Win7 native GPU accelerated video transcoding

[akugami]

Interesting tidbit from bit-tech.net.

COMPUTEX 2009: Microsoft?s Murray Vince yesterday revealed that Windows 7 features native support for GPU accelerated video transcoding.

Vince announced this interesting development during Nvidia?s pre-Computex shindig, where he joined Drew Henry on stage.

The implementation was seamless as Henry simply dragged and dropped the high definition video file onto a Sony Walkman portable media player in Windows Explorer, where it automatically started transferring the file onto the Walkman.

The demo included an Ion-based machine and another, similarly specced machine without Nvidia integrated graphics ? the Nvidia-based machine finished the transcoding task around five times faster than the Atom-based PC with Intel integrated graphics, taking just over 1 minute 30 to complete.

Vince said that you can treat media as easily portable content with Windows 7 ? ?it?s conveniently capable across multiple devices,? he added.

With native support for features like this, it?s easy to understand why Huang is so excited about Windows 7. ?DirectX Compute is the most important API in Windows 7,? proclaimed Henry as Vince returned to his seat in the audience..

The major question is, will this work on ATI cards? One would assume yes since the technology is called DirectX Compute.

If it supports all video cards then I wonder if this will hurt nVidia's CUDA efforts. ATI's STREAM tech is still in its infancy so I can't really say this will hurt them (ATI). Heck, more like it would benefit them (ATI).

 

[BenSkywalker]

Likely they are using CS which is part of Win7/DX11. I would assume that nV likely already has a reasonable amount of CS functionality up and running while ATi may be a bit behind(maybe not, just guessing based on general GPGPU efforts to date). I would assume that ATi parts will accelerate as much as they can when they have driver support, but I'm not sure they will be able to do as much as nV's current parts. All those extra transistors on nV's GPUs actually are used for something We won't be able to get a good idea of exactly how the GPUs stack up per se because CS will split the load between the CPU and GPU and we are likely to see performance impacted by both the speed of certain ops being completed and what ops each part can complete. They may all end up tied, for very different reasons(although I think it is safe to say that is rather highly unlikely ).

 

[nitromullet]

This is definitely an interesting find. While I don't think GPGPU performance/ability would influence my decision to purchase one gpu over another, it would be nice to be able to use existing horsepower that would otherwise remain unused when not gaming.

 

[Idontcare]

If it's anything like the MPEG2 codec MS used in their free Movie Maker app then the IQ is probably going to be so piss-poor (to anyone who cares about IQ that is) that it won't be worth using...free is free for a reason, if it was worth something then it'd be sold for extra revenue.

But for zillions of consumers who could care less that their crappy looking transcode is due to the transcoder and not the 2" LCD screen they are using to view their videos it will be nice and convenient.

 

[BenSkywalker]

OTOH, VC1 wmv transcoding produces some extremely nice results.

 

[lopri]

Well, Windows Media Encoder 9 is free, isn't it?

 

[Idontcare]

Originally posted by: lopri
Well, Windows Media Encoder 9 is free, isn't it?

Yeah and I wouldn't touch it with a 10ft pole either when compared to my TMPGEnc 4 encoding when it comes to IQ/bitrate and filesize.

But I am not trying to argue that this matters to many people, as surely >90% could care less about IQ/bitrate efficiency.

My point was just that for anyone who does care about such things they probably aren't going to be overly excited about a native implementation of GPU transcoding, and if you aren't going to care about quality then you could probably get away with current free transcoders without GPU processing anyways as you can just turn the IQ way down and get your video fast anyways.

But I don't mean to undervalue the convenience factor, it will be convenient (and thus value-add) to many many people provided it does a "good enough" job in terms of balancing IQ/bitrate versus processing time.

 

[AzN]

I though MS and ATI had it tight. I guess it changed soon as AMD owns ATI now.

 

[ronnn]

not sure I see the excitement? Dragging and dropping files seems like very old news. Think I can live without seamless file movements for mini screen movies. :laugh:

The biggest difference for me would be actual creative games that have the excitement of something new.

Another shooter with a couple of new effects is about as interesting as terminator 25.

 

[wlee15]

AMD demonstrated their seamless transfer as well in their press conference as well.

http://www.tweaktown.com/popIm..._amdconfctx0916wtr.jpg

from
http://www.tweaktown.com/news/..._the_future_is_fusion/

 

[tuteja1986]

Yeah and I wouldn't touch it with a 10ft pole either when compared to my TMPGEnc 4 encoding when it comes to IQ/bitrate and filesize.

Well VLC is better but for my potable device i use ati encoder since its fast. Quality is bad but you can't tell much difference when watching on a 3.5" LCD screen.

 

[tuteja1986]

I want to decent fast encoder that will encode 40min tv show for my IPOD TOUCH in less than 3mins. Doesn't have to have the best video quality since i will be watching on small screen.

 

[lopri]

Originally posted by: Idontcare
But I am not trying to argue that this matters to many people, as surely >90% could care less about IQ/bitrate efficiency.

You really can't. We're talking about 2"x3" screen here. No sophisticated upscaling, upsampling, etc. necessary.

But there is truth to what you're saying, from a different angle. Because if this is real, many small companies (like Badaboom) can go out of business. I'll believe it when I see it.

 

[ViRGE]

Originally posted by: BenSkywalker
Likely they are using CS which is part of Win7/DX11.

I'm not so sure. MS's GRA-T585 slides from WinHEC don't mention DXCS at all when discussing Media Foundation Transcode, and features small shots of several products such as the Quartics QV1721, which is a video acceleration chip but not a GPU(and hence doesn't use DXCS). As best as I can tell, DXCS is not part of MFT in any way.

