[Toaster]

To make big gains in video throughput, the AGP port has to go.
This would mandate a "true" CPU bus to video bus. In the days of yore, this was called "the local bus". While the idea was sound, the actual use was far from ideal. AGP went a step further by increasing the clock rate and more closely coupling the video bus to the CPU bus.
IMHO, they have this backward.

Lets take SGI (silicon graphics) for an example.
In thier entry level 24 bit adaptor of yore, they used similar approaches as now on the Pee-Cee. However, to gain REAL bandwidth, they redesigned the video bus to work along side the CPU. Now, the CPU and the video interfaces act as one at the speed of the CPU. Couple this with a 64bit data path and performance becomes obviously fast. Next, they dedicated several "Geometry Engines" or "local processors" if you will so that the vid card is now fully independant of the CPU. The video even has the ability to control the system bus to move its data.
This could be seen as a sort of "bus master" but actually the system is totally alien to Pee-Cee`s. Now, its as if the system has a SBC (single board computer) that does video duties. This SBC has multiple "X" and "Y" and multiple "plane" acceloration. The main CPU is almost completely out of the picture as everything is handled locally by the SBC video adaptor. So, in essence, a computer within a computer.
Nvidia is on the right track so far. Thier NV20 now boasts more transistors then the Pentium 4. Still, thier graphics chipset relies on a 32/16 bit transfer bus operating at 66mhz. Reinvent the video bus because from here on out, you are looking at 5-10% increases at best.