ontheverge wrote on Dec 14^th, 2010, 12:41pm:


what makes you think that the requirement 1/(2^N) is too restrictive?

If you consider the systematic DNL introduced at the boundary of each quantization interval, then you will see that a finite settling accuracy dv will appear as +dv on one side and -dv on the other, giving you DNL = 2*dv, just due to finite settling errors (combination of finite DC gain + settling errors).

And that would be just the DNL from one single stage. If all your stages were identical, you would get input-referred DNL:

= (2*dv + 2*dv/2 + 2*dv/2^2 + ...)/2, where the assumption is that you are using a stage gain of 2. You can work out what dv needs to be in order to get whatever target DNL you want due to systematic errors alone, not counting other error sources.

Vivek

ontheverge wrote on Dec 16^th, 2010, 4:26pm:


The key is not the "per stage resolution" but only the resolution remaining after the stage in question.  The reason we can scale stages is because the number of remain stages decreases as we go through each pipelined stage.

So, a large per stage resolution reduces settling requirements because there are few bits to resolve after the first stage.  On the other hand, the feedback factor of the MDAC goes down with larger resolution, so you may need to burn more power to meet even this reduced requirement.