Evaluation of HEVC decoders (SW, Hybrid and HW)

[NikosD]

I didn't bring the conversation regarding new manufacturing process to mock Intel being second to Samsung and TSMC.

The inability to go to 10nm and the "nm war" are irrelevant to my discussion.

I mentioned the fact that Intel needs a new process to bring a new architecture, because it has stuck to its latest Skylake's architecture since 2015.

So, the problem of manufacturing process is huge for Intel not because it could loose the "nm war" but because it needs a new process to bring a new architecture, like all CPU companies actually, but with one little difference.

Intel has its own fabs and can't give their CPUs to GloFo, Samsung, TSMC etc.

If they can't fix the broken 10nm process they can't deliver a new CPU architecture. It's as simple as that!

[el Filou]

I mentioned the fact that Intel needs a new process to bring a new architecture, because it has stuck to its latest Skylake's architecture since 2015.

If they can't fix the broken 10nm process they can't deliver a new CPU architecture. It's as simple as that!
Why? Can't there be other reasons for the fact they used the same architecture for 3 years? Like, slowing down in development?
Do you know secret stuff about their next architecture that says they absolutely need a newer process to introduce it?

[NikosD]

The questions you asked me are like:

"Why 1+1 equals 2 ? "

[NikosD]

According to techpowerup.com VEGA supports these:

The HDMI port is version 2.0, and DisplayPort has been updated to 1.4 HBR3, MST and HDR.

This enables support for 4K @ 120 Hz, 5K @ 60 Hz, or 8K @ 60 Hz.

Supported HDR configurations are 4K60, 4K120, and 5K60, whereas Polaris supported only 4K60 HDR.

GPU-accelerated encoding is now capable of doing HEVC/H.264 at up to 1080p240, 1440p120 and 2160p60.

Hardware-based decode works up to 4K60 for H.264, H.265 and VP9.

[NikosD]

We have to wait for Navi or maybe a VEGA refresh for mainstream, like VEGA 11, could probably have VP9 hardware decoding inside.

AMD is a little (?) slow adopting non mainstream formats (actually the last one of the big three)

[gagmani]

I have just joined this forum to thank edwdevel for detailed review of playback capabilities of gt 1030. I bought it instead of 1050 and I am happy with my purchase. 1050/1050ti ti is very expensive here. And I am not into gaming. Plays even 8k smoothly with W10 Media player. But so far I am not able to use Nvenc or Cuda successfully in Mediacoder,Staxrip and any video converter. I do not know if it is due to hardware limitation or driver problem.

[NikosD]

But so far I am not able to use Nvenc or Cuda successfully in Mediacoder,Staxrip and any video converter.

I do not know if it is due to hardware limitation or driver problem.

What is your OS and drivers ?

[el Filou]

I am not able to use Nvenc or Cuda successfully in Mediacoder,Staxrip and any video converter. I do not know if it is due to hardware limitation or driver problem.
NVIDIA lists the 1030 as not supporting GameStream, which needs NVENC: https://www.geforce.com/hardware/desktop-gpus/geforce-gt-1030/specifications
It looks like the 1030 doesn't have NVENC, either in hardware or they disabled it through the drivers/BIOS.
The 1030 is absent from this page too: https://developer.nvidia.com/cuda-gpus even though they just may have forgotten to update it.
If you run GPU-Z, is the CUDA box in the Computing section checked?

[peca89]

I've tested 4K Netflix streaming on GT1030 and it does not work, as expected. Test setup:

Hardware: MSI GeForce GT 1030-2GH-OC
OS: Windows 10 Pro, Creators Update, version 1703, build 15063.540
Drivers: 384.94 WHQL and 385.28 WHQL
TV: Sony 850C, directly connected to HDMI, nVidia Control Panel shows that HDCP is enabled
App: Latest Netflix app from Microsoft Store
Output mode: 3840x2160@60Hz, RGB, 8bpp

Can anybody with 1050 and better confirm that it works without Insider Preview version of Windows and "special" driver?

