Evaluation of HEVC decoders (SW, Hybrid and HW)

[pirlouy]

at least i know now that my PSU is not efficient anymore...
I wasn't aware of PSU being less efficient with time. Are you sure about this ? FWIW, my first wattmeter was wrong.

[huhn]

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

good question maybe power settings or PCIe speed...

I wasn't aware of PSU being less efficient with time. Are you sure about this ? FWIW, my first wattmeter was wrong.

capacitor age there is no question so the efficiency curve changes too. no one really cares about this topic and it takes a long time/heat to kill a good capacitor. this PSU in question is 5+ years old. so with a new PSU i can get an guess how much it is in this single case.

i have currently ~80 watt idle and that has to come from somewhere.

if your wattmeter isn't broken than the number is correct +-5% which doesn't matter because we want to see a difference and that will work totally fine.

[aufkrawall]

good question maybe power settings or PCIe speed...

I suspect you are simply just not testing with demanding 4k files.
I have have the same rendertimes with native vs. copyback for 1080p 30fps video, but not 4k 60fps.
Not very surprising, since the overhead for copyback likely increases with amount of data.

Oh, and the PSU definitely can mask differences in power draw.

[huhn]

no i'm testing with UHD.

[aufkrawall]

Native:
https://abload.de/thumb/nativehksw6.png (http://abload.de/image.php?img=nativehksw6.png)

Copyback:
https://abload.de/thumb/copybackj2spk.png (http://abload.de/image.php?img=copybackj2spk.png)

And of course PCIe isn't a problem. Got a 6700k with fast RAM and games run fine.

[huhn]

sandybridge are common and the PCIe 2.0 is totally fine for games.

in this case copyback "wins" for some reasons higher GPU usages but lower powerstate. if i would run it for a longer time native would most likely look the same:
d3d11 native: https://abload.de/img/d3d11nativehyuu3.png


d3d11 copyback: https://abload.de/img/d3d11copyback56uyr.png

the only clear difference is the bus utilization.

here another one: https://abload.de/img/d3d11native2qkkfd.png

the part with the bus interface load is copyback you can see again that for what ever reason voltages is lower with copyback but GPU load is higher with this current set up. it'S very hard to get good number with pascal GPU they are changing the voltage clock.
as you can see the card is capping because of power usages with both native and copyback (green marker). just changing some settings in madVR and it will look totally different again...

[aufkrawall]

Could you test average values for at least ~30s period?

Native:
https://abload.de/thumb/nativeers1e.png (http://abload.de/image.php?img=nativeers1e.png)

Copyback:
https://abload.de/thumb/copyback60s3t.png (http://abload.de/image.php?img=copyback60s3t.png)

Edit: Stuff like browser etc. should be closed.

[huhn]

i first need to know why your bus load is so different.
and i will try a different none HDR UHD file.
and i may use a different PC too.

500 mhz and 0.25 volt difference...

edit: my other PC with a 960 4% bus load with UHD 8 bit and 6% with UHD 10 bit. i don't have time for more right now.

[huhn]

another day and another bunch of totally different numbers.

i have BUS load with native this time and and up 18 % with copyback Vram goes totally crazy and what not.

but most important the TDP it is relative constant.
960 2 gb UHD 10 bit, random madVR settings to get the highest possible rendertimes without drops.

copyback: https://abload.de/img/d3d11copybackvmj8k.png
don't ask me how i manged to take this screen without video engine load...

native: https://abload.de/img/d3d11nativeipkpt.png
the random 0% usages(error readings or what ever) and the start let this look better than it is.

on the other PC Vram usages was about the same in both mode it is clearly higher with copyback on this system.

i will edit in power consumption later and that's it clocks are more constant on this GPU so maybe i can see 1 or 2 watt more power drain with copyback.

edit:
again a4-5 year old PSU and again gold rated PSU
idle: 57.3 watt
native: ~150 watt highest spike i saw 171 lowest 140
copyback: ~150 watt highest spike i saw 157 lowest 141

[aufkrawall]

Well, I'm out of clues why it would be different. I just checked my PCIe bandwidth with a cuda tool and it seems to be totally normal.

