3dmark03 On Nvidia Cards Redux

[GlowStick]

I have an intresting observation.

After playing aroudn with diffrent AA and AF settings and taking screenshots useing 3dmark03 to compare results (for me to see what they do) i found something distrubing.

Though the thing at the bottom shows the same time, and the same frame, 800. The leaves are in totaly diffrent positions for some of my screens.

It seems to be all the leaves connected to the Left side Closeist tree, it is very noticeble if you put it in slideshow mode (a windows XP thing)

Im to lazy and dont have any hosting left to post them someware. they are about 4mb compressed with RAR.

Anyhoo, if you dont belive me email me at glowstick@insightbb.com and ill send them to you.

[jbirney]

I believe the leaves postions are determined by a pusedo random seed. Thus they are not suppose to be in the same spot all this time. Worm over at B3D (he works for Futuremark) hinted this.

[GlowStick]

Originally posted by jbirney
I believe the leaves postions are determined by a pusedo random seed. Thus they are not suppose to be in the same spot all this time. Worm over at B3D (he works for Futuremark) hinted this.

It could be ture, but that would make it a crappy benchmark

but the leaves on the other trees are the same? I think it could be some kind of bug.

[marcocom]

what about the issue that nvidia raised regarding a 3dfx-esque post-buffer that makes screenshots not exactly represent the realtime-rendered image (i agree, its totally shady. i just wanted to know if its being considered in this comparison)

what about the new cinematic rendering methods (like focus-blurring and other T-buffer 3dfx technologies supposedly infused into this new chip) could that be the reason theres a blur on backgrounds?


Im all for giving nvidia a bit more time to get a completely new GPU and platform out the door and thats going to require a little time (we gave it to ATi, their first R3xx drivers werent much better)- but i agree that creating workarounds for each application will create a huge driver and make for shaky backwards compatibility.

its something to consider but could we be missing whats happening here?? nvidia pulled out of the 'madonion 3dmark' program saying that paying to be in alliance with them seemed shady (while ATi stayed in the program as an alliance and paid futuremark for it and worked with Futuremark hand-on while nvidia didnt - this was over a year ago.)...and boy did they pay in the end. is it possible that Futuremark understands how to optimize against nvidia on a level we can never dissect or know about? thus nvidia balks and tries to tell us what futuremark with ATi is conspiring towards and hopes we see through it? because you cant just accuse them outright. its an alliance. its common practice.

cant wait for Aquamark. we need another party in this to keep this honest and on the up and up.

[GlowStick]

AquaMark is avalible to some reviewsites, its done done though, but the GeforceFX rips up the 9700pro in it. The AquaMark is a byproduct of a REAL game though.

[AnteP]

Originally posted by GlowStick
AquaMark is avalible to some reviewsites, its done done though, but the GeforceFX rips up the 9700pro in it. The AquaMark is a byproduct of a REAL game though.

Are you referring to Auquamark 3 or 2.03?
Because in 2.03 there's certainly no "ripping" going on except in 1600x1200 and higher.

If it is version 3 you are referring to could you provide a link where it's used?
(BTW shouldn't it be available by now, I'm very much looking forward to using it for my reviews!)

[GlowStick]

Originally posted by AnteP
Are you referring to Auquamark 3 or 2.03?
Because in 2.03 there's certainly no "ripping" going on except in 1600x1200 and higher.

If it is version 3 you are referring to could you provide a link where it's used?
(BTW shouldn't it be available by now, I'm very much looking forward to using it for my reviews!)

Couldnt tell you what version they were useing, i was looking at some review of the gffx i clicked on that was linked from the front page.

If you go to the devloper of AquaMark's webpage they say 'E-Mail us for the most current version of AquaMark, we will review yo ur credentals' or something like that, but their is some 'zips' of it floating around on the internet.

[marcocom]

i may eat these words (most likely gleefully fed by cowboy), but im wagering that six months from now, the 9800 vs nv30 may have some closer-than-you-think scores in all the benchies.

this is a new kind of GPU. programmable pipelines and shaders means that this chip is much more sensitive to driver refinement than any in history. AA and AF are math algorithms thus driver tweakable forever and ever. (stop talking like all of this is built in stone just because it has a name)

gosh... i wonder if nvidia can improve their drivers??

/me breaths inward in horror and suspense

[AnteP]

Originally posted by marcocom

this is a new kind of GPU. programmable pipelines and shaders means that this chip is much more sensitive to driver refinement than any in history. AA and AF are math algorithms thus driver tweakable forever and ever. (stop talking like all of this is built in stone just because it has a name)

NVidia can tweak their drivers all day long and they'd still only have up to 4x MSAA, only 2x Rotated Grid MSAA and only 15,6 GB/s memory bandwidth on the other hand..

[AnteP]

Originally posted by GlowStick
Couldnt tell you what version they were useing, i was looking at some review of the gffx i clicked on that was linked from the front page.

If you go to the devloper of AquaMark's webpage they say 'E-Mail us for the most current version of AquaMark, we will review yo ur credentals' or something like that, but their is some 'zips' of it floating around on the internet.

[marcocom]

raw memory bandwidth.

but memory holds textures, textures are a collection of colors, and colors are compressable by algorithms and (according to nvidia) they say they can really get that 1ghz DDR2 working for them with up to 32GB theoretically (its all theoretical) but thats definitely saying 'drivers can make this puppy sing'

maybe worth the wait.

i gotta tell you my last card, nv20 ( the Geforce3, bought it for 400$ the day it released. amidst alot of critisizm from Geforce2Ultra owners that schooled it in the benchmarks) got me two years of really nice gameplay. i get almost double my first scores in 3dmark2001 almost two years later! such an example of scale-ability at its best. it played BF1942 at 1024x768 with 4xAA (always no AF of course) at very smooth framerates.

so we got an interesting year ahead of us.

sweet.

[AnteP]

Originally posted by marcocom
raw memory bandwidth.

but memory holds textures, textures are a collection of colors, and colors are compressable by algorithms and (according to nvidia) they say they can really get that 1ghz DDR2 working for them with up to 32GB theoretically (its all theoretical) but thats definitely saying 'drivers can make this puppy sing

the 32 GB/s refers to compression which is hardware limited
with compression ATi could also reach silly bandwidth almost reaching 100 GB/s but there's no point in touting such figures as there's no such thing as constant compression here
(While on the topic, actually ATi offers higher compression than nVidia, at least when using 6x AA)

edit: of course I know that nVidia probably have substantial performance to gain by further driver development
I just think that many of the results we are seing are due to pretty much hitting the hardware limitation wall