Nvidia GPU Encode Anomalies in PD21

JL_JL

Senior Contributor Location: Arizona, USA Joined: Oct 01, 2006 20:01 Messages: 6091 Offline

Solved by

Aug 15, 2023 17:47

I’ve had to do a lot of H.265 60p encoding of some long duration timelines recently. I’ve come across some severe anomalies with PD21. Basically, modifying either H.265 default profiles of 4K 3840x2160/30p (37Mbps) or 4K 4096x2160/30p (37Mbps) to 59.94 fps results in very poor encode performance time with newer generation Nvidia GPU’s. This anomaly appears to only occur in PD, if I use another editor or a basic file transcoder that supports NVENC, I don’t see these issues. I suspect the PD behavior is probably due to encoder presets internally set. Most users probably don’t have the suite of Nvidia GPU’s to perform a test so I’m just putting the data out there that newer not necessarily better or on par with older GPU's with PD and H.265.

Source video for this test was a 2m10s clip of 4K H.264 4096x2160/60p 50Mbps but it appears source does not matter much, the produced 60p H.265 profile does.

GPU	3840x2160/60p (37Mbps) Encode Time (sec)	4096x2160/60p (37Mbps) Encode Time (Sec)
GTX 1070	105	112
RTX 2070	191	204
RTX 4070	150	160

Basically, I put in an old GTX 1070 to complete these PD21 H.265 encodes about 2x faster than my current general use RTX2070. A RTX4070 was better than a RTX2070 but significantly short of GTX1070 performance. File size and visual quality were unchanged.

From what I observed, H.264 encoding in PD21 is not affected like this with newer generation Nvidia GPU's, a RTX4070 with 60p profile was ~1.7 times faster than a GTX1070 for encoding time.

Jeff

Warry

Senior Contributor Location: The Netherlands Joined: Oct 13, 2014 11:42 Messages: 853 Offline

Aug 16, 2023 05:47

This is very interesting. When moving between PD versions and PC versions, I did some testing too.
I used from the sample files: The mountainbiker.mp4 and the Skateboard 01.mp4 and put a transition: Overlapping shape 02 inbetween
In took a number of output formats (working with PAL).
For my system it gave me the impression that it would take me less time to get the work done.
I am a bit puzzled about the file sizes, but I always get different file size when I rerun the export with same parameters.

JL_JL

Senior Contributor Location: Arizona, USA Joined: Oct 01, 2006 20:01 Messages: 6091 Offline

Aug 16, 2023 10:23

Quote This is very interesting. When moving between PD versions and PC versions, I did some testing too.
I used from the sample files: The mountainbiker.mp4 and the Skateboard 01.mp4 and put a transition: Overlapping shape 02 inbetween
In took a number of output formats (working with PAL).
For my system it gave me the impression that it would take me less time to get the work done.
I am a bit puzzled about the file sizes, but I always get different file size when I rerun the export with same parameters.

Some of the small file size changes are to be expected as output is variable bitrate vs constant. A tall order going from ~4.5Mbps sample files to 37 and/or 50Mbps output.
The large 38% difference for your MPEG2 case is odd indeed, but nothing GPU related there, as it is not supported. Output obviously didn't achieve 25Mbps avg setting.
I couldn't decode your naming convention, for instance hevc-4k-27-37-fast-gpu to me seems like the 27 should be 25 for 25fps perhaps, similar with avc-m2tsxxxxx names.

Jeff

Warry

Senior Contributor Location: The Netherlands Joined: Oct 13, 2014 11:42 Messages: 853 Offline

Aug 17, 2023 05:20

Quote

Some of the small file size changes are to be expected as output is variable bitrate vs constant. A tall order going from ~4.5Mbps sample files to 37 and/or 50Mbps output.
The large 38% difference for your MPEG2 case is odd indeed, but nothing GPU related there, as it is not supported. Output obviously didn't achieve 25Mbps avg setting.
I couldn't decode your naming convention, for instance hevc-4k-27-37-fast-gpu to me seems like the 27 should be 25 for 25fps perhaps, similar with avc-m2tsxxxxx names.

Jeff

Yes, sorry for the confusion. I have included more clear format specs below and also reran the test. I used all the PD sample files, images, sound etc. , selected them all and dragged them to the time line. The clip is 00:05:36:16 long, using PAL 50 FPS.
The yellow % are the times faster between PD20 and PD21. The green % is the difference between pd21.3 and previous PC, and PD21.5 and current PC.
What strikes me when I compare the export times with previous 21 versions, 21.5..3125.1 is overall a bit faster. I know that these results are only valid for comparison within my own environment.
For what it is worth:

JL_JL

Senior Contributor Location: Arizona, USA Joined: Oct 01, 2006 20:01 Messages: 6091 Offline

Aug 17, 2023 08:21

Quote Yes, sorry for the confusion. I have included more clear format specs below and also reran the test. I used all the PD sample files, images, sound etc. , selected them all and dragged them to the time line. The clip is 00:05:36:16 long, using PAL 50 FPS.
The yellow % are the times faster between PD20 and PD21. The green % is the difference between pd21.3 and previous PC, and PD21.5 and current PC.
What strikes me when I compare the export times with previous 21 versions, 21.5..3125.1 is overall a bit faster. I know that these results are only valid for comparison within my own environment.
For what it is worth:

The GPU status in the name in prior table was clear, but this clears up your fps, obviously it would appear just a typo in prior table.

But you're right, these results would be good for you. This appears to be a totally different test than prior. Since ~80% of the timeline is music with no video, bitrate will be very low there in produced file as well as in the pic area of timeline, as such it's really hard to even calculate if you got the profile details you specified from the file size. I was looking at the effect of multiple GPU's same PD version all in the same platform. This appears to be a rather different evaluation and not to comparable.

Jeff