Yeah, as much bandwidth as the big dogs have today (16 and 20 GB/Sec), it still is not the "max possible" with today's technology: pair the fastest ram with a 256 bit bus. Today, that would mean 32 GB/Sec. This fall, we're likely to be looking at 600 or 700 Mhz DDR, pulling theoretical bandwidth well over 40 GB/sec if a company can build a card that can successfully run the fastest memory available on a 256 bit bus.
It will be interesting to see which company gets there first.
The General consensus seems to be, that with 8 pixel pipelines, that is enough to satisfy a 256 bit wide bus at a synchronous frequency. So, this means either an 8 pipe chip running at 500Mhz +, or a 16 pipe chip running at half the frequency. Of course, with drastic changes in architecture, all bets are off.
ATI has the following challenges:
* Move to 0.13u, and build a chip that has the fillrate to warrant that kind of bandwidth. (500Mhz + 8 pipe, or 300 Mhz+ 16 pipe.)
* Build a PCB that can handle those frequencies. (500 Mhz+)
nVidia has these challenges:
* Design a 256 bit bus for their chips.
* Redesign their 500 Mhz PCB to handle that bus.
I personally think ATI has the easier road at this point. Mostly because the R-300 has given me confidence in their ability to design complex chips with high clock frequencies for a given process. nVidia has not nearly impressed me as much with their ability to control power consumption and general PCB design.
I think the R-350 will fall short. I don't expect any faster than 400 Mhz DDR for ther R-350. NV35 (if MuFu is to be believed) has a chance, though I think R-400 has the best chance.
