I was thinking about something the other day, and after reading some of the issues getting a proper boost using SLI or Xfire setups for gaming, I was wondering why it wouldn't be possible for devs to use ONE of the GPUs for actually scene rendering, while using the "slave" card for other tasks, like Physics, OpenCL computation, AI, etc..
Any good technical reason why a devs couldn't use those types of setup to bring more things to a game environment (better RBD and/or soft body physics, smarter AI, pathfinding, etc...)