Sandy Bridge Reviews

Steve Underwood (steveu@dis.org) on 1/10/11 wrote:
---------------------------

>I don't think anyone suspects that Fermi chips miscalculate. I think they are highly
>suspicious that shortcuts have been taken in the calculations, in order to achieve
>a speedup. If it was merely a bug, the issues should have gone away long ago.

It's more fuzzy than that -- there is not just a single way a frame can be encoded for h.264 -- the encoder has (very) many options and it is encouraged to make it's own decisions in making the output look as close to the original as possible while fitting it in the smallest possible space.

The output of the various cpu-based encoders vary a lot, with some being optimized for PSNR, some for SSIM, some using complex psychological optimizations to make the scene look good (while possibly making it look less like the original than some others would), and some just plain suck, producing completely valid h.264 without using most of the options available and making the output look like crap.

The accusation is that the cuda-based encoders belong in the last camp -- while completely legal and valid, they just suck. Then again, so does almost everything else when it's compared with x264.