madshi bug tracker

View Issue Details Jump to Notes ] Issue History ] Print ]
IDProjectCategoryView StatusDate SubmittedLast Update
0000232madVRbugpublic2014-08-31 18:542014-09-11 10:37
Reporterboday 
Assigned To 
PrioritynormalSeveritymajorReproducibilityalways
StatusclosedResolutionno change required 
PlatformNUCOSWindowsOS Version8.1
Summary0000232: Unreasonably high rendering time with HD 5000
DescriptionJust bought a Intel NUC D54250WYK powered by i5 4250U with HD 5000. Installed Windows 8.1 with all system updates, all drivers are recent versions (July 2014).

Playing any video with MPC-HC with madVR gets unreasonably high rendering time, and thus tons of dropped frames.

For example, playing a normal 10-bit 1080p video without any image scaling gives roughly 40 ms of average rendering time, which on my laptop with HD 4000 this should be well within 10 ms.
Steps To ReproduceMPC-HC version 1.7.6
LAV filters 0.62.0
madVR 0.87.10
XySubFilter 3.1.0.697

madVR using default settings.

Open any video file with popular format (MKV etc.) will get a high rendering time and very likely excessive frame drops.
Additional InformationI have tried the following but won't help:

 - Use different versions of graphics driver as old as 10.18.10.3412 in Feb 2014 to the latest version;
 - Change all scaling algorithms as low as bilinear or even nearest neighbor;
 - Disable smooth motion;
 - Enable/disable "windowed overlay", "fullscreen exclusive mode", "use separate device for presentation/DXVA processing";
 - Enlarge CPU/GPU queue size;
 - Any settings under "trade quality for performance".

Video playback with madVR disabled (using EVR-CP), or using other video players (mpv and VLC) seems perfectly fine.

This is a very strange issue since this NUC model has been out for more than half a year, and I've seen quite a few people using this as an HTPC without complaining. Reviews from AnandTech and MissingRemote also report it works reasonably well with madVR.
TagsNo tags attached.
madVR Version0.87.10
Media Player (with version info)MPC-HC 1.7.6
Splitter (with version info)LAV 0.62.0
Decoder (with version info)LAV 0.62.0
DecodingSoftware
Deinterlacingnone (progressive)
DXVA2 Scaling Activeno
Aero / Desktop CompositionOn
Problem occurs with modeall modes
GPU ManufacturerIntel
GPU ModelHD 5000
GPU Driver Version10.18.10.3643
Attached Files

- Relationships

-  Notes
(0000659)
cyberbeing (reporter)
2014-09-08 02:16
edited on: 2014-09-08 13:53

The problem you're describing sound similar to what occurs when there is not enough GPU memory bandwidth for baseline functionality in madVR. For discrete GPUs, this usually occurs around 15-17GB/s. For iGPUs and APUs, you'll need to aim higher than that, since bandwidth is shared with system memory.

What is the system memory bandwidth like on your NUC? I believe you need a bare minimum of 20GB/s to not severely bottleneck an Intel iGPU with madVR.

A simple test is to see if render times improve significantly with 8bit 1080p video compared to 10bit 1080p video. There should be only a negligible difference between the two, unless you are bandwidth constrained.

(0000660)
boday (reporter)
2014-09-08 12:21

@cyberbeing

Oh, that makes a lot of sense.

I only bought a single 4 GB DDR3L 1600 ram for the NUC as I thought that is more than enough for just a HTPC, but the problem is in this way I can only get 12.8 GB/s memory bandwidth. It seems that I need to bought another ram to form a dual-channel and get memory bandwidth doubled.

So I tested several 8-bit and 10-bit files and it looked like the case: rendering time for 8-bit files are mostly within 10 ms, but for 10-bit files they are about 25 ms. (Again, no imaging scaling, just chroma upsampling here)
(0000661)
boday (reporter)
2014-09-08 20:59

@cyberbeing

Now I think this is not a madVR bug, so I don't know if it's still OK to discuss it here, but just out of curiosity, I tested some files again on another laptop with i7-3520m and HD 4000, with- or without the second memory chip installed. The laptop is using DDR3 1600 MHz memory, so with single-channel I get 12.8 GB/s bandwidth, and with dual-channel I get 25.6 GB/s.

I can see the performance difference between the two conditions, but I can still see the rendering time are quite different (like 4 ms vs 10 ms), not like what you described "only a negligible difference", though I can see the GPU load is reduced to half in some cases thanks to the higher bandwidth.

And what really bothers me is that even working on a single-channel with only 12.8 GB/s bandwidth, madVR still works better on this laptop than the NUC: it can play most of my files without problem if I lower the settings a bit, but the NUC can't even smoothly play with the lowest setting of madVR. I think HD 5000 on the NUC should be significantly better than HD 4000, so I don't know how this could be explained, and it really made me hesitated to buy another memory chip for the NUC.
(0000662)
cyberbeing (reporter)
2014-09-09 03:03

It's not abnormal for faster GPU hardware to have higher requirements for memory bandwidth, in order to keep those extra resource properly fed. Though that is just speculation, as I've never used a Intel HD5000 myself.

At this point, I think there is no question your HD5000 performance is being constrained in some way by memory bandwidth. Adding another memory module would be the obvious next step, but of course that would raise the overall cost of your NUC quite a bit...

Since you have a HD4000 laptop handy, you could do some 3D gaming benchmarks against your HD5000 and see if the relative difference matches up to what review sites report.

You can also try running the following legacy tool from madshi, which measures various texture upload/download speeds:
http://madshi.net/madNV12Test.zip [^]

Or this similar 3rd party tool could also be useful for direct comparisons:
http://www.kegetys.net/dl.php/D3Dbandwidth.zip [^]


The 'Missing Remote' reviewer is Doom9 member 'andyvt':
http://forum.doom9.org/member.php?u=177082 [^]
You could try PM'ing him and see if he still has his NUC hardware, and is willing to perform some additional tests comparing single and dual-channel madVR performance. That should rule out one way or another if your results in single-channel mode are 'normal'.
(0000664)
boday (reporter)
2014-09-09 19:34

That is very helpful. Thanks very much!

So I bought another memory module anyway, and the performance just got much better. However,...it's still inferior compared with my laptop.

I ran the tools you suggested and got very close results, while the NUC is marginally better. Then I ran the 3DMark Ice Storm benchmark and all look OK: my laptop actually did better than the NUC, but the scores are all in line with other similar systems as suggested by 3DMark website and in Notebookcheck benchmark list.

So I think I just set my expectation too high for this HD 5000. At least for video playback with madVR, it's not that much better than HD 4000.

Thanks a lot for your information, though. I really appreciate it.

- Issue History
Date Modified Username Field Change
2014-08-31 18:54 boday New Issue
2014-09-08 02:16 cyberbeing Note Added: 0000659
2014-09-08 12:21 boday Note Added: 0000660
2014-09-08 13:53 cyberbeing Note Edited: 0000659 View Revisions
2014-09-08 20:59 boday Note Added: 0000661
2014-09-09 03:03 cyberbeing Note Added: 0000662
2014-09-09 19:34 boday Note Added: 0000664
2014-09-11 10:37 boday Status new => closed
2014-09-11 10:37 boday Resolution open => no change required


Copyright © 2000 - 2012 MantisBT Group
Powered by Mantis Bugtracker