Neo: Hehe, I'm always up for a flame war
I'll have to disagree with you on the PCIE comments though, although I know where you are coming from.
From the basic specs, PCIE x16 will have significantly higher throughput than AGP8x (max 2.5Gb/s per lane, but protocol overhead brings realworld throughput lower). However the big difference is that PCIE is bidirectional and has the same link characteristics going both ways. On AGP, you can transfer from host->card very quickly, but the other direction is extremely slow. With PCIE, you can not only transfer in both directions at the same speed, but both can happen simultaneously!
The big benefit here is for things like video editing where the GPU processes the image using shaders and then transfers the video frame back to the system for writing out to disk or whatever. All of a sudden it's possible to do this with HDTV frames in real time without breaking a sweat. There are other uses as well, such as many games now use rendering to texture heavily and often read the pixels back onto the host with glReadPixels() or whatever. Similarly, going forward to longhorn the OS can make use of the fast transfers for texture page-offs and so forth.
re: your comments about no significant delta between AGP 4x and AGP 8x. The real issue there was that most games/systems have been largely render bound, not bus bound. ie when you crank up the resolution and put on a fast CPU, the graphics card's fillrate was the limiting factor to how fast you could go. So it didn't really matter if your bus was 1GB/s or 2GB/s because the game wasn't sending data at even 1GB/s.
However, as we move forward, pure fillrate is not going to be as important as games move to more complex shaders and really start stressing the geometry pipe by sending up very highly detailed models. This will cause the bus to become a more significant factor, especially as games move from eg. 5k poly models for the characters to 60k or 100k per character, just because they can.
As far as proliferation of PCIE, it's going to be a pretty huge switch. I think our forecast is something like 50% of our volume shipping by years end will be PCIE. The big OEM's are pushing PCIE heavily and by end of next year AGP is not really going to be a significant factor. It's not necessary to jump on the bandwagon right away, but it's going to be a fast transition. Both our and nVidia's roadmaps quickly phase out AGP chips going forward, although we'll both continue to support them via bridge chips.
