Benchmarks of a GF3 with 28.90, 30.87, and 42.01 drivers

[NickSpolec]

I have done some piliminary testing of 3 different driver versions from three different driver generations to see what works the best on a GF3 (and my system).

While these benchmarks are not entirely thorough (yet), they do tell a little bit about the drivers and what might work best.

First, my system...

Athlon XP 1600+ @ 1.95ghz (1700+)
MSI KT3 Ultra 2 @ 166FSB
512MB OCZ PC2700 Ultra Rev 3.2 @ DDR 350
VisionTek GeForce 3 @ 250/540
Sb Live! 5.1
Windows 98SE
Dirext X 8.1 (For Drivers 28.90 and 30.87, DX9 was used for 42.01)

Next, the settings...

All drivers were used with Anti-Aliasing Off, Antisopic Filtering set to 2x (OpenGL was set to PERFORMANCE). VSync was OFF. The 42.01 Drivers were set to BALANCED in the Performance setting. The LOD Bias set to 0*. And DX9 was insta

*It should be noted that each driver set has dramatically different LOD Bias implamentation.

28.90 has the least aggressive, with also the inability to set the LOD to a negative bias setting without patching the driver first.

30.87 has the most aggressive LOD Bias of the bunch, as do all the 3x.xx drivers. LOD Bias of 0 is pretty shimmery, and after testing the drivers at the default of LOD 0, I changed it to +1, because 0 is too shimmery.

42.01 seems to be slightly better then 28.90, but not as aggressive as the 3x.xx drivers.


Again, these are preliminary tests, actual gaming tests will be added.



3DMark was used for basic D3D testing. It was used with the default testing settings


3DMARK (330)


TOTAL

28.90 (DX 8.1) = 7836
30.87 (DX 8.1) = 7841
42.01 (DX9) = 8028






******
28.90
******

3DMark Score 7836
Game 1 - Car Chase - Low Detail 124.5 fps
Game 1 - Car Chase - High Detail 50.3 fps
Game 2 - Dragothic - Low Detail 111.0 fps
Game 2 - Dragothic - High Detail 63.6 fps
Game 3 - Lobby - Low Detail 123.1 fps
Game 3 - Lobby - High Detail 61.1 fps
Game 4 - Nature 37.5 fps
Fill Rate (Single-Texturing) 513.4 MTexels/s
Fill Rate (Multi-Texturing) 1006.6 MTexels/s
High Polygon Count (1 Light) 26.1 MTriangles/s
High Polygon Count (8 Lights) 6.4 MTriangles/s
Environment Bump Mapping 126.0 fps
DOT3 Bump Mapping 86.5 fps
Vertex Shader 64.9 fps
Pixel Shader 73.3 fps
Advanced Pixel Shader 59.5 fps
Point Sprites 18.5 MSprites/s



*****
30.87
*****

3DMark Score 7841
Game 1 - Car Chase - Low Detail 123.8 fps
Game 1 - Car Chase - High Detail 50.7 fps
Game 2 - Dragothic - Low Detail 114.2 fps
Game 2 - Dragothic - High Detail 63.9 fps
Game 3 - Lobby - Low Detail 123.4 fps
Game 3 - Lobby - High Detail 61.1 fps
Game 4 - Nature 35.7 fps
Fill Rate (Single-Texturing) 517.8 MTexels/s
Fill Rate (Multi-Texturing) 1021.7 MTexels/s
High Polygon Count (1 Light) 26.5 MTriangles/s
High Polygon Count (8 Lights) 6.5 MTriangles/s
Environment Bump Mapping 127.8 fps
DOT3 Bump Mapping 87.7 fps
Vertex Shader 65.5 fps
Pixel Shader 74.0 fps
Advanced Pixel Shader 60.3 fps
Point Sprites 18.8 MSprites/s



