why do fx cards perform so badly in 3dmark05

[zakelwe]

Eh? Not even sure what your saying. Yes its both pixel shaded and vertex limited. But vertex performance plays a substancial Roll in overall performance as well.

Surely Pixel performance is important in 3dmark. But so its Vertex performance.

OK then, let me put it this way, looking at your figures and aassuming you had a 6800 non ultra and Unwinders rivatuner would you rather have an extra VP or an extra 4 pipelines ?

Increase from 12 to 16 pipelines = > 400
Increase from 5 to 6vp => 100

The word limited is, to be honest, also incorrect as it implies it is stopping the score going up. This is not the case, even with 4VP at 16 pipelines you get 600 points more than 6vp at 8 pipelines .

It would be more accurate to say that 3dmark05 is both pixel and vertex dependent but with pipeline quads being more important.

Regards

Andy

[ChrisRay]

OK then, let me put it this way, looking at your figures and aassuming you had a 6800 non ultra and Unwinders rivatuner would you rather have an extra VP or an extra 4 pipelines ?

Increase from 12 to 16 pipelines = > 400
Increase from 5 to 6vp => 100

The word limited is, to be honest, also incorrect as it implies it is stopping the score going up. This is not the case, even with 4VP at 16 pipelines you get 600 points more than 6vp at 8 pipelines .

It would be more accurate to say that 3dmark05 is both pixel and vertex dependent but with pipeline quads being more important.

Regards

Andy

Oh jeez, you're getting bent out of shape over a techicality? The whole point of the chart was to show that 3dmark is sensitive to vertex performance and its not entirely a pixel shaded limited demo, In which case, This hasnt changed a bit,

of course 4 pixel pipelines is going to have a dramatic increase on effect, you gain 4 shader dedicatedt shader units +4 shared shader units with 1 quad activation, You only lose 1 vertex unit by disabling vertex performance.

And no this is not a 6800NU, If you look at the chart it clearly labels it a GT.

[zakelwe]

Oh jeez, you're getting bent out of shape over a techicality? The whole point of the chart was to show that 3dmark is sensitive to vertex performance and its not entirely a pixel shaded limited demo, In which case, This hasnt changed a bit,

of course 4 pixel pipelines is going to have a dramatic increase on effect, you gain 4 shader dedicatedt shader units +4 shared shader units with 1 quad activation, You only lose 1 vertex unit by disabling vertex performance.

And no this is not a 6800NU, If you look at the chart it clearly labels it a GT.

A technicality ? I didn't comment on the cart but your wording.

To quote you

"But geometry plays a huge role "

this is an exageration.

" But its just as much vertex limited as it is pixel limited."

no it isn't.

I didn't say anything about your chart being a 6800nu, I said assuming you had a 6800nu because that is the card most people will be trying to increase the pipes and vertex units on.

I keep seeing these comments about 3dmark05 being vertex limited and they seem to be coming from B3d it seems. If you look at the top scores for the 6800nu though you will see that a prereqisit is that the card is a 16 pipeline card. I have to say that if someone gave me a 24 pipeline card with only 6 vertex shaders I would not be too unhappy :D

Regards

Andy

[ChrisRay]

A technicality ? I didn't comment on the cart but your wording.

To quote you

"But geometry plays a huge role "

this is an exageration.

" But its just as much vertex limited as it is pixel limited."

no it isn't.

I didn't say anything about your chart being a 6800nu, I said assuming you had a 6800nu because that is the card most people will be trying to increase the pipes and vertex units on.

I keep seeing these comments about 3dmark05 being vertex limited and they seem to be coming from B3d it seems. If you look at the top scores for the 6800nu though you will see that a prereqisit is that the card is a 16 pipeline card. I have to say that if someone gave me a 24 pipeline card with only 6 vertex shaders I would not be too unhappy :D

Regards

Andy


Geometry does play a large role. Did you actually "look" at the chart? Comparing 2 vertex units to one quad, There is a substancial loss in performance. Sure one quad has a larger impact than 1 vertex unit, But 1 quad implements 4 dedicated shader units and 4 shared units, 1 vertex unit is exactly 1 vertex unit.,

In Some cases the vertex performance lost by 2 units is > than the loss of 1 quad, Therefore it shows 3dmark is also quite dependent upon geometry, Some tests, paticularly test 2 are unrealistically dependent upon it.

