Did ATI hold back on R520 to make room for R580?

I don't know enough about that to credibly comment. (Although that applies to most of my posts, it applies especially to this "narrow and deep" argument against NV30.) It's a question of how many of which operations you can perform per clock, and as R300 proved more flexible than NV30, so G71 seems more flexible than R580. I suppose this inflexibility is narrow and deep.

Hugely uneducated guess here (see above qualification), but it seems ATI took the huge advance R300 offered and gambled it on a new architecture. They coasted on R300 for three years (granted, 90nm design delays wouldn't have made this "coasting" as pronounced), and now R520 seems midway to their end-goal gamble of getting to unified shaders first. This whole gamble theory will be shot to hell if NV wins again with G71 and then competes well against the unified shader core ATI's due to release later this year (which I'm still doubting) with one of their own.

 

Yeah, that's what I meant by "unnecessary," in a roundabout way. ATI spent all its extra transistors on shaders, whereas NV increased both shaders and texture units.

Yep. Still can't figure why they'd only apply this to RV530 (and maybe RV560), yet leave it out of R580--unless you're right in that future shadow algos won't lean on Z ops as much.

Of course, but that doesn't stop G70 from competing very well against R520 with equal bandwidth. Bandwidth might become the predominant bottleneck (if it isn't already), but it seems that current games that still favor G70's extra texturing units will continue to show that in G71 benchmarks. My seat-of-the-pants wild guesstimate may be wrong, though, and you may be right in thinking that R5x0's ostensibly more efficient memory controller will give it an advantage as we run against RAM speed constraints. Or maybe bandwidth is already a huge bottleneck and (thanks to that) G71's rumored extra pipes will prove useful only with shading, not texturing. But why, then, move up to 24 ROPs as well? For that matter, why does the 7800GT prove faster than the X1800XL in most cases?

I'm deliberately excluding AA, as I'm sure most reviews will show both non- and AA benchmarks, so NV may still show a conspicuous lead w/o AA and merely remain competitive with AA (losing by 10% doesn't catch the eye as much as winning by 30%).

Sorry, just workin' with what we've been given. It's just that the GS is so available and only $20 more than either the X1600XT or 6600GT (both also sporting 256MB 500+MHz RAM). At this price point, an extra 10% for much higher framerates returns seems eminently wiser. You're right about RV530's intended 128-bit competition, but it's hard to ignore the 6800GS's ubiquity. (We'll see what happens to NV's profit margins--surely AIBs aren't paying NV $20 more per GPU and then eating into their profits with the more complex PCB.)

 

Depends where you are in the world. Over here are asking price for the 6800GS is >20% more than the X1600XT and >40% more than the 256MB 6600GT. X1600Pros and 128MB GTs are close to half the price of the GS.

 

Amazing how many sites have published detailed reviews - only to pull them a few hours later when they realise they've broken embargo or NDA...

http://www.hardwarezone.com/articles/view.php?cid=3&id=1808