Performance numbers calculated (from Anandtech's reviews)

[Hellbinder]

Dude your FSAA math is WAAAAAAY off....

Gf4 ti 4600 41 FPS and the R9700 is 2.5x faster or 250% increase you get 46 FPS??? get serious. EDIT never mind i see where i went wrong)

also is it just me or are there some issues with these numbers...

Jedi Knight 2
'demo jk2ffa' 1024x768x32
Radeon 9700: 122.5
GF4 Ti4600: 125
Parhelia: 90.5
Radeon 8500: 123.5

Jedi Knight 2
'demo jk2ffa' 1280x1024x32
Radeon 9700: 123.2
GF4 Ti4600: 124.4
Parhelia: 74.9
Radeon 8500: 116.9

Jedi Knight 2
'demo jk2ffa' @ 1600x1200
Radeon 9700: 124.3
GF4 Ti4600: 113.0
Parhelia: 65.9
Radeon 8500: 93.2

 

[mboeller]

http://www.anandtech.com/video/showdoc.html?i=1656&p=23

Come on people, we all can do basic math here. 2.5 times faster != 250% faster.

fixed, but it does not alter the math!

 

[Oats]

I don't think jk2 is cpu limited. I think the R300 is not using all available bandwith when aniso and FSAA are not turned on. On quake3 engine games, few passes, the 4600 with the same multi-texture fillrate has the 9700 can keep up. With the new UT engine, the R300 seems more efficient (or makes use of some the additionnal bandwith). With FSAA and aniso we really see the extra bandwith kick in (I doubt ATi's FSAA method is really that more efficient than nVidia's).

 

[mboeller]

deleted

 

[Dave Baumann]

I don't think jk2 is cpu limited. I think the R300 is not using all available bandwith when aniso and FSAA are not turned on.

Look at the fillrate graph above (and the FPS table on the linked review) - that clearly shows CPU limation on a GF4 Ti 4600.

 

[Oats]

Yeah, you are right, definetly cpu limited. Didn't look at the performance from res to res, but only compared each chip...

 

[Tagrineth]

erf... Anand F*cked up.

We have explained this in the past but the way ATI’s supersampling (Rotated Grid Super-Sampling – RGSS) works is by rendering multiple copies of the scene and shifting each copy by a certain amount off the center. These copies were then blended together to produce the final image.

That'd be 3dfx, not ATi. And it's wrong anyway.

 

[DemoCoder]

The Radeon 9700 can do 300+ million POLYS per, not VERTICES per cycle. How so? Well, since it has 4 Vertex Engines, each can do a Vertice right?

Correct me if I am wrong.

No, each vertex takes 4 vertex ops to process. So 4 vertex engines may be able to process 4 vertices at once, but each vertex takes 4 clocks then, giving a throughput as 1 vertex per clock.

Another way to look at it is all 4 vertex engines could be combined to deal with a single vertex, executing the 4-op shader in 1 cycle.

 

[Nagorak]

Am I wrong, or were all past Ti4600 scores on Anandtech from AthlonXP systems, but the new comparison scores with a P4?

I don't think the math works in that case.

ta,
-Sascha.rb

I don't claim that the numbers are exactly representative, but I think it can be generally assumed the % difference will be pretty close across both platforms (you'd hope at least). So the calculations give you a ballpark figure, take them with a grain of salt (If a P4 system had been available I'd have used it, but this was all there was).

 

[Nagorak]

WHERE ARE THE AA-NUMBERS???? Please!

Has someone benchmarked the GF4 Ti4600 with 4xAA in UT 2003 already?

I would like to see the numbers so we can compare them with the numbers from the Radeon 9700. With an (up to) 2.5 times higher speed with AA enabled it seems (for me) that the Radeon 9700 does 4xAA nearly for free (in comparison with the competition)

I was only able to spot an estimate based on anandtechs reviews of the parhelia and doing a little bit of math :

GF4 Ti4600 4xAA @ 1024x768 [ DM antalus ] = 42,8 fps [parhelia review ]
GF4 Ti4600 noAA @ 1024x768 [ DM antalus ] = 94,5 fps [UT2003 benchmarks ]
GF4 Ti4600 noAA @ 1600x1200 [ DM antalus ] = 41,1 fps [UT2003 benchmarks ]

Radeon9700 = 2,51 x (42,8/94,5) x 41,1 = 46,7 fps @ 1600x1200 with 4xAA!!
Radeon9700 without AA @ 1600x1200 = 63,3 fps

so the fps goes only down ~25% with 4xAA enabled. If they really have fixed the bilinear anisotropic filtering too, then you can have all the glory with only ~50% performance penalty. If it is true (my speculation based on the preview from tomshardware) that the Radeon9700 supports RGMS then this chip will/should give really outstanding image quality.

Yeah, I didn't post any AA numbers because there was nothing even remotely close to compare it to.

 

[3dcgi]

Anand needs to throw a little variety into his reviews and test something in addition to First Person Shooter games.

 

[nggalai]

Hi Nagorak,

I don't claim that the numbers are exactly representative, but I think it can be generally assumed the % difference will be pretty close across both platforms (you'd hope at least). So the calculations give you a ballpark figure, take them with a grain of salt (If a P4 system had been available I'd have used it, but this was all there was).

Thanks for your reply. It might have been better, though, if you had rounded the numbers to at least 1fps, perhaps even in steps of fives and wrote a large ESTIMATED NUMBERS or something next to them, too.

