Does Chrome9 HC support 2-texel/clock multitexturing?

[overnite81]

Hi, I'm having a lot of work trying to understand this:

http://www.m3fe.com/fillratetestweb/ViewResult.php?id=492

Card Mem. bandwidth (MB/s)
1222
Single texture fillrate (MT/s)
248
Multi-texture fillrate (MT/s)
248

Other integrated graphics like the SiS300 series are known to perform single-texturing in a single clock cycle and multi-texturing in two clock cycles, but VIA markets its IGP as "the industry's only simultaneous usage of single-pass multitexturing and single-cycle trilinear filtering", which tips at a more efficient (at least 2 texels/clock and close to 500 MT/s fillrate) multitexturing engine than SiS older IGPs.

Where's the truth?

 

[Fiki959]

I have tested mine, and my results ate 249/250 Mtexels. It's intresting that the IGP is dual pixel pipeline, 250 MHz one texture per pipe that equals to 500 Mpixel, 500 Mtexels. My bandwith is around 600MB a second, because of the high latency setting on the memory. It a notebook so the are set little high.

The question should we ask are there 2 pipelines ?

 

[CarstenS]

This test yield about 2900/5800 MTexel on an OCed GTX280, so I'd take those results with a ton of salt.

Maybe the fillrate tests from 3DMark would be a more accurate indicator in this case.

 

[Fiki959]

OK, 3d mark 2001/2003 results are even worse: 149 MT for single texturing, 232 for multi texturing.

But is an IGP from Via, probably it need 2 clock cycles, or it has one pipeline, or the drivers aren't optimized. Either way the filtrate performance is low.

 

[overnite81]

Would be of usefulness to check this with a real-world 3d app. I recall that Serious Sam does have a ST/MT fillrate benchmark.

Prosavage is the classic Savage4 core with one pipe and 2 texels/pipe. Chrome9 is not a such rehash and no one knows for certain what VIA messed into, beside PS2.0 support.

Interstingly enough, embedded chipsets with little RAM and very low-end specs are enough for 3D visualization, as much as the CPU is capable of occlusion detection, mesh and vertex optimization along with S3TC compression

 

[MrMilli]

The Chrome9 HC is based on the Deltachrome.
If history serves me well, i remember that the 8 pipe Deltachrome didn't get anywhere near to it's theoretical fillrate as soon as one texture was applied. It wasn't until the Gammachrome that S3 fixed it's fillrate issues. The 4 pipe Gammachrome was way more efficient at that department.

I also found these references:
http://www.xbitlabs.com/articles/video/display/s3-deltachrome_7.html#sect0
http://www.xbitlabs.com/articles/video/display/s3-gammachrome-s18-2_5.html#sect0

So it seems that when one texture is applied, the 2 pipe Chrome9 HC will likely lose like 50% of it's fillrate. Add to that the probably even smaller caches then in the Deltachrome and very limited memory bandwidth (from what i remember, it's still connected through an AGP bus internally since Deltachrome is AGP based. Normally you can even find the AGP controller listed in your device manager), then the 250MT/s seems feasible.