The reason that people are saying its geometry dependent too because it is. And it helps describe some disparity people are seeing with 3dmark in regards to pixel pipeline/vertex performance between ATI and Nvidia, Where the actual vertex performance is playing a large role.

I'm glad you like your quads, But it no way makes vertex performance any less important,.

And no, You said you assumed I was using a 6800NU, if you read the graph it clearly says I was using a GT with clock speeds illustrated.


Edit:

To reiterate, 1 quad "Should" have a larger impact on performance than 2/1 vertex units, Because it has 8 shader units per quad, But the fact is 2 vertex units are causing disparity between 1 quad, shows that the test relies on vertex data quite a bit.

[zakelwe]

"OK then, let me put it this way, looking at your figures and aassuming you had a 6800 non ultra and Unwinders rivatuner would you rather have an extra VP or an extra 4 pipelines"

I was talking about people having a 6800nu and uprating otherwise I would have said

"OK then, let me put it this way, looking at your figures would you rather have an extra VP or an extra 4 pipelines"

sorry if that was not clear.

The fact is that for both modded X800's and nvidia 6800nu's getting a big gain when going to 16 pipelines means to me that the pieplines are what is important to have and the vertex is a secondary benefit, as you say because you get 4 pipelines at a go.

We'll have to disagree on this oen I guess.

Regards

Andy

[ChrisRay]

"OK then, let me put it this way, looking at your figures and aassuming you had a 6800 non ultra and Unwinders rivatuner would you rather have an extra VP or an extra 4 pipelines"

I was talking about people having a 6800nu and uprating otherwise I would have said

"OK then, let me put it this way, looking at your figures would you rather have an extra VP or an extra 4 pipelines"

sorry if that was not clear.

The fact is that for both modded X800's and nvidia 6800nu's getting a big gain when going to 16 pipelines means to me that the pieplines are what is important to have and the vertex is a secondary benefit, as you say because you get 4 pipelines at a go.

We'll have to disagree on this oen I guess.

Regards

Andy

And what are we disagreeing on? once again 1 quad "Should" reflect a larger gain over 2 vertex units, The entire purpose of this test was to show how vertex performance can effect the overall score, And when compared to a pixel pipeline. it's much more dramatic, If you could disable quads, 1 pipe at a time. You could disable shader units 2 at a time. Which would be reflective of how each Pixel shader unit reflects overall performance compared to each vertex unit.

The purpose of the test is to show they scale. Nothing more. Nothing less. In this case, 8 pixel shader units (on 1 quad) Can lead to less performance increases (for example Game test 2) than 2 vertex units.

The fact that pixel pipelines provided a higher boost was never disputed, It was merely a means to an end of showing how much 3dmark05 also emphasizes on vertex performance.

[zakelwe]

Disagreement on your wording "huge" and the word " limited "

I prefer "some" and " dependent ". Only two words but I think that it better reflects real life behaviour of this benchmark.

Regards

Andy

[ChrisRay]

Disagreement on your wording "huge" and the word " limited "

I prefer "some" and " dependent ". Only two words but I think that it better reflects real life behaviour of this benchmark.

Regards

Andy


a technicality, it doesnt change the purpose of this test or the data it represents, And it shows that you completely mis understood the nature of the test. The fact that 2 vertex units can offer a similar loss to 8 pixel shader units, Show a rather large emphasis on vertex performance.

Obviously the Pixel and vertex units are not designed the same at or with the same function in mind. But when you see how close disabling 1 quad to 2 vertex units is, It shows to me that 3dmark05 is unreasonably vertex dependent.

[kahloq]

To answer the topic starters post. The 9800 pro is a good card. So are the 5900/5950. 9800 beats 5900/5950 in D3D, but fx series wins in opengl However these cards are still priced out rather high after the new gens came out and are still priced to high. If you can wait, get a 6600GT AGP. OR, wait for the 6700 series. There is also the possibilty of getting a 6200 and soft modding it to a 6600. There was an article on the front page.
http://translate.google.com/translate?sourceid=mozclient&u=http%3A//www.hardwareluxx.de/story.php%3Fid%3D1225

[Fahim]

Sounds like a silent war is going in this thread........

