Intel's QS based, Skylake powered hevc encoder beats x265 says MSU

[sophisticles]

http://compression.ru/video/codec_comparison/hevc_2016/

It's taken a while but in the end you just can't compete with deep pockets. According to the latest codec comparison done by MSU Intel's Skylake and one other encoder I have never heard of beat both x265 and x264 and convincingly at that.

I knew the day would come when hardware encoders would beat their software counterparts, we are at he point where it doesn't make any sense to spend the money needed to build a monster encoding machine, you're better off buying a cheap i3 powered $300 laptop and encoding via quick sync, way faster, better quality and much lower power consumption.

If anyone has any links to the any excuses being made by either the x265 or x264 developers I would love to see them.

[poisondeathray]

Unfortunately SSIM is only a moderately accurate predictor of "quality". It "fails" pretty badly in certain types of content, just like PSNR. I'd love to see some accompanying subjective assessments

I agree with the faster, lower power consumption part - and that's a move in the right direction, but not necessarily the "quality" part based solely on SSIM testing.

This is not an excuse, this is the main complaint when x264 (or any codec) supposedly "wins" testing in the past - those tests only show a limited subset , and are "tweaked" to score higher "points" but yet usually show lower visual quality. The problem is with the measurement. ie. a higher SSIM or PSNR score does not necessarily correlate with what a human being "sees" as higher quality or more resembling the source. If it's not clear, I can post some examples to demonstrate

Even so, that's great news for Intel. And I've never heard of the other encoder "Kingsoft", anybody know anything about them ?

[KarMa]

They only used x265 in ABR/Single Pass mode. And then oddly used x264 in 2 Pass mode. They only disclose this in the commands given to the codecs at the bottom of the pdf.

[sophisticles]

They only used x265 in ABR/Single Pass mode. And then oddly used x264 in 2 Pass mode. They only disclose this in the commands given to the codecs at the bottom of the pdf.

They asked the developers of the various codecs what settings they should use, the respective developer's made the decisions. I'm assuming the x265 crew chose single pass mode in order to be competitive speed wise, and the x264 developers probably chose 2 pass because the tests were being done at various bit rates and they wanted the highest possible quality.

[sophisticles]

Unfortunately SSIM is only a moderately accurate predictor of "quality". It "fails" pretty badly in certain types of content, just like PSNR. I'd love to see some accompanying subjective assessments

I agree with the faster, lower power consumption part - and that's a move in the right direction, but not necessarily the "quality" part based solely on SSIM testing.

This is not an excuse, this is the main complaint when x264 (or any codec) supposedly "wins" testing in the past - those tests only show a limited subset , and are "tweaked" to score higher "points" but yet usually show lower visual quality. The problem is with the measurement. ie. a higher SSIM or PSNR score does not necessarily correlate with what a human being "sees" as higher quality or more resembling the source. If it's not clear, I can post some examples to demonstrate

Even so, that's great news for Intel. And I've never heard of the other encoder "Kingsoft", anybody know anything about them ?

The thing is that subjective tests were also done, were they had various folks look at the encodes and pick which one they thought looked the best.

More importantly, if you look at the command lines used, for the "ripping" test, for x264 they used x264 --preset placebo --me umh --merange 32 --keyint infinite --tune ssim --pass 2 and for x265 they used -p veryslow --tune ssim and Intel's MSS gpu encoder still beat them in the SSIM metric!

Did you ever think you would see the day when x264 + placebo + tune ssim would lose a test where SSIM is the metric?

[jagabo]

Did you ever think you would see the day when x264 + placebo + tune ssim would lose a test where SSIM is the metric?

Of course. Did you think encoding technology would sit still?

[poisondeathray]

The thing is that subjective tests were also done, were they had various folks look at the encodes and pick which one they thought looked the best.

It would be nice to see all the data, encodes, without "paying" for the pro version

More importantly, if you look at the command lines used, for the "ripping" test, for x264 they used x264 --preset placebo --me umh --merange 32 --keyint infinite --tune ssim --pass 2 and for x265 they used -p veryslow --tune ssim and Intel's MSS gpu encoder still beat them in the SSIM metric!

