I made a research to compare Cycles and Arnold in January and some people mentioned that GPU rendering with Cycles can be optimized and will be much faster with a different and bigger tile size.
• Research ARNOLD vs. CYCLES
But this has been changed for a while in the daily builds.
From my latest information the Blender dev team is working on a solution that tiling doesn't matter anymore and a multiple CPU/GPU support for tiles to close the bottleneck with the final tiles of a rendering.
I rendered my Still Life scene with different tiles with 2.79.5 and there's not a big difference anymore. Rendering the same scene with the official 2.79b is extremely slow with 16x16 tiles.
Path Tracing Samples: 256
05m 35s @ 16x16
05m 40s @ 32x32
05m 44s @ 64x64
05m 38s @ 128x128
05m 29s @ 240x216 (Auto Tile Size)
05m 34s @ 256x256
06m 11s @ 512x512
The video has been accelerated to 300%.