Why does CUDA rendering not work well for me??

[BD231]

Why is it that every time I try and encode/compress a video using CUDA it comes out looking horrible?? Typically I'll download specific episodes of TV shows and compress them but sadly I cannot find any combination of settings that would make for a decent compression. Video comes out looking horrendous, to the point where I'm wondering why CUDA is even a feature given the terrible results?? I'm using a GTX460 and MediaCoder.

So far, from a compression standpoint, CPU based compression via FFmpeg gives me the best results. MEncoder comes in second because it reduces the amount of time I need to compress a video by a factor of three but you do sacrifice a little quality.

What's your favorite encoder of choice and why? And, Is there any way to get decent video quality out of CUDA compression?

 

[imaheadcase]

If you are downloading a video..its already compressed. Compressing a compressed video results in garbage..

 

[DaveSimmons]

CUDA is a data-parallel processing toolkit, not an encoding scheme.

CUDA has no rendering or encoding quality. It can be used to write encoders for any format including a lossless one.

 

[1h4x4s3x]

That's the nature of GPU rendering. The image quality is crap.

 

[mosox]

AMD's APP is better lately (ati stream used to be worse than CUDA) but the CPU encoding has no rival.

http://www.tomshardware.com/reviews...d-app-nvidia-cuda-intel-quicksync,2839-9.html

Try X264, it's open source and the best there is. That is if you have a good source.

 

[nenforcer]

File format transcoding speed is supposed to be greatly increased by using GPGPU parallel techniques.

However, that says nothing about its correct implementation. Elemental Technologies finally released Badaboom 2.0 almost a year after announcing it.

I don't know the reason but it may have something to do with the difficulties transcoding between two different lossy formats.

 

[darckhart]

afaik, it's because you're transcoding and not actually encoding.

 

[Sheninat0r]

That's the nature of GPU rendering. The image quality is crap.

I can't tell if you're serious or not. Elaborate?

 

[Idontcare]

Checkout this comparison done by Anand: http://www.anandtech.com/show/4083/...-core-i7-2600k-i5-2500k-core-i3-2100-tested/9

 

[BD231]

If you are downloading a video..its already compressed. Compressing a compressed video results in garbage..

One would assume you'll lose some image quality of course, but my CPU still does a much *MUCH* better job.

Checkout this comparison done by Anand: http://www.anandtech.com/show/4083/...-core-i7-2600k-i5-2500k-core-i3-2100-tested/9

... that explains it.

 

[WelshBloke]

Checkout this comparison done by Anand: http://www.anandtech.com/show/4083/...-core-i7-2600k-i5-2500k-core-i3-2100-tested/9

I was looking for that article for ages... in the gpu section .

 

[Idontcare]

I was looking for that article for ages... in the gpu section .

Same here, even trying to use the search function on the home page was not getting me where I wanted to be.

Finally just broke down and started manually skimming through recent articles to find out where Anand stuffed that nugget of goodness.

 

[Phynaz]

Checkout this comparison done by Anand: http://www.anandtech.com/show/4083/...-core-i7-2600k-i5-2500k-core-i3-2100-tested/9

How does this compare to the output you get using CUDA with TMPEGEnc?

 

[Idontcare]

How does this compare to the output you get using CUDA with TMPEGEnc?

TMPGEnc implements all their CUDA accelerated transcoding features such that they are IQ identical.

The tradeoff is that you don't get the hyped 10x speedups in transcoding times, and the implementations are limited to filters like video noise reduction and so on for the transcode process because those are the only places in the process where you gain a speedup benefit to using Cuda over the CPU when requiring identical IQ output.

So in TMPGEnc if I am doing just a straight change in bit-rate transcode, say I want to take my 7500 kb/s DVD quality movie and reduce the bit-rate to 3500 kb/s so the filesize is smaller...TMPGEnc will run this entire transcode job on the CPU and not use any CUDA.

But if I am taking a 7500 kb/s movie and it is grainy as all hell (maybe a home video shot with cheap low-quality vidcam or is a TV show from the 70's) so I want to reduce the video noise as well as reduce the bit-rate to 3500 kb/s then TMPGEnc will use CUDA for the video noise reduction and the CPU for the bit-rate reduction.

In this case the result is an identical bit-rate output and IQ as verse to doing the same transcode job with pure CPU and no CUDA but the time required for the CUDA-assisted transcode job will be around 25-40% faster (depending on the specific GPU and specific CPU) than the pure CPU job.

But you pay for the privilege of not having a craptastic implementation of CUDA-assisted transcoding, TMPGEnc is around $100 and it doesn't come with DVD/Blu-ray authoring capability so you need another $100 app for decent authoring capabilities.

 

[metalmania]

CUDA is not bad. Just use the right software and right settings, you can still get good IQ.

 

[Phynaz]

Thanks IDC.

I've been a TMPEGEnc user since about 2004. My current project is converting 44 DV tapes to DVD compliant mpeg system streams for authoring.

I was kicking around the idea of picking up a cheap used CUDA capable card to help things along, but I'm not doing any filtering at all. No use for it then.

Quality optimized encodes are running at 4x real time on cpu anyway.

 

[VirtualLarry]

TMPGEnc is around $100 and it doesn't come with DVD/Blu-ray authoring capability so you need another $100 app for decent authoring capabilities.

It's payware? I thought it was open-source. I've never used it.

 

[paperwastage]

Thanks IDC.

I've been a TMPEGEnc user since about 2004. My current project is converting 44 DV tapes to DVD compliant mpeg system streams for authoring.

I was kicking around the idea of picking up a cheap used CUDA capable card to help things along, but I'm not doing any filtering at all. No use for it then.

Quality optimized encodes are running at 4x real time on cpu anyway.

maybe some Sandybridge Quicksync or physical hardware (USB h.264 acceleration, not sure how well it works)

http://www.amazon.com/Elgato-Turbo-2.../dp/B000PCVIEU

but GPU acceleration vs CPU, so far I still prefer a more powerful CPU for more settings adjustability over GPU

 

[1h4x4s3x]

I can't tell if you're serious or not. Elaborate?

I've not read the linked article yet, so, I might be wrong but from what I know GPU rendering is slower than CPU, thus the lower IQ to make up for it.
Video encoding can apparently not be done in parallel or only partially and therefore only a certain portion of the available stream processors will be used.

This is way beyond my knowledge though, just repeating what I've read some time ago.

 

[Phynaz]

maybe some Sandybridge Quicksync or physical hardware (USB h.264 acceleration, not sure how well it works)

Thanks for the suggestion, but...

DVD compliant streams are mpeg2, not h.264.

Also, raw DV is 30mbs. Using an external device would actually slow me down - a lot.