I'm going to have a lot of static cubes with the default orientation in a scene. Would I get a performance boost if I split each cube into 6 face squares, with faces being destroyed and created based on the camera's location so that a maximum of 3 faces exists at any time? Or are faces that aren't visible to the camera automatically ignored by the engine?
It will be more expensive to split it up because you would be adding 6 times the object overhead in order to save on 3 faces inside a mesh. This would affect view frustum culling, collision detection, have much higher memory overhead etc. The back faces aren't rendered anyway, so just leave the cubes alone.
For a performance boost with rendering static cubes, see this thread, where I reduce draw calls from the thousands to a handful.
