AMD: R8xx Speculation

So basically you have to get up in the middle of the night to read the reviews?

 

I really, really hope ATI allow us to have full high quality AF without any under filtering without forcing us to disable AI, it's pointless having best quality AF if we have to disable bugfixes/optimizations to get it.

When we're paying for the top end card we should be able to get the max image quality possible, not under filtered and flickering because of it.

Looks like it's still going to be worse than nvidia though, wish ATI would fix this.

http://www.hardware-infos.com/news.php?news=3208&sprache=1

 

KonKorT!

Fuad with his usual spin and mistakes:

On the HD5850:

Yep, as neutral as Charlie.

 

Not sure if this has been posted (didn't see it). Looks like a Gigabyte marketing presentation about the 5870. One thing that I noticed, which I haven't seen before, is the mention of HDR texture compression...

EDIT: There is also one for the 5850 here (mentions 1440 stream processors).

 

HDR Texture Compression is part of the DX11 spec.

 

Exactly. Far too many people have this wrong assumption.

There is the gradient calculation and the filtering part which takes the samples along the longer gradient. The first one was already good enough since the HD 2000 series, but they are not sampling very well.

In contrast NV there is no visible difference between my reference filtering which takes all samples and NVs HQ filter, so I'm assuming they learned their lesson with NV40/G70.

I hope ATI will come to their senses in the future as well. I think this should be a DirectX requirement. We have full IEEE754 precision in the ALUs with D3D11, but they can do what they want with texture filtering.

 

So, how many samples does it use at say 45 degrees?

 

http://www.scribd.com/doc/20009711/Gigabyte-GVR587D51GDB-graphics-card

Powerpoint presentation for the HD5870 from Gigabyte.

Edit: another excerpt from RV870 presentation, too...

 

Can you please qualify the impact of this alleged subpar filtering? What would be the visible result?

 

Ehm, suggesting you could wait and see what the competition has in store, isn't that generally sound advice?

Seems pretty harmless compared to the usual Charlie diatribes.

 

Man I'm years out of it. And I though ATI had the better when in motion.

 

Texture crawl/flickering I think.

 

Yes, when was the last time he suggested that at an NV launch? Hmm. Seems the traffic goes only way, not the least bit surprised.

 

RefRast for DX9 is pretty horrible though so I'm not sure you want that. Also, IMHO, this is more likely to be a software limitation rather than a hardware one. It does not mean it'll be solved tomorrow, or ever (I have no idea), but it just points the finger in the right direction.

 

I've said this before, but I think we should really have a reference rasterizer which simply used a ridiculous supersampling ratio (gaussian averaging filter) and no texture filtering at all (no mipmapping either). Do a MSE comparison against that to see which does better (or SSIM, or A/B). Would provide a better way to judge all the exotic AA methods too.

 

The number of samples does not depend on an angle, but on the distortion of the texture that is reevaluated for each sample.

Look at this diagram:


The parallelogram is defined by the rate of change of the texture coordinates in the X and Y direction in screenspace.

For a perfect line of anisotropy the GPU then needs to calculate the major and minor axis of the enclosed ellipse. The major axis is the desired LOA.

The level of anisotropy is then determined by the quotient of the length of these two axes and clamped to the AF level the App requested (16:1 maximum).

After that the GPU should take enough samples along the major axis as necessary to fulfill the nyquist shannon sampling theorem. But ATI doesn't.

Even with A.I. disabled ATI is still undersampling. I don't think the hardware is even capable of filtering perfectly. I think they do this because they can save on buffer sizes, because the maximum latency goes down or something like that

 

Just like sex. You spend all that time building it up only to blow your load in 5mins. Atleast it makes for comfortable sleep.

 

So, how many? (At a surface angled above max. anisotropy and at 45 degrees, which is nastiest for memory access.)

Or to put it another way, how many samples do they skip?

 

Solution: don't blow your load that easily. Err, I mean reviewers should write better reviews

 

Thanks for the explanation. So why then does the flower show a perfect pattern if they're undersampling? Is it because the angles used by that particular pattern aren't affected?