Hardware H.264 decoding - do I need it?

I thought I'd share my experience, in case anyone is on the fence about whether to buy a discrete video card for their Linux HTPC build. YMMV.

I finally got around to upgrading to a threaded mplayer build from svn (the most recent release was not threaded). Doing a couple of benchmarks on 1080p H.264 content, I've discovered that software decoding is now about 3 times faster than libvdpau!

My processor is nothing special; it's an Athlon II X3 running at 3.5 GHz. The video card is a cheap NVidia GT220.

Point is, the only reason I'm using a discrete video card in this box is because I wanted hardware H.264 decoding. With the threaded mplayer prerelease, though, it seems that hardware decoding is now obsolete for some applications.

Obviously, if I were gaming, I'd still want a video card, but I'd consider an ATI/AMD card more seriously than I would have before.* If I were building a dedicated HTPC today, I'd strongly consider omitting it. I will, however, note that while an Athlon II X3 is not an expensive processor, it will run hot at 3.5 GHz; I needed an aftermarket CPU cooler. I don't know if a cooler, quieter, slower solution would give the same performance advantages.

Multitasking is another consideration. If you're running MythTV, and continually transcoding video, sharing the playback workload with the GPU makes plenty of sense.

*As a longtime desktop Linux user who occasionally plays games, I have some loyalty to NVidia. They have a long history of providing outstanding driver support for Linux. Their drivers are closed-source, unfortunately, but so are ATI's.

Interesting. What processor/gpu do you have? I can't imagine that 3 cores makes that much difference over 2.

I did my benchmarks using 10 seconds' worth of "Sucker Punch," pulling from 1000 seconds in, more or less arbitrarily.

mplayer -nosound -benchmark -vo vdpau -ss 1000 -endpos 10 Sucker\\ Punch.mp4

vs

mplayer -nosound -benchmark -vo xv -lavdopts threads=3 -ss 1000 -endpos 10 Sucker\\ Punch.mp4

I saw the vdpau version finishing in about 3 seconds, and the threaded xv version finishing in under one. The -benchmark option, combined with -nosound, will have mplayer rush through the frames as fast as it can render them, ignoring the appropriate frame duration and audio sync.