The SV_ViewportArrayIndex semantic allows you to select which viewport from this array to use on the fly, instead of constantly switching viewports in the API, note that the viewport bounds you choose must eventually find their way onto the graphics card for the rasterizer to use, so passing a new viewport is just like any other state change. Apparently it's a semantic applied to the geometry shader output, but I would imagine you can apply it to the vertex shader output if you have no geometry shader (I may be wrong on this).

I haven't tried it myself, nor am I sure how it would work, but perhaps it could be used for rendering stereo images - i.e., split-screen with slightly different views to simulate distance between the eyes. I.e., one set of objects into the geometry shader, two sets (one to each viewport) output.

The main example I've seen it used for is things like cubemap rendering, where the GS can send each triangle to n/6 of the viewports.

Might also be useful for doing stereo/VR in a single pass, using a GS to duplicate all the triangles into the 2nd viewport.

I think it's also used for 3d/volumetric rendering. Render targets have to be two-dimensional, meaning you can't render directly to a 3d volume... but you can represent a 3d volume as a stack of 2d viewports. The GS can then send triangles to all of the slices/viewports that they cover.

I guess you could also do stuff like try to render shadowmaps for a large number of light sources to an atlas in one pass. The geo shader could decide which triangles are visible to which lights, and output them to the right views within the shadow atlas.

Apparently it's a semantic applied to the geometry shader output, but I would imagine you can apply it to the vertex shader output if you have no geometry shader (I may be wrong on this).