[el Filou]

This thread by someone who got it to work on a Windows 10 Public Release build and driver 384.76 shows it's still flaky: https://www.reddit.com/r/nvidia/comments/6kjbra/nvidia_driver_38476_enables_netflix_4k_without
AFAIK NVIDIA still hasn't updated their Knowledge Base to mark that feature as generally available, so no guarantee yet that it works with public W10 builds and/or public release drivers.

Maybe try the specific driver v384.76 and make sure to follow all the guidelines in that thread?

Oh, and btw: from those screenshots, Netflix uses ~450 MB buffer at 4K, which is massive (= 220 seconds, YouTube only uses ~20-25 secs), and it wouldn't surprise me if the Microsoft DRM forced all the buffer to be kept in VRAM, which could explain the VRAM requirements.

[NikosD]

If someone buys a VEGA 56/64 card, please put a DXVA Checker screenshot here.

Or even if someone finds one posted anywhere on the Net.

[gagmani]

What is your OS and drivers ?

Windows 10 x64 (15063.540), Nvidia drivers 385.28
CPU - Pentium G3220
I saw in some Linux forums that nvenc works in Ubuntu. If it is indeed so, then nvidia disabled it in drivers. Otherwise the hardware is simply not there like previous GT cards. I will regret my decision to not buy a 1050 ti if hardware is absent. At least I have Ubuntu to try, but I don't know how to test nvenc in Linux.

[gagmani]

The 1030 is absent from this page too: https://developer.nvidia.com/cuda-gpus even though they just may have forgotten to update it.

I saw that developer page too. Maybe they are lazy. If they had posted about GP108 there, I would have made the right decision.
In GPU-Z it says CUDA 6.1 supported. Also the card runs at max at PCI -E 3.0 x4 only(written on the box).

http://imgur.com/a/sGY5w

[huhn]

make sure you are not using the driver from microsoft update.

i haven't done the math but 4x could be a huge problem.

[gagmani]

make sure you are not using the driver from microsoft update.

i haven't done the math but 4x could be a huge problem.

No, I downloaded from Nvidia website. I just did not pay attention while buying.

[NikosD]

Windows 10 x64 (15063.540), Nvidia drivers 385.28
CPU - Pentium G3220


All of my readings regarding that issue across the Net say that usually Nvidia disables HW encoding on GT cards (but not always)

I can't believe they just removed the NVENC HW from GP108 chip, it is easier for them to disable it, what a pitty.

There is a direct question asked in devtalk Nvidia forum, I don't know if it was you.

I'm not sure if Nvidia replies officially, but the same question asked before in devtalk Nvidia forum regarding GT 1030, the reply was that NVENC is not supported by that card.

[gagmani]

I think if we need nvenc we will have to to pay the green tax by buying a GTX. Also same applies for Netflix.

[aufkrawall]

Is here some mpv user with a GT 1030? I briefly had a chance to play around with a 1030 and noticed that hardware decoding wasn't working for VP9 10 bit, while it was working with LAV Filters.
Could someone please report this for mpv at Github with a logfile like described there?

Edit: GTX 1050 would likely show the same behavior.

[el Filou]

i haven't done the math but 4x could be a huge problem.For playback? Or encoding? For playback, with Copy-Back the bus load would be around 80% with 2160p60 10 bit, needs testing. Of course now you have the option to use D3D11VA Native, with which the PCI Express interface is barely loaded at all.

[rcksh]

I've given up waiting for relevant information to show up on somewhere on the internet so I decided to buy the GT 1030 and test it myself. Let me first answer my own questions:


Can it upscale 1080p60 and output 2160p60 using EVR? Yes, absolutely!
Can it do it using madVR? Did not test properly yet.
Can it decode AVC/HEVC/VP9 2160p60 SDR and output 2160p60 SDR? Yes, absolutely! No problem at all.
EVR or madVR? Tried only EVR so far
Can it decode AVC/HEVC/VP9 2160p60 HDR and just pass it through to a HDR compatible TV, therefore skipping resource-hungry 4K scaling and HDR<->SDR conversion? I managed to set up madVR to trigger HDR mode on TV. However, colors are completely wrong in fullscreen. I need more time figuring out how madVR works


Everything I threw to the card (up to 4K res), it managed to decode and render properly. I've tried most of the HEVC test clips mentioned in this topic and everything works fine except higher than 4K resolutions. VRAM usage at 4K 60fps 10bit HEVC output at 4K@60Hz is up to 1500MB when using EVR.

