While certainly correct, I find this graph misleading in practice.
In practice, coherent (or time-domain) averaging
reduces the FFT baseline noise floor by 10*log(N), N=number of averages, just the same way the FFT size itself reduced the baseline noise. However, it does not make the noise floor smoother unless a very large number of averages is used (1000 or so).
Spectrum magnitude averaging does not reduce the FFT baseline noise floor at all, it just makes it
smoother and cover a smaller level span. A smooth noise floor makes it much easier to identify components peaking out.
Without overlapped FFT-processing the benefit of coherent averaging is minimal/nonexistent. Aquiring 32 coherent averages of a 64k FFT takes just as long (and gives almost the exact same result) as a 2M FFT (64k *32) without any averaging.
However, with large FFT block overlap, like 16x (93.75%), the coherent averaging is much faster as the all the frames except the first are aquired in 1/16th of the time. Overlap is most useful when using narrow windows where large sections of the block have only little impact on the result, basically throwing away the information that would be available there.
But, with 32 coherent averages the smoothness of the noise floor is not great so it is harder to seperate real components from random spikes. I seldom see an improvement larger than 10dB improvement for the "safe" baseline noise floor where components stick out clearly to not get confused with noise. What's missing in current software (
@JohnPM) is an option to apply additional magnitude averaging on the output of a silding window of coherent averages. That would really help a lot to reduce the roughness of the noise floor (down low in frequency it will remain rough compared to an equivalent full-size magnitude average, though)
For precision level measurement of components I fount coherent averaging having error bands of +-0.2dB or so whereas magnitude averaging always gives larger than real result (not a bad thing, actually), the closer the component is to the noise floor the larger the error (several dB easily).