The human factor is the best benchmark there is.
1) Benchmarks just time-demo, they don't measure the game play under load with keyboard, mouse, and inability to predict the right data to fetch more often than is realistic. (When time-demo's are run the game engine executes more streamlined code paths, and todays processors can pre-fetch what will be needed in advance and not get surprise pipeline stalls anywhere near as often (for example)).
2) If the framerate goes under 44fps to 48fps that is bad.
3) If you get micro-stuttering over 125fps (Vsync disabled, Gsync, or the like) then that is bad. Even if performance is 'high'.
4) You can still get irregular frame render times (fast/slow alternating frame render times, averaging high) at over 64fps for all sorts of reasons. (e.g. more than say 25% variance in every 2nd frame when it's running well... or appears to... until you start gaming and using the mouse at what is reported at > 64fps in game).
5) Other factors I haven't covered.
All that said, Futuremark / 3DMark products tends to predict 'future' game performance on previous, current and next gen GPU's rather well.
SiSoft SANDRA also has some good L2/L3 cache latency metrics (and lots of other stuff), which are oft overlooked in gaming benchmarks.
In all honesty though, any system with a true Windows Experience Index of 7.0 or greater for CPU, and 7.5 or better for RAM -with a modern video card with a 'max' score for the two graphics metrics- is probably a good system.
3D Mark is good as you can get a lot of variance between otherwise identical systems. It can really highly the small improvements that add up. (1.03 ^ 32 is a lot more than just +96% better, if you get my gist. So long as the improvements don't have overlapping performance gains).
SiSoft SANDRA is good for nutting out the reasons why.
Ever since the Z77 chip-set, today's systems are getting very close to each other in performance though; it is nowhere near as exciting in the hardware and benchmarking world as it used to be!