Single Pass Stereo provides performance improvements by allowing all of the data for a frame to be sent to the graphics card once, after which it can use that same data to draw both the left and the right eye in sequence, without us having to send the data for the two eyes separately.
You can try out Single Pass Stereo on Pascal generation NVidia cards - this only requires one GPU and as an example of its benefit, it can improve the performance in the Visionary Render start-up environment by approx. 25%.
Parallel Rendering provides further improvement when using two cards, by allowing us to send all of the draw instructions to both cards at the same time, allowing them to draw the left and the right eye in parallel.
How could this benefit a scene?
To understand whether these would benefit a scene it helps to consider whether the scene is "cull-bound" or "draw-bound".
Cull-bound scenes are scenes with very complicated tree structures but not necessarily a lot of visible geometry - the bottleneck for these scenes is on the CPU, where we decide which parts of the scene are visible and need to be sent to the GPU.
Draw-bound scenes are scenes with less complicated tree structures but a lot of visible geometry - the bottleneck for these scenes is the GPU, where it has to perform significant work to draw the data we send to it.
You can see the time we spend in cull vs draw for any given view by using the Diagnostics window
Single Pass Stereo and Parallel Rendering have the most effect on draw-bound scenes. As you can see in my screenshot above, the thing I am looking at can be considered draw-bound, because the time spent drawing is significantly more than the time spent in the cull.
The state of these settings is enabled by default, if supported by the hardware.