9700 benchmarks

[Doomtrooper]

The antialiasing quality between the 9700 and the 4600 are essentially a dead wash (though the GF4 seems do a better job cleaning up the tent support lines. At least in that shot). The 9700 though does seem to have a noticeably higher LOD.

Mipmap boundaries are odd looking to my eye in both the 9700 and 4600 shots.

Not the 6X mode

 

[Sabastian]

Yeah you can really see the Geforce 4 ti 4600 from them shots.. look at the right hand side behind the chopper..... I find this is an annoying aspect of Geforce 4 cards. My Xbox has the same troubles with grass and tree leaves. I find this very irritating. Looks like ATI has hammered nvidia here. Every review that I have read ..... solid drivers, superior performance and positive conclusion in every one of them. Kudos to ATI for their efforts on the Radeon 9700.

Dave great in depth review.. best I read so far.

 

[Bambers]

I haven't read the review yet but to me it looks like some kind of AF is being used on the 9700.

For AA its harder to tell (jpegs smooth the detail unfortunately) but I would say the 9700 is better, the non ordered pattern shows up well along the flat hill to the left of the flag and along the bottom of the rotor blades.

 

[Xmas]

Yeah you can really see the Geforce 4 ti 4600 from them shots.. look at the right hand side behind the chopper..... I find this is an annoying aspect of Geforce 4 cards. My Xbox has the same troubles with grass and tree leaves. I find this very irritating. Looks like ATI has hammered nvidia here.

Huh? None of these three cards does anything to improve the grass/leaf edges. It looks bad on each card. Hopefully ATI is able to implement something like "adaptive" supersampling, ie. supersampling for alpha tested textures while rendering the rest with multisampling.

 

[Bambers]

Hmm just looked closer at the leaves and grass and the gf4 appears to be using a very low LOD.

 

[Dave Baumann]

Huh? None of these three cards does anything to improve the grass/leaf edges. It looks bad on each card. Hopefully ATI is able to implement something like "adaptive" supersampling, ie. supersampling for alpha tested textures while rendering the rest with multisampling.

I made a note of what they could do in the review:

http://www.beyond3d.com/reviews/ati/radeon9700pro/index.php?page=page4.inc

 

[Tahir2]

Dave in your next article can we see some of the videoshader effects?

Me is most interested in that aspect of the card assuming it is enabled in the drivers

 

[Colourless]

Could it be that the game is setting a 'high' LOD Bias using D3DTSS_MIPMAPLODBIAS, and ATI isn't supporting it (cap = D3DPRASTERCAPS_MIPMAPLODBIAS)?

 

[Johnny Rotten]

Not the 6X mode

That indeed is a lovely image.

Must... resist... upgrade... until spring...

 

[Xmas]

Hmm just looked closer at the leaves and grass and the gf4 appears to be using a very low LOD.

Yep, that's true. Or the other way round, the R9700 is using a very high LOD. It's a matter of taste what level of detail seems "right", and luckily it's quite easy to adjust the LOD bias.

I made a note of what they could do in the review:

That's an interesting possibility, although it has some (really) minor flaws with overlapping alpha-tested textures.
But I thought that's part of the ARB_multisample spec, so shouldn't this be possible on other cards, too?

 

[KimB]

This coming from someone who doesn't have (and claims to have never had) past ATI products?

I've said again and again: I've seen texture aliasing in the screenshots. If I can see it in screenshots, I know I'll see it in motion.

That one possible thing that could account for the difference in texture quality, is that GeForce's is not aggressive enough. GeForce's could be more aggressive, and not exhibit significant aliasing, but it is lowered to help boost performance

Having owned a GeForce DDR and a GeForce4 Ti 4200, I can say that that is not true in the least. I have, in fact, played around with the LOD settings, and I have found that making the LOD only slightly negative results in significantly increased aliasing. By contrast, positive values don't have much effect on aliasing. The only problem has been fairly recently, where it appears that recent drivers have been using, by default, a much more aggressive LOD setting than I would prefer. This appears to only be an issue when anisotropic filtering is disabled (the main reason why I'm not overly-concerned....since I rarely run without aniso).

 

[hax]

I've said again and again: I've seen texture aliasing in the screenshots. If I can see it in screenshots, I know I'll see it in motion.

It's possible that game developers are using the NV line as their development platform, and with the pushed LOD inherent in their drivers, it is making a card with no LOD tweaks (read that as 0 default bias and using the app settings and geometry to properly calculate LOD) appear as being too aggressive (too negative).

An LOD bias should be adjusted by the app, not the drivers, as it should be up to the artist to determine that.

 

[Nagorak]

I don't know about you, but I prefer some aliasing over an overall blurry picture.

 

[KimB]

It's possible that game developers are using the NV line as their development platform, and with the pushed LOD inherent in their drivers, it is making a card with no LOD tweaks (read that as 0 default bias and using the app settings and geometry to properly calculate LOD) appear as being too aggressive (too negative).

To my knowledge, game developers usually don't ever touch LOD settings, and the hardware/drivers absolutely must set some sort of LOD.

What most certainly should happen is that every video driver out there should adhere to the exact same default LOD settings. Unfortunately, there's a large problem with this. That problem is that every video card does not select MIP maps in the same way. How do you standardize MIP LOD when one piece of hardware uses a semi-circular MIP selection algorithm, and another uses a linear algorithm? The truth is that you can't.

And then you have to take into account anisotropic filtering. Given the fact that the standard semicircular or semicircular approximation boundaries are based on bilinear/trilinear filtering, it stands to reason that such patterns may not be correct when considering anisotropic filtering. For example, I know that the GeForce4 sometimes instead uses a hyperbolic MIP map boundary when anisotropic is enabled. For what reason, I'm not certain. Nor am I certain that it is the correct boundary, but it does beg the question.

