4.7 KiB
nVidia hardware decoder (NVDEC)
Certain nvidia cards include a hardware decoder, which can greatly improve the performance of video decoding. In order to use NVDEC, a special build of ffmpeg with NVDEC support is required. The special docker architecture 'amd64nvidia' includes this support for amd64 platforms. An aarch64 for the Jetson, which also includes NVDEC may be added in the future.
Docker setup
Requirements
nVidia closed source driver required to access NVDEC. nvidia-docker required to pass NVDEC to docker.
Setting up docker-compose
In order to pass NVDEC, the docker engine must be set to nvidia
and the environment variables
NVIDIA_VISIBLE_DEVICES=all
and NVIDIA_DRIVER_CAPABILITIES=compute,utility,video
must be set.
In a docker compose file, these lines need to be set:
services:
frigate:
...
image: blakeblackshear/frigate:stable-amd64nvidia
runtime: nvidia
environment:
- NVIDIA_VISIBLE_DEVICES=all
- NVIDIA_DRIVER_CAPABILITIES=compute,utility,video
Setting up the configuration file
In your frigate config.yml, you'll need to set ffmpeg to use the hardware decoder. The decoder you choose will depend on the input video.
A list of supported codecs (you can use ffmpeg -decoders | grep cuvid
in the container to get a list)
V..... h263_cuvid Nvidia CUVID H263 decoder (codec h263)
V..... h264_cuvid Nvidia CUVID H264 decoder (codec h264)
V..... hevc_cuvid Nvidia CUVID HEVC decoder (codec hevc)
V..... mjpeg_cuvid Nvidia CUVID MJPEG decoder (codec mjpeg)
V..... mpeg1_cuvid Nvidia CUVID MPEG1VIDEO decoder (codec mpeg1video)
V..... mpeg2_cuvid Nvidia CUVID MPEG2VIDEO decoder (codec mpeg2video)
V..... mpeg4_cuvid Nvidia CUVID MPEG4 decoder (codec mpeg4)
V..... vc1_cuvid Nvidia CUVID VC1 decoder (codec vc1)
V..... vp8_cuvid Nvidia CUVID VP8 decoder (codec vp8)
V..... vp9_cuvid Nvidia CUVID VP9 decoder (codec vp9)
For example, for H265 video (hevc), you'll select hevc_cuvid
. Add
-c:v hevc_covid
to your ffmpeg input arguments:
ffmpeg:
input_args:
...
- -c:v
- hevc_cuvid
If everything is working correctly, you should see a significant improvement in performance.
Verify that hardware decoding is working by running nvidia-smi
, which should show the ffmpeg
processes:
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 455.38 Driver Version: 455.38 CUDA Version: 11.1 |
|-------------------------------+----------------------+----------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 GeForce GTX 166... Off | 00000000:03:00.0 Off | N/A |
| 38% 41C P2 36W / 125W | 2082MiB / 5942MiB | 5% Default |
| | | N/A |
+-------------------------------+----------------------+----------------------+
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| 0 N/A N/A 12737 C ffmpeg 249MiB |
| 0 N/A N/A 12751 C ffmpeg 249MiB |
| 0 N/A N/A 12772 C ffmpeg 249MiB |
| 0 N/A N/A 12775 C ffmpeg 249MiB |
| 0 N/A N/A 12800 C ffmpeg 249MiB |
| 0 N/A N/A 12811 C ffmpeg 417MiB |
| 0 N/A N/A 12827 C ffmpeg 417MiB |
+-----------------------------------------------------------------------------+
To further improve performance, you can set ffmpeg to skip frames in the output, using the fps filter:
output_args:
- -filter:v
- fps=fps=5
This setting, for example, allows Frigate to consume my 10-15fps camera streams on my relatively low powered Haswell machine with relatively low cpu usage.