Why is it suggested to power XOs with LDOs?

Question

I've read in various places that crystal oscillator ICs require very clean power (as opposed e.g. to your usual 3.3V rail from some switching regulator) and one way to achieve that was via LDOs. (Ref1)

However, the only thing that LDOs actually achieve is to stabilize the voltage against low frequency variations, perhaps out to several 10s or 100 of kHz. In contrast, XOs run at 10s of MHz typically, where most LDOs are completely inefficient at cleaning anything. Capacitors or LC filters are anyway used around XOs to provide sufficiently low supply impedance at the switching frequency. To be clear, yes LDO datasheet do claim high rejection at 10 MHz with suitable output caps, but this is not due to the active rejection, but because of the passive filter formed by the output cap. A passive LC filter likely achieves better rejection at 10 MHz than LDO + output cap.

The only reason in favor or powering XOs via LDOs would be, if their frequency depended appreciably on the supply voltage. However, this doesn't seem to be the case as shown e.g. in this datasheet from Abracon. Something like 100 mV low frequency ripple (which LDOs can clean well) would cause only something like 0.1 ppm frequency drift, if at all.

I am a bit puzzled by reference 1 above. It shows how low frequency variations of Vdd cause corresponding low frequency phase noise, but how? Wouldn't the oscillator frequency have to change to cause phase noise ? Or is that couple of ppb frequency drift enough to cause the excess phase noise ? 0.1 ppm of 10 MHz could only cause phase noise up to 1 Hz offset if I understand correctly (probably don't).

The application I have in mind is for a precision ADC (not RF). Therefore, close-in phase noise between 10 Hz and 100 kHz is especially relevant.

So what is the benefit of powering an XO via an LDO (plus capacitors around the XO) instead of a basic LC filter that lets all the low frequency ripple through to the XO ?

Answer 1

Your power supply only needs to be as clean as your application demands.

If your XO is driving an MCU, then you can get away with practically anything. If it's part of a low noise RF synthesiser, then you have to get your hands dirty.

All oscillators have a 'pushing' performance, which may or may not be specified. It's the delta(freq)/delta(Vcc). A good oscillator can tolerate a noisy power supply, and a poor oscillator may require a quiet one.

'LDO's are very different in their noise performance, and it's not always easy to interpret the data sheets to see what they are really specified for. Hint - if noise isn't specified, it's probably bad. Be prepared to trawl through a lot of data sheets to find a quiet one, there are some out there.

There is no LDO that beats a 'big C' final filter of the power supply before it gets to the oscillator, at least for offsets of 1 kHz +.

As a general rule, to avoid over-designing, build your prototype with bog-standard LDOs and minimal filters. Then test it. Then fix the problems that need fixing. If you start out with the lowest noise available LDO and big caps in your power supply line, then you may succeed, but you'll never know whether you could have gotten away with something cheaper. But it depends whether you're making one, or one million, whether that route is going to be cheaper overall.

Answer 2

With respect to your edit, as @user_1818839 pointed, unless you have a perfect (infinite bandwidth) square wave, changing the amplitude also changes phase even at constant frequency. This is because the time when you cross the threshold voltage between 0 and 1 will change if the amplitude changes. If you could have infinite bandwidth, than the edge of the square wave is vertical and this is not the case. You always have some finite bandwidth though, so you need to keep your voltage as stable as possible. Otherwise if your amplitude is jittering up and down, so is the phase you detect.

With respect to why you would use an LDO, their ability to reject low frequency noise can be very high, 50, 60, 70 dB is claimed in many datasheets. Conversely, an RC filter with reasonable size will not be very good at this. Hence you really want both, at least in applications where a stable clock is important.