I've used SLI with 7800 GTXs, 7900 GTXs and now the 280s. It has certainly come a long way over the years but is by no means perfect. My biggest remaining issue is the frame jitter, which is what many call "micro-stutters". It is easily demonstrated in the following example.
This is a benchmark of TF2 run with vsync enabled where 18,000 frames were captured from an online demo of a few minutes in length. Frametimes were logged using fraps, converted into individual frame display times, and finally converted to hertz. This final value is essentially instantaneous framerate, or the constant framerate you would achieve if all frames were displayed in the time it took for a given sample.
Instead of plotting a bunch of useless squiggly lines, this data was then converted into a histogram. Only framerates between 50 and 70 were considered to remove loading issues or whatever, and it was then binned at 50 mHz to provide adequate resolution.
The resultant plots illustrate how many frames fell at a given instantaneous framerate. The ideal trace in this example would be an impulse function occuring exactly at 60Hz with no additional frames falling anywhere else. As can plainly be seen in the SLI case, the jitter that occurs when the frames are recombined causes "spreading" about the target vsynced rate. Significantly more than the single card case.
Keep in mind the total frames are the same, and if you integrate each curve you get roughly the same answer. The disabled case provides 1.022 M framecount*Hz, while the enabled is 988K framecount*Hz, or great efficiency at only 3% overhead.
The variance in displayed framerate, induced by the jitter caused by SLI is what people have termed "micro-stuttering". Granted not everyone can see it, just like some people can't tell the difference between 30Hz and 60Hz, but if you can it's rather distracting.