Short version of my SF/GPU performance notes:
1.
Main! The sweet spot of the Bucket size setting is HEAVY scene-dependent (GPU memory-dependent I suppose). Render FireFly script is your friend for the tests before the final render.
2.
Critical! (Drivers-dependent?) Sometimes your render just
crawls.
If the first number (after Sample, not after Path tracing sample) does not increase at least at 10 in a second, than probably something is wrong. Try decrease the Bucket size, even to 20px for GPU sometimes works and can give your better speed than CPU. (Or not. Test!) It can also helps with GPU/BPT combination.
Switch from OpenGL to SreeD for the time of the render.
Reload Poser also could help. (Some drivers glitches I suppose.)
Basic optimisations:
smaller HDRI, blurred HDRI, blurred HDRI on the object larger than the Yon of the camera (gives you transparent background) with different render with the crisp/larger HDRI as a background;
render size = Bucket Size * integer;
hide/delete or make diffuse all unnecesarry...