performance degradation about nvidia-patch

[roderrooder]

Just to chime in,

while the GTX 1070Ti has 2 NVENC chips only one is active, in my tests the 1070Ti is very comparable with the 1050Ti.

EDIT: That is unless this patch also gives users access to the second NVENC chip also.

[Snawoot]

It is possible decoding is bottleneck in your case. Check CPU utilization when converting 4 streams. If CPU is utilized about 100%, consider using -hwaccel option in order to speedup decoding and nvresize filter in order to speedup rescaling (if it takes place in your pipeline).

Anyway, all encoding jobs share same NVENC chip and GPU is not expected to provide unlimited performance for arbitrary number of simultaneous sessions.

[Ivanezko]

Thanx for the swift answer
Here is an example of the command I use
ffmpeg -y -hwaccel cuvid -c:v h264_cuvid -resize 1280x720 -i origin.mp4 -c:v h264_nvenc test1.mp4
CPU usage is very low

[Snawoot]

Could you replace h264_cuvid with h264_nvenc in order to use same encoder for all outputs?

[Ivanezko]

h264_cuvid is a decoder, not encoder
Looks like you are right, no magic and nvenc chip is limited.
in my case 1 process takes 60% of nvenc according to "nvidia-smi dmon"
2 processes take 100% and FPS of each drops a lite
3 processes take 100% and FPS drops heavily
So using more then 2 processes is logical only for light tasks

[Snawoot]

h264_cuvid is a decoder, not encoder

Sorry, my bad

FFmpeg users report 60 realtime SD channels with GTX 1070Ti. It's hard to say whether GTX 1050Ti has same NVENC chip or not.

[vans163]

Lol, really? Did you expect to run 1000 sessions on 1 GPU and have them all compete in 10 seconds too? If so NVIDIA would be out of business as they would only need to sell 1 GPU, and ofc the laws of physics would be broken and Nvidia will rule over all due to their secret technology.