Turing NVENC HEVC doesn't worth,it was just marketing.Wasn't it? TESTS

[Cumberiron]

Hi!

Do you remember when RTX cards appeared? All of the famous magazines and experts were full with the importance of HEVC B-frame feautre on the new Nvidia Turing cards.

They all write about the huge leap of the video quality , which they measured by SSIM and PSNR.

However, when I tested them (PSNR and SSIM), there is not much better quality as the experts song about it in their online magazines. It was often called the best quality HW encoder on this planet etc....



TESTS:

All options are the same, except the B-frames.








The original video:
https://sendgb.com/IiPWx675tUC


The 0 zero B-frame sample video:
https://sendgb.com/5QvuJ4qH8NA

The 4 Bframe Version.
https://sendgb.com/mdZ4wpGCMlB



Please share your opinion!

[Cumberiron]

Sorry, I forgot to post the 2 b frame version.

Here you can download it.
https://sendgb.com/4QaAW8LPVFy

[sophisticles]

I haven't looked at the test encodes but a couple of things to remember about ALL encoders, including the "x" cousins:

The all default to lower quality B frames than P frames and in turn the P frames of a lower quality than the I frames. This is primarily because they all use a higher quantizer for P and B than they do for I frames.

Wuith x264(and x264 IIRC) there are 2 settings named I/P ratio and P/B ratio, NVENC and Quick Sync have similar settings, that set the quantizer used for B frames relative to either I frames or P frames (I don't remember which).

The above is done to "improve compression" but compression is the enemy of quality, lossy compression works by throwing details away, and the details discarded are determined by the encoders algorithms based on what they "think" will be less noticeable or less important to the big picture.

If you want higher quality encodes, regardless of the encoder used, do not be afraid to set the I/P ratios and P/B ratios to 1, that is have each frame type use the same quantizer level and up the bit rate accordingly.

Bit rate starving encodes is bush league, strictly amateur hour.

[Selur]

Bit rate starving encodes is bush league, strictly amateur hour.

Then let us know how someone who thinks he's not an amateur devices what bit-rate starves a content.
When encoding intermediate content and size isn't an issue lossless encoding is fine, but otherwise any file size restriction might require a 'Bit rate starving encode'.
Personally I think that producing a good quality with hard restrictions is way more challenging than simply boosting the bit rate till most encoders of a format look the same.
(as a side note: tweaking the encoder and pre-filtering the content in a good way usually requires a lot of knowledge,...)

Cu Selur

[sophisticles]

Then let us know how someone who thinks he's not an amateur devices what bit-rate starves a content.

I'm assuming you are asking what I consider not bit rate starving an encode, correct?

I adhere to industry standard encode rates, such as closed GOP length equal to 1 second, i.e. GOP=23 for 23.97 fps, GOP=29 for 29.97 fps and a bit rate between 20-30 mb/s for 1080p.

For UHD 4k, I have no problem with 100 mb/s.

It always amazes me, people will complain about the quality of You Tube, Net Flix, Vimeo, iTunes, Hulu, commercial Blu-rays and 4k Blu-rays and then they will turn around, take that content and re-encode it at 1/5 of the bit rate.

It's mind boggling, they will use long open GOPs, with as many B frames as they can squeeze in there, as many reference frames as they can away with, with settings that take their encodes way out of spec and not decodable with a HRD and then complain that their encodes don't look as good as they have been led to believe they would if only they used some "magical" "miracle" software encoder.

In the old days I could almost understand, hard drives were expensive, but today, when you can get a 1 TB Intel 660P for $100, then what the hell are they doing.

These people sit there and wait for the latest processor release from AMD and Intel, looking to spend hundreds, if not thousands, of dollars to put together a new system so they can use more aggressive settings with their favorite "x" encoder and or faster encoding speed when they can just up the bit rate with a hardware encoder that is already present on their system and call it a day.

[Zero-11]

However, when I tested them (PSNR and SSIM), there is not much better quality as the experts song about it in their online magazines. It was often called the best quality HW encoder on this planet etc....

well of course they were lying, the world is not a planet, knowing this could have saved you a headache.

[Selur]

I'm assuming you are asking what I consider not bit rate starving an encode, correct?

correct.

So:
- fixed GOP with a one second max gop (I hope it's just max GOP and not a fixed GOP size)
- 20-30 MBit/s for 1080p
- up to 100MBit/s for 4k content
that's it?
Have you never read the specifications of H.264 and H.265?
You do know that open GOP doesn't really limit the reference frame count with H.264 and H.265, right? I mean that's why they differentiate between Key- and IDR- frames.


Cu Selur

[KarMa]

Bit rate starving encodes is bush league, strictly amateur hour.

It's a great way to easily see the short comes of a given encoder, which I guess is why you are against such things.

[Cumberiron]

Hello!

Did anybody download the video samples?

[sophisticles]

Bit rate starving encodes is bush league, strictly amateur hour.

It's a great way to easily see the short comes of a given encoder, which I guess is why you are against such things.

Yes, I am against using settings that one would normally never use when they were creating content just to show a difference between 2 or more encoders. I don't see the point, are you ever going to encode a 4k video at 5 mb/s? If the answer is "no" then what do you care which encoder looks best at those settings?

[sophisticles]

You do know that open GOP doesn't really limit the reference frame count with H.264 and H.265, right? I mean that's why they differentiate between Key- and IDR- frames.

I use closed GOP in order to avoid software players introducing artifacts during seeking, as is sometimes the case with open GOP.