It's my understanding that 3dmark05 will not be as GPU limited as 3dmark03 is. In that case if I'd want to waste enough money on a dual GPU setup, there would be theoretically little to nothing holding back from also going for a dual CPU config also.
Whether it makes any sense or not is an entire story of it's own.
I think the ultimate SLI realization will be to make supersampling perform well. 8xS and 16xS will probably be feasible.
8xS is already feasable in 1280*960 at least in many flight/racing/space sims where one really doesn't need framerates in excess.
However it's usually a no go for demanding FPS games and if then only in mediocre to lower resolutions. Now if a single sollution gets let's say 25fps with 8xS+ some AF in 16*12 and the dual setup gets 50fps, then honestly I don't see much point in it either. It still makes far more sense to use a MSAA+AF combination and if possible to crank up the resolution even more.
On the other hand I don't think 16x has anything more to offer than 8xS, other than the added 2*1 SSAA; I can't get FSAA Viewer to work with 16x to see what the sampling grid exactly looks like.
That said for a flight simulator fan as an example where cost is not an issue it will be the the best affordable sollution he can get.
The risk, of course, is that they've let slip that SLI probably isn't going to achieve the kinds of performance gains they were projecting earlier (i.e. 1.8x). At 7229, that means a $1000 6800 Ultra SLI setup (probably closer to $1500 if you count the motherboard and PS upgrades that will almost certainly be necessary) is less than 25% faster than a $500 X800 XTPE. And it pretty much kills the argument for using SLI on anything slower than a 6800 GT. Especially when you account for the fact that 3DMark05, being a graphics-focused benchmark, is probably going to be less CPU bound than a typical game (which has AI, collision detection, physics, sound, networking, etc. to worry about).
I recall myself when I first heard those 80% estimating that it should be considered as the highest threshold and NOT the average.
Average performance predictions are hard to make due to the complicated nature load balancing gets handled. Drivers expose three available modes so far:
1. Scene gets split up in strips where each GPU renders a group of strips.
2. Alternate frame rendering in combination with option (1) above.
3. Auto mode where the driver itself decides whether (1) or (2) is better.
Democoder had a point though; hybrid AA modes will most likely scale by a factor of 1.8 or sometimes even more. The question though still remains under which situations exactly.
I personally would love to have a sollution where I could use 8xS with 8xAF in 16*12 in games like Rallisport Challenge; yet the cost is way too high for me.