What I feel needs to happen is that somebody (MS and/or OpenGL) needs to create a program that gives an aliasing measurement, designed to operate on nearly all possible surface angles. One possible way to do this would be to use a texture whose MIP map levels are all black and white checkerboard patterns (not standard pre-generated MIP maps), where the textures are placed in such a way that no texel is larger than approximately 1/4 the size of a pixel, which should result in each pixel being close to grey. The deviation from grey should give a count of the amount of aliasing (Disclaimer: This isn't a particularly well thought-out idea, and may have other considerations that I have no taken into account, but I don't see why something similar can't work).

 

[Ailuros]

HotHardware used 29.80 WHQL. Does someone know to what LOD settings those default to in D3D? I´m positive that 27.50 defaulted to a positive value (+0.5 or +1.0) as in "best image quality".

Other than that in my eyes the 9700 does a much better job on horizontal and vertical edges in 4x mode due to it´s rotated grid nature. I can see a significant difference on the flag pole. It used to look crappy in motion even with 2x2 OGSS.

 

[Althornin]

HotHardware used 29.80 WHQL. Does someone know to what LOD settings those default to in D3D? I´m positive that 27.50 defaulted to a positive value (+0.5 or +1.0) as in "best image quality".

Other than that in my eyes the 9700 does a much better job on horizontal and vertical edges in 4x mode due to it´s rotated grid nature. I can see a significant difference on the flag pole. It used to look crappy in motion even with 2x2 OGSS.

Exactly.
I never understood why RGMS is said to look better than OGMS.
One looks better on lines closer to horizontal and vertical, due to rotated samples) and the other looks better on rotated (near 45 degrees) lines, because the samples are "rotated" in relation to the line being AA'ed.
This is why i think the GF4 shot looks better on the tent lines and the R9700 on the flagpole.

 

[KimB]

I never understood why RGMS is said to look better than OGMS.
One looks better on lines closer to horizontal and vertical, due to rotated samples) and the other looks better on rotated (near 45 degrees) lines, because the samples are "rotated" in relation to the line being AA'ed.
This is why i think the GF4 shot looks better on the tent lines and the R9700 on the flagpole.

The supposed reason is that it's more noticeable when the near horizontal/near vertical lines are not AA'd. I can't say I disagree, but what I'd really like to see is a pseudo-random sampling pattern (I'm not sure I'd like an algorithmic sample pattern based on what is being rendered), though I'm not certain a pseudo-random sampling pattern would be useful with as few as four samples per pixel.

 

[jjayb]

Chalnoth, something tells me that Ati's next card could create fully realistic 3d hologram worlds like the holodeck on star trek and you would still find fault with it. "that's not the correct way to create a holodeck, it should be done with less lod like Nvidia. The textures are to clear. All the colors are to vivid, they should be more washed out so it looks like my gf4" A valid complaint is a valid complaint, but sometimes you are really, really having to reach for your trivial complaints.

This is killing you isn't it? Ati has delivered the goods. A great card with no show-stopping problems. And Nvidia is at least a couple months away from having an answer. You've just got to find something wrong with the card to make it all better. Just got to find something......Got to find.....Something. Must...Find.....Something....There has to be...Something. Can't be good....Noooooooooo!!!!!!!!!

 

[Bigus Dickus]

The supposed reason is that it's more noticeable when the near horizontal/near vertical lines are not AA'd. I can't say I disagree, but what I'd really like to see is a pseudo-random sampling pattern (I'm not sure I'd like an algorithmic sample pattern based on what is being rendered), though I'm not certain a pseudo-random sampling pattern would be useful with as few as four samples per pixel.

Yes, and as someone that uses CAD on a daily basis and deals with wireframe display day in and day out, I can attest that the difference is striking.

A 45-degree line has essentially no aliasing, and there are a few other angles that are equally pleasing. In general, the farther a line is from 0 or 90 (closer to 45) the less offensive it is.

Since monitors are going to continue to have basically the same pixel layout, I think rotated grid AA is the only valid choice (other than "jittered" or pseudo-random sample placement that is more useful with >4 samples).

 

[OpenGL guy]

To my knowledge, game developers usually don't ever touch LOD settings, and the hardware/drivers absolutely must set some sort of LOD.

This is absolutely incorrect. Hardware and drivers provide a baseline LOD, but applications can tweak that to suit their needs.

What most certainly should happen is that every video driver out there should adhere to the exact same default LOD settings. Unfortunately, there's a large problem with this. That problem is that every video card does not select MIP maps in the same way.

Another problem is that there are generally performance benefits to making things blurrier.

What I feel needs to happen is that somebody (MS and/or OpenGL) needs to create a program that gives an aliasing measurement, designed to operate on nearly all possible surface angles.

Why don't you make one? Then you'll see that what is pleasing to you, isn't pleasing to everyone. Also, when you write yours, you can run it on the D3D ref rast and compare results to the output of different video cards and different drivers.

One possible way to do this would be to use a texture whose MIP map levels are all black and white checkerboard patterns (not standard pre-generated MIP maps), where the textures are placed in such a way that no texel is larger than approximately 1/4 the size of a pixel, which should result in each pixel being close to grey. The deviation from grey should give a count of the amount of aliasing (Disclaimer: This isn't a particularly well thought-out idea, and may have other considerations that I have no taken into account, but I don't see why something similar can't work).

I don't like this idea because it doesn't check if things are too blurry.