ATI 9800 VS NV FX 5950

[Tridam]

ShaderMark v2.0 - DirectX 9 HLSL Pixel Shader Benchmark - ToMMTi-Systems (http://www.tommti-systems.com)

video mode / device info
(1024x768) X8R8G8B8 (D24X8) vram used 58720256
HAL (pure hw vp): NVIDIA GeForce FX 5950 Ultra (Anti-Detect-Mode: off, gamma correction: DAC)



options

pixel shader version: 2_0
partial precision: off
number of render targets: 1

results:
shader 2 ( Per Pixel Diffuse Lighting): 143 fps 6.9832 mspf 716 rendered frames
shader 3 ( Per Pixel Directional Light Shader (Phong)): 91 fps 11.0132 mspf 454 rendered frames
shader 5 ( Per Pixel Spot Light Shader (Phong)): 79 fps 12.7326 mspf 393 rendered frames
shader 6 ( Per Pixel Anisotropic Lighting): 88 fps 11.4073 mspf 439 rendered frames
shader 7 ( Per Pixel Fresnel Reflections): 83 fps 12.0482 mspf 415 rendered frames
shader 9 ( Per Pixel Car Surface Shader): 41 fps 24.3902 mspf 205 rendered frames
shader 10 ( Per Pixel Environment Mapping): 172 fps 5.8275 mspf 858 rendered frames
shader 11 ( Per Pixel Environment Bump Mapping): 140 fps 7.1179 mspf 703 rendered frames
shader 12 ( Per Pixel Bump Mapping): 73 fps 13.6346 mspf 367 rendered frames
shader 13 ( Per Pixel Shadowed Bump Mapping): 43 fps 23.4010 mspf 214 rendered frames
shader 14 ( Per Pixel Veined Marble Shader): 56 fps 17.7093 mspf 283 rendered frames
shader 15 ( Per Pixel Wood Shader): 55 fps 18.3052 mspf 274 rendered frames
shader 16 ( Per Pixel Tile Shader): 43 fps 23.3827 mspf 214 rendered frames
shader 17 ( Fur Shader With Anisotropic Lighting): 6 fps 153.8826 mspf 33 rendered frames
shader 18 ( Per Pixel Refraction and Reflection Shader with Phong Lighting): 33 fps 30.7467 mspf 163 rendered frames

 

[micb]

Thank you very much Tridam, those are very helpfull.

Those scores are a fair bit lower as well.

 

[Tridam]

Thank you very much Tridam, those are very helpfull.

Those scores are a fair bit lower as well.

The core clock of his card seems to be ~550 MHz.

 

[micb]

550 / 1050 memory is the clock rates of the Gainward CoolFX PowerPack! Ultra/1800 XP 256M Golden Sample when in enhanced mode.

I can't wait to get a reply about the card from him now.

 

[nelg]

Are you sure that that is not the clock speed of the is the "Asusward HotXP-ouse Super Duper DandyFX-lnmop Nuclear Power Pack XT ultra 512 MB Golden Urine Sample" card?

God, that card must come in a big box with a name like that.

 

[Pete]

FX5950 is never 8x1. It's always 4x2. The only exception is the "Quad 8 Mode". The chip works on two quads in the same clock cycle. The back end interpolates the depth (8 per cycle or 2 per pipeline) and stores 4 results on the depth data output and the 4 other results on the color data output. After, the combiners pass color and depth output unchanged to the ROPs. This is not 8x1. This is 8x0. No color calculations (shader or texturing) can be done in this mode as the color output is used for the 4 more depth datas.

Thanks for that slightly technical but very comprehensible explanation, Tridam. Merci beaucoup. Now all I have to do is learn exactly what a quad is....

So my next Q is, how often does Doom 3 use the FX's 8x0 mode? Judging by these synthetic numbers, am I to expect a +60% framerate advantage for the 5950, or will real framerates be compromised by the -20% texturing performance as well? Basically, will nV cards show as much of an advantage as Anand's 5800U review benches showed?

 

[Tridam]

FX5950 is never 8x1. It's always 4x2. The only exception is the "Quad 8 Mode". The chip works on two quads in the same clock cycle. The back end interpolates the depth (8 per cycle or 2 per pipeline) and stores 4 results on the depth data output and the 4 other results on the color data output. After, the combiners pass color and depth output unchanged to the ROPs. This is not 8x1. This is 8x0. No color calculations (shader or texturing) can be done in this mode as the color output is used for the 4 more depth datas.

Thanks for that slightly technical but very comprehensible explanation, Tridam. Merci beaucoup. Now all I have to do is learn exactly what a quad is....

So my next Q is, how often does Doom 3 use the FX's 8x0 mode? Judging by these synthetic numbers, am I to expect a +60% framerate advantage for the 5950, or will real framerates be compromised by the -20% texturing performance as well? Basically, will nV cards show as much of an advantage as Anand's 5800U review benches showed?

A quad is a block of 2x2 pixels.

I think that Doom 3 will be able to use the Quad 8 Mode for the first rendering pass but remember that this pass is really fast to render and that Radeon and GeForce will have the same Z output (8 per clock).

The Quad 8 mode is unusable with FSAA 4X (I don't know if it can work with FSAA 2X).

