So what can we conclude from this analysis?
- The performance of this converter is very good in terms of SFDR, which is the key specification in its intended application. After many months of refinement, we have been able to measure an SFDR that is actually slightly better than the manufacturer’s specification!
- The component that limits SFDR is our good friend the second harmonic, as usual. Maybe something else can be done to reduce distortion such as matching impedances better, improving power-supply bypassing, or improving grounding and shielding.
- Low-frequency noise is getting into the system somehow: more grounding, shielding, bypassing, etc. might be in order.
- There is phase noise around the fundamental which is detectable but is at a low enough level so that it is probably acceptable.
Our A/D converter has already received extensive attention in these areas, so this is probably about as good as it gets. To get to this point, we have:
- Put together a high-performance test setup capable of generating a signal whose distortion is less than the converter produces. (One quickly becomes a connoisseur of signal generators!)
- Improved the grounding and shielding of the A/D converter circuit.
- Captured data signal samples from the A/D converter and analyzed their performance with ScopeDSP™.
- Repeatedly iterated on the above over a period of months!
“But wait a minute! We did all that analysis without any Data Windowing! I thought whenever you did an FFT you were supposed to window the data first”!
(There’s one in every crowd…)
OK, let’s window the data. We’ll use the famed “Blackman-Harris” window:
As any self-respecting Data Window should, this one reduces the amplitude of the data more and more toward the ends so that it eventually became zero. An intuitive explanation of this is that it makes the ends of the data ” match up”. That is sometimes important before doing a transform because the “spectrum” of the signal is derived from an “infinite” signal in which copies of the Time Data sample are repeatedly pasted together, ad infinitum. If the beginning and end of the signal do not match up, there is a sudden discontinuity when the end of one sample copy is concatenated to the beginning of the next, which results in a “distorted” spectrum. The fact that the window zeroes the data at each end prevents any discontinuities from occuring as the (repeated) ends come together.