SLI issues

[mito]

3dmark2006

with SLI score = 12996
w/o SLI score = 11846

Is 3dmark2006 use SLI?

Now games running at 1920x1200 and everything maxed out:


Stalker: on some areas I get 70fps, other areas the system crawls at 15fps.

Call of Juarez: the first level, when you steal the gun from the old man: 100fps.
But after I reach the town, frames drop to 20fps...

FEAR: Extraction Point (1600x1200).
it flies: 120fps...


SLI is enabled in nVidia control panel. Is SLI still that immature?

 

[ChrisRay]

SLI primarily improves performance in areas where the graphics throughput is your primary bottleneck. I'm not overly familiar with stalker. But 3dmark06 believe it or not is primarily a CPU limited bottleneck. ((For various reasons such as vertex setup)) Your scores tend to reflect that.

SLI will always perform most optimally when pixel performance is the primary bottleneck of the engine. ((Which is why higher resolutions are preferred)) vertex limited titles ((Even under AFR)) dont tend to scale as well in my experience. Possibly due to the way Nvidia handles vertex setup data. No idea how the geometry shader is going to play into SLI scaling. I have seen very little data on how it scales so far.

Chris

 

[OpenGL guy]

Try running at higher resolutions if you want to check 3DMark06 scaling.

 

[Rainbow Man]

Why is 3Dmark 06 so heavy on the CPU since it's a graphics benchmark? I can't see how that makes any sense.

If the argument is that games stress the CPU also, well, 3Dmk06 obviously ISN'T a game. So why bother?
Peace.

 

[Silent_Buddha]

3dmark 06 wasn't created as a purely graphics benchmark. It's supposed to be a benchmark that tries to see how well your computer will be running a game.

As such it tries to stress graphics AND CPU. Except that in the case of 3dmark 06 it's a little more CPU bound than it is GPU bound. With some of the older benches you have the exact opposite case where they are generally more GPU bound than CPU bound.

It's still a good tool to use just as long as you realize that in todays 3D gaming world. There is no single benchmark (gaming or synthetic) that can possibly tell you how your rig will run in the majority of games.

Regards,
SB

PS - sorry for the offtopic.

 

[mito]

I learned something yesterday regarding SLI and poor perfomance.

I crank things up (1920x1200 details = MAX) and do not OC.

The following storyline reflects what happened to me since the beginning:

1. Installed Win XP and forceware 158 (comes with evga cd);

2. avg fps: stalker=75 / juarez=50 / scda=60 / fear=150;

3. Installed forceware 162.18, nightmare begins;

4. avg fps: stalker=10 / juarez=19 / scda=60 / fear=150

5. Downgrade to forceware 158;

6. same fps

7. go to nVidia Control Panel -> 3d settings -> Manage 3d settings -> Program Settings

8. set SLI Perfomance Mode to Force Alternate Frame Rendering 2

9. voilá!!!

 

[ChrisRay]

3dmark 06 wasn't created as a purely graphics benchmark. It's supposed to be a benchmark that tries to see how well your computer will be running a game.

As such it tries to stress graphics AND CPU. Except that in the case of 3dmark 06 it's a little more CPU bound than it is GPU bound. With some of the older benches you have the exact opposite case where they are generally more GPU bound than CPU bound.

It's still a good tool to use just as long as you realize that in todays 3D gaming world. There is no single benchmark (gaming or synthetic) that can possibly tell you how your rig will run in the majority of games.

Regards,
SB

PS - sorry for the offtopic.

Eh it's not offtopic from my PoV. And its personally onkey with my issues with the benchmark. 3dmark06 at the high end is a toy benchmark toy for CPU performance and CPU overclock junkies. It's been this way since 3dmark05. It's almost the exact oposite of 3dmark03 which is still almost completely GPU bound. It's funny testing 3dmark03 on modern SLI setups and seeing the insane scores its pumps out by just upgrading the GPU or adding SLI with rather rundamentory pentium 4 or low end athlon 64 SLI systems.

I really look forward to the next futuremark iteration where the bottleneck gets shifted back to the GPU some. That way I dont have to explain it to every new member on the SLIzone forums who didnt get double the 3dmark score they had hoped for.

Chris