pipeline and texture units

[diuby]

Hello guys,

I would like to know some G400 chip spec. In an Ace's Hardware article (Ultimate 3D Video Deathmatch part2!) they said that G400 have 3 pipeline! So I suppose the chip has three texture units and it can render 3 texture per pixel/s (3dmark 2000 also said that it can apply 3 texture with a single pass - you can see this in the system info -)! But if the chip have to render a dual textured pixel the third is used or not?
I think it is otherwise i don't understand why with bump mapping (3 texture)there is a strong performance hit at all the resultions.
But if so than the G450Max in theory can render 450M single textured pixels and not 300M as many reviews I have read said!

Can you give me your opinion please.

K6-III 400@428 MHZ
VIA MVP3 chipset 128Mb RAM
G400MAX 150/200
Win 98

[Humus]

The G400 has only two texturing units and two pipelines. Thus it'll need two passes for the EMBM. However, the driver will accept three textures in a single pass from the API, but internally it'll two passes.

[xortam]

You can find some specs here.

[C=64]

However, the driver will accept three textures in a single pass from the API, but internally it'll two passes.

What the hell!?! I'm glad someone mentioned this. I'm was just about to get started with multitexturing, and was going to take whatever OpenGL told me on faith.

That's just friggin' unbelievable that Matrox would be willing to do something so... stupid! Instead of giving the programmer a chance to find out the TRUE number of texture units and have his code act accordingly (i.e. minimize state changes), Matrox has decided, in it's infinite wisdom, to cripple the performance of its card.

Wonderful...

C=64

[diuby]

I quote the original article (Ultimate 3D Video Deathmatch, part 2 - image quality) and what they said about trilinear filtering:

------------------------------------------
TRILINEAR FILTERING: FILL-RATE HOG

....how much of a performance degradation do video cards implementing trilinear filtering experience? The Matrox G400, nVidia's GeForce, and the Savage series all implement real trilinear filtering.

Here's an interesting comment from our messageboard about the G400 (GLDM):

Matrox's approach is interesting. They seem to have realized that most 3D games use multitexture, texturemaps, and lightmaps (which are alternate mode texture maps). By using 3 pipelines, they can apply the lightmap, then render 2 versions of the texture to create a trilinear pixel, all in one cycle. An alternate use for this is to do lightmap, texturemap, bumpmap.
Matrox seems to have found a way to limit the fillrate decrease (only 50% instead of 100%), while S3 claims its Savage chips do trilinear filtering for free. We decided to investigate the matter using 3DMark 99 (1024x768x32):

Chipset Trilinear Filtering
(Bilinear = 100%)
ATI Rage Fury Maxx 85.6%
GeForce SDR (Asus) 95.2%
GeForce DDR (Guillemot) 94.7%
Savage 2000 74.7%
TNT2 150/183 (Xentor) 90%
G400 71.4%

.....Both the G400 and the Savage 2000 suffer the most. It is clear that those two implement REAL trilinear filtering. I've got a hunch that in the case of the Savage 2000, this is due to bandwidth limitations. Trilinear filtering might be done for free, but it still sucks up bandwidth. The Savage 2000 has the slowest memory interface (155 MHz versus 166 MHz for the GeForce). The G400, however, probably suffers from the fact that it needs three pipelines instead of two in order to perform trilinear filtering. This reduces the fill-rate of the G400 by 50%. So, the G400 performs better in high quality (32-bit, trilinear filtering) than we think. Probably much better than the TNT2, in fact, if it were to use real trilinear filtering.
---------------------------------------------

So they believed that a there is a third pipeline in the G400 chip!!

Boh, I don't know....

[Humus]

Originally posted by C=64:
What the hell!?! I'm glad someone mentioned this. I'm was just about to get started with multitexturing, and was going to take whatever OpenGL told me on faith.

That's just friggin' unbelievable that Matrox would be willing to do something so... stupid! Instead of giving the programmer a chance to find out the TRUE number of texture units and have his code act accordingly (i.e. minimize state changes), Matrox has decided, in it's infinite wisdom, to cripple the performance of its card.

Wonderful...

C=64

They have to do it this way to make support for EMBM possible since it need three textures. A correctly implemented driver would be able to handle this without a performance decrease.

[Humus]

diuby:
We can do the math ... G400 has 250Megatexels fillrate, and a clockfrequency of 125MHz.
250/125 = 2

The G400 has two texturing units which are shared between the two texturing pipelines, thus it doesn't gain more than 10-20% by using multitexturing while other cards would gain 30-40%.

[Humus]

I finally have 100 post at this forum, and have become a MURCer! Yeehaw!

Too sad this happends now that I have replaced my G400 with a Radeon.

[paul]

yeah, but your last 3 posts all show '101'

[Humus]

Yes, the last one was my 101, now I'm even higher ...