Running the tests with the various block sizes revealed that the optimal block size is 16 x 16, timings below. The timing for the kernel has gone down from 179 ms to 106 ms. Observation: Although block sizes of 16 x 16 and 32 x 32 both have 100% thread occupancy, the 16 x 16 turned out to be optimal. This is probably due to the fact that less warps are competing for access to global memory.