CUDA acceleration of H.264 in Sorenson Squeeze

Question

I was checking out the Squeeze forums, and I found the following article: Sorenson Media Forum - Mainconcept H.264/AVC CUDA. I suspect it's a "whoops" posting since Squeeze 7 hasn't been released, much less announced on Sorenson's website. The article describes the implications and requirements to achieve GPU-accelerated H.264 encoding.
Here are a few interesting snippets, in case the article disappears:
Mainconcept H.264/AVC CUDA Requirements
In order to take advantage of GPU acceleration of MainConcept H.264/AVC, the following is required:
Squeeze 7 or later.
nVidia video card with CUDA support (any GeForce 8, 9, 100, 200-series GPUs, Fermi, Tesla and Quadro with a minimum of 512 MB.
CUDA architecture 1.1 (any card except of 8800 GTS, 8800 Ultra, 8800
GTX and some Teslas and Quadros). CUDA architecture 1.0 to 1.3 implies a
Tesla video card. CUDA architecture of 2.0 implies a Fermi video card.
Windows, XP, Vista, Windows 7 (32-bit/64-bit).
Minimum nVidia Driver version v196.21 (no Fermi support)
Minimum nVidia Driver version v196.47 (Fermi support)
OS X 10.5.7 or later.
Minimum nVidia CUDA Driver version 3.1.14
Minimum nVidia GPU Driver version 19.5.9f02
The memory requirements are unfortunately not as
straightforward as they would seem.  Each resolution requires a
different amount of memory, but not all of the memory on the card is
necessarily available to the Mainconcept H.264/AVC CUDA encoder. If the
encoder is not able to allocate enough memory, the encoder will fail.
Because of this, meeting the minimum memory requirements is not a
guarantee that a particular video card is capable of encoding in the
specified resolution. Therefore, it is recommended that the minimum
memory requirements should be exceeded when purchasing a new video card.
Selecting an nVidia CUDA-Enabled Video Card
The most important factor is choosing a CUDA card is the number of cores
that the GPU has. The more GPU cores a card has, the better the
performance will be. Although the minimum required memory size is 512
MB, it is recommended to buy a card with more memory.
nVidia provides a list of all CUDA-enabled products at
http://www.nvidia.com/object/cuda_gpus.html.
When choosing an nVidia card with CUDA support, remember that the
“Compute Capability” must be 1.1 or greater and the memory size must be
at least 512 MB.
When choosing an nVidia card with CUDA support for a Mac, make sure that
the Mac’s model identifier is compatible with the nVidia card. For
example, the “Quadro FX 4800” will only work with a Mac with a model
identifier of “MacPro3,1” or “MacPro4,1”. You can find the model
identifier by doing the following:
Go to “About This Mac”
Click on the “More Info” button
Select “Hardware”
Some nVidia video cards actually contain two nVidia video cards
(GeForce GTX 295 or GeForce 9800 GX2 for example). The Mainconcept
H.264/AVC CUDA encoding will use only one half of such video cards.
Multiple, simultaneous Mainconcept H.264/AVC CUDA encodes in Squeeze can
share the same nVidia video card. If there are several nVidia video
cards in your system, multiple simultaneous Mainconcept H.264/AVC CUDA
encodes can use different video cards.
When using two or more NVIDIA cards in the same system, the cards need
to use the same NVIDIA driver to avoid conflicts. For this reason,
GeForce and Quadro cards should not be used in the same system. When
using two or more NVIDIA cards in the same system, it would be best if
they were the same card.
H.264/AVC CUDA Encoder Feature Restrictions
Mainconcept H.264/AVC CUDA Encode has the following restrictions.
Only supports One and Two pass VBR – no CBR or Multi-pass.
Does not support “Interlace Mode”.
Only supports one B-Frame.
Does not support “Use B-Slices as Reference”.
Does not support “Reference Frames”.
Does not support “Multiple Slices”.
Does not support “Black Normalization Level”.
Does not support HRD – no Blu-Ray.
I think the most interesting part is the list of restrictions. As far as I'm concerned, those restrictions make this feature largely pointless, since most of them are what make H.264 so effective at low bitrates and at large frame sizes. I wonder if these restrictions are due to the Mainconcept encoder itself, or just that Sorenson rushed to get this "feature" into the upcoming version of Squeeze without implementing it fully.
In any event, it's interesting to see how the GPU can actually be leveraged for encoding, and not just rendering as it is with hardware-accelerated MPE in CS5.

Answer 1

Tmpgenc uses cuda when frameserving from CS5.
However GPU-z shows a lower percentage then TMPGenc does during encoding.