AMD: R8xx Speculation

And you can confirm this with Rv870 as well.

 

It seems the HD 5000 series is filtering a bit better, but it's still worse than what NV GPUs can do with HQ - which simply take ALL the samples with no underfiltering at all.

 

"it seems" isn't quite a definite statement. Could you show us screenshots of your tests?

Excuse my reluctance to accept this as fact since the fud-o-tron seems to be stuck in overdrive.

 

The difference is in the range of variations of any two driver versions (new vs old, optimized vs unoptimized, catalyst 9.x vs catalyst 9.x+1) so yes it is plausible. But it wouldn't look so good if they started the graphs from 0 instead of 80.

 

I'm sorry, but I can't. But I can assure you that my sources are quite credible (at least two big german online sites), because they use my tool and I gave them some advice how to use it etc.

The gradient calculation on HD 5000 is perfect, but sadly they still offer no usable solution for the user that wants maximum IQ when it comes to sampling the gradients.

 

It's probably not something that can be verified with just screenshots.

 

That's true indeed. You need to look at it in motion or else undersampling just produces an overly sharp image.

 

even four sequential captured frames should be enough to create a clear image for comparing them to a reference rasterizer.

As long as it's hearsay -> fud-tactics.

 

http://img1.abload.de/img/linuxb14l.jpg
Both fake imo.

Look at two times Farcry2 4xAA Bench

 

I don't think something counts as FUD if it's true. It's not "hearsay", I have seen the videos. I would have been very pleased if ATI would have made the step and give the user an option to filter without cheating - but they didn't.

Most people don't notice, but I don't think it's a bad idea to have an option if you want to.

Also there are no transistors to save here, because the sampling is just talking n samples along a calculated line.

 

It's more the pavlov denial reflex at the merest suggestion that anything should be less than optimal which seems stuck in overdrive.

 

I'm not convinced there's enough "expensive" shaders that need to run for tiny triangles to justify special-casing this, particularly given the additional complexity in the API.

I think deferred rendering and "just tessellating down to ~4 pixel triangles rather than ~1" provide a good enough solution to expensive shaders on tiny triangles in practice.

 

It would be even better if we could generate and rasterize (flat) quads as well..

 

LOL, PR twist gone mad:



I wonder how Atom would compare here.

 

Well sample patterns are pretty uninteresting for dissecting new hardware, as they are completely driver controlled. You'd think if AMD figures out a better pattern they'd apply it to all cards not just the newest generation - but it's probably optimal (for the average case) anyway .

 

It's not, you just don't have the single HD5870 performance there at all, and you have GTX295 bar blocking what would be HD5870 CFX bar

 

i'd like to see these benchmarks re-done after vista's Dx11 patch/SP3/w/e is out.

 

Since I don't think that patch adds WDDM 1.1 support to Vista, the performance still should be the same.

 

I tought WDDM 1.1 brings only performance when the application knows and supports it. Or do ALL Apps automatically benefit from the new driver model?

 

There is no additional complexity, don't use the difference instructions in your shader ... that is all.