Crossfire scaling (got more then 100% scaling)

[MarcVenice]

I just published a review, and in that review this happened:

HD 5870: 17,6min / 22,3avg
HD 5870 X2: 31,5min / 41,8avg
HD 5870 X3: 59,3min / 69,0avg

This was at 2560, without AA, but 16xAF and pretty much all settings maxed.

It's rather weird that the X3-solution has minima 4 times higher then a single HD 5870. Could this have something to do with the way CF works. Because when AFR is used, this would be really hard to explain. When Scissor-mode or Supertiling is used, it could make sense, right?

Because, let's say Supertiling is used, each card basically renders 1/3th of the frame. Since only 1/3th of a frame passed through the gpu, there shouldn't be any bottlenecks, that might be there when a full frame has to be done by a single gpu. Does this sound plausible?

Btw, I've yet to find out what kind of rendering mode CF uses in Clear Sky.

Also, could this explain why some games scale rather well, like FC2 and Clear Sky, yet Crysis Warhead plain sucks?

 

[thilanliyan]

It would be cool if they started using more rendering modes...why did they stick to AFR anyway?

 

[v8envy]

These kinds of anomalies make me think the 58xx has severe driver issues, intentional or otherwise. Been analyzed and plenty of people are surprised roughly 200% the hardware of a 4890 is only yielding a 30% real world improvement.

We've seen CF seeming to scale better than a single GPU with comprable hardware before (read: 5770 CF vs 5870) and it's very puzzling indeed.

 

[lopri]

Was it tested on X58?

 

[deimos3428]

Originally posted by: MarcVenice
each card basically renders 1/3th of the frame. Since only 1/3th of a frame passed through the gpu, there shouldn't be any bottlenecks, that might be there when a full frame has to be done by a single gpu. Does this sound plausible?

Pure conjecture on my part:

The minimum frame rate is going up because you're chopping the frames into successively smaller pieces as you add GPUs. This reduces (exponentially) the probability that any single GPU gets a long sequence of "tough" pieces. These pieces get distributed progressively more evenly as you add GPUs, so that's why you're seeing a non-linear improvement.

 

[MarcVenice]

Originally posted by: lopri
Was it tested on X58?

Yes, tested with X58, core i7 920 @ 3,8GHz. EVGA X58 Classified which provides 3 x 8 pci-e 2.0 lanes.

 

[T2k]

Originally posted by: MarcVenice

Also, could this explain why some games scale rather well, like FC2 and Clear Sky, yet Crysis Warhead plain sucks?

Well, that's easy - Crytek engines, in general, suck big time. They are getting better but still nowhere close in terms of optimization to any half-decent game engine out there...

 

[Fox5]

Did you have vsync enabled? Vsync can cause super-linear scaling since the gpu refresh is synchronized to the refresh rate of the display.

 

[MarcVenice]

No, vsync disabled.

Also, I checked with AMD, ALL games use AFR, supertiling and scissor-mode are a thing of the past. I'm still baffled though. Maybe it was a fluke ... Maybe I'll retest it on Win 7

 

[edplayer]

Originally posted by: MarcVenice
I just published a review, and in that review this happened:

Where can we see it at?