GPU Transcoding Slow

[bradford 9]

Emby is detecting my Quadro P400 and when I start watching a 4k HEVC source the dashboard shows the following:

 

 

16 fps is what I would expect from a CPU transcode, and I am seeing one of my 8 cores pegged at 100% with ffmpeg. 

`nvidia-smi` shows the ffmpeg process, but it's not using the GPU:

 

 

Logs are attached. Any ideas what's going on? It worked a bit ago, with the 418 driver, I'm on 430 now. Maybe that is the difference? 

 

Log.txt

[Luke 42079]

And the ffmpeg log?

[bradford 9]

That would be helpful wouldn't it - here it is!

FFMPEG Log.txt

[bradford 9]

According to the Nvidia Video Encode and Decode Matrix the P400 should work: https://developer.nvidia.com/video-encode-decode-gpu-support-matrix

 

But maybe it doesn't have the power for a 4k transcode? I guess I assumed that if it was capable of it it could do it faster than 16 fps. The weird thing is, it's not even using the GPU in my case even though ffmpeg is being reported in nvidia-smi. 

[Luke 42079]

Well this is transcoding 4k to 4k. Maybe that's just as fast as it can go? Can you try lowering the in-app quality setting to 1080p?

[bradford 9]

Well this is transcoding 4k to 4k. Maybe that's just as fast as it can go? Can you try lowering the in-app quality setting to 1080p?

I did try that and the performance was the same.

 

I dug a little deeper into nvidia-smi output. I was wrong about the process not using the GPU, I was reading the output of `nvidia-smi` wrong - utilization is per GPU and is listed there instead of per-process. Some benchmarks:

 

4k (HEVC) -> 4k transcode: ~14fps, GPU utilization from 45-52% via `nvidia-smi`, one core spikes up to 100% regularly

4k (HEVC) -> 720p transcode: ~20fps, GPU utilization from 25-33% via `nvidia-smi`, one core spikes up to 100% regularly

1080p (high bitrate AVC) -> 720p transcode: 140+ fps, GPU utilization spikes around 16% via `nvidia-smi` but heavily throttled based on setting, one core spikes up to 100% occassionally

 

So more taxing loads (4k to 4k hevc) use more of the GPU but neither can match the source framerate. I wonder if there's a single-thread process that's throttling me, like audio transcoding (since the 4k source requires audio and video transcoding to play in the browser) - I recall from another thread that audio transcoding can't be offloaded to the GPU and is single threaded so my relatively weak Atom 2785 single core performance might be the limiting factor here.

Edited June 19, 2019 by bradford

[Luke 42079]

Ah yes that is possible.

[bradford 9]

Doing a little more reading it looks like there's not many multi-threaded or GPU enabled audio codecs. Too bad. Looks like I have another reason to upgrade my main pc and repurpose my Ryzen 1800x for my server. Sorry for wasting your time, hope this helps someone in the future. 

 

An enhancement request could be representing the -stats output of ffmpeg in the server dashboard, or maybe you're already doing that.

[Luke 42079]

Thanks for the feedback.