Radeon9500 O/Cing o_O

[BoddoZerg]

Anand got a Radeon9500 to O/C to 351/635!!!!

I wish my GF4Ti4400 could do that!

 

[Hyp-X]

The review cards have faster ram than the final boards will have...

 

[Joe DeFuria]

Anand is slipping, big time IMO.

He AGAIN compared GeForce4 at 8X aniso, vs. ATI at 16X...gee...that makes sense.

And why bother with overclocking at all? The review sample is not in any way representative of the final product. Hard OCP did the right thing by not bothering with overclocking at all. Don't expect any sort of relevance with Anand's overclock success with respect to the actual product.

 

[John Reynolds]

ATI's very efficient adaptive ansiotropic filtering algorithm yields the best quality/performance ratio out of the bunch and thus puts all of the Radeon cards ahead of even the fastest GeForce4.

This sentence also implies that NV's AF isn't adaptive, which isn't true.

 

[Hyp-X]

Anand still has no clue on how anisotropic filtering works...

Edit: neighter does Lars Weinand ... at least he does 8x versus 8x

 

[arjan de lumens]

This sentence also implies that NV's AF isn't adaptive, which isn't true.

One question: Can the Nvidia anisotropic filtering method ever adjust the number of samples taken, or does it only adjust the sample locations being used?

 

[ERP]

It does adjust the samples taken.
If it didn't the performance hit would be significantly more than it currently is. My understanding of the way NV2X's pipelines work is that there is no fixed alotment of samples to pipes, but there is a limit to the total number of texels that can be fetched from the cache/clock by the combined pipelines. There is also a limit to the number of texels that can be read into the cache/clock. The exact numbers are optimised to give trilinear free most of the time.

 

[Bambers]

Something is wrong with the gf4 core , though, that reduces the fillrates when aniso is enabled even when viewing polygons that don't need any aniso.

I don't think the gf3 has this problem.

 

[Hyp-X]

Something is wrong with the gf4 core , though, that reduces the fillrates when aniso is enabled even when viewing polygons that don't need any aniso.

I don't think the gf3 has this problem.

The GF4 can't do anisotropic filtering on the second TMU.
So it's reduced to 4x1 in that case.

There are driver hacks to enable aniso on the first TMU only to avoid the fillrate loss. (Most games has their base texture on the first texture stage.)

IMHO that was the worst move of the GF4.

If they removed this ability in trade off for the second vertex pipeline - I'd say it didn't worth it.

 

[arjan de lumens]

Just wondering, since it appears that the NV2X cards generally take a much harder and more consistent performance hit from enabling AF than the R9xxx cards do.

 

[horvendile]

He AGAIN compared GeForce4 at 8X aniso, vs. ATI at 16X

Well, yes, but isn't it generally agreed that the two methods of aniso are a bit on the incommensurable side anyway?

And isn't 8X the maximum quality for GF4? (Could be wrong there though, not actually having a card capable of such feats as aniso.)

If so, I don't think comparing max quality vs max quality is more unfair than any other configuration. Sure, it's not exactly apples to apples, but then again, it'll never be.

 

[Fuz]

It would be a fair comparison if the nv 8x aniso resulted in similar image quality as ati 16x.

 

[John Reynolds]

Well, yes, but isn't it generally agreed that the two methods of aniso are a bit on the incommensurable side anyway?

And isn't 8X the maximum quality for GF4? (Could be wrong there though, not actually having a card capable of such feats as aniso.)

If so, I don't think comparing max quality vs max quality is more unfair than any other configuration. Sure, it's not exactly apples to apples, but then again, it'll never be.

Well, if you're going to break a fundamental rule of comparitive benchmarking (a level playing field), what's to stop a reviewer from going even further astray? Beyond just benchmarking at 8x AF vs. 16x AF, Anand is also comparing 4x OGMS against 4x RGMS, which isn't fair either. His AA + AF benchmarking is simply a waste of time.

 

[Joe DeFuria]

If so, I don't think comparing max quality vs max quality is more unfair than any other configuration. Sure, it's not exactly apples to apples, but then again, it'll never be.

Of course it's unfair...if the actual quality differences are not as close to identical as possible.

Are you saying it would be fair if one card "maxed out" at 1024x768 resolution, and that was tested against another card "maxed out" at 1600x1200?

It would be a fair comparison if the nv 8x aniso resulted in similar image quality as ati 16x.

Agreed. And from everything I've read / seen, if anything ATI's aniso is a bit higher quality at the same setting. Though most people would say that at the same setting, the quality is comparable. (ATI looking better in some instances, Nvidia in others.)

What Anand has not done, is show a good image quality comparison to justify his reason for comparing the two different levels of aniso.

Anand is also comparing 4x OGMS against 4x RGMS, which isn't fair either. His AA + AF benchmarking is simply a waste of time.

Exactly.

 

[Hyp-X]

Well, yes, but isn't it generally agreed that the two methods of aniso are a bit on the incommensurable side anyway?

No, that was the 8500 and the 9000.

The 9500 and the 9700 are comperable with the GF3/GF4.

OTOH, while 16x vs 8x is not fair it doesn't affects performance or image quality much. The Radeons still takes less hit in 16x than GF4 in 8x.

 

[Nagorak]

Come on guys, he even said in his review that it wasn't a final card and thus the overclocking results wouldn't necessarily compare! The point was he showed how the card performed when overclocked (and there's no reason to think the core won't overclock just as well on the R9500 PCB).

Overall I found the R9500 to be somewhat disappointing, but I guess it fits well within its niche. The R9700 non-pro on the other hand...

 

[BoddoZerg]

Is there ever a situation where the GeForce3 actually beats the GeForce4 with aniso applied?

 

[Bambers]

I think I've seen cases where the gf3ti500 is faster than a ti4200 with 8xaniso. Can't remember where though.

 

[Xmas]

Something is wrong with the gf4 core , though, that reduces the fillrates when aniso is enabled even when viewing polygons that don't need any aniso.

I don't think the gf3 has this problem.

The chance to find a polygon that doesn't 'need' anisotropic filtering ('needing' in the 'mathematical sense', not taking into account how much difference is actually visible to the human eye) in an in-game situation is very low. And the GF4 like the GF3 is perfectly able to save ressources and not take any fillrate hit in these rare cases.

The GF4 can't do anisotropic filtering on the second TMU.
So it's reduced to 4x1 in that case.

This is only partly true, as it implies that the second TMU isn't used for anisotropic filtering. It is, but only for trilinear aniso.

Bilinear and trilinear aniso result in the same fillrate on GF4 (half the trilinear fillrate in case of 2xAF), while on GF3 and in the no-aniso case, bilinear filtering is twice as fast as trilinear (pure fillrate).

 

[Hellbinder]

Overall I found the R9500 to be somewhat disappointing, but I guess it fits well within its niche. The R9700 non-pro on the other hand...

yeah if you go by anands numbers.. But Hardocp shows the 95009pro trading blows with the Ti 4600 and beating it handily with AA+AF.