I hate these stupid articles, where a laughing author wants to be smart and tells with a few lines things which make no sense, at least in the way he did. What he told is only partially correct, but in this shortness it's not helpful at all! The consequence will be that thousands of engineers will run arround and tell, we must do this so and so, because this heroe from analog devices has advised it...
The author hadn't told which ADC or DAC he meant, are these 8bit converters or 24bit, are these converters accessed by a slowly serial bus or by a fast parallel ECL bus? The digital section, is there only a slowly microcontroller or is it fast digital board containing lots of FAST-TTL-, ACMOS- or event ECL-chips? Points which are extremely relevant when handling these converters, though. No, he pours his soup over the whole issue and the reader feels like an idiot...

First, why are the manufacturers of converters distinguishing betwen an analog and digital ground at all? Well, some chips don't have these two grounds, others do. But why?

Very easily: To prevent digital noise from eroding analog signal integrity. Each digital signal applied to a digital input of converter contains steep edges, which charges the stray capacitance within the chip. As consequence a spike current will flow back to the microcontroller which accesses the converter. That's not the only noisy digital current circulating within a converter. If the converter contains oscillators or sequencers additional noisy currents are flowing.
All these currents will flow somwehow through the ground pin of converter and right there trouble will occur, because these noisy currents produce a voltage drop across unavoidable impedances, namely substrate resistances and bond wire inductivities.

As consequence the internal ground on die is shifted by these voltage drops and even if the analog signal to be converted outside of chip is totally clean, the analog section on die will see a noisy analog signal, because it does not know that the noise is a result of voltage drops across ground bond wire.

That's the reason, why high precision converters distinguish between an analog and digital ground pin, just to allow the noisy digital currents to flow along their own path, no furtherly contaminating the analog ground.

This noise is common mode noise and therefore also a technique can be used which is helpul to suppress common mode noise elsewhere, namely the use of symmetrical signal routing, where the signal is the difference of two lines. So, high precision converters often provide a symmetrical input, yes even a symmetrical output, often in addition to a separate analog ground pin.

Very low precision converters often don't need a separate analog ground pin, because the voltage drop across ground bond wire due to ground noise can be less than one LSB. A 5V 8bit converter, for instance, has an LSB of about 20mV and if the converter does not contain noisy sequencers or oscillators ground noise might be negligible.

High precision converters, on the other hand, are so sensible against ground noise, that they nearly always contain a separate analog ground pin, especially if one knows, that symmetrical signal routing loses its advantage at higher frequencies, frequencies where digital ground noise is dominating, unfortunately. One LSB of 2.5V range 16bit converter is only 38µV!! So, even if the converter contains only the least noisy sequencers the ground noise along a common ground bond wire will nearly always be too high.

Facing these numbers it seems to be impossible to build high precision converters at all. There will be always so much digital noise that analog section might always be contaminated. But there's a trick: There are differently dangerous ground noise sources: Ground noise resulting from accessing the converter before and after the conversion don't play a significant role on the result of conversion. So, it's sufficient to keep the noise within converter minimal at least during the actual conversion. This is also the reason, why some people get proper results with a unproper grounding management and others do heavily fail: To disturb the conversion the digital noise spike must arrive at the "right" moment. A bit earlier or later and it will not affect the result of conversion! Most converters are built to be immune against digital noise which hits the converter outisde of the actual window of conversion and, of course, they are immune against the noise fabricated by themselves.

This separation of analog and digital ground is often confused, that these two pins can arbitrarily float against each other, they can not, of course, because they are sharing the same circuitry and are not allowed to float to more than the voltage drop of one forward biased pn-junction. More, any noise between analog and digital ground pins would affect the result of conversion due to charge injection mechansims, better known as "crosstalk". (But again, only if the noise spike comes at the right moment!)

That's the reason, why manufacturers always recommend in their datasheets to connect the analog and digital ground pin directly near the package.

Manufacturers often demonstrate the performance of their converters by a recommended layout, which divides the ground plane into an analog and digital section, with a connection of both directly at the package. They even have separate supply voltages for the analog and digital section, mostly.

But make no mistake, this is only an idealization. There are no separate things in reality, there's always stray capacitance, interwinding capacitance of transformers, input to output capacitance of optocouplers, etc, everything, which seems to look seperated but works as plain short circuit at high frequencies! A good example is these PC-based pico scope series, which offer "full" isolation from PC by the help of opto-couplers. But what you can see in the 5mV range is lots of nasty ground noise, coming from the nowhere!

Also, what, if you have to connect a chip on digital section and a chip on analog section to outer world? How to keep the both grounds separated then?? These simple recommended layouts can never actually be used in reality, where everything has to be powered by a common transformer, has to sit in a common enclosure and where cables from analog and digital section have to leave the board to say hello to outer world. In these cases you have to solve the situation differently: Either by using a common solid ground plane, even if the precision of converters suffers, or by using an additional chassis plane, where the signals which leave the board are referred to, or by distinguishing between a digital ground plane and a common analog/chassis plane, as I do it very often. Even digital signals can leave the board from the analog section, but need to be common-mode filtered, when crossing the gap between digital and analog section or when entering or leaving the chassis plane.

Real applicatons are always so much comlicated, that you will never find a ready solution in a datasheet of a converter, to think this is just an illusion.

So, finally, how to connect the analog and digital ground pins of converters in a real application, means when more than only one converter is used?

First, there's no golden rule how to proceed! It just depends on actual application! To find the right way can be very hard and painful, believe me!

Some manufacturers recommend to connect all the ground pins of converters to the anaolg ground plane and all the digital ground pins to the digital plane and to connect both ground planes together at only one point, near the regulators.
This can work, if the digital section is rather quiet. Then, there will only be rather little ground noise between each pair of analog/digital ground pin of involved converters. Remember, the more digital circuitry is used, the higher the probability is, that a nasty ground noise peak will arrive during the critical window of conversion and affest the result of conversion.

Others recommend to put everything on the one and only solid ground plane but neverheless keeping the analog and digital sections strictly separated, like Russell mentioned. This will only work, if at all, when a multilayer board is used, means when the magnetic coupling of signal lines to solid ground plane is so high, that ground retun currents (ground noise!) is forced to circulate (nearly) only in the according section.
This methode will probably fail with the highest precision converters. But this again heavily depends on whether the nasty disturbing noise spike will arrive at the "right" moment, which is the more probable the faster and complexer the digital section is.
This scheme is advantageous, when many signals have to leave the board, equally whether analog or digital ones. This scheme is more and more used, maybe because less and less engineers have learned how to route analog signals properly.

Another methode is to use a separate analog and digital ground plane and to connect the digital ground pins of all converters to the analog ground plane, like above author recommended. But this will make ALL the digital ground return currents to flow across the analog ground plane back to the microcontroller. This CAN work, if low precision converters are used, or if the digital section is so quiet, that no disturbing noise spike will arrive during the critical window of conversion. But normally, using this scheme means heavy filtering of all digital stages of converters and of all digital lines running from the microcontroller to the converters, in order to keep the digital ground return currents flowing back along the analog ground plane minimal. Special line filters are needed to round the knees and smooth the edges of digital pulses and this for each line! So, this scheme has a good chance to work if the converters are accessed by a rather slow, slew rate limited serial bus. But with a fast digital board and being accessed by parallel buses I dare say, that it will be impossible to save the precision of high precision converters.
Also, the needed supply and signal filtering and layouting is probably beyond the capabilites of most digital engineers, at least to my experience.

Kai