[pirlouy]

FWIW, I've tested again my power consumption. My PSU is 5 years old, and there are no differences despite years.
My result is with classic AVC @4.1 and not HEVC (sorry)

Idle: 67 W
Software: 90-95 W
D3D9 DXVA Native: 85 W
D3D9 DXVA Copy Back: 80 W

I have GPU-Z screenshots if needed, but I know this thread is dangerous so I prefer to avoid problems.

[el Filou]

i first need to know why your bus load is so different.
I also get a 15% bus load with 2160p60 10 bit on copyback (PCIe 2.0 x16), but speed is system limited so it should be higher if it was smooth, something like 20-25.

[el Filou]

Just noticed Chrome has now blacklisted VP9 decode acceleration for all Windows versions before Windows 10 1703 (was 1607 before) due to crashes when seeking. :rolleyes:
Well I guess it's finally time to update to 1703 before 1709 comes out...

[CruNcher]

Interesting im testing GP104 myself now and i currently measuring a difference of 20W on Nvidias Video Decoder depending of the Bitstream Complexity 2160P@60 FPS HEVC

Nvidias Dynamic Boost 3.0 clocks between 60-80W (TSP) depending on the tested bitstream complexity.

a delta of 20W more for certain complexity level (TSP).

Also overall the clock adjustment isn't really fast i rememberd times the clock adjustments reacted faster on Maxwell overall Nvidia seems to have highered the reaction latency for Video complexity changes.

You can see this very well when Seeking it takes a lot of time for the Clocks to lower again, they jump up on the Seeek to full and then it takes around 1-2 minutes on a higher power consumption state before they lower again and slowly stabilize to the bitstream complexity and smaller changes again, in this "turn around" time period Nvidias Engine can draw those 20W more pretty inefficiently.

I tested Nvidias Optimal Power,Adaptiv and Maximum Power states.

they show no real difference in these reaction time but clocking behavior differences very well.

Not sure if this is maybe only a visible issue with this specific Card Design im testing here and Bios and overall driver or upper 3rd Party Software Level i keep an eye on it doing some tests with Nvidias API directly (newer drivers im currently testing with 375.96 pre SDK 8) and see if it is the same overall latency behavior.

It feels really slow currently in Adapting and stabilizing especialy in seek situations from this test (to much overhead power way to long).


Must be even more catastrophic with the Browser HTML5 overhead depending on the OS.

LG/ATEME Encoder Test Production

Stabilized Power Draw (Adaptiv)

http://i1.sendpic.org/i/mB/mB4W6aJojnEyr616Ipx0zjAKZPm.png

Turn around Power Draw (Adaptiv) Seek

http://i1.sendpic.org/i/tZ/tZVLURSgXZPs9TUl00BjOY8iXBf.png

You can also see how the Engine tries to even trigger a Higher Power Draw but luckily it runs into it's overall Voltage Limit.

[aufkrawall]

Boost clock after seekigns seems to occur only with bilinear scaling. Use Windows 10 video app for such tests or wait until madVR supports D3D11VA scaling.

[mariush]

good question maybe power settings or PCIe speed...



capacitor age there is no question so the efficiency curve changes too. no one really cares about this topic and it takes a long time/heat to kill a good capacitor. this PSU in question is 5+ years old. so with a new PSU i can get an guess how much it is in this single case.

i have currently ~80 watt idle and that has to come from somewhere.

if your wattmeter isn't broken than the number is correct +-5% which doesn't matter because we want to see a difference and that will work totally fine.

Capacitor aging would NOT affect the efficiency of the power supply in any significant way. It would just make the voltages less "stable", they'll have higher ripple .. for example instead of 12v +/- 50mV you may have 12v +/- 100-200 mV

In theory that could affect the efficiency of the VRM on the graphics card (the DC-DC converter that takes 12v and produces 1v..1.4v) and motherboard (for cpu and memory) but the variation would be extremely small, maybe less than 0.1%

-

Video cards are pretty smart when it comes to saving power. If there's little load, they'll switch from pci-e x16 to x8 or x4 and even switch from pci-e 3.0 to pci-e 2.0 to save maybe 1-2w of power, and in some cases they may even fall back to pci-e v1.1

Use some tool like GPU-Z to see in which modes the card works , when you're using various settings. It may just turn out that when you're using some setting, the card keeps the DDR5 memory at highest frequency or that the pci-e slot runs at full x16 v3.0 using a bunch more watts.

