OpenCL 'non-blocking' reads have higher cost than expected

Question

Consider the following code, which enqueues between 1 and 100000 \'non-blocking\' random access buffer reads and measures the time:

#define __CL_ENABLE_EXCEPT