The good news is I've got my mythtv box up and running and its recording and playing HDTV and pretty much everything looks great, with only a few minor issues.

My system specs:
Abit NF-M2 (socket AM2) motherboard
AMD X2 3800 64-bit
512MB DDR2 800MHz RAM (1 stick)
WD 7200RPM SATA Harddrive
Kworld ATSC-110 HDTV tuner card
Knoppmyth (most recent release /w Mythtv 0.19)

When I was researching the specs I would need, people had reported that the AMD X2 3800 is more than fast enough to decode 1080i in realtime, and that its performance was actually better than using XvMC. I find, however, that the software decoding is a little jumpy (barely noticeable) and XvMC is visually flawless. I would be perfectly happy with just using XvMC, but it has a minor audio problem where every minute or two I hear a minor pop in the audio stream (using AC3 over S/PDIF). This audio anomaly goes away if I disable de-interlacing, but it just doesn't look as great without it. (on a side note, the on-screen display is now gone with using XvMC as well. not a huge deal, but kind of strange).

So my main questions are: Has anyone had this sort of audio "pop" noise when using XvMC (and how do you get rid of it)? and...
What is the bottleneck when running w/o XvMC? 'top' shows the CPU is way less than 100% either way (usually around 55% for software decoding).

I am using the on-board video chipset, which shares from the main system RAM, so I was thinking maybe it is competing with the CPU to access the memory bus, but I am using the fastest DDR2 memory available - so not sure how likely that is.

Any suggestions?

I have some more relevant information regarding CPU utilization:

I did some CPU usage profiling while mythv was doing other tasks such as commercial flagging and transcoding, and I noticed that while commercial flagging, once again only 50-60% of the cpu's resources are being used. Transcoding, however, is able to use a the full 100% the entire time. Does anyone know what the key differences are between these two processes (in terms of resource requirements)? Could the commercial skip detection be using the on-board graphics somehow, which might indicate that is where the bottleneck is?

At the time of these activites, I was not watching any shows, nor were any recording in the background. Any information would be appreciated. Thanks!

Check the CPU usage option in the backend setup. It's probably set for "low CPU". If you look in "top", you'll see a "nice" value (process priority) of 17. See what the transcode runs at, and make them the same. 0 is normal. 17 is being very "polite". The minimum value is 19, which is "only if you don't have anything better to do, otherwise I'll just wait here and not bother anyone".

If you find that commflag still uses less CPU than transcode when the "nice" values are equal, then I expect it's due to the commflag being more I/O intensive. Or rather, less CPU intensive. Transcode has to do more transformation. It's going to be manipulating every frame, so it's CPU-intensive by nature. Commflag is just searching through, looking for breaks. It's going to run out of work pretty quickly, and need to get more from the disk.
I HAVE NOT LOOKED AT THE CODE!! So take this with a grain of salt.
But I think I've got an idea of what's going on here.
Suppose you've got two people looking through a manuscript of "gone with the wind". The pages are in a filing cabinet in the next room, and they can only have one page in the reading room at a time.
They each have a kid running to the filing cabinet for them.
One of the readers is tasked with highlighting the parts having to do with sex.
The other is tasked with translating the book into French.

As you look at them, the French translator is busy all the time. But the sex censor speeds through his pages, but then spends a lot of time waiting for the kid to fetch the next page. Plus, his kid seems to always be in the way, blocking access to the file cabinet, and making it hard for anyone else to use it. He's only 50% busy, but he's pretty much monopolizing the filing cabinet.

That's how I envision all of this, anyway. As Dennis Miller says, "I could be wrong".

I agree that the processes must be different in the way they perform, but I can say with 99% certainty that the hard-drive is not the limiting factor in the case of the commflag. The reason I'm looking with interest at the commflag process is because it has similar characteristics (at least in terms of CPU usage) to the mythfrontend playback functionality. Both top out around 50-60% CPU utilization. I realize that the commflag is marked as 'nice', but even a nice process should use 100% if there is nothing using the CPU anyway (and this is what I see with transcode process). I'm wondering if perhaps the commflag is using video card resources to do its 'blank screen' detection, and that its bottleneck... or maybe with that type of decoding the memory speed is holding it back. Has anyone successfully used the 6150 on-board Nvidia with HDTV decoding?

Well, I've done some more testing. I bought a couple sticks of DDR2-667 RAM (dual channel) to see if that would improve my speed and I noticed no discernible difference. I then put in a PCIe x16 Nvidia 6200 LE graphics card /w 128MB ram and disabled the on-board video. Again, this also didn't improve the software performance (I put this in at the same time as the faster RAM). However, when using this alternative video card and XvMC, my audio popping problem went away. So this is the setup I'm going with now. XvMC gives me smooth enough video and without any wierd audio anomalies, I'm satisfied for now.

I'm still not sure why my CPU won't go above 60% when decoding mpeg2, but as I try more things, I'll keep posting to this forum to let everyone know.