When you run a benchmark with Fraps in the background, you end up with a long list of rendering times. Putting these in a graph gives you a good visual overview of how the game runs on that graphics card. The frame that takes the most time is the indicator of how smooth a game will run.
The render times of the first 1,500 frames in our Battlefield 3 benchmarks with an Nvidia GeForce GTX 680
Again there can be other reasons, however, that a single frame takes much longer to process, reasons that have nothing to do with the graphics card or the video game. That's the reason why we look at the 99th percentile, the time it takes in which 99 percent of the frames can be processed. That means we disregard the slowest 1 percent of the frames. This should avoid incidental peaks, but should still include structural, re-occurring peaks.
If the 99th percentile is lower than 30 ms then the game runs smooth. If it's significantly higher, you will notice stuttering in practice.
Credit where credit is due: Tech Report started this discussion.
We have to give credit to Tech Report, they are the ones that started this discussion back in September, 2011 with an article that looked into this issue in-depth. When yours truly spoke to Tech Report chief editor Scott Wasson in mid-2012, he expressed his surprise that the frametime method had not been adopted yet by other websites despite the proven added value.
For Hardware.Info the reason is quite simple, using that method takes a lot more time. And since we tend to test large numbers of graphics cards, it's not an insignificant time investment. You also get so much data you end up with quite a few spreadsheets worth of it.
However, Tech Report is right. To get an accurate comparison of video cards, just comparing the average frames per second is really not enough. Frametime information paints a much better picture.