Millisecs() is perfectly fine if you measure over a long enough time.
No it isn't - not at > 2500 fps. It's like trying to measure quantum leaps using a track stopwatch. For each discreet unit of time you have 2.5 occurrences of whatever it is you're trying to measure. That gives you a nice 150% margin of error - which grows linearly proportional to the time it's been running to the point where the numbers become meaningless. It's not entirely by chance that the FutureMark guys design benchmarks that run at low framerates. There must be a reason for the difference.
Poor Benchmarking.