*****
42.01
*****

3DMark Score 8028
Game 1 - Car Chase - Low Detail 120.7 fps
Game 1 - Car Chase - High Detail 46.9 fps
Game 2 - Dragothic - Low Detail 118.4 fps
Game 2 - Dragothic - High Detail 64.1 fps
Game 3 - Lobby - Low Detail 121.8 fps
Game 3 - Lobby - High Detail 59.6 fps
Game 4 - Nature 50.3 fps
Fill Rate (Single-Texturing) 513.9 MTexels/s
Fill Rate (Multi-Texturing) 1007.6 MTexels/s
High Polygon Count (1 Light) 28.0 MTriangles/s
High Polygon Count (8 Lights) 6.4 MTriangles/s
Environment Bump Mapping 127.7 fps
DOT3 Bump Mapping 88.4 fps
Vertex Shader 50.3 fps
Pixel Shader 73.6 fps
Advanced Pixel Shader 61.5 fps
Point Sprites 18.5 MSprites/s






As we can see, in the Car Chase, both 28.90 and 30.87 are just about identical, with 42.01 falling behind by about 4 FPS. It's interesting that even withs it's far more aggressive LOD Bias that 30.87 beats 42.01 and comes next to 28.90 (both these drivers, again, are not as aggressive).

Dragothic Low belongs to 42.01, topping the 28.90 drivers by 7, and the 30.87 by 4. The High version of the bench is also tops on the 42.01 by 1 frame, while 28.90 and 30.87 are neck and neck.

The Lobby benches again show the 42.01 falling behind a bit, by 2 frames on both High and Low, 30.87 and 28.90 are the same.

Nature no doubt is the 4x.xx drivers high point, with it crushing both 28.90 and 30.87 13 and 15 points, respectively. IT should be noted that it looks like LOD bias really matters a lot in the Nature bench, with the 30.87 driver being hurt by it's aggressive LOD.

The fillrate bench goes to the 30.87 drivers, as both Single and Double are tops on it, with the 28.90 and 42.01 drivers being the same.

The 42.01 drivers are the best in the Low Polygon count bench, beating the other two by about 2 million, with the High Polygon bench the same on all three sets.

Enviormental Bump Mapping is lowest on the 28.90 drivers, falling behind both 30.87 and 42.01 (which are the same) by 1 FPS, but even that is within the margin of error. The DOT3 Bump Mapping is best on 42.01, which gets progressively better with each newer driver set.

The Vertex Shader bench shows a serious flaw in the 42.01 drivers. It plummets beind both 28.90 and 30.87 by 14 FPS.\

And while it could be said that the 4x.xx drivers are best for Pixel Shader performance, the two Pixel Shader benches don't show it. All score about the same. 4x.xx drivers are only optimized for Nature, I guess..

Finally, Point-Sprites remains the same on all drivers.

That does it for 3DMark. Next post will contain GL Excess.

[NickSpolec]

GL Excess is a great little benchmark. Many don't like it, but I do.

I ran the benchmark at 1024x768 and 32Bit color, again, with the same settings used for D3D (No AA, 2xAF, LOD 0, Vsync off).

Lets get right to it.



GL Excess 1.2

TOTAL

28.90 = 5161
30.87 = 5233
42.01 = 5696




Scenes 1/2/3 (Fillrate/CPUFPU/Polygon Count)

28.90 = 4087
30.87 = 4113
42.01 = 4773



Scenes 4/5/6 (Vram/Polygon Count/Polygon Count)

28.90 = 5673
30.87 = 5735
42.01 = 6633




Scenes 7/8/9 (CPUFPU/CPUFPU/Fillrate)

28.90 = 3444
30.87 = 3634
42.01 = 3858




Scenes 10/11/12 (Vram/Vram/Fillrate)

28.90 = 4961
30.87 = 4967
42.01 = 5012







*********
Scene 1, Fillrate Test
*********

28.90: Min = 90, Avg = 129, Max = 143
30.87: Min = 90, Avg = 130, Max = 143
42.01: Min = 100, Avg = 152, Max = 168