[huhn]

my power consumption is pretty stable between native and copyback aufkrawall has a huge difference.

here is a german test for aging power supplys.
https://www.computerbase.de/2013-11/alte-netzteile-test/4/#diagramm-effizienz-bei-230-volt-eingangsspannung

as you can easily see there can be a huge difference in low load efficiency.
don't put to much through into that could be a faulty power supply at my side i just take a different PC to check it.

[aufkrawall]

I got a Corsair AX860 Platinum power supply. However, I don't think that the PSU can fully eliminate the differences shown by the card's sensors.

[el Filou]

Is the NVIDIA API for reading power to be trusted when the card is in power save?
My 1050 Ti never shows anything lower than "47,6% TDP" even when clocks, loads, voltages and bus speed are all at the lowest possible values, which I find highly improbable (that would be 36W at idle :confused:).
When there is a small load on the GPU using madVR, the reported power barely goes up to 50-52% TDP, so the idle value can't be right and it's making me doubt the reporting.

[huhn]

you are not using high performance or stuff like this?

my 1060 and 960 fall below 10% TDP all the time.

[el Filou]

I tested with both Adaptive and Optimal Power, and rebooted after each setting change just to be sure, but it doesn't change the reading.
GPU 139 MHz 0.675V 0% load, memory 202 MHz 4% controller load, bus PCIe v1.1 0% load; and it still says 47,6% TDP.
If you correctly see 10% on yours then I think something must be wrong on my card either with the sensor or the calibration in the BIOS or something. I'll report this to Palit just in case.
Thanks for the information!

[huhn]

i would say it is GPU-z in this case.

GPU 139 MHz 0.675V 0% load, memory 202
nothing more i need to know that your card knows how to save power.

[CruNcher]

Boost clock after seekigns seems to occur only with bilinear scaling. Use Windows 10 video app for such tests or wait until madVR supports D3D11VA scaling.

The boost jump itself isn't a real problem the time it needs to recover from it is the overall culprit (the time it lasts in the higher p-state before lowering again, switch latency).

Maybe it's just the driver gonna check this now was very happy overall in all tests with this driver on Maxwell though but pre SDK 8 is rather suboptimal now with Paxwell overall.

The PSU

http://www.game-debate.com/psu/index.php?ps_id=1623&psu=be%20quiet!%20Dark%20Power%20Pro%20P6%20530W

And before anyone asks yes it works stable ;)


Also interesting some visible pattern for the clocks behavior with this bitstream test.

http://leakimg.com/qwW0u53/image.png

The Fan stuff is a known problem with the cards Design semi passive Hysterese config, though not really optimal in this case it stresses the fans to much.

Overall the test shows a lot of inefficiencies for the complexity it is overall way to high clocking, even with all the OSD rendering 80W (TSP) is to much 60W (TSP) should be pretty enough for everything on Screen.

Also the sudden GPU Usage spikes could be Keyframe related, which is interesting.

http://www.tomshardware.com/reviews/best-nvidia-geforce-gtx-1080-graphics-cards,4725-10.html

Either the Driver was still suboptimal for Paxwell (375.95) or something else in the Cards overall configuration is suboptimal.

I try the 2nd Bios Results again and see how those react.

Also trying full DXVA Mode could be that the overall slightly higher stress is not well tuned in Nvidias Adaptiv Boost metrics, so it just overreacts here expecting the worst brakedowns to come (more Interactive Game tuned) ;)

The coil whine of the card is relatively sane i would have expected much worse for the -50 compared to MSIs Design though if you think about those some cents more it would have cost to get the overall Design even quieter it still pains a lot why such decisions are being taken ;)

Maybe i gonna rebuild the card with a friend of mine and show those what can be done for cheap and send it to Tomshardware Germany for another Exclusive Review, some seem to dumb to get 99% Efficient cheaper PSU Designs done and also to dumb to get quieter relatively cheaper Graphiccard Designs done as well. ;)

[CruNcher]

Switch Position 1

Palits OC Bios (I still try to understand the higher purpose of this Bios configuration)

86.04.3B.00.34
1645/1785
1251 mem


Switch Position 2

Palits Default Bios (A not so bad configured overall Bios)

86.04.3B.00.31
1607/1734
1251 mem

The OC Bios is a total disaster i really dont get the purpose of it, it's a total Stability Failure on the Render output side
It is what it is sher bruteforcing on the Cost of overall Stability only to be ahead within Benchmarks and shit Synthetic tests.

