Agreed. This is pretty much a worthless benchmark. There's nothing complex taking place, no hard memory thrashing, no difficult calculations.
It's just taking 2 arrays with 5000 numbers in them and adding them together into a new array of 5000 numbers. Simple atomic add operations of 2 numbers over and over and over. The more cores you have the more you can split the array up (4 cores = each core processes 1250 items, 32 cores = each core processes 156.25 items.) and the faster clock means that each item gets processed faster.
This is NOT AT ALL indicative of a real OpenCL app that will be doing hundreds of thousands of difficult computations, with dependancies between the dat and working across huge datasets.
I'd say ignore every result we get out of this app. It's not at all indicative of real-world performance in the least.
I found the phoronix.com benchmark (and/or reviews) a lot more interesting than this thread