboy1990]

Yeah dude!!! im happy as fuck :-) Maybe it will also help you to get your 3070 even more chilled

[NickM-27]

Unfortunately I run on UnRaid and some of these utilities are not available for running, but I am sure others will find it very useful

[bananaboy1990]

Its also boosting to max Clock when using my Gen AI.

[bananaboy1990]

Im not that familiar with unraid (im on the dark side with Broadcom ESXi now). but its still linux based. The binary should work if there is a nvidia driver, cuda and the 2 mentioned libs installed.