Did you ever think you would see the day when x264 + placebo + tune ssim would lose a test where SSIM is the metric?

x264 was already "beaten" by x265 in metrics , and x265 only "won" by a "hair" last year (Intel was already nipping at its heels). So congrats to Intel and "Kingsoft" (who? )

SSIM tests are of limited value, nobody uses --tune ssim for real encoding, they look worse most of the time. What I would like to see is real encodes, real situations, real usage settings, real bitstreams

[ezcapper]

To the OP, what do you think about buying a $300 HEVC encoder that encodes in real time. Is it good or a piece of junk? One plus seems to be 1080p60 real time encoding upto 12mbps. The skylake tests were all 30fps and lower.

HDMI input HDCP decoding. Capture through command line vlc:

https://www.aliexpress.com/item/Mini-H-265-encoders-HEVC-hardware-encoder-HDMI-input-f...f-5a0949ca9dea

[sophisticles]

Seems too limited to me for all but the most narrow usage scenarios.

[ezcapper]

Seems too limited to me for all but the most narrow usage scenarios.

Thanks

[jagabo]

Seems too limited to me for all but the most narrow usage scenarios.

Thanks

Just because it's an h.265 encoder doesn't mean it's any good.

[KarMa]

you're better off buying a cheap i3 powered $300 laptop and encoding via quick sync, way faster, better quality and much lower power consumption.

This is a GPU accelerated encoder, which I don't believe has to do with quick sync, nor does the PDF say anything about quick sync. Also take note that the CPU used in the test is a $300 Corei7 6700K, which comes with a Intel HD Graphics 530. So this was tested on an Intel APU, and not any Intel Chip will do.

They asked the developers of the various codecs what settings they should use, the respective developer's made the decisions. I'm assuming the x265 crew chose single pass mode in order to be competitive speed wise, and the x264 developers probably chose 2 pass because the tests were being done at various bit rates and they wanted the highest possible quality.

Guess I would rather see CRF used instead of ABR, as they both should have about the same speed but that would take longer for MSU to test. Pretty much the only time ABR is a good idea is if you are steaming to Twitch or something. Just about no one recommends ABR here.

[sophisticles]

you're better off buying a cheap i3 powered $300 laptop and encoding via quick sync, way faster, better quality and much lower power consumption.

This is a GPU accelerated encoder, which I don't believe has to do with quick sync, nor does the PDF say anything about quick sync. Also take note that the CPU used in the test is a $300 Corei7 6700K, which comes with a Intel HD Graphics 530. So this was tested on an Intel APU, and not any Intel Chip will do.

They asked the developers of the various codecs what settings they should use, the respective developer's made the decisions. I'm assuming the x265 crew chose single pass mode in order to be competitive speed wise, and the x264 developers probably chose 2 pass because the tests were being done at various bit rates and they wanted the highest possible quality.

Guess I would rather see CRF used instead of ABR, as they both should have about the same speed but that would take longer for MSU to test. Pretty much the only time ABR is a good idea is if you are steaming to Twitch or something. Just about no one recommends ABR here.

You can not use ABR for any meaningful codec comparison because you are testing the quality of the encode at various bit rates, thus you have to be able to control the bit rate used.

Skylake i3's do come with the same integrated gpu as do Skylake Celerons and they all support quick sync.

The MSU test used the Intel MSS encoder, which is the quick sync hardware accessed through the Intel Media Server Studio:

https://software.intel.com/en-us/intel-media-server-studio

You're implied objection is correct though, while the MSU tests show what the QS hardware is capable of it does not follow that applications available to the average user, such as Staxrip, TMPG's software, Sony Vegas, Handbrake, anything based on ffmpeg or that other freeware QS encoder (I forget it's name) will deliver the same results, it just means IF someone codes an app properly that makes full use of of QS they should get similar results.

I would like to see someone with a Skylake do some tests with Staxrip, maybe someone will stumble on this thread and do some tests.

[KarMa]

You can not use ABR for any meaningful codec comparison because you are testing the quality of the encode at various bit rates, thus you have to be able to control the bit rate used.

I'm going to assume you mean CRF instead of ABR, and you can use CRF, just have to do mulitiple encodings to get the bitrate where you want it for the test. Which is why I said it would take MSU longer to do that.

Or they could just get the bitrate close enough and use their "relative bitrate" metric, because none of the encoders give you exactly the bitrate you ask for. The Intel bitrate was off by ~5-25% above target bitrate, so to make up for that they use their own "relative bitrate" metric. x265 was off the target bitrate also but by less than half as much.

Skylake i3's do come with the same integrated gpu as do Skylake Celerons and they all support quick sync.

