Could it be because of a tradeoff in quality vs speed? The more you rely on CUDA the more quality suffers?
CUDA transcode quality compared with Intel's Quick Sync: http://www.anandtech.com/show/4083/the-sandy-bridge-review-intel-core-i5-2600k-i5-2500k-an d-core-i3-2100-tested/9
Please note that the transcode quality from the GTX 460 (using CUDA) appears to be the worse of the lot.
This 100 Cuda Core myth is so out of control. There is not a limitation to the amount of CUDA cores the MPE uses or any other CUDA based application. CUDA does not work that way. Think of it like a sprinkler. You have 1 stream of water coming in and a water droplets going to the first available opening. Cuda throws out threads to any Core available. It is not like standard CPU based applications where it has to be multithreaded correctly to thread out to multiple cores. Some benchmark site did a test and came to an incorrect assumption because they did not understand how CUDA really works. There use to be a very extensive white paper on Nvidia's website that explained CUDA memory management and threading. Last time I looked for it , it was no longer there but it may be back up again. Sheesh this has become as bad as someone putting something up on Wikapedia.
Heed what Eric says.