SFDR and SNR

I meant to say, the 40dB SNR, would be required if only a 1e-5 bit error rate was acceptable on the 6-bit PAM.

SFDR and SNR really aren't related much. SNR has to do with the signal energy versus quantization noise. Some ADCs are noisier than others so they have an effective number of bits that is less than the actual number of bits per sample to represent that the noise is higher than what one would expect for just quantization, e.g. an 8 bit ADC may only have 7.2 ENOB. The SNR then should be given by ENOB*6 or 43.2 in this case.

SFDR however has to do with spurs or images that are generated from a sinusoid at the input of the ADC. You see when a sinusoid is input, the quantization of some ADCs isn't really random it has a periodic structure that creates another sinusoid at a different frequency and much lower amplitude. In other words the quantization noise isn't spread across the whole bandwidth it collects in tiny little "spurs" that track the frequency and level of the input sinusoid. The SFDR then is the difference between the input sinusoid level and the highest spur, e.g. an 8-bit ADC may have a sinusoid input at -1 dBFS and a spur at -60 dBFS so the SFDR would be 59.

The energy in these spurs is very tiny and rarely matters for single signal systems. The two cases when SFDR does matter is wideband digitization and spectrum monitoring. If you want to take you ADC and digitize the entire HF band for example such that you want to be able to tune within the ADC data to pull our large and small signals the SFDR is important because the big signals in the band will produce spurs that you will not be able to seperate from real (but weak) signals. The other case is spectrum monitoring where you want to see the entire band on an FFT display, for example. Again the spurs created by the strong signals will appear like real but weak signals. In that case you have to decide that every signal that is less than -SFDR full scale is not reliable. Some systems even add wideband noise to wash out everything below the SFDR just to make sure the operator can't confuse spurs with real signals.

-Clark

OK, I think I got that. I am interested in wideband digitisation. In that case we wont have a single spur, we will have noise that is spread across the band. Is that noise flat in frequency? In that case is the SFDR not just the same as the SNR? If everything below the SFDR is washed out, then this would seem to me to be equal to the SNR.