Are you disabling the HD3200 on-board graphics in BIOS or trying to run with both GPUs active ?
If you are running with both GPUs active then fglrx is your best choice; just make sure you initialize it properly -- something like :
aticonfig --initial --adapters=all -f
You can get the details from the aticonfig --help output.
I don't think the open source drivers can support acceleration on two GPUs at the same time right now, so I'm guessing that's why you're seeing performance problems with radeonhd. If you are disabling the HD3200 in BIOS then please post your xorg log with radeonhd and let's see what we can do about your performance.
The HD3200 is disabled (it disables itself if PEG is detected, and forcing it to enabled disabled the PCIe GPU). So only one card at a time. HDMI works with the HD3200 on fglrx. HDMI does not worth with the 4650 and fglrx.
RadeonHD performance is fine with the HD3200. Just that it's terrible with the 4650. I can yank the xorg.log when I get home, assuming it actually works (about 50% of the time, the system hardlocks with RadeonHD on the 4650 when it tries to initialize X -- no SSH, nothing, though xorg.0.log looks fine after this).
If you are seeing bad performance on 4650 but good performance on HD3200 I suspect your drm driver is too old and is not initializing properly with the 4650. That would result in no acceleration and very poor performance. The log file will be very useful.
Uh, I was pretty sure that radeonhd only supported 2D acceleration on R7xx series cards (experimental 3D support that involves compiling mesa and radeonhd from git with lackluster performance doesn't interest me). At any rate, it's a Jaunty system for now, which should be good on the drm driver. This doesn't solve the fglrx HDMI problem, anyway.
I thought this was for an HTPC? If so, the only thing you NEED is 2D, and 2D accel works out of the box on most new distros using open source drivers (I never much liked radeonhd btw, it always did weird things like you talk about, try just "radeon").
And for that matter, for an HTPC, the performance bottleneck will be the CPU anyways -- the media decoder implementation, what reason do you even have for not just using the onboard 3200?