Tip: CPU Only, GPU Only, or Both

Unless your system has a problem distributing heat (it's very hot and it crashes), most people should use all rendering devices. Or at least test for the best performing configuration.

Here are some example numbers for a sample frame from this scene, rendered at 1920x1080 on a Core i7 8750h, and an NVidia GTX 1060. My CPU has 6 cores, which means it can render 12 threads at once. The GPU can render one thread at a time. Combined means 13 tiles simultaneously. Which is the best configuration of hardware and tile size?

GPU Only: 

  • 26.56 (16x16) 
  • 21.43( 30x40*)
  • 21.22 (32x32)
  • 19.30 (64x64)
  • 18.38 (128x128)
  • 17.90 (256x256)
  • 18.31 (300x300*)
  • 17.50 (512x512)

* not powers of 2

35% difference fastest to slowest

CPU Only: 

  • 49.49 (16x16)
  • 46.79 (30x40*) 
  • 45.77 (32x32)
  • 46.27 (64x64) 
  • 47.37 (128x128)
  • 50.69 (256x256)
  • 49.60 (300x300*)

10% difference fastest to slowest

Overall, twice as slow

Both:

  • 19.54 (16x16)
  • 16.50  (30x40*) 
  • 16.60 (32x32)
  • 15.99 (64x64) 
  • 15.70 (96x96*)
  • 16.11 (128x128)
  • 20.87 (256x256)
  • 26.22 (300x300*)

40% difference fastest to slowest

10% faster than the fastest GPU time

I chose 96x96 only because when using both devices, the time is highly dependent upon whether the CPU and GPU can be effectively used and not get in each other's way. If the tile is too big or too small, the GPU will be underutilized. Anywhere from 64x64-128x128 is an optimal balance for my setup that could save 10% over GPU only, which will be more important for complex scenes later on.

Of course, you can and should test this on your own system and see which mix of devices and tile sizes work fastest for your system.