*********
Scene 2, CPU/FPU Test
*********

28.90: Min = 89, Avg = 116, Max = 143
30.87: Min = 89, Avg = 116, Max = 143
42.01: Min = 89, Avg = 116, Max = 143

*********
Scene 3, Polygon Count Test
*********

28.90: Min = 122, Avg = 168, Max = 201
30.87: Min = 123, Avg = 172, Max = 201
42.01: Min = 133, Avg = 210, Max = 253





*********
Scene 4, Vram Test
*********

28.90: Min = 55, Avg = 72, Max = 185
30.87: Min = 62, Avg = 74, Max = 198
42.01: Min = 64, Avg = 81, Max = 200

*********
Scene 5, Polygon Count Test
*********

28.90: Min = 83, Avg = 129, Max = 237
30.87: Min = 83, Avg = 131, Max = 249
42.01: Min = 90, Avg = 159, Max = 319

*********
Scene 6, Polygon Count Test
*********

28.90: Min = 98, Avg = 126, Max = 163
30.87: Min = 99, Avg = 128, Max = 165
42.01: Min = 110, Avg = 147, Max = 183





*********
Scene 7, CPU/FPU
*********

28.90: Min = 66, Avg = 92, Max = 112
30.87: Min = 66, Avg = 94, Max = 112
42.01: Min = 71, Avg = 103, Max = 126

*********
Scene 8, CPU/FPU
*********

28.90: Min = 83, Avg = 100, Max = 113
30.87: Min = 89, Avg = 101, Max = 112
42.01: Min = 97, Avg = 112, Max = 126

*********
Scene 9, Fillrate
*********

28.90: Min = 86, Avg = 103, Max = 112
30.87: Min = 88, Avg = 103, Max = 118
42.01: Min = 98, Avg = 119, Max = 143




*********
Scene 10, Vram
*********

28.90: Min = 49, Avg = 71, Max = 125
30.87: Min = 50, Avg = 72, Max = 126
42.01: Min = 51, Avg = 77, Max = 128

*********
Scene 11, Vram
*********

28.90: Min = 89, Avg = 111, Max = 144
30.87: Min = 89, Avg = 113, Max = 144
42.01: Min = 99, Avg = 129, Max = 168

*********
Scene 12, Fillrate
*********

28.90: Min = 109, Avg = 115, Max = 133
30.87: Min = 109, Avg = 116, Max = 142
42.01: Min = 123, Avg = 133, Max = 166





Obviously, 42.01 is the best here. Weather this shows real world OpenGL performance is unknown, but it should be a good indication. 28.90 is the worse, with 30.87 bettering it slightly.

[Blakhart]

Thanks for doing this. Is there a way to set LOD manualy without something like rivatuner or gtu? I would like to set my lod on this 3082.
Again, thanks.

[NickSpolec]

Well, inside the Nvidia control panel should be an option called Mipmap Detail Level.

[Blakhart]

So this is a d3d-only adjustment?

[NickSpolec]

Yes --- The Nvidia Control panel only lets you adjust D3D.

[Blakhart]

Thanks bro.

[g0at]

very interesting, thanks for the info m8

[Blakhart]

After looking at the gf4ti die, it seems that the electronics given to dx7/legacy games that do not use pixel shaders is perhaps 1/10th the die. This may explain why tribes(1) looks so crappy on my gf4ti, It (gf4ti) has plenty of bandwidth of mem and fps, but a horrible image, with buildings being transparent in some views, and lots of jaggies where say a structure meets the ground. This is regardless of res/af/aa and color depth. Tribes2, a different engine, looks fine.

Do you suppose that my gf3 will devote more die area/processing power to dx7/legacy games than the gf4ti? I have seen that the 3082 driver gives best image quality with the gf4 in dx7 games.
Perhaps it is just tribes, dunno, but since it is the game I play the most, I may pitch my gf4 to get the nice image I remember from my gf3.


Anyone else have iq probs in dx7 games with a gf4ti?