If anyone has an idea how to set up madVR to do basically nothing and just pass HDR through, please write here. I know that's not what madVR was made for, but it is probably only thing possible using low end video card. Also, I can do some more tests if someone is interested.

I'm using haswell IGP (intel hd 4400)
and one day it died.
I see your post, and I decided to buy gt 1030.
(my budget is on low, and I read we could use two different nvidia GPU in one system for additional multiple display )

yes thanks for your review.
I'm happy gt 1030 could play 4k60fps 5k60fps even hevc smoothly :thanks:.
also I could play popular pc games with decent quality.
but later I'm a little bit dissapointed after I Found if gt 1030 not support for NVDIA surround :(
and sadly gt 1030 can't play 8k 60fps video, it capped at 8k @ 24fps.
but it doesn't matter because for this time 4k60fps is enough for me :D

overall gt 1030 is not bad choice.

[huhn]

For playback? Or encoding? For playback, with Copy-Back the bus load would be around 80% with 2160p60 10 bit, needs testing. Of course now you have the option to use D3D11VA Native, with which the PCI Express interface is barely loaded at all.

for encoding you are usually uploading uncompressed 4:2:0 images which quite big compared to an encoded video stream.

d3d11 or DXVA native are not that different.

[NikosD]

I've given up waiting for relevant information to show up on somewhere on the internet so I decided to buy the GT 1030 and test it myself. Let me first answer my own questions:


Can it upscale 1080p60 and output 2160p60 using EVR? Yes, absolutely!
Can it do it using madVR? Did not test properly yet.
Can it decode AVC/HEVC/VP9 2160p60 SDR and output 2160p60 SDR? Yes, absolutely! No problem at all.
EVR or madVR? Tried only EVR so far
Can it decode AVC/HEVC/VP9 2160p60 HDR and just pass it through to a HDR compatible TV, therefore skipping resource-hungry 4K scaling and HDR<->SDR conversion? I managed to set up madVR to trigger HDR mode on TV. However, colors are completely wrong in fullscreen. I need more time figuring out how madVR works


Everything I threw to the card (up to 4K res), it managed to decode and render properly. I've tried most of the HEVC test clips mentioned in this topic and everything works fine except higher than 4K resolutions. VRAM usage at 4K 60fps 10bit HEVC output at 4K@60Hz is up to 1500MB when using EVR.

If anyone has an idea how to set up madVR to do basically nothing and just pass HDR through, please write here. I know that's not what madVR was made for, but it is probably only thing possible using low end video card. Also, I can do some more tests if someone is interested.

Could you try DXVA native mode (since copy-back is choppy and memory hungry) using EVR-CP of MPC-HC or similar video player instead of plain EVR ?

Because EVR-CP can use external subtitles and it's a little heavier than plain EVR.

[tony359]

Hi there,

I've landed here looking for info on a GTX1030.

I am putting together a new HTPC with a Ryzen 3. My display setup is currently just 720p but planning to upgrade to 1080p soon. 4K is not on my roadmap yet.
I am using MPC-HC with MadVR and 3DLUT calibration. I am looking for a GPU capable of dealing with some 'must have' MadVR features - not sure what to be honest as my current setup cannot handle anything but the basic settings.

I could go with a 1050 but the only passive model I found may be too tall for my case. The 1030 is also much cheaper!

Most of the comments on this thread are for 4K and HDR. Any thoughts on regular 1080p content with madVR and some good filters enabled?
Thanks!

[peca89]

Could you try DXVA native mode (since copy-back is choppy and memory hungry) using EVR-CP of MPC-HC or similar video player instead of plain EVR ?

Because EVR-CP can use external subtitles and it's a little heavier than plain EVR.

My bad! Whenever I said EVR before, I actually meant EVR-CP.

I tried native mode, but only with madVR and only with 4K HDR content in which case the colors are all messed up. In madVR thread I've read that that only happens with 10bit content, so I have to do some additional testing to be sure.


Most of the comments on this thread are for 4K and HDR. Any thoughts on regular 1080p content with madVR and some good filters enabled?
Thanks!

