AMD Radeon 6000 Series Image Quality Analysis
Anisotropic Filtering (HQ)
Shortly after the 5000 series launched I wrote an article highlighting the flaws in AMD’s anisotropic filtering. It was one of the first (if not the first) published articles to do so. The topic of AF differences amongst the vendors flared up again on the tech forums shortly before the 6000’s launch, so it deserves to be revisited.
To refresh your memory, the main problems with AMD’s 5000 series’ AF are harsh filtering transitions, dead zones that lack texturing detail, and additional texture aliasing over nVidia’s parts. These problems were later confirmed to be a hardware limitation by AMD, caused by an uneven shifting of kernels within a MIP level. So without further ado, here are the filter test results for all three cards set to HQ:
5770 HQ
6850 HQ
470 HQ
Starting with the 5770 you can see it looks exactly as it did in my original 5000 series article, confirming it’s a hardware problem.
Moving to the 6850, we can see several improvements over the 5770. The filtering bands in the grey areas have reduced (but not disappeared) in number and intensity, the texture aliasing at 90 degree angles closest to the camera has been reduced, and the untextured gray zone outside the red circle now has texturing detail in it.
Compared to the 470, the 6850 has angle invariance (the colored mip-maps are perfectly circular) and it has less texture aliasing at the 45 degree angles touching the outer edges of the red border. The 470 meanwhile has smoother texturing transitions in the grey areas before the red border, and it also has less texture aliasing at the 90 degree angles touching the outer edges of the red border.
Anisotropic Filtering (Q)
Another topic that flared up at the 6000 series’ launch was the accusation that AMD’s default driver quality was lower on this series than it was on the 5000 series, thereby invalidating the benchmark scores to some degree. Here are the filtering images with all three cards set to Q:
5770 Q
6850 Q
470 Q
Starting with the 5770, the only change from Q over HQ is the mip-maps moving closer to the camera, indicating a higher sampling bias is being used. In gaming this could result in slightly blurrier textures at long distances. Interestingly enough there are absolutely no changes to the texture aliasing, indicating that a low LOD bias is probably not the cause of it.
Moving on the 6850, we see the mip-maps are brought forward like they are on the 5770, though not as much. This again indicates a higher sampling bias being used. We also see more texture aliasing appear at 90 degree angles touching the outer red circle area. Thus when using default driver settings the 6000 series can cause more shimmering than the 5000 series, showing evidence that image quality took a step backwards.
And lastly, we take a look at the 470. Some people have the mistaken impression that nVidia’s default driver settings do not degrade image quality over HQ, but this is absolutely false. You can see the mip-map colors become more solid with less blending in the gradients due to “brilinear” optimizations being used. The trilinear optimization setting is responsible for this, and can result in harder texture transitions at long distances in gaming. There are no other changes to AF quality, however.
Anisotropic Filtering (HQ & 8xSSAA)
From the earlier examples you can see that all three cards suffer from visible texture aliasing in the gray areas, and in fact perfect 512xAF rendered on the ALUs has the same problem. Fortunately we have something that helps this problem: super-sampling. Here are the screenshots of all three cards running HQ settings with 8xSSAA:
5770 SSAA
6850 SSAA
470 SSAA
All three cards show a visible reduction in texture aliasing with 8xSSAA, though both Radeons are far superior to the 470 in this regard, even with their negative LOD adjustment. Yes, AMD now automatically adjusts the LOD bias when SSAA is being used which is helpful in avoiding the slight blurring that SSAA can occasionally cause. nVidia doesn’t do this for their rotated/sparse modes, but does it for the old xS modes.
Compared to the 5770, the 6850 also keeps a little more texture detail in the gray area outside the red area, but both Radeons still have the same problem of losing far too much texture detail at long distances.
After looking at those images for a long time my head hurts *dizzy*
Great article.
Although “Both vendors degrade image quality at default driver settings” may sound a bit missleading. Because AMD lowered default IQ historicaly speaking. I mean compared to previous drivers and previous cards.
ysondurr, the comment referred to the fact that both vendors’ image quality increases when running HQ compared to their default settings.
excellent review
i’m translating another article from Rage3D regarding the same subject – image quality comparision between AMD & NVIDIA’s latest graphics. i read both articles, they are both great but i still feel something lacks. besides the AA & AF and few zoomed game image comparision, i think it better to add more games to compare the overall image feeling. because there are a lot of arguments in my country stating that AMD provides much better image quality in both games and video-playback than nVidia. but for my personal experience i think nVidia graphics provide much more natural images. so i search the web and found Rage3D’s article last month and translating it. haven’t finished the translation due to recently increased payload in work. the AA & AF quality comparision would be a great proof to “who’s better”, but people will have another quesion. it seems that sometimes AMD offers a more colorful image while nVidia’s is a bit flat, is this a type of image quality issue? i read some reviews but they never regard this as an image quality issue. however people in my country think it as an image quality issue. what’s your opinion dear Mr. BFG10K?
and finally, can i put this article in my translation works?
thx
kiss4luna,
Though I’ve seen many such claims over the years, I’ve never seen any evidence of this so called “extra vibrance”, aside from nVidia’s old digital vibrance setting. In 3D gaming the colors have always looked identical to me with both vendors. I think it’s a placebo effect similar to how some individuals keep stating that each new driver is “smoother” and has “better IQ” than the old one when in reality nothing has changed.
Feel free to translate the article but give full linkage and credit to ABT. We want the traffic coming here.
Thank you, BFG10K. You have my words 😉
What I find surprising is that you haven’t noticed that the 470 can’t render a circle!
Real Life, the term for the circle is “angle invariance”, and I mention it several times in my article.
I don’t recall AMD ever claiming MLAA to be faster than MSAA. I do recall them saying it offers performance that of edge detect aka CFAA.
If you reviewed this with the thought that MLAA was intended to be an end all AA feature, then you have the wrong mind set and your conclusion is going to reflect that.
In my finding thus far, it works brilliantly in games with half AA or no AA(differed render’s) AND with a limited color spectrum. High contrast color edges will give a nasty result like many of the games you tested.
Hey very nice site!! Man .. Excellent .. Amazing .. I will bookmark your site and take the feeds also…I am happy to find numerous useful information here in the post, we need work out more strategies in this regard, thanks for sharing. . . . . .