r520 in xbox2?

[Geo]

And you'd be very wrong.

That "very" is practically two and a in Dave-speak (he's British, y'know).

 

[kemosabe]

Yeah, I would've expected a simple "No."

 

[ShootMyMonkey]

You certainly do not need the kind of fillrate we have in the X850 PE to fill a screen that is approx 1 million pixels in size 60 times a second even if it applies some sort of anti-aliasing - but you could always do with more horsepower to compute those shaders.

Well, I wouldn't entirely say that. Assuming that there isn't any such mechanism to take an interlaced 1080i frame and render it at 60 fps in 1920x540 alternating fields (which I'd say is a fairly safe assumption), that means you're rendering 1920x1080, or about 2 megapixels at, let's cut back and say, 30 fps. Now if you take that as just meaning 60 MPixels fillrate, that's basically what we call oversimplifying.

First of all, it disregards a little thing called overdraw where you end up eating through fillrate over the same pixels over and over again. Even if we assume something like 10x overdraw (which, btw, is extraordinarily unimaginably incontrovertibly generous), that already puts us at 600+ MPixels/sec. Now then you have things like alpha blended polygons and particles where you really don't WANT to avoid overdraw. Rather, overdraw is actually necessary. And these types of polys will locally have very high fillrate demands. Even now, local clumps of big alpha blended polys eat through framerate so fast, it's not funny.

Then we also have to consider the demands for multipass rendering, which would include things like shadow casting (whether it be maps or volumes), subsurface scattering, and other miscellaneous transgressions of the soul. That easily multiplies our fillrate demands by much more than the number of passes, because the latency between passes is bad enough to drive you to sacrifice goats in a thunderstorm to get those extra few polys down the pipeline.

That said, though, I wouldn't say we'd need something like 5x the fillrate as badly as we'd need 5x the shader power. It's all too often that we get limited by the complexity of vertex or pixel processing and the practical limits are hit. Also, more important than raw pixel fillrate would be texel fillrate, because we're bound to be using up all 8 texture channels on damn near everything for next-gen.

 

[Tahir2]

Now that is a nice post ShootMyMonkey.

It is apparent that even at 1920*1080 and with let us say MSAA of 4x @ 30fps and we are limited by fillrate.

But before we are limited here we are going to be hit a wall with pixel shading computations of certain polygons that may also be alpha blended.

Firstly to target a resolution of 1280*720 might help as MS has stated in some of its announcements should alleviate some fillrate issues.

We have overdraw but then of course we have overdraw efficiency techniques like rendering in tiles and HyperZ IV or whatever it is going to be called. It will not solve all problems as you have stated and with ever increasing number of polygons and PS effects onscreen this problem will most probably increase rather than decrease.

Everyone assumes that there will be no large eDRAM other than the normal caches found in GPU's currently so no local storage cache for the Z Buffer and VRAM Memory bandwidth is still unknown.

EDIT: OK MAYBE THERE IS eDRAM AFTERALL!

But what is a good compromise for a resolution like that if we were to take current gen technology and multiply it by a conservative 1.5X in general performance?