Xabre400 performance

[CoolAsAMoose]

Found some 3DMark 2001 scores for the Xabre400. The total score for the Xabre400 clocked at 250/500 MHz is 7317, compared to 6984 for GF3Ti200, 6077 for GF4MX440 and 6727 for the GF4MX460. I must say that this is impressive: I mean, you don't have that high expectations on SiS.

But a few of the numbers are strange. For single-texturing the Xabre400 only get 450.5 MTexels/s despite having 4 pipelines and running at 250 MHz (=> 1 GTexels/s theorethical).

The EMBM performance is also low: 43.5 fps compared to 103.6 fps for the GF3Ti200.

High polygon count and vertex shader numbers are better on the Xabre400, but the pixel shader ones are less than half of the GF3Ti200 ones!?!


Can this be explained with early drivers? Any other explanation?

http://www.vr-zone.com/cgi-bin/vb/showthread.php?s=1b6971a9b91ba42b7662bb897019072b&postid=36655#post36655

[pcchen]

To think positively, it may be a driver issue regaring pixel shader performance. A future driver revision will solve this problem.

To think negatively, it can be due to the lack of efficiency of memory bus for dependent texture access. For good dependent texture access performance, a long FIFO is required for buffering the memory requests. Without this, performance can be very bad when a lot of dependent texture access is required.

Anyway, if a 3D chip has no good dependent texture access performance, its value as a pixel shader compliance chip will be very small, since many interesting effects come with pixel shaders are related to dependent texture access.

[Joe DeFuria]

High polygon count and vertex shader numbers are better on the Xabre400, but the pixel shader ones are less than half of the GF3Ti200 ones!?!

Note that this was on a top-of-the-line CPU: P4 2.53 Ghz. IIRC, the Xabre doesn't have hardware vertex shaders, but passes it off to the CPU. We may be looking at a case where the CPU vertex shading outclasses the performance of the GeFore3 vertex shaders in these synthetic tests.

[Tygrus_Artesiaoa]

Well pixel shaders and bump mapping performance seem to not be so keen...dot3 mapping seems up to par but seeing as many older games and new ones still like embm and it's good for some materials...it's embm performance tanks.
As for single and multitexturing performance...does anybody actually know officially what sis has laid out for the chip...Looking at their chips support pages at sis.com.tw...they never have really alluded to how many pipelines there were and texture combiner units or tmu's-whatever you wanna call it...
Game performance seems decent in the scenes...

Think about it
at 250/500

450mtexels during single texturing almost makes me think 2 pipes
and 1660mtexels in multi texturing makes me think 4 tcu/tmu's per pipe
or when at 270/540 it did 488/1800
This could be a great thing...that many tcu per pipe could mean good trilinear/anisotropic performance as it could fetch many texels
Perhaps less pipes was easier to design for as well?

the sis 315 had/still has? not very good filtering though...it does approxiamated trilinear and not very good bilinear in some cases...could just be a driver thing...though xbit-labs in their review of one...got the AA to work...and it was good thing I guess...but not enuff performance to justify it.

[Tygrus_Artesiaoa]

Could just be the drivers are extremely bugged to the point or the version of the chipset is

that seemingly they lose 2 of the pixel pipe theoritical performance in single texturing

I still think though barring any serious driver or revision bugs

that
the 2 pipe 4tcu makes sense according to the numbers.

[CoolAsAMoose]

the 2 pipe 4tcu makes sense according to the numbers

According to this (http://www.vr-zone.com/Home/news142/news142.htm#2332) table, the Xabre400 has 4 pipes with 2 TMUs each.


Note that this was on a top-of-the-line CPU: P4 2.53 Ghz. IIRC, the Xabre doesn't have hardware vertex shaders, but passes it off to the CPU. We may be looking at a case where the CPU vertex shading outclasses the performance of the GeFore3 vertex shaders in these synthetic tests.

That could explain a lot. Because most new 3D cards sit inside 1GHz+ computers, it does make sense leaving out the hardware vertex shaders on a budget/mainstream level chip and use the gates for the pixel part instead - allthough it seems SiS may have failed there.

[Nappe1]

you must remember that when we are talking about PS performance on Xabre chips, those chips are targetted against GF4MX. They have about 30 million transistors when GF3Ti200 has what? something like over 60 million?

IMO, to Radeon 7500 and GF4MX class, it looks like pretty good so far, when we are talking chip performance with pre-shipping drivers.

I thnk Xabre is one of the major factors causing early die to GF4MX series, or at least it going to Lowend instead of Mainstream.

[Tygrus_Artesiaoa]

So I read that...still maybe I'm stubborn but I'm still wondering if the source is correct

That or as I said the revision is severely needin' some bug fixing in single texturing as far as pixel pipe performance

or it's as I said...

Hmm what the hell...maybe an email to Sis...well find everyting all out

[CoolAsAMoose]

Found some slides describing the Xabre400 here (http://www.tbreak.com/forums/showthread.php?postid=86642). Confirms vertex shaders run on CPU, and pixel shaders are 1.3.

[Tygrus_Artesiaoa]

...one of the slides says it has 4 pipes 8 textures
aka 2 tmus per pipe...


gah...my friend is supposed to get one...maybe he can get down to the nitty gritty

sent off an email to the sis rep for us...still dunno if I'll get anytin back

[Tagrineth]

VSA-100 based cards do the same thing...

My Voodoo5 5500 did ~300Mpps and ~630Mtps in T&LMark2001.

It's a driver thing IIRC.