You might wonder why I stress this point. Well, guess what--some guys in our forums posted your "exact" numbers and now claim that you used figures from two P4 systems that ran side-by-side, and that the calculations are completely accurate. Others have picked them up, and I am just counting the hours until somebody takes them for real benchmark numbers and they make their way into the low-quality, high-volume print magazines of Germany. You get the idea.

but, eh. no worries. Real figures will be around in one month's time anyway.

ta,
-Sascha.rb

 

[Nagorak]

I'm kind of confused as to why Anand suddenly switched back over to the P4 system really. I remember when he first benchmarked something with the P4 (was it the R8500?) everyone freaked out, so it's just kind of funny that he just suddenly switched back over to P4 (although I guess it makes sense now that Intel is back on top again).

I must admit I will be somewhat amused if 'my' numbers (they're not really mine, hehe) somehow find their way into a low quality magazine. I mean the whole idea of basing something off a post in a forum just strikes me as ridiculous. Does that stuff really happen?

 

[nggalai]

I mean the whole idea of basing something off a post in a forum just strikes me as ridiculous. Does that sutff really happen?

You'd be surprised, dismayed, or delighted--depending on your tastes.

have a nice day,
-Sascha.rb

 

[demalion]

The Radeon 9700 can do 300+ million POLYS per, not VERTICES per cycle. How so? Well, since it has 4 Vertex Engines, each can do a Vertice right?

Correct me if I am wrong.

No, each vertex takes 4 vertex ops to process. So 4 vertex engines may be able to process 4 vertices at once, but each vertex takes 4 clocks then, giving a throughput as 1 vertex per clock.

Another way to look at it is all 4 vertex engines could be combined to deal with a single vertex, executing the 4-op shader in 1 cycle.

When you say vertex, it sounds like "poly" to me. I guess I'm just not familiar with the way the 3D information is encoded. I'm pretty sure this was explained before, and I may have understood a bit better after that explanation, but perhaps enlightening me here will be convenient for clarifying the point above.

Ok, a 3d position value is a: ? (to me it is a point, and also a vertex if there are other points for reference)
A 3d position + scalar is a: ? (if a direction is calculated based on other points, this is enough data for a vector for that point)
A vertex is: ? (I thought the above "point" could satisfy this with other points in relationship...atleast going by the math definition)
A vertex op is: ? (Hmm...you say 4 sets of 3d positions + scalar is necessary for a vertex op...my current understanding then is that the 4 points in question are the actual point of the vertex + 3 other points used to give it meaning in the context of being a vertex of a 3D object in a triangle, as opposed to quad, based rendering system).

Well, putting that into words seems to have clarified the issue for me atleast...the question is am I on the same page now? If so, what are the other 3 scalar values used for in the context of the one vertex? I can see how the "vector" calculated for the vertex points of a triangle can be used to calculate a normal vector for the entire triangle, but I'm a bit hazy as to what the 3 scalar components would actually signify for that surface, and for an operation on the one vertex. Perhaps that just depends on what you are trying to do? Or perhaps I should sleep on it?

 

[Joe DeFuria]

I'm kind of confused as to why Anand suddenly switched back over to the P4 system really.

From what I can tell, all of the bencharks (from Anand and elsewhere) have come from machines that ATI set up. So, ATI chose the platform, not Anand.

From what I can tell, ATI set up several pairs machines, and then gave several reviewers "free reign" to test whatever they wanted. (ATI did not run the tests themselves, only supplied the hardware).

 

[mboeller]

Hi Nagorak,

I don't claim that the numbers are exactly representative, but I think it can be generally assumed the % difference will be pretty close across both platforms (you'd hope at least). So the calculations give you a ballpark figure, take them with a grain of salt (If a P4 system had been available I'd have used it, but this was all there was).

Thanks for your reply. It might have been better, though, if you had rounded the numbers to at least 1fps, perhaps even in steps of fives and wrote a large ESTIMATED NUMBERS or something next to them, too.

You might wonder why I stress this point. Well, guess what--some guys in our forums posted your "exact" numbers and now claim that you used figures from two P4 systems that ran side-by-side, and that the calculations are completely accurate. Others have picked them up, and I am just counting the hours until somebody takes them for real benchmark numbers and they make their way into the low-quality, high-volume print magazines of Germany. You get the idea.

but, eh. no worries. Real figures will be around in one month's time anyway.

ta,
-Sascha.rb

Reminds me of the good old concept3D days when I posted an rumour about the GF3, even giving the link to the forum where the rumour was posted and saying in the text that it was only an rumour, but...
I found this "news" all over the net after a few days. One guy in the news-forum responded with an even more absurd rampage-rumour and this was spread all over the net too. Sometimes the net is really an strange place, and yes you are correct we will find this "benchmarks" in many magazines and online too.

 

[Bambers]

erf... Anand F*cked up.

We have explained this in the past but the way ATI’s supersampling (Rotated Grid Super-Sampling – RGSS) works is by rendering multiple copies of the scene and shifting each copy by a certain amount off the center. These copies were then blended together to produce the final image.

That'd be 3dfx, not ATi. And it's wrong anyway.

Thats what ATis 9700 techdemo says thats what the first SV did and I agree with you, its wrong. It doesn't explain the noise given in d3d 4x adn 2xq in the 7206 drivers for a start.