The MSU test used the Intel MSS encoder, which is the quick sync hardware accessed through the Intel Media Server Studio:

The speed of the Intel encoder was not amazingly better, especially considering it had a built in GPU at its disposal. Making me think that it's still heavily CPU based and bottlenecked at the CPU still. So dropping down to a much cheaper i3 APU would probably take a massive fps hit. Would help if they showed the CPU and GPU usage during the Intel tests.

Looks like any of the Intel APU chips would work with quick sync, info on quick sync seems harder to come by. I know OSB and I think AsVideoConv support quick sync.

[sophisticles]

More than the speed for me the big selling point is the reduced power consumption. If all you have to encode is 1 or 2 videos then the heat and power consumption and the noise from having the cpu fan (assuming air cooling) running at full tilt can be ignored but if you regularly encode and/or want to encode dozens of Blu-Rays or dozens of 4k videos, then a low power, quite solution becomes very attractive.

[KarMa]

More than the speed for me the big selling point is the reduced power consumption. If all you have to encode is 1 or 2 videos then the heat and power consumption and the noise from having the cpu fan (assuming air cooling) running at full tilt can be ignored but if you regularly encode and/or want to encode dozens of Blu-Rays or dozens of 4k videos, then a low power, quite solution becomes very attractive.

But it was not that much faster, nor do we know what the CPU/GPU usage was. So you don't really know what the power savings were.

[stonesfan129]

Ok, so are there any programs that can do this yet? Can it be done on my i7-4790k? If not, this isn't news to me.

[KarMa]

Ok, so are there any programs that can do this yet? Can it be done on my i7-4790k? If not, this isn't news to me.

Your chip has the HD Graphics 4600, so I would believe the answer would be yes.

[RazorBurn]

As I always have my Original DVD and BD on my storage room.. I have no problem using QuickSync for all of my encodes.. Quality differences are soo minimal that the gain from the speed and power consumption is Highly Recommended..

[sophisticles]

Ok, so are there any programs that can do this yet? Can it be done on my i7-4790k? If not, this isn't news to me.

Media Coder, Stax Rip, Hybrid, Handbrake, FFmpeg, TMPG, Sony Vegas, the list of software that uses quick sync is decent; QS has existed on Intel cpu's since Sandy Bridge but with the Skylake version it's the first time it's been shown to produce better quality than software solutions.

The only thing is, as I already pointed out, the test were done with Intel's Media Studio Server software, which makes full use of all the capabilities of QS, I do not know if any of the above are coded the same way though from experience it does seem like Stax Rip most likely would come closest to fully exploit QS.

Be that as it may, the test in question applies to Skylake cpu's, the i7-4790k which was used for the last test version didn't score as high.

Still, I think it speaks volumes that Nvidia this time decided not to participate in this test, nor did AMD include their VCE encoder, almost as if they knew from internal testing that they couldn't compete.

Makes me wonder what Kaby Lake will bring to the table.

[poisondeathray]

4970 doesn't support HEVC encoding through QS, only skylake and newer . The freely available version is QSVEnc by rigaya, all the front ends like staxrip, mediacoder, hybrid erc... use it

Makes me wonder what Kaby Lake will bring to the table.

Intel has improved speed and quality every generation, so I would assume improvements on both fronts again

They say proper Main10 support, but it's unclear if that's decode only or whether encode is supported as well
http://www.anandtech.com/show/10570/intel-teases-mobile-kaby-lake-hevc-main10-profile-...ng-this-autumn

[sneaker]

Intel has already confirmed HEVC Main10 encoding (+decoding) support for Kaby Lake. (source)

Still, I think it speaks volumes that Nvidia this time decided not to participate in this test, nor did AMD include their VCE encoder, almost as if they knew from internal testing that they couldn't compete.

It's not like anyone needs to apply for a test. You can just walk into a store and buy the GPUs if you want to. (At least with Nvidia Maxwell. The new AMD GPUs might have simply been released too late for testing?)

[sophisticles]

It's not like anyone needs to apply for a test. You can just walk into a store and buy the GPUs if you want to. (At least with Nvidia Maxwell. The new AMD GPUs might have simply been released too late for testing?)

That's not the way MSU does testing, they reach out to various developers and hardware vendors and ask them if they want to participate in the test, if they do they supply the hardware (and software) and they also supply how they should be configured for various test scenarios.

I would have liked to have seen Nvidia's newest nvenc compete, they added SAO to their HEVC engine which is supposed to bring substantial quality improvements.