The Hysterese is also a pure catastrophe it is inside of both Bios the same nothing was fixed there no official update available on their support side either at 52°C it sets in.
In this state currently i don't want to compare it against any Custom RX Vega 56 even if the non OC Bios is overall pretty sane and has way fewer fluctuations and a more sane idle setup on top and behaves much nicer in a lot of different situations.

http://leakimg.com/b6c2hov/image.png

http://i1.sendpic.org/i/7q/7qnkUjm5sRyKDHCL8LlLF3VGIKH.png

Funnily the 2nd Bios seems to be also used on the Jetstream (sister card)

https://www.techpowerup.com/vgabios/190761/190761

Hmm officially that Bios doesn't exist

https://www.techpowerup.com/vgabios/184340/palit-gtx1080-8192-160601-3

Unverified but this is the Bios Palit has done after it most probably the previous had even bigger hysterese issues and was the release one so when igor found them in the GameRock Premium and they fixed them igor most probably didn't checked every case this could happen but only Gaming related Loads.

https://www.techpowerup.com/vgabios/186892/186892

Hmm

Looks like worth a try aside of the stable bios on switch position 2 putting it on switch position 1 and testing it

https://www.techpowerup.com/vgabios/194017/194017

Bios is in the same TDP Range as the OC Bios of the GameRock (not as new as the super jetstream though) looks promising as a test candidate

https://www.techpowerup.com/vgabios/186322/gainward-gtx1080-8192-160830-2

[Someone]

If You Want 8K 60 Fps Wait Nvidia Volta or AMD Navi
Don't know if it was posted already, but the Win10 1703 App "Movies & TV" CAN play YouTube 8k VP9 60fps (https://www.youtube.com/watch?v=FEpvgVJN33w&index=3&list=PLg1oOJ486INkrllAl_jQ4u7cleHHSucXp) downloaded files very fine with a Nvidia Pascal GC (GTX 1070). BUT it doesn't work in any current chromium based Browser, BUT in Edge with normal playback speed (VPU = ~80%, or playback speed 1.25 = 100%).
DXVA Checker Playback Benchmark = max. fps with this file = ~72fps. VPU = 99%.

The App also plays UltraHD Bluray HDR 10 bit (http://abload.de/image.php?img=2017-09-18-00-51-09ufs78.png) files "correct".
(Sorry for the bad screenshot. There's something wrong with screenshots, since my Win 10 Upgrade, with DSR@1440p.)

[aufkrawall]

The win 10 Video App can't show 10 bit gradients without banding, it distorts colors in general and it doesn't compress highlights for HDR video, which will look like a North Korean nuke test for bright scenes (at least with SDR output, dunno if it supports HDR output). ;)

[el Filou]

My 1050 Ti barely reaches 65 fps in DXVA Checker with 8K VP9 (playback or decode), this is not really usable in practice for 60 fps videos. It would be fine with 24p though.

Interesting: DXVA Checker can decode that video https://www.youtube.com/watch?v=sLprVF6d7Ug in 8K in hardware with LAV in DXVA2 Native mode but in MPC-HC, LAV won't use hardware decoding except with CUVID.
Using latest nightly I tested D3D11 Native + madVR, DXVA2N + madVR, DXVA2N + EVR, and it always falls back to avcodec even though in DXVA Checker it works. :confused: (nevcairiel?)

Anyway, 8K is useless for consumer use now and will be for quite a long time.

[loekf]

Stupid question perhaps, maybe it's somewhere hidden in the 107 pages of this thread.

