A Basic Overview of GPU HD Video Processing

CHARTS UPDATED January 7, 2011!

With all kinds of questions about what GPU can accelerate what codec? and how much acceleration does it offer? I’ve been working on helping to answer these questions in various revisions of this guide off and on since 2008…

Updated with Flash 10.1 and HDMI 1.4 information!

The Table and Notes

UPDATED January 7, 2011

Current GPUs

Moderate = Most of the steps in the decocde pipeline are offloaded, however the last steps (which can be some of the most intensive) in the decode process are not. These GPUs lack acceleration for the IDCT [Frequency Transform] and the VLD or CAVLC/CABAC [Entropy Decode] portions of the decode process, they only offer Motion Compensation [Pixel Prediction] and Deblocking.

Advanced = These GPUs accelerate the IDCT portion of the codec, but stop short of full hardware acceleration of the entire decode process.

Full = As the name implies the decode process is fully accelerated, leaving only minimal overhead for the system’s processor to handle.

Advanced HD Deinterlacing = this refers to two technqiues for 1080i sources:
1. An enhanced motion adaptive deinterlacing for video sources.
2. Proper inverse telecine (3:2 cadence) detection for film based sources.

Advanced Deinterlacing Product Names:

• ATI has two modes of advanced deinterlacing: the highest quality is ‘Vector Adaptive’ followed by ‘Motion Adaptive’. Vector Adaptive (VA) is the level needed for proper adaptive deinterlacing of all types of 1080i content.
• NVIDIA’s advanced deinterlacing is referred to as ‘Spatial-Temporal Deinterlacing’ in various PureVideo documentation.
• Intel’s advanced motion-adaptive deinterlacing is folded into their ‘Clear Video’ moniker.

If deinterlacing for a type of content is not working it is noted in the column. Driver updates may enable advanced deinterlacing features at a future time.

Legacy GPUs

Performance and Misc. Notes

Standard Definition notes:

Generally, most common graphics chips available today are more then capable of decoding 480i material, deinterlacing it, and post-processing it. Even integrated video chipsets like Intel’s G35 scores near perfect on the HQV tests, ATI’s 690G scores in the 80s, and the newer 780G scores towards the top with a score of 113 [out of 130].

MPEG-2 Decoding Notes:

With MPEG-2 being pretty easy to decode on modern CPUs neither NVIDIA nor ATI were in a hurry fully accelerated the MPEG-2 decoding process. Both vendors did the acceleration in their pixel shaders up until very recently. With introduction of the ATI Radeon HD 4000 series and its new UVD 2 engine did ATI move to dedicated silicon to handle MPEG2 decode much like the H.264 and VC1 support the earlier Radeon HD cards had.

GeForce 7-series H.264 Decode Notes:

In the GeForce 7-series H.264 decode performance was all done in the pixel shaders and so it is highly dependant on GPU speed, ~500MHz core clock or faster is best for 1080p content.

Post-procesing Notes:

Both ATI and NVIDIA appear to do advanced deinterlacing and 3:2 pulldown detection in the pixel shaders. This means that underpowered graphics cards may not be able to utilize the advanced features. There is some level of correlation between shader power, memory bandwidth, and the ability to deinterlace 1080i properly and/or layer on additional post-processing such as noise reduction. For example the Radeon HD 2400 Pro was widely known to not have enough horsepower to deinterlace 1080i properly.

Noise reduction and image sharpening, while less useful in HD sources, are tested by the HD HQV test disc. NVIDIA’s processing is done in the shaders while ATI’s Radeon HD 2×00 and newer have dedicated silicon borrowed from ATI’s portfolio of Xileon digital TV products.

NOTE: NVIDIA currently doesn’t have noise reduction/image sharpening of HD sources enabled in their Windows XP drivers and likely never will.

More Information:

Wikipedia now has a very through article on NVIDIA PureVideo and the different generations of the hardware and what decoding they support, it also discusses Linux support for video acceleration

http://en.wikipedia.org/wiki/Nvidia_PureVideo