According to VR-Zone, the rumours that the GeForce 8600 (G84) will be using a 256-bit bus are incorrect; instead the chip will still be using a 128-bit bus, like its mid-range predecessor (G73), a configuration which matches what we have heard. They further claim there will be two versions upon initial release; one at 650/900, and the other at 550/700, both with 64 shader processors.
Assuming the architecture did not change, this would imply 16 samplers and 32 filtering units, and the equivalent of 8 ROPs (but with "free" 4x MSAA in terms of depth tests). Practically, this is rather similar to the figures of the Radeon X1950 XTX (higher in terms of filtering; lower in terms of ROPs, arguably), but this card would have less than half the bandwidth of that SKU.
Furthermore, it would be assumed that a mid-range card tends to be used at lower resolutions than a high-end one. As you move up in resolution, the overall bandwidth requirements go up, but not as fast as the fillrate increase because the memory coherence also improves (triangle size goes up, relative texture sizes go down, etc.)
However, G80 has significantly more texture cache than any part that's come before it (128KiB of L2, and a fair bit of L1 too). While an efficient texture cache mechanism and hopefully quite efficient framebuffer compression techniques can't hurt, they are also unlikely to do miracles. That's why we speculated NVIDIA would go with 1.1GHz GDDR4 for the higher-end model, but perhaps the prices on these memory chips are still too high. Either way, it remains to be seen how bandwidth limited G84 will be, should these specifications be accurate.