Cycles render speed: how to optimise GPU rendering?

Question

I am making my way (sometimes painfully) through an online "intro to Blender" course and have got to the Lighting section, in which the end-of-class exercise is to light a scene using Cycles. I am running Blender (Steam version) under El Cap, GTX970, latest WebDriver and CUDA driver from NVIDIA. Based on Heaven and other benchmarks (3d sims), this a fairly capable graphics platform.

However, I'm finding Cycles incredibly slow to render just one frame of a (not too complex) scene. So slow, in fact, that I don't think I can afford the time to do the exercise! Blender Render will render the scene in about 25 seconds, whereas Cycles was getting only a small fraction rendered in 25 seconds, less than .10 of the total image area. I did enable CUDA and I did select GPU for the Cycles render hardware, and nothing seemed to help.

I have read a lot of conflicting info online, much of it obsolete (old posts), and am confused. Was wondering if anyone here has a pointer to a really definitive reference on speeding up Cycles on NVIDIA hardware. I did find and install "blenchmark" but when I tried to run it w/o GPU, to compare my CPU vs GPU render times, it crashed :-(

In the meantime I "de-noded" the meshes in the scene and tried to light it with BI, but the results were not nearly as impressive as other people achieved with Cycles. I would like to try again with Cycles if I can find out how to speed it up a bit! If I reduce "samples" from 2000 to about 100 I can get Cycles to render the scene in 30 sec or so, only a little slower than BI, but the quality is awful (grainy); I won't get an acceptable image to submit for grading that way!

Answer 1

Samples:

2000 is far to many samples for a "(not to complex)" render. I would use 100 for previews, and 200-400 for final renders (depending on the complexity of your scene). Images involving lots of glass, complex shadows, indoor scenes, or depth of field may need more samples.

Tile Size:

Do some tests with your specific hardware, but, generally smaller tile size is faster for CPU, bigger tile size is faster for GPU.

An alternative to increasing samples for less noisy renders is to try some of the techniques mentioned here: How to reduce fireflies in Cycles?