RV770 - more than meets the eye?

[Andreas G]

RV770 is the heart of the Radeon HD 4800 series. It's based on the R6xx architecture, but is supposed to add additional performance and features. The R6xx series of chips, both the R600 and RV670, were substantially different from previous generations as ATI had decided to move the antialiasing processing from the ROPs to the shaders. It was suggested that this was the cause for the cards' lackluster performance in heavy antialiasing and anisotropic filtering environments, even though the latter is handled by a completely different section of the GPU, texture units, and that the AA is not the problem but the AF.

Thus, it's not really fair to blame the shaders for the drop in performance. Somehow stories have started going around that AMD has "fixed" the performance when using antialiasing, which assume is really about the anisotropic filtering. It has been suggested that AMD has reverted back to making the ROPs handling the AA, but that doesn't really make much sense if AF is the problem and not AA, right? Instead, we simply suggest that the fix is the increased number of texture units, something which has been known for a very long time. 

There might be more to it though, we will just have to wait and see. It's not impossible that AMD has made more changes to how AA and AF are handled to improve performance.

[reaperrr]

That's what I was suspecting all along since I saw some 38x0 benchmarks on xbitlabs.com. The game Gothic 3 has no AA support, so they tested with AF only, and guess what, the R6xx cards still suffered heavily once 16xAF was turned on.
There was one strange thing about it, though: The performance penalty on the X1950 XTX cards was not nearly that heavy, even though it has the same amount of TMUs and actually lower core clock.

So the low number of TMUs cannot be the only reason for the weak AF performance, there must be something else wrong on top of that.
I mean, at least on paper the ROPs and TMUs of the R6xx should be better than R580s, yet once you enable AA/AF the whole R6xx based cards suffer so heavily, that a 3850 sometimes falls behind X1950XTX, despite higher core clock and far superior shader architecture.

Also, I remember some benchmarks somewhere else where they enabled only AA and not AF, and things didn't look much better, either.

Now what I suspect is that the whole R6xx architecture has some serious design flaw somewhere in it that results in a heavy performance hit once you enable either AA or AF (or, in worst case, both). That's the only logical conclusion, I see no other possible reason why the R600 and RV670 show worse performance per clock than R580 in some cases.

[Delph1]

I couldn't agree more with you there.

The "AA bug" simply doesn't exist, alas it's a bit late to point this out...

//Andreas
_________________
"I'm sort of a postmodern vegetarian, I eat meat ironically"
"You are our Swedish superhero "

Forget everything, knowledge is your enemy, it's only going to make you second-guess yourself."