aaron_do wrote on Jan 5th, 2014, 5:37pm:Hi carlgrace,
thanks for the tip. Assuming my comparator threshold is well defined, this problem would only occur if my signal could change fast enough to eat up my entire redundancy overhead within the time difference between the subADC comparison instant and the MDAC hold instant right? So as a back-of-the-envelope calculation, for a 1.5b subADC,
Vref*sin(2πfnyquistterror) = Vref/8
terror = 0.02/fnyquist
which works out to 0.2ns for a 200MHz sampling rate. Does that sound right?
There are several publications where the subADC threshold is modulated with a PN sequence as part of a background calibration. I suppose this would make the problem worse?
So if I want to go the matching route, I need to make sure that the subADC comparator decision instant and the MDAC hold instant is well aligned right? And if I want to avoid matching, I need a front-end S/H. In your experience, is this the main source of high-frequency error?
That sounds about right. As you can see the matching of the MDAC and subADC paths becomes pretty tight at high frequency.
You're also correct that a background calibration like you described would make this problem worse. That's because these algorithms depend on using the correction range to inject their test signals. I imagine it is somewhat involved to specify the comparator offset in this case.
If you have a front-end SHA then this matching issue goes away. But you have to pay with a lot of power.
However, keep in mind that if you eliminate the SHA you will increase the required power in the comparators because you will need to keep the offset down. Dynamic offset typically dominates the offset of a CMOS comparator in deep submicron technology. It is very difficult to simulate accurately so you often over-design with a class-A preamp and dedicated sampling caps for the subADC. It's a delicate tradeoff to be sure.
I'm not sure what you mean by "main source of high-frequency error". What error specifically are you talking about? Distortion in the sampling switch limits the SFDR at high frequency. Incomplete MDAC settling limits the SNDR. If you don't have a SHA my guess is mismatch between the MDAC and subADC would limit your sampling frequency.
If you're going for high accuracy too, don't forget jitter in your sampling clock can also limit your sampling rate. Pipelined ADCs are fun.