Yesterday Apple released High Sierra (macOS 10.13). One of the features is HEVC decode backed into the OS. Until 10.12 (Sierra) I couldn't play a 4K HEVC file without severe frame drops in e.g. VLC.

After the update, VLC mostly works and specially IINA is working great. I didn't try Quicktime, because I was too lazy to remux my MKVs into a MP4 container + convert DTS-HD tracks to AAC.

My iMac is a Late 2014 Haswell based 5K one. AFAIK, Haswell can only do "hybrid" decode, so some parts of the GPU assist in decoding (Intel Quicksync). Don't think there's anything done with the standard AMD GPU by Apple. But... only 8-bits HEVC according to:

https://communities.intel.com/thread/59216

The files I tried were all ~30 Mbps, main 10, 4K, 23.976 fps files.

Can Haswell do more than we think or is Apple doing something special ?

I didn't try the test files from this forum yet. I also still have to play with the new AppleTV 4K. An new beta of MrMC is pending approval with Apple. I'm curious what the A10X is capable of.

[huhn]

the first AMD GPU with a hardware HEVC decoder was polaris and i don't see a reason why a true i5 can't decode a HEVC 30 mbit video in software.

[aufkrawall]

Fiji "already" had native HEVC 8 bit decoding. But I agree that 8 bit shouldn't be a problem for the CPU and so neither for a hybrid decoder. Unlike 10 bit, which kills my Skylake IGP.

[NikosD]

The files I tried were all ~30 Mbps, main 10, 4K, 23.976 fps files.

Can Haswell do more than we think or is Apple doing something special ?

That could be weird, if your players uses GPU assisted acceleration for 10bit HEVC for 2+1 reasons:

1) Haswell using Windows drivers can only use GPU for 8 bit HEVC, IIRC

2) Even 8bit GPU assisted HEVC acceleration should be slow, very slow for 4K HEVC at 30Mbps

3) Even CPU decoding could be slow for Haswell i5 when decoding 10bit 4K HEVC, but it's the only case I could think of with some decent performance compared to GPU assisted and the only one offered for 10bit HEVC encoding using Haswell iGPU.

Try to monitor your CPU usage during 10bit HEVC 4K decoding.

Probably Apple has managed to use more optimized CPU HEVC codec than ffmpeg and LAV.

[NikosD]

Did Intel remove 8K support of all codecs for Coffee Lake ?

Also, removed VP9 10 bit ?

Someone with a Coffee Lake, just post a DXVA Checker screenshot please.

https://s1.postimg.org/4dvioem0v3/gen9_lp_video_decode.png

[P.J]

Is the NVIDIA API for reading power to be trusted when the card is in power save?
My 1050 Ti never shows anything lower than "47,6% TDP" even when clocks, loads, voltages and bus speed are all at the lowest possible values, which I find highly improbable (that would be 36W at idle :confused:).
When there is a small load on the GPU using madVR, the reported power barely goes up to 50-52% TDP, so the idle value can't be right and it's making me doubt the reporting.

I had the same problem too. Could you fix it?

[el Filou]

GPU-Z devs say they use NVIDIA's API, so I contacted Palit technical support by email - both Germany and Taiwan - but never received a response, not even an acknowledgment. :(

[huhn]

there is a simple thing i would here.

i would measure the powerconsumption at the power supply and see if it is let's say over 50 watts if not i would ignore it.
and i would compare the powers consumption between full GPU load and idle if the delta is bigger than 50 watt you should be fine.

[aufkrawall]

It shouldn't matter for decoding. ;)

[el Filou]

But it will probably matter for bugs. :p

i would measure the powerconsumption at the power supply and see if it is let's say over 50 watts if not i would ignore it.
and i would compare the powers consumption between full GPU load and idle if the delta is bigger than 50 watt you should be fine.I'm not worried about absolute power consumption really. It's just it would be nice to be able to compare the power consumption of the card with different graphics settings/drivers/power plans.

[huhn]

software is not accurate for that.
just to give you an idea.
voltages
GPU-Z 0.9620
HWmonitor 0.875

it's not a trust worthy source. now take different bios version of each brand and it get's totally out of hands.

i only see this useful for major differences like nvidia is using a really "high" powerstate for hardware decoding.