Should I post my scores? My XT PE score & the X800 PRO score @ 12 pipe and 16 pipe??

[zakelwe]

a technicality, it doesnt change the purpose of this test or the data it represents, And it shows that you completely mis understood the nature of the test. The fact that 2 vertex units can offer a similar loss to 8 pixel shader units, Show a rather large emphasis on vertex performance.

.

From your table

2 vertex shaders offer a loss of 300 to 500 points
8 pixel shaders offer a loss of 900 to 1400 points

So your conclusion of "The fact that 2 vertex units can offer a similar loss to 8 pixel shader units" does not match your own table. Your table shows 2x -3x greater loss when losing 8 pixel pipelines than 2 vertex units.

Like I said at the very beginiing, better to gain 4 pixel pipelines when modding a 6800 than the 1 vertex unit for 3dmark05.

Regards

Andy

[tyx]

OK, i see where this is going :)

I have decided to get a 5900XT by leadtek.
Thanks guys!!

[ChrisRay]

From your table

2 vertex shaders offer a loss of 300 to 500 points
8 pixel shaders offer a loss of 900 to 1400 points

So your conclusion of "The fact that 2 vertex units can offer a similar loss to 8 pixel shader units" does not match your own table. Your table shows 2x -3x greater loss when losing 8 pixel pipelines than 2 vertex units.

Like I said at the very beginiing, better to gain 4 pixel pipelines when modding a 6800 than the 1 vertex unit for 3dmark05.

Regards

Andy

You realise there are 8 pixel shader units in 1 quad? 2 quads is 16 shader units. You arguing a point you simply dont seem to understand.

1 quad should have a higher effect than 1 vertex unit, because of the pixel shader architecture of a quad. The graph here shows how 2 vertex units can offer a similar performance deficit to 1 quad, (IE 4 shader units and 4 non dedicated shader units, 8 Shader units per Quad, Including a 25% reduction in Fillrate)



http://members.cox.net/omega1979/nvnews/vertex.png

[Fahim]

It seems that difference between 16x1:6VP & 16x1:4VP is higher than difference between 16x1:6VP & 12x1:6VP.

Difference between 16x1:6VP & 16x1:4VP = 495 3DMarks
Difference between 16x1:6VP & 12x1:6VP = 407 3DMarks

That illustrates that 2VP is important than one Quad, atleast in 3DMark05.

[kev13dd]

To answer the topic starters post. The 9800 pro is a good card. So are the 5900/5950. 9800 beats 5900/5950 in D3D, but fx series wins in opengl However these cards are still priced out rather high after the new gens came out and are still priced to high. If you can wait, get a 6600GT AGP. OR, wait for the 6700 series. There is also the possibilty of getting a 6200 and soft modding it to a 6600. There was an article on the front page.
http://translate.google.com/translate?sourceid=mozclient&u=http%3A//www.hardwareluxx.de/story.php%3Fid%3D1225

Have you seen some of the benchmarks for Doom 3? 5900 vs 9800?

It's a pretty dang good tie, with the 5900 pulling ahead in some lower resolution, and the 9800 comming ahead with the eye candy

K

[Fahim]

There are some rumors with pretty tight proof that with upcoming 70.16 driver, FX Series PS 2.0 performance will be increase upto 60%. Check HERE (http://forums.guru3d.com/showthread.php?s=&threadid=112093&perpage=40&pagenumber=1)

[Fahim]

Have you seen some of the benchmarks for Doom 3? 5900 vs 9800?

It's a pretty dang good tie, with the 5900 pulling ahead in some lower resolution, and the 9800 comming ahead with the eye candy

K

I have the results...

ABIT Siluro GeForce FX 5900 OTES 128MB (400,850) w/ 66.31

1024x768x32 ~ High Settings = 38.1 FPS

ATI RADEON 9800 PRO 128MB (380,680) w/ Beta 8.07

1024x768x32 ~ High Settings = 48.4 FPS

1024x768x32 ~ High Settings = 43.6 FPS w/ CAT 04.8

[THE VIKING]

The FX series has never been particularly good with shader-intensive programs.