[poisondeathray]

I would have liked to have seen Nvidia's newest nvenc compete, they added SAO to their HEVC engine which is supposed to bring substantial quality improvements.

nvenc still no b-frames. Can't compete

[ezcapper]

The Kaby Lake HD Graphics 620 (new Quicksync) laptops are now here for $350. They claim to be able to transcode to 4k30 HEVC at 2x realtime speed. Does anyone have one to test and post?
http://www.ebay.com/itm/New-Laptop-Acer-Aspire-E-15-15-6-Full-HD-7th-Gen-Intel-Core-i3...4AAOSwYIxX8kc8

I'm very interested in the speed and quality of these new chips at encoding RAW video.

[sophisticles]

The Kaby Lake HD Graphics 620 (new Quicksync) laptops are now here for $350. They claim to be able to transcode to 4k30 HEVC at 2x realtime speed. Does anyone have one to test and post?
http://www.ebay.com/itm/New-Laptop-Acer-Aspire-E-15-15-6-Full-HD-7th-Gen-Intel-Core-i3...4AAOSwYIxX8kc8

I'm very interested in the speed and quality of these new chips at encoding RAW video.

$350 is a great price, but I don't trust buying anything from Ebay, especially computer related stuff. I'll wait until Best Buy gets the newest models (they still have the Skylake based ones) and then I'll definately be picking one up (my current laptops are at least 10 years old).

[ezcapper]

The Kaby Lake HD Graphics 620 (new Quicksync) laptops are now here for $350. They claim to be able to transcode to 4k30 HEVC at 2x realtime speed. Does anyone have one to test and post?
http://www.ebay.com/itm/New-Laptop-Acer-Aspire-E-15-15-6-Full-HD-7th-Gen-Intel-Core-i3...4AAOSwYIxX8kc8

I'm very interested in the speed and quality of these new chips at encoding RAW video.

$350 is a great price, but I don't trust buying anything from Ebay, especially computer related stuff. I'll wait until Best Buy gets the newest models (they still have the Skylake based ones) and then I'll definately be picking one up (my current laptops are at least 10 years old).

Thank you sophistocles!!! Amazon has them if you like them. But I definitely understand if you prefer Best Buy.
https://www.amazon.com/Acer-Aspire-i3-7100U-Windows-E5-575-33BM/dp/B01K1IO3QW/ref=sr_1...er+Aspire+E+15

p.s. When you get the laptop, can you test quicksync hevc to see if it can record the free-to-copy 12 min movie "Tears of Steel" playing off your 1080p60 computer screen in real-time using OBS at a 30 mbps setting?

[VHUser]

I would like to see someone with a Skylake do some tests with Staxrip, maybe someone will stumble on this thread and do some tests.

I can do some testing with the Intel HD 530 I have as well as CPU encoding for comparing encoding time and quality.

Software:
QSVEnc 2.58
StaxRip 1.3.7.0

Source:
Tears of Steel - I downloaded the 4K 3840 pixels wide (6.3 GB, mov) clip

Let me know if you have any other source you'd like me to test for you.

Let me know what settings you'd like me to try with StaxRip, QSVenc CLI

I'm very curious about Quick Sync performance/Quality, but I'd like some help with reviewing the results.
I could keep all my records in a spreadsheet format available to anyone.

[sophisticles]

I would like to see someone with a Skylake do some tests with Staxrip, maybe someone will stumble on this thread and do some tests.

I can do some testing with the Intel HD 530 I have as well as CPU encoding for comparing encoding time and quality.

Software:
QSVEnc 2.58
StaxRip 1.3.7.0

Source:
Tears of Steel - I downloaded the 4K 3840 pixels wide (6.3 GB, mov) clip

Let me know if you have any other source you'd like me to test for you.

Let me know what settings you'd like me to try with StaxRip, QSVenc CLI

I'm very curious about Quick Sync performance/Quality, but I'd like some help with reviewing the results.
I could keep all my records in a spreadsheet format available to anyone.

I don't recommend the 6gb version of ToS, it's previously been compressed with x264 and that will throw the results off.

I have the y4m uncompressed version sitting on my hard drive, I'm cut a piece of it, compress it losslessly (sans audio) and upload that for testing purposes (I have to run a few experiments to get a clip that fits within the 500mb attachment limit of this forum).

Check back in the morning.

Thanks.

[VHUser]

I actually downloaded the 4K DCP 4096 x 2160 (14 GB, Digital Cinema Package) also, so I do have that available.