Funky, you bring up a good point above. I didn't reboot my phone prior to running the tests. To be a bit more scientific about it, I should have rebooted each time prior to running a test under the same conditions. Rebooting is a good best practice, though, and helps to clear any non-persistent stuff in cache and thereby eliminate "warm cache" side effects, as well as knock down any background processes.
Your second test comes in somewhat above my Power Save Enable run, but approximately 22% lower than the Power Save Disable run. Given both our S3's had the full #Core running for the Power Save Disable, then the delta could fall into a few areas like: JellyBean is more resource intensive, Vellamo may be more optimized for ICS and thus run faster and yield better results, or that some of the workloads in the Metal suite perform better when more system memory is present, or the memory subsystems between our phone versions vary in bus speed (certainly do in capacity). Its likely a combination of all four areas noted (and probably more). Take Stream for example. Stream is all about measuring and understanding memory bandwidth (summarized from
STREAM Benchmark Reference Information). Stream puts sufficient load on the system so that as work moves beyond the CPU local cache and into main memory, systems with less main memory may struggle in such cases.
So, the above may account for some of the differences. I'm also not clear at this point if the Metal Stream (is that a rock band by chance?) is executing as a single or multithreaded process. In either case it may be that for the memory intensive parts of Vellamo (or any micro benchmark or synthetic workload mixture) the International S3 may not have enough system memory to support all four cores. This is probably only a worry in cases with more extreme utilization patterns and multi-threaded memory intensive workloads.