Run on arch linux with blender version 2.79
CPU: Ryzen 7 1800x
GPU: GTX 980
Time: 1 min 40 seconds (GPU - 128x128)
Time: 1 min 37 seconds (GPU - 512x512)
Time: 1 min 12 seconds (GPU - 960x540)
Time: 1 min 32 seconds (CPU - 64x64)
Time: 1 min 29 seconds (CPU - 32x32)
Time: 1 min 27 seconds (CPU - 16x16)
It is a close call, if you adapt tile size to the concrete output resolution then GTX 980 wins; otherwise CPU seems to be faster with reasonable tile sizes.