Yes, I've focused on 4K HDR. I've tried 1080p with some scalers other than bilinear or DXVA2, but render times were to high even for 24p. I'm currently not sure if high render times are caused by 7-year-old low-power CPU (Athlon 240e). I'm experiencing high CPU usage whenever I use madVR. Also, render times are about 10 times higher when I use D3D11 comparing to D3D9, although it should be the other way round...I need more time for investigation...

[tony359]

Thank you - that's bad news then. If you have a moment, what is the CPU/GPU load when you run 1080p content with all default settings on MadVR?

[huhn]

this thread is about hardware decoder speed not madVR filter performance.

ask in the madVR thread or what i recommend you create a new thread for your question.

because i'm writing right now i can give you a a little bit of help but i will not answer more question of this type in this thread.

a 1030 is ok for 1080p output but the very low ram speed could be a problem not sure 48 GB/s is really low.
24 hz ngu low/mid or super xbr /should work.
because i don't know what you understand as a "must have feature" or a "good scaler" i can't be more precise.

you may run into a huge problem with a 3d LUT:
http://bugs.madshi.net/view.php?id=462
my 960 show problem with it and it has much faster Vram.

Also, render times are about 10 times higher when I use D3D11 comparing to D3D9, although it should be the other way round...I need more time for investigation...
are you talking about hardware decoding or madVR presenation?

madVR d3d11 is not known to be faster but know to use more Vram it can output 10 bit so that's that.

as i may said before UHD playback at UHD is possible with 2Gb but you are limiting your self and you can run into problems anytime.
Could you try DXVA native mode (since copy-back is choppy and memory hungry) using EVR-CP of MPC-HC or similar video player instead of plain EVR ?
so copy-back is choppy? please back this up with something.

a modern system has no problem with copyback what so ever if real playback is what you are aim for.

[tony359]

thank huhn, I'll open a new thread.

[NikosD]

so copy-back is choppy? please back this up with something.

a modern system has no problem with copyback what so ever if real playback is what you are aim for.

Here you are...

Since madVR requires copy-back (thanks huhn) it's no surprise that 4K 60fps does not work.

Even SDR 4K 60fps using EVR does not work if copy-back is used (memory gets maxed out).

[huhn]

and it is hard to see that you point to a single case pages back.

[aufkrawall]

Copyback can have a significant performance penalty:
https://forum.doom9.org/showpost.php?p=1815449&postcount=22110

Btw: The Windows 10 video app seems to have by far the lowest VRAM consumption and it doesn't seem to show banding anywhere.

[huhn]

you should test this with real measurement.

but the copyback operation is of cause not free.

copyback has clearly no real effects on the rendertimes of madVR at elast with PCIe 3.0.

BTW. in my short tests d3d11 decode uses more Vram than DXVA native and copyback.

so it is to early to make a clear statement.

[aufkrawall]

you should test this with real measurement.

Do you mean VRAM consumption or banding?
As for banding: I watched several 10 bit SDR videos and couldn't spot any banding which wasn't there with madVR/mpv as well.


copyback has clearly no real effects on the rendertimes of madVR at elast with PCIe 3.0.

That's probably because you aren't close to the limit of your GPU.


BTW. in my short tests d3d11 decode uses more Vram than DXVA native and copyback.

With 4k HEVC 10 bit, DXVA2 copyback consumes more than D3D11VA native, but with VP9 8 bit, it seems the opposite applies.


so it is to early to make a clear statement.
Not regarding GPU load and power consumption. ;)

[huhn]

i mean power measurement.

and what should i say i have a meter and well it's about the same with copyback or native.

[aufkrawall]

I've connected a wattmeter. It shows exactly the same like MSI Afterburner, even the aggressive jumping.
With DXVA2 copyback, I have spikes to 150W, while with D3D11VA native, they are hardly above 115W.
With all the jumping, the average is a bit hard to approximate, but it's very much like:
D3D11VA native: ~90W
DXVA2 copyback: ~120W

Edit: Ok, Windows 10 video app doesn't have issues with banding. However, it's messing up colors in general and thus should be considered as garbage :D :
madVR:
https://abload.de/thumb/madvr1fseo.png (http://abload.de/image.php?img=madvr1fseo.png)

Win 10 video app:
https://abload.de/thumb/win10videoapp71sw5.png (http://abload.de/image.php?img=win10videoapp71sw5.png)
(slightly different frames because video can't be paused without overlay)

[huhn]

well your wettmeter isn't lying but if you need ~30 watt to do a copyback how is a 1030 even able to do it at all?

well it is not that easy but you know what i mean.

