Entropy said:
And what exactly guarantees that 2*64 bit vs. 4*64 bit doesn't have better granularity overall? (***edit: replace with any other possible split values, the sense of the question remains the same).
RV530 might be based on the R580 characteristics but in my mind it was designed as a mainstream part to the R520. The "real" mainstream GPU I would compare to X1900 will be RV560 and there you roughly have 1/2 * R580. In my mind for today's games the RV530 is too fill-rate constricted and by the time multitexturing fillrates will matter way less, it will be most likely dated. I'm not implying that RV530 isn't an excellent design one bit; it's just that the concept behind it is a tad ahead of it's time.
What I'm trying to say here is quite simple: each IHV targets within a product line X amount of performance for a high end GPU and normally X/2 performance (more or less) for the mainstream part of that line.
Depends if they're targetting half the X1900 performance or more. I'm generally against concentrating on specific factors on each GPU, one has to view each GPU as an entire design.
The way I see it ATI is most likely targetting the G73 both in terms of performance as in final street price with the RV560 and I don't see why a 128-bit bus would stand in it's way for achieving that. In the meantime the X1800/X1900GTOs will in a relative sense aim to fill that gap. For which theoretical ALU throughput ignored, the X1800GTO doesn't seem to slaughter the G73 exactly either, despite it having 43% more raw bandwidth and 34% more Pixel fillrate.
What I'm claiming is that we can see that RV530 to R580 performance scales with bandwidth and not processing power, for most applications. Whether this has to do with inherent limitations difficult to assess from the outside (such as actual bus efficiency vs. theoretical bandwidth) or that usage patterns do not match your assumptions, I just don't know, and make no claims. 
