CUDA Accelerates Transcoding for Pegasys

[ShaidarHaran]

Press Release (http://www.techpowerup.com/69536/NVIDIA_CUDA_Delivers_446_Speed_Increase_to_Pegasys _Video_Processing_Solution.html)

Up to 446% performance increase over CPU transcoding with TMPGEnc 4.0.

Now that's useful!

Oh yeah, the option to transcode to VOB is a MUST. Badaboom is a joke compared to TMPGEnc 4.0.

[MfA]

"technology demonstration"

"beta version"

"video decode and processing speed"

Whether it's useful or not remains to be seen ... at the moment in quality for speed you can't beat CPUs, I doubt this is going to change that.

I'm afraid this whole PR push for GPU encoding is just going to hurt GPGPU in the end, because it fundamentally can't live up to the hype. Apart from ME, encoding just doesn't suit the GPU at all ... and even at that it will struggle to compete because of it's lack of efficient 8-bit integer support and the fact that the mostly serial search you would perform on CPUs is a lot more efficient than the mostly parallel search you would perform on a GPU.

[ShaidarHaran]

"technology demonstration"

"beta version"

The fact that it's in beta means it's clearly close to productization. It has a version number, for crying out loud!

Whether it's useful or not remains to be seen ...

Isn't that the entire point of a public tech demo? :lol:

at the moment in quality for speed you can't beat CPUs,

Only because this product is still in beta and has yet to be launched ;) It's just a matter of time.

I doubt this is going to change that.

You doubt that GPGPU is a viable alternative to multi-core CPUs for transcoding? Why?

I'm afraid this whole PR push for GPU encoding is just going to hurt GPGPU in the end, because it fundamentally can't live up to the hype.

Few things do live up to the hype. We can thank human nature and sensationalist marketing and complicit media for that.

Apart from ME,

ME?

encoding just doesn't suit the GPU at all ...

:shock: Come again?

and even at that it will struggle to compete because of it's lack of efficient 8-bit integer support

Uh-huh... When does having too much precision hurt?

and the fact that the mostly serial search you would perform on CPUs is a lot more efficient than the mostly parallel search you would perform on a GPU.

Could you elaborate?

[cho]

This is not using CUDA to do transcoding but just some video post-processing .

[aaronspink]

I'm afraid this whole PR push for GPU encoding is just going to hurt GPGPU in the end, because it fundamentally can't live up to the hype. Apart from ME, encoding just doesn't suit the GPU at all ... and even at that it will struggle to compete because of it's lack of efficient 8-bit integer support and the fact that the mostly serial search you would perform on CPUs is a lot more efficient than the mostly parallel search you would perform on a GPU.

As has been pointed out other places, the GPU is primarily good for the decoding part of the transcoding pipeline. This has been effectively verified over at avsforum from what I recall.

It appears that the encoding side on the actual GPU has significant limitations and minimal speed increase over CPU only encoding. What would be nice to see is a transcoding package that took advantage of this and used the GPU features for the decoding portion of the transcoding pipeline and something like x264 for the encoding portion.

[silent_guy]

ME?
motion estimation

What MfA is trying to point out is that there are a number of operations downstream from motion estimation that are very hard to parallelize.

[ShaidarHaran]

motion estimation

What MfA is trying to point out is that there are a number of operations downstream from motion estimation that are very hard to parallelize.

Ah, thanks.

[pcchen]

For video compression, transform (e.g. iDCT) and motion estimation have more parallelism than other tasks. However, for off-line encoders (do not have live streaming requirements) it's possible to parallelize on GOPs, though it would require a much larger memory footprint.

Currently TMPGEnc beta only uses CUDA to accelerate some filtering tasks, such as denoise and color corrections. It also supports using GPU's video decoding capability to accelerate transcoding, which should be quite useful for, say, AVCHD editing and transcoding, but it's not directly related to CUDA (although I suspect that they do not use DXVA, but the NVIDIA provided cuvid SDK for decoding on GPU).

[cho]

ArcSoft's Upscaling Technology, SimHD, Is Being Demonstrated at NVISION 2008

http://www.reuters.com/article/pressRelease/idUS131446+25-Aug-2008+BW20080825