Wow, a lot of unfounded rumors and "I had a dream that the R520 will totally own the G70"... 🙂
We know what the G70 is capable of, we know r520 is 90nm and has been delayed again and again. And we know that yields still aren't very good (if they were, they wouldn't need to delay to stockpile cards). And finally, we know the R520 will use a dual-slot design. (Admittedly, that might have changed, since this information dates back to before the G70 launched)
My guess, based on this:
1: 90nm is tricky. Look at the Prescott. Pioneering smaller processes with high-end GPU's is tricky. Look at the NV30. There's a reason they usually introduce new processes in low-end/mainstream cards first. In other words, ATI is taking a lot of risk here. (One might wonder why, since they've always been more cautious than NVidia in this respect). But the question is, will it pay off for them?
2: The fact that they're taking so long to stockpile to avoid a paper launch most likely means that they're currently only picking the top cards. In other words, an "average" r520 is not enough for them. They obviously feel that they need to launch the r520 with all pipelines working or with extremely high clock speeds. That alone means that it won't be head and shoulders above the G70. If it was, they could get good yields by launching a card with pipelines disabled or the speed lowered to begin with. After all, at the moment, they only need to catch up with NVidia. Once they have a card out that comes near the G70, they can always start worrying about making something even faster. It's possible that the R520 will end up faster, but it won't be a lot faster.
The dual-slot fact seems to support this too. NVidia is actually taking it easy, they've launched the chip at speeds where a single-slot solution suffices. ATI seems to feel they need a dualslot solution just to keep up. Wouldn't it make sense then for NVidia to clock the hell out of the G70, pop a dualslot cooler on it, and release it? What is ATI going to do to respond? Invent triple-slot? 😉
Now, another piece of speculation. The G70 is absolutely huge, with over 300 million transistors. What do they need all that for? Surely, it's not just for the extra 8 pipelines and those minor improvements? Some sites have speculated that it's actually a 32 pipe design. If that's the case, I wouldn't bet on R520 being the faster architecture. But of course, this is just speculation based on the transistor count.