Here’s some results for the BMW scene test, although I still can’t test more than 1 core on my CPU without Blender crashing to desktop mid-render every time.
CPU: Intel i5-2500k (4 cores) 4.5Ghz
GPU: Nvidia GTX 1070
OS : Windows 10 64bit
960x540 Full: 1 min 4 sec (GPU - CUDA)
480x270 Quarter: 0 min 59 sec (GPU - CUDA)
256x256 Tile: 1 min 4 sec (GPU - CUDA)
128x128 Tile: 1 min 21 sec (GPU - CUDA)
64x64 Tile: 2 min 37 sec (GPU - CUDA)
960x540 Full: 20 min 26 sec (CPU)
480x270 Quarter: Blender Crashes (CPU)
256x256 Tile: Blender Crashes (CPU)
128x128 Tile: Blender Crashes (CPU)
64x64 Tile: Blender Crashes (CPU)
The crashing is probably related to me overclocking my CPU. I’m not sure. Either way, even if it worked perfectly fine, it looks as though CUDA is still the way to go by a considerable amount so I’m not going to bother troubleshooting CPU renders (especially if it would mean reducing the overclocking as no other program or process crashes, only Blender).
Interesting to see that is appears dividing the GPU render up into 4 equal parts shaved off a few more seconds than just rendering the entire section at once. And as the tiles got much smaller, the render time starting increasing significantly on GPU.