Some new findings on Quake 3 with TurboGL drivers!!

[Ricardo]

Hi Manoj,


Thanks for the link. Very poor results indeed. Do you know how to make the crunsher benchmark ?


Regards,
Ricardo

[franksch3]

Hey Ozymandis,

there is a performance hit between vertex and lightmap lighting. Let me explain:

The Voodoo2 when it launched had support for multitexturing. When a multitextured polygon is drawn both pipelines are used, together pumping out one pixel per clock. However, when only one texture is used only one pipeline is used and the other is idle, so one pixel is outputted per clock.

However, the TNT1/2 and the G400 MAX and probably all other cards can use both pipelines when single texturing, resulting in two outputted pixels per clock. When multitexturing they also only output one pixel per clock.

So on a Voodoo2/3, you don't have a performance difference between vertex and lightmap, other than that extra bandtwidth is needed for the lightmap. This bandwitdh requirement isn't that high since the lightmap isn't detailed at all.

On all other cards the fillrate with vertex lighting is twice as high as when using a lightmap. But because of other overhead (game physics, transformation, music, sounds, etc) this performance hit is much smaller. But there IS a performance hit.

For me I rather have a performance hit and have a beautifully lighted world than put up with more frames and a ugly non inspiring, non atmospheric enviroment. But others have their own opinion about that, which I also respect.

Hope this explains some things.

Frank

[This message has been edited by franksch3 (edited 10-12-1999).]

[ManojM]

Ricardo:

I forgot where I downloaded the crusher demo for Quake III but I have it. I haven't tried it yet. May try it later on.

It is about 1.6megs in size. there is even a crack file which is very small. In total it should be about 1.7 megs. If you want I can e-mail it to you.

Regards,
Manoj Mahtani

[Ricardo]

Hi Manoj,

That would be great. My email is ricardo.oliveira@optimus.pt.

Thanks,
Ricardo

[Deuge]

Just to add some pees to the soup, I thought I'd post my results using the same settings as above:
Demo - q3demo1
Color Depth - 32bit
Full screen - on
lighting - lightmap
geometric detail - high
texture detail - full
texture quality - 32bit
texture filter - bilinear
sound enabled & quality: high

System specs: P3 450 o'c to 580; 128Mb PC133 RAM; G400 vanilla o'c 126%; SB16(ISA); Aopen AX6BC Pro MB.

results:
16bit Z buffer 32bit Z buffer
640 - 75.1 69.6
800 - 57.2 49.1
1024 - 37.2 30.9

Its interesting to note the biggest drop by using the 32bit Z is at the 800 rez - maybe more performance can still be sqeezed out of the TurboGL...

[Ozymandis]

Yeah, actually you're right. But in real world terms it is only 2fps on my G400max. Still though, thank you reminding me.

[This message has been edited by Ozymandis (edited 12 October 1999).]

[Himself]

On all other cards the fillrate with vertex lighting is twice as high as when using a lightmap. But because of other overhead (game physics, transformation, music, sounds, etc) this performance hit is much smaller. But there IS a performance hit.

Vertex lighting up to lately was done on the cpu side, and hence wouldn't have a fillrate influence on the video card from what I understand. The way Q3Test uses vertex lighting is also done with the cpu and it's why it doesn't fully support the G256's T&L processing, just the transformation bits. That is the whole point of cards like the G256, they can do vertex lighting on the card since they have the information to calculate it from the transformation step. I am not sure exactly what format the data sent to the card is in, perhaps it could just be vertex generated lightmaps. In that case the fillrate load would be the same. With the G256 for instance you wouldn't have nearly as much fillrate load, since the lighting would be calculated, less video ram access.