I don't get that, whats wrong with the 256bit bus? Its no worse than a G80 384bit bus as long as the frame buffer vram isnt saturated. That's only a big issue with high res and antialiasing. Bandwidth is still limited on both to 64bit per channel at core clock speeds, 4x64 for G92 and 6x64 for G80. Overall bandwidth the way they like to quote it, is less important. Modern game rendering produces pixels that are resampled and looped, moves much of the storage to on-chip registers and caches. Couple that with the G92's increased number of texture address units TMU, twice the sampling per clock of the G80. Also the G92-GTS if it exists, will have another 8 TMU from the 8th shader cluster, 64 in total.
Just do the math on the texture sampling rates and shader ops.
G82-GTS
Sampling 650 x 64 = 41.6 GTexels/s
Shader Ops 128 x 3 x 1625 = 624 GFlops
G80-ULTRA
Sampling 612 x 32 = 19.6 GTexels/s
Shader Ops 128 x 3 x 1500 = 576 GFlops