 

[Creig]

Originally posted by: ViRGE

Originally posted by: BenSkywalker
Likely they are using CS which is part of Win7/DX11.

I'm not so sure. MS's GRA-T585 slides from WinHEC don't mention DXCS at all when discussing Media Foundation Transcode, and features small shots of several products such as the Quartics QV1721, which is a video acceleration chip but not a GPU(and hence doesn't use DXCS). As best as I can tell, DXCS is not part of MFT in any way.

Congratulations! You've just won today's "Most acronyms used in a single post" award!

 

[Scali]

Originally posted by: akugami
The major question is, will this work on ATI cards? One would assume yes since the technology is called DirectX Compute.

From what I understand, this is mostly a feature of Windows 7, probably built into Explorer or Media Player or something. Sounds a bit like the sync-feature you have on mp3 players.
It 'knows' when you have a video player connected, and it 'knows' what video formats the device desires, so it can fire up a transcoding session on-the-fly.
That's how I interpret this.
In which case you can probably still install whichever codecs you want, for the transcoding. So nVidia, ATi, DirectX Compute or OpenCL should all work with this system...

 

[SlowSpyder]

Originally posted by: BenSkywalker
Likely they are using CS which is part of Win7/DX11. I would assume that nV likely already has a reasonable amount of CS functionality up and running while ATi may be a bit behind(maybe not, just guessing based on general GPGPU efforts to date). I would assume that ATi parts will accelerate as much as they can when they have driver support, but I'm not sure they will be able to do as much as nV's current parts. All those extra transistors on nV's GPUs actually are used for something We won't be able to get a good idea of exactly how the GPUs stack up per se because CS will split the load between the CPU and GPU and we are likely to see performance impacted by both the speed of certain ops being completed and what ops each part can complete. They may all end up tied, for very different reasons(although I think it is safe to say that is rather highly unlikely ).

It looks like AMD includes this in their current Catalyst 9.5 drivers for Vista and Win7.

It looks like the AMD transcoder balances different loads to the CPU and GPU, depending on what would be faster. Is that to hide some of the inefficiencies of their GPU architecture? Or is it done because it really is the better route to take?

My guess is that Nvidia's solution will use less CPU power, but will it be slower? Do we care about 1 minute vs. 1:30? Do we care even if you use the CPU and it takes 4 minutes? I guess I don't get the hoopla over transcoding, but it's brought up a lot here so I guess some people really do care about it.

Since AMD offers this for free in their driver, and it'll be built in to Win7, does that mean Nvidia users won't have to pay $30 for Badaboom soon? I guess AMD is making due with out those extra 400+ million transistors.

 

[WelshBloke]

Thing is as portable media players get more powerful and come with more storage the need to transcode anything is reduced.

Most of the video files on my PC (mostly AVI's) will play on my mobile phone or PDA using coreplayer.

The new parts coming out from VIA/NV should handle HD video no probs.

 

[Scali]

Originally posted by: SlowSpyder
It looks like AMD includes this in their current Catalyst 9.5 drivers for Vista and Win7.

It looks like the AMD transcoder balances different loads to the CPU and GPU, depending on what would be faster. Is that to hide some of the inefficiencies of their GPU architecture? Or is it done because it really is the better route to take?

I think that depends a lot on the underlying GPU architecture and drivers.
I recall a discussion about Folding@Home on Beyond3D, where they said that the ATi client was limited by the CPU (at some point a faster GPU no longer increased the scores), while nVidia didn't seem to suffer the same problem.
So I guess the "better route" depends on the technology you're using. What is better for ATi may not be better for nVidia, and vice versa.

Originally posted by: SlowSpyder
My guess is that Nvidia's solution will use less CPU power, but will it be slower?

I'm afraid we'll never see a true 1:1 solution with proprietary transcoders.
We've seen that the first Avivo for the 4000-series was fast, but the quality was very poor.
Badaboom didn't offer 'perfect' encoding quality either, but it was considerably better than Avivo, and quite acceptable.
Someone at Beyond3D did some experiments with very low quality settings on a CPU transcoder, and found that you could get very close to the performance of GPU transcoders as long as you were willing to sacrifice enough quality. There are a lot of corners that can be cut with lossy compression.

I think if you REALLY want to compare, you'll need to wait until there's a third-party transcoder based on OpenCL or DX Compute, so you know that both GPUs will run the exact same code, and produce the exact same results.
It's not all about speed.

Originally posted by: SlowSpyder
Do we care about 1 minute vs. 1:30? Do we care even if you use the CPU and it takes 4 minutes? I guess I don't get the hoopla over transcoding, but it's brought up a lot here so I guess some people really do care about it.

I don't care about video encoding much in general, but I find it interesting as it is one of the first 'mainstream' applications for GPGPU, and it can be an interesting benchmark... which GPU is the fastest?

 

[WelshBloke]

Originally posted by: Scali

I don't care about video encoding much in general, but I find it interesting as it is one of the first 'mainstream' applications for GPGPU, and it can be an interesting benchmark... which GPU is the fastest?

At something you dont care about?

 

[Scali]

Originally posted by: WelshBloke
At something you dont care about?

No, I care about processor performance in general. I'm an academic and enthusiast, see.
I like to know about performance characteristics, strong points, bottlenecks etc.
So I don't just want to know which one is fastest, but also why. The benchmark simply serves as an investigative tool, that is its purpose. The fact that it also produces a transcoded video as output is just a side-effect for me.