[pacuro]

Did Intel remove 8K support of all codecs for Coffee Lake ?

Also, removed VP9 10 bit ?

Someone with a Coffee Lake, just post a DXVA Checker screenshot please.

https://s1.postimg.org/4dvioem0v3/gen9_lp_video_decode.png

NikosD take a look at screenshot I have got from colleague. In Coffe Lake there are 8K decoders to 8/10bit, hevc/vp9:

http://static.frazpc.pl/board/2017/10/b89c3b5e23099b35.jpg

[P.J]

GPU-Z devs say they use NVIDIA's API, so I contacted Palit technical support by email - both Germany and Taiwan - but never received a response, not even an acknowledgment. :(

I had that problem with Asus, then it's Nvidia's fault.
Have you tried other apps except GPU-Z?

Edit: It seems TDP has been removed:

http://phota.me/mHHu.gif

[el Filou]

I had that problem with Asus, then it's Nvidia's fault.
Have you tried other apps except GPU-Z?

Edit: It seems TDP has been removedYes they all show the same value.
My v2.4.0 still has it. Maybe it's unchecked in your settings?

Anyway huhn is right and I've given up on trying to get any reliable value that way.

[NikosD]

NikosD take a look at screenshot I have got from colleague. In Coffe Lake there are 8K decoders to 8/10bit, hevc/vp9

Thanks, looks reasonable.

The iGPU of Coffee should be exactly the same like Kabylake's and it is.

[NikosD]

Well, if you don't believe in miracles, this is the time to reconsider it.

Microsoft managed to measure the Video Decoding performance of the AMD ASIC inside my RX 470 card, something that ATI/AMD never revealed in public!

Yes, it is true.

Using latest Windows 10 Fall Creators Update (possibly the least interesting name for an OS) and the new Task Manager, there is a new GPU monitor tool.

I did a lot of tests using an AMD RX 470 card and the latest beta Crimson Drivers 17.40 for Fall Update (WDDM 2.3) and the Video Decoding usage reported by GPU Monitor tool seems very accurate.

I tried H.264, H.265 8bit and H.265 10bit with various decoding difficulty in normal and benchmark mode (DXVA Checker) and the results were reasonable.

Next system will be my Core i3 Haswell and the MFX video decoding engine (QuickSync) which I'm pretty sure Microsoft managed to measure that too.

Very impressive, at least.

[aufkrawall]

Was that not possible before via GPU-Z?
Nvidia driver does report VPU usage to tools who want to read it out for ages.

[huhn]

like he said AMD doesn't do that. not a huge deal to be honest

[aufkrawall]

Does it also work without WDDM 2.3 driver? If not, it won't work on Intel.
And I fail to really say what would be impressive about it. It's one metrics of many, accessible or not by stupid proprietary drivers on Windows.

[NikosD]

It is very impressive because AMD has never exposed officially such info in their drivers.

Tools like GPU-Z and others need to just read a suitable parameter from Nvidia drivers to report that info.

But how Microsoft did that without using AMD drivers, meaning without access to the specific ASIC ?

[el Filou]

Maybe that's part of a list of standard utilization counters that WDDM 2.3 forces GPU drivers that adhere to it to report to the OS in a standard way?

[NikosD]

Does it also work without WDDM 2.3 driver? If not, it won't work on Intel.
And I fail to really say what would be impressive about it. It's one metrics of many, accessible or not by stupid proprietary drivers on Windows.

Maybe that's part of a list of standard utilization counters that WDDM 2.3 forces GPU drivers that adhere to it to report to the OS in a standard way?

Just tried my WDDM 2.0 Intel drivers for Haswell v.4624 and it works fine.

My HD 4400 has a lot faster H.264 hardware video decoding ASIC than RX 470, as we already know from the benchmarks and I can confirm that using GPU monitor of Fall Creators Update.

Intel has already released a new driver v4703 but with no difference.

So, it's not WDDM version but some other low-level routines that can achieve this accomplishment.

[nsnhd]

On my i5-4440/ HD4600 (v.4703), GPU monitor of Fall Creators Update only shows 3D chart, the rest of charts (video decoding, processing etc.) shows nothing.