(Radeon's results showed here are Radeons 9800 numbers, not Radeon 9800 XT numbers !)

[edit] : it works with FSAA 2X but not with FSAA 4X as expected

 

[Xmas]

I think that Doom 3 will be able to use the Quad 8 Mode for the first rendering pass but remember that this pass is really fast to render and that Radeon and GeForce will have the same Z output (8 per clock).

The Quad 8 mode is unusable with FSAA 4X (I don't know if it can work with FSAA 2X).

(Radeon's results showed here are Radeons 9800 numbers, not Radeon 9800 XT numbers !)

[edit] : it works with FSAA 2X but not with FSAA 4X as expected

Doom3 will be able to use the Quad 8 Mode in the initial Z-only pass (or is it Z and ambient? Then it won't work) and in every shadow volume pass (one per light source)

So it seems like NV30/35 have 16 ROPs of which 8 can only do Z/stencil ops and 8 can do either color or Z/stencil ops

R3x0 can do 16 Z/stencil and color ops, IIRC.

 

[digitalwanderer]

It does look more like my mate in America has fed me some tweeked stats, I hope to get a reply from him on this soon.

No offense, but that's why I don't trust other peoples stats for direct comparisons. I really feel you get a LOT more accuracy running it on the same PC and just switching cards, it eliminates a bunch of variables.

Interesting stuff none-the-less though, thanks.

 

[micb]

Yes I agree for offical reviews same host machine, loads and testing methodolgy are required to be used.

This was just an informal comparision posted to a message board to see what poeple on here think.

I am glad I posted this thread, as allways (on forums) dicussion evolves and expands. In this case about NV architecture and Doom 3.

If we do find the american card was clocked higher, then this post shows that even upclocked it is not as fast as ATI 9800.

 

[micb]

Here is his reply:

I can re-run it.
The clock rates I used were as reported by powerstrip, and confirmed in the clock setting page in the display settings panel.

The thing is my card is straight from nVidia, and was selected specificly for my next project.
So I can't say that it is exactly the same as the review models.

I will run the test both at non-overclocked and overclocked to see if it works out with the fill rate.

Here are the results at default settings:

Fillrate Tester
--------------------------
Display adapter: NVIDIA GeForce FX 5950 Ultra
Driver version: 6.14.10.5216
Display mode: 1024x768 A8R8G8B8 85Hz
Z-Buffer format: D24S8
--------------------------

FFP - Pure fillrate - 2132.696533M pixels/sec
FFP - Z pixel rate - 2139.774902M pixels/sec
FFP - Single texture - 1069.791382M pixels/sec
FFP - Dual texture - 1069.757324M pixels/sec
FFP - Triple texture - 713.167664M pixels/sec
FFP - Quad texture - 534.848206M pixels/sec
PS 1.1 - Simple - 1069.770386M pixels/sec
PS 1.4 - Simple - 713.116821M pixels/sec
PS 2.0 - Simple - 534.862122M pixels/sec
PS 2.0 PP - Simple - 713.156494M pixels/sec
PS 2.0 - Longer - 267.415985M pixels/sec
PS 2.0 PP - Longer - 427.852631M pixels/sec
PS 2.0 - Longer 4 Registers - 267.400665M pixels/sec
PS 2.0 PP - Longer 4 Registers - 534.845459M pixels/sec
PS 2.0 - Per Pixel Lighting - 106.967201M pixels/sec
PS 2.0 PP - Per Pixel Lighting - 76.402199M pixels/sec



And at 550/1020:

Fillrate Tester
--------------------------
Display adapter: NVIDIA GeForce FX 5950 Ultra
Driver version: 6.14.10.5216
Display mode: 1024x768 A8R8G8B8 85Hz
Z-Buffer format: D24S8
--------------------------

FFP - Pure fillrate - 2144.551514M pixels/sec
FFP - Z pixel rate - 2139.699219M pixels/sec
FFP - Single texture - 1069.731812M pixels/sec
FFP - Dual texture - 1069.768311M pixels/sec
FFP - Triple texture - 713.191650M pixels/sec
FFP - Quad texture - 534.835327M pixels/sec
PS 1.1 - Simple - 1069.810181M pixels/sec
PS 1.4 - Simple - 713.350769M pixels/sec
PS 2.0 - Simple - 713.132202M pixels/sec
PS 2.0 PP - Simple - 713.233276M pixels/sec
PS 2.0 - Longer - 305.655029M pixels/sec
PS 2.0 PP - Longer - 534.843689M pixels/sec
PS 2.0 - Longer 4 Registers - 305.634247M pixels/sec
PS 2.0 PP - Longer 4 Registers - 534.854431M pixels/sec
PS 2.0 - Per Pixel Lighting - 118.876770M pixels/sec
PS 2.0 PP - Per Pixel Lighting - 152.842331M pixels/sec

This is really strange, the stock numbers are different to the last set.

The Z values have gone down in both new tests.

The Pixel Shading is the same be the same in new tests, but different in the older test.

The per pixel lighting was high in the older test, low in the stock new test and high in the overlocked test.

His numbers don't seem reproducible and therefor look rarther invalid.