Interesting find: R9700 AA benches without ColorCompression

[Indiana]

Some guy at 3DCenter.de found a way to disable the Radeon9700s ColorCompression.
Here are my results with 3DMark2001SE in 1024x768x32, 4x FSAA, 8x "quality" anisotropic:


The yellow bars show the results with Color Compression enabled and the red ones without.
Translating this into percentage gains due to Color compression:
Total Score: +20%
Car Chase low: +18%
Car Chase high: +8%
Dragothic low: +32%
Dragothic high: +21%
Lobby low: +22%
Lobby high: +11%
Nature: +29%
Fill rate (ST): +205% (!!!)
Fill rate (MT): +25%
High poly (1 light): +8%
High poly (8 lights): +2%
EMBM: +7%
Dot3 BM: +23%
Vertex Shader: -1%
Pixel Shader: +9%
Advanced PS: +14%
Point Sprite: -14% (???)

Quite a healthy gain in spite of the high bandwidth the R9700 has. This is with only 4x FSAA, apparently it's even more with 6x.

Makes you wonder what performance the Parhelia would've been able to show if there had been some bandwidth-saving techniques included....

[Greebe]

Moved to Benchmarks forum

[Pace]

What is ColorCompression actually used to compress?

And doesn't the Parhelia have texture compression and maybe some other compressions? (vertices?) (ignoring DM/DAT).

P.

[Indiana]

The Color Compression is used to compress colour data. It was not talked much about at the release of the R9700Pro, but came more into light when NVidia hyped this thing with their GF FX and people suddenly found that the R300 had this as well.

And no, unfortunately the Parhelia has near to no bandwidth-saving techniques, which is a MAJOR fault, imo and might be the culprit for it's lower than expected performance.

I did another test, this time with 6x FSAA and now the hit due to the missing color compression becomes dramatic:



Total Score: +36%
Car Chase low: +40%
Car Chase high: +4%
Dragothic low: +58%
Dragothic high: +40%
Lobby low: +50%
Lobby high: +20%
Nature: +48%
Fill rate (ST): +303%
Fill rate (MT): +36%
High poly (1 light): +9%
High poly (8 lights): +4%
EMBM: +32%
Dot3 BM: +61%
Vertex Shader: +1%
Pixel Shader: +50%
Advanced PS: +48%
Point Sprite: -11%

That's a performance plus of 40-60% in many cases on the exact same gfx-card just en- or disabling this feature!! So now imagine the Parhelia had this plus some early Z-culling...

[borat]

bandwidth is not really an issue with parhelia as you can take a bulk card and o/c the memory from 500 to over 600 mhz ddr and it makes little difference to the performance, overclock the core without o/c the memory and you get a huge difference showing its the core not the memory which holds things back. needs another 100mhz on it.

[mmp121]

As borat said, BW is not a huge factor...

http://www.tech-report.com/reviews/2002q4/radeon-9700s/index.x?pg=1

Scroll down and take a look at the table 2/3 of the way down.

[Indiana]

mmp121, in your link there are only R9700 oc'ing results?!?
And if you mean the table with the plain raw bandwidth comparision, the R9700Pro has higher bandwidth than the Parhelia and still suffers A LOT in high FSAA modes if bandwidth saving techniques are turned off.

Borat, coukd you provide a link? I'm sure that those oc results were without AA, which is the whole point about this.
Of course the Parhelia has enough bandwidth when operated without AA and anisotropic filtering, but if both AA and aniso are turned on, I do think that it's a bit struggling because of too low badnwidth in spite of its 256Bit bus.

So what we need is a Parhelia user doing a test (take UT2003, e.g., or 3DMark2001SE) with 16x FAA and 2x aniso and then first oc'ing only the memory then only the GPU.

[GNEP]

FAA is not a real bandwidth hog compared to FSAA.

[Indiana]

Yes, I know - but it still raises bandwidth requirements and together with anisotropic filtering it might get over the limits of even a 256 bit wide bus, at least in higher resolutions.

So any Parhelia user with an overclocking friendly card can provide 16x FAA + aniso benches default and the with only the Ram and only the GPU oc'ed?

[Indiana]

Originally posted by Indiana

So any Parhelia user with an overclocking friendly card can provide 16x FAA + aniso benches default and the with only the Ram and only the GPU oc'ed?

Never mind, there actually is already such a bench done by noone less than Ant himself. He gives results with the GPU operating at 240 vs 220 MHz (9% oc) and the Ram operating at 305 vs. 275 MHz (11% oc).
And his results seem to hint that the Parhelia DOES get bandwidth limited in 16x FAA.
While there is in fact nearly no increase with higher RAM but with higher GPU clocks when FAA is turned off, the situation changes when FAA is activated:

Taking Ants values with some of the apparently more bandwidth hungry tests (the ones with the most gain due to Color Compression):

Car Chase low + Dragothic low&high:
gain from GPU oc: +4%
gain from Ram oc: +6-7%

Nature:
gain from GPU oc: +3%
gain from Ram oc: +7%

Dot3 BM:
gain from GPU oc: +1%
gain from Ram oc: +9%

Advanced PS:
gain from GPU oc: +2%
gain from Ram oc: +8%

So with the Parhelia the gain through oc'ing of the memory is higher than when oc'ing the GPU as long as FAA is activated - especially in the more complex scenes like Nature, advanced PixelShader and the astonishingly bandwidth dependent Dot3 BM.
A 7-9% increase due to a 11% increase only in Ram speed with the exact same GPU clock is indeed quite good scaling - which means the Parhelia could do well with even more bandwidth.

[knirfie]

how does the 9700 do in the parhelia shark bench?

[Guru]

Originally posted by knirfie
how does the 9700 do in the parhelia shark bench?

First y9ou need to find someone who has the sharkbench!

[Novdid]

Speaking of sharks, did anyone try to ad something like " i promise to buy a parhelia" to the reefdemo config and got it to run on a 9500/9700!?! (just kidding)

[3dfx]

lol ..good one

[Indiana]

The purpose of this thread was more to show that Matrox maybe crippled the Parhelias performance A LOT more than most people think by not including bandwidth-saving techniques in the chip.
I first thought that with the 256Bit bus those would not be that effective, but when seeing this test I was really stunned about the perfomance-hit the Radeon9700 takes - despite it's 256bit wide bus.
The performance of the R9700 is much less impressive with just ColorCompression disabled. Deactivate the rest of HyperZ and you'll loose even more performance.

Knirfie, if you can provide a dl link for Sharkmark and the archive is not too big, I can do a bench.