My experience, the only graphics use case right now that benefits from multple GPUs is non-realtime rendering. In realtime, you have to worry about added latency from SLI, and also the drop to 8x speed if you have a Z-series motherboard. Neither of these is an issue with non-realtime renderering.Â
For that, you're mainly looking at just two specs: sum of all your cards' cores = rendering speed, and max VRAM on the smallest card = scene complexity.