[aufkrawall]

I was using 4k downscaled to WQHD, HDR -> SDR, deband, Jinc AR chroma + bicubic AR LL luma and "scale chroma seperately". It seems to be some kind of worst case for the difference between copyback and native.

[huhn]

even with everything at bilinear the copyback cost the same it'S a fixed cost.

so the only difference is that it pushes the GPU in a higher powerstate?

[aufkrawall]

Why do you think copyback should have the same cost?
Lots of data need to be copied from GPU to RAM and back, before the actual computation for renderer stuff begins.
Copying huge amounts of data -> eats energy and adds lantencies. This can't be free.

[huhn]

but it can be done with a iGPU/entry level GPU and no matter what else you use it cost the same.
edit:there are buffers in use so latency is not an issue.

[aufkrawall]

edit:there are buffers in use so latency is not an issue.
Latencies can also occupy parts of the GPU and prevent it from being efficiently utilized. Think this more from a physical side, not from "frame is done and next" software perspective.
Reported GPU load is higher with copyback, and it even switches to higher clocks at the same time compared to native, which is a clear sign for GPU not being utilized efficiently with copyback.

[huhn]

with a 1060 i can't really measure a difference what so ever.

[aufkrawall]

I assume you were not using the same configuration as me?

To "ultimately" prove it:
I set madVR up to hit ~90% GPU usage, which usually is the highest value which is possible without drops.
D3D11VA: no drops
DXVA2 native: no drops
DXVA2 copyback: drops

-> copyback steals performance from madVR.

[huhn]

sorry i can't create such a close case where this is happening.
i have nearly 100 % the same power consumption and GPU load with or without copyback.

and for fair comparison we should either compare dxva native with dxva copyback or d3d11 with d3d11 copyback.

[aufkrawall]

D3D11VA copyback drops frames as well.
The problem likely is the copyback on the same GPU, I wouldn't expect it to drop frames when using a different GPU for decoding than which is used for rendering. However, I can't test this since my Skylake IGP is too slow for HEVC 10 bit decoding.

Or I'd have to create a new testcase for 4k 60fps H.264. Can do that.

Edit: Using IGP via D3D11VA copyback shows similar behavior:
Software decoding: no drops
D3D11 VA native: no drops
DXVA2 native: drops
D3D11VA GTX 1070 copyback: drops
D3D11VA Skylake IGP copyback: drops
QuickSync: no drops

[nussman]

I was using 4k downscaled to WQHD, HDR -> SDR, deband, Jinc AR chroma + bicubic AR LL luma and "scale chroma seperately". It seems to be some kind of worst case for the difference between copyback and native.
30W for copyback? Come on ...
Thats not some kind of worst case, you measurement is just wrong ...

[aufkrawall]

Wattmeter and software tools show the same consumption, copyback prevents using higher quality settings.
Good reasons to assume the numbers are accurate. What do you have to back up your statement? It seems nothing.

[huhn]

my wattmeter shows ~192 for d3d11 native and ~192 for d3d11 copyback.

at least i know now that my PSU is not efficient anymore...

[aufkrawall]

Inefficient PSUs can mask huge power drawing differences. ~192W really is very bad, I got a OC 1070 which doesn't run with very low voltages.
It's no miracle that I got this huge difference between copyback and native in power drawing, since, as I already stated, copyback increases both usage and clock (and thus also voltage).

[huhn]

i have 80 watts in idle that's a lot on that system.

well it's a good efficient PSU at the time i got it and no it can't hide power draw differences.
if something needs more power it will show it it can not simply use more power and show the same number.

for me copyback doesn't increase the clock and it doesn't increase GPU usages.

well of cause it does but it is so small i can't see the difference in the graph.

[aufkrawall]

And why would it increase it for me and not for you?