[HW Accel Support]: GPU HW accel vs CPU

[ryanmsteed]

Describe the problem you are having

Hey team,
I guess my first question is: is CPU HW acceleration preferred for decoding (VPP) vs NVDEC? My understanding is VPP is more efficient from a power consumption standpoint. If so, it appears my GPU is doing the decoding and I'm not sure how to fix it. I tried to use the vaapi hwaccel_args under the camera itself but it appears the GPU is still being used according to the nvtop console. Sorry, i'm only a week into my Frigate experience here..

Version

0.15.0-6cb5cfb

Frigate config file

mqtt:
  enabled: false

detectors:
  tensorrt:
    type: tensorrt
    device: 0

ffmpeg:
  hwaccel_args: preset-nvidia

model:
  path: /config/model_cache/tensorrt/yolov7-tiny-416.trt
  labelmap_path: /labelmap/coco-80.txt
  input_tensor: nchw
  input_pixel_format: rgb
  width: 416
  height: 416

go2rtc:
  streams:
    rear_entry:
      - rtsp://login:ps@IP/h264Preview_01_main

cameras:
  rear_entry: # <------ Name the camera
    enabled: true
    ffmpeg:
      inputs:
        - path: rtsp://login:ps@IP # <----- The stream you want to use for detection
          roles:
            - detect
        - path: rtsp://login:ps@IP # <----- The stream you want to use for detection
          hwaccel_args: preset-vaapi
          roles:
            - record
    detect:
      enabled: true # <---- disable detection until you have a working camera feed
      width: 1280
      height: 720
    record:
      enabled: true
      retain:
        days: 5
      alerts:
        retain:
          days: 5
      detections:
        retain:
          days: 5
    motion:
      mask: 0,0.059,0.274,0.107,0.24,0.887,0,0.897
      threshold: 30
      contour_area: 10
      improve_contrast: true

docker-compose file or Docker CLI command

services:
  frigate:
    container_name: frigate
    privileged: false # this may not be necessary for all setups
    restart: unless-stopped
    stop_grace_period: 30s # allow enough time to shut down the various services
    image: ghcr.io/blakeblackshear/frigate:stable-tensorrt
    shm_size: 1024mb # update for your cameras based on calculation above
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities:
                - gpu
    volumes:
      - /etc/localtime:/etc/localtime:ro
      - ./config:/config
      - /mnt/nfs_cam:/media/frigate
      - type: tmpfs
        target: /tmp/cache
        tmpfs:
          size: 1000000000
    ports:
      - 8971:8971
      - 5000:5000 # Internal unauthenticated access. Expose carefully.
      - 8554:8554 # RTSP feeds
      - 8555:8555/tcp # WebRTC over tcp
      - 8555:8555/udp # WebRTC over udp
    environment:
      - YOLO_MODELS=yolov7-tiny-416
      - USE_FP16=false
networks: {}

Relevant Frigate log output

NA

Relevant go2rtc log output

NA

FFprobe output from your camera

NA

Install method

Proxmox via Docker

Object Detector

TensorRT

Network connection

Wired

Camera make and model

NA

Screenshots of the Frigate UI's System metrics pages

Any other information that may be helpful

No response

[inkeep-ai-assistant[bot]]

Frigate supports hardware acceleration for video decoding using the GPU, which is highly recommended to reduce CPU usage(1). The preset-nvidia is intended for NVIDIA GPUs (using NVDEC), while preset-vaapi is for Intel/AMD GPUs (using VAAPI)(1). There is no documentation suggesting that CPU-based hardware acceleration (like VPP) is preferred or more efficient than GPU-based decoding—Frigate is designed to offload video decoding to the GPU, not the CPU, for improved performance and lower power consumption(1).

If you're seeing GPU usage for decoding with preset-nvidia, that's expected for NVIDIA hardware. Setting hwaccel_args: preset-vaapi under a camera will only work if your hardware and drivers support VAAPI; otherwise, the GPU (with NVDEC) will continue to be used(1).

To summarize:

• GPU hardware acceleration (NVDEC for NVIDIA, VAAPI for Intel/AMD) is the recommended and supported approach in Frigate(1).
• There is no Frigate-supported method to force CPU-based hardware acceleration (VPP) for decoding.
• Your configuration is correct for NVIDIA hardware; no further action is needed unless you switch to Intel/AMD hardware, in which case you would use preset-vaapi(1).

You can review more details at Video Decoding.

📚 Sources:

1. Video Decoding

---

Have another question? Just tag @inkeep.

[NickM-27]

It's unlikely that your camera is VPP and in general using a GPU is highly recommended

[ryanmsteed]

OK sounds like i'm doing the right thing. thanks for the feedback, humans and AI!