[Zarich]

Tribes 1 was written in Glide. Opengl was added as sort of an extra side item. Your never going to get good IQ in tribes with an Opengl based card. If you want good IQ in tribes.. go buy a used voodoo.

[EciDemon]

Thanx NickSpolec this is exelent :)
There are not many tests like these around nowdays with GF3 and a fast comp. You saved me the troubble as I was going to test my GF3 with different drivers to se what was best for me. I didnt really want to change drivers below 4x.xx as the video out might break again :( According to your tests Ill just use my current drivers as they seems to be better overall :)

Thanx mate :)

[Blakhart]

Hee heee, found a voodoo, a slot A mobo, some sdram, and now to add a tbird 850. Retro tribes goodness will be mine.


As to the benches, I did not see any reference to iq, even if subjective, was there better/more iq in the 3xxx and 2xxx drivers than the 4xxx? My proposal is that to get speed, nvidia must be cutting corners, as far as the api and any game itself are concerned. You are still dealing with the same hardware, just leaving room for software trickery.

Anyone?

[NickSpolec]

As I said, the 30.87 drivers have by far the most aggressive LOD Bias.

A LOD Bias of 0 on the 30.87 drivers = an LOD Bias of -2 or -3 on the 28.90 and 42.01 drivers.

Not only that, even though it has a more aggressive LOD Bias, it still gets the same speed in most benchmarks.

But the LOD Bias on the 30.87 drivers is too aggressive at 0, as textures/polygons/effects in the distance will shimmer too much, a side effect to low LOD Bias. On 30.87, I put it at +1, and it still looks better then 42.01 and 28.90 at 0.

[NickSpolec]

You could try and use a GLIDE-Wrapper to emulate Glide on your GF4 to play Tribes in it's native API.. I know that it does not always work that great (as Glide-Wrappers were primarily built for The UltraHLE emulator), but it's worth a try. You could, of course, get far better speed. Image IQ, not sure.

[Blakhart]

Thanks for the ideas, Nick.
See ya after some sperimentation, put a gf3 back in and gonna try tribes again. If it looks good, then it stays. If not, the gf4ti will go back in and tribes1 will only be played on the voodoo box.


The wrapper can be had from where?

[NickSpolec]

Just recently had to reinstall Windows, and instead of using DX9, I used DX8.1 with the 42.01 drivers.

The results are a 150 point increase, see for yourself.


************


RESULTS
3DMark Score 8187
Game 1 - Car Chase - Low Detail 122.2 fps
Game 1 - Car Chase - High Detail 49.5 fps
Game 2 - Dragothic - Low Detail 119.7 fps
Game 2 - Dragothic - High Detail 65.4 fps
Game 3 - Lobby - Low Detail 123.4 fps
Game 3 - Lobby - High Detail 61.3 fps
Game 4 - Nature 50.5 fps
Fill Rate (Single-Texturing) 513.4 MTexels/s
Fill Rate (Multi-Texturing) 1007.6 MTexels/s
High Polygon Count (1 Light) 27.9 MTriangles/s
High Polygon Count (8 Lights) 6.4 MTriangles/s
Environment Bump Mapping 120.8 fps
DOT3 Bump Mapping 88.3 fps
Vertex Shader 50.4 fps
Pixel Shader 74.0 fps
Advanced Pixel Shader 61.7 fps
Point Sprites 18.5 MSprites



Increases across the board against DX9 and 42.01, except for the Enviormental Bump Mapping, which fell 7 FPS.



Blackheart,

You can find a Glide wrapper here...

http://ngemu.com/n64/plugins.php?page=wrapper


You can also look on the web for more.

[Blakhart]

Thanks, I am going to try this wrapper right now. cu

[NickSpolec]

That is actually an outdated page.. There is an actively updated wrapper somewhere..

You may have to search the web for "Glide wrapper".

[NickSpolec]

Oh, here :)

http://www.voodoofiles.com/type.asp?cat_id=14

Lots of them

[Blakhart]

Tried the emu one and no go for tribes, will try the voodoofiles jobbies. Thanks again.