6600GT/6800GT SLI Test

[Evildeus]

Interesting figures:
http://www.anandtech.com/video/showdoc.aspx?i=2258&p=1

 

[Dave Baumann]

Greater than 100% results at high res / AA are coming from the fact that each board is rendering half the screen meaning the frame buffer requirements are lower hence fewer textures are being addressed from system RAM. However, this doesn't help out the 6600 GT SLI as much as its only got 128MB.

Seems to suggest there is no AFR mode in the drivers at the moment either.

 

[Mendel]

and meanwhile xgi has hard time taking any advantage at all out of its two chips on one card compared to one chip...

Bravo! Nvidia engineers have done a great job here, don't you agree? Scrap that, >100% improvement is better than great, it's just unbelievable, achievement unlike anything, better than voodoo 2 SLI!

Now only they need to make it available and dual pci-e 16x slots needs to get more common and maybe, just maybe 6800 gt sli will be a worthy option to consider!

edit: I'm still not sure it's even theoretically possible... how fast was that link between the cards again?

 

[Dave Baumann]

Scrap my earlier reply - frame size will be altering per frame, so reallocating textures between the system and local RAM on a per frame is not likely to occur. The frame buffer memory requirements is still likely to be set at initialisation which means either the space for a full frame buffer is set per board (in which case, its unlikely that you'll be seeing greater than 100% gains) or there are limits set on the split before rendering occurs.

 

[AlNom]

so... >105% increase is fud

 

[Evildeus]

The random error i suppose

 

[Dave Baumann]

It might be the case that there is ma maximum split allocated - i.e. set the allocated space for the frame buffer at, say 70% of the overall resolution. In this case the maximum split differences could only ever be 70/30 or 30/70 but you can guarantee to save some frame buffer space hence have more textures in local RAM. It could be that limits are set for high res / high AA because the savings gained from not addressing across the bus would be greater than the gains from a fully variable split (those limited could also be removed for non-memory space intensive rendering resolutions).

 

[Rys]

Scrap my earlier reply - frame size will be altering per frame, so reallocating textures between the system and local RAM on a per frame is not likely to occur. The frame buffer memory requirements is still likely to be set at initialisation which means either the space for a full frame buffer is set per board (in which case, its unlikely that you'll be seeing greater than 100% gains) or there are limits set on the split before rendering occurs.

There's no limit on the split before rendering occurs. It can (and will) shift from the full spectrum (100-0 to 0-100) of load types at any point in the rendering process, with load-balancing enabled.

And again, there's no AFR mode in any SLI supporting driver up to just and I've just received another build.

Rys

 

[Dave Baumann]

What about at 1600x1200 with 4x AA?

 

[Rys]

I can't be certain there's no hard limits at 1600x1200 with 4x AA, since the loads hardly shift at that setting, and each card is worked as hard as possible. I only see 'wild' load fluctuation at relaxed settings. So I can't be 100% sure for the 16x12 4AA case, but I'm not sure that's something implemented, at least with the less buggy of the driver sets I have.

You should pop up and see for yourself

Rys

 

[KimB]

There will be just as much load fluctuation at high resolution as there is at low resolution.

 

[Rys]

There will be just as much load fluctuation at high resolution as there is at low resolution.

If that's the case, there's locked load limits at that setting

Rys

 

[Pete]

Any reason why the 500MHz, 8 pipe x 2 6600GTs aren't anywhere close to handing the 350MHz, 16 pipe 6800GT its buttocks? "Drivers," or RAM limitations in the high res + AA situations where I assume there's the greatest chance of approaching theoretical limits?

 

[KimB]

If that's the case, there's locked load limits at that setting

Rys

No, it's just that the system as a whole is more CPU-limited (or geometry-limited, since AFR is apparently not being used) at lower resolutions.

 

[KimB]

Any reason why the 500MHz, 8 pipe x 2 6600GTs aren't anywhere close to handing the 350MHz, 16 pipe 6800GT its buttocks? "Drivers," or RAM limitations in the high res + AA situations where I assume there's the greatest chance of approaching theoretical limits?

Between the two 6600 GT's, they still have half the ROPs of a single 6800 GT. That's probably a fair part of it, particularly since the tests are being run at 4x AA. The memory size limits are also a concern, and are surely the reason why the 6600 GT falls further behind at higher resolutions.

 

[Pete]

Dammit, I keep forgetting that NV43 has half the pipes and a quarter the ROPs. I'm still equating NV40's eight alpha blends per clock to meaning it has eight ROPs.

Heck, my memory may be failing me on the alpha blends, too.

 

[Graphics_Krazy]

What about automatically lowering texture filtering or levels when switching to SLI mode - technically you could get more than the regular boost.

 

[trinibwoy]

What about automatically lowering texture filtering or levels when switching to SLI mode - technically you could get more than the regular boost.

Why would they do that? Talk about bad PR.