IDK, let's ask Nvidia: http://developer.download.nvidia.com/assets/events/GDC15/GEFORCE/SLI_GDC15.pdf
In the best case:
- Non-uniform flip intervals (microstutter)
- *Interframe dependencies
- Input latency does not reduce with increased performance
First one is well known and mitigated by driver applying back pressure.
Second one hurts performance scaling with generic temporal algorithms, like SMAA T2x/4x. Some game engines really go wild with this and AFR doesn't work at all.
Third one means that 2x 980 has same FPS as a single 1080, but has almost twice the input latency.
These problems can be avoided with split frame rendering... but at that point scaling alone would be bad enough that it's better to buy a single large card.