Within a single second you may have an increase of 2-3 Frames, however, if you total up the speed at which frames have been rendered within one minute you will determine that far more frames have actually been rendered.
20FPS x 60 = 1200 Frames per Minute
23FPS x 60 = 1380 Frames per Minute
This difference can be noticed in 3DMark 06 if you look closely, as the scores are based on frame rendering over a set period of frames.
Take Quake 4 as an example, as I have seen coding from the Doom engine used within the application itself.
Quake 4 does not use the full array of SIMD Level Instructions (Single Instruction Multiple Data, SSE-SSSE3+). Even without using this full array it can obtain very decent performance as the engine is extremely efficient.
If the full array of SIMD instructions are not used then the application will not address the full 4096KB as there are not enough instructions to fill the instruction sector of the internal cache data array.