Adding the PVR-250 to my SN41G2 system cut CPU utilization during Live TV down from about 70% to around 41% on average, which is great, but I was puzzled about why it CPU use didn't drop further. Turns out that Snapstream does both a real-time MPEG2 encode and a real-time MPEG2 decode in order to make its time-shifting feature work. Makes sense, when you think about it. CPU utilization bounces between 0% and 2% with the PVR-250 just recording a show to disk. Then I found this article explaining how to use an external decoder in order to lower CPU utilization in Snapstream. MPEG2 decoders from ATI and NVIDIA know how to use the hardware acceleration features in GPUs to assist with the decode task. In no time, I had NVDVD 2.0 installed, and my GeForce2 MX card was helping out with the MPEG2 decode. CPU utilization dropped by 20%, to about 20% even, when watching Live TV.
Of course, this piqued my curiosity, so I switched back to the Personal Cinema briefly to check CPU utilization with the NVDVD decoder in action. Again, CPU utilization was down about 20%, this time from 70% to 50%. In my book, that change alone makes my SN41G2 perform acceptably without help from the PVR-250's MPEG2 encoder. Should be even better on the SN45G with the 400MHz bus, but I wouldn't know, because I fried the Windows install on that box. Doh.
I played around some with NVDVD's hardware deinterlacing, as well, which the move to the NVDVD decoder enabled. That works, but it does eat up a little CPU time, especially at higher resolutions. I was curious to see how an ATI card would handle MPEG2 decoding and deinterlacing, so I tried to install a Radeon 9600 Pro in the SN41G2. After a brief visit to driver hell (this box has had lots of different video drivers installed and uninstalled on it) I gave up and went back to the GeForce2 MX. Next stop: trying the Radeon on the SN45G with a fresh install of Windows. Perhaps I'll have better luck that way.
Anyhow, the SN41G2 with the PVR-250 and the GF2 MX works well. The last bit of experimenting I did was with screen resolutions. I've seen docs that suggest cutting the screen resolution to 800x600 in order to cut CPU utilization, but having played with it some, I think that's a mistake. You do lose detail, at least with the PVR-250, at 800x600. Going to 1280x1024 at 85Hz with NVDVD's "combine fields" deinterlacing does generate some overhead (about 10% of CPU time on the SN41G2), but it's worth it. After using that ultimate full-motion TV test pattern, NFL football, to test different settings, I'm convinced. This is the best you can do to repair an NTSC signal. Now, the story is probably different when using a composite or S-Video output for display on a TV, but on a computer screen, higher resolutions seem to matter quite a bit.