Cutting through the BS of low noise supply design

[ezalys]

Hey all,

I'm interested in low-frequency low noise low distortion signal acquision and generation, mostly for audio and control loop characterization. 130 dB of dynamic range up to 400 kHz let's say on a +/- 1 volt signal. I realize that given I'm even asking these questions I've got a long way to go but I've gotta start somewhere, but I do have most of the equipment to measure such things... so that's maybe a good start. What I just can't seem to cut through is how to think about low noise power supply design. Most modern high end instruments that I've torn into seem to use off-the-shelf switching power supplies like those from Meanwell. I notice a lot of audio folks still using linear power supplies... which at this point I'm convinced is totally unnecessary. I can imagine that careful design of a switching power supply from the AC mains input to the voltage fed into the LDO can reduce switching harmonics, which the LDO no longer needs to work to suppress... which is good. How then do these high end manufacturers get away with these COTS power supplies? Can you clean up the output of these COTS supplies? What are the general strategies? At what point is it necessary to design your own supply instead of filtering like hell? I have to think it's never worth it if I just see COTS supplies in T&M equipment!

What I understand is that with a linear supply you're starting with a noisy AC signal containing mostly 60 Hz plus harmonics of this from rectification plus tons of other garbage from poorly designed electronics also plugged into the wall that barfs noise back onto the line. Is it always better to step down rectify and filter this as with a linear supply in comparison to starting with the garbage that comes from a COTS supply and filtering this? These seem to me to be equally hard problems. That seems to be the comparison I have in my mind.

Second... Is there a purely filter-based way to think about power supply design? I've always understood power supply design as trying to design some sort of high pass filter with as low as possible an output impedance. I've always imagined that an SMPS is a high-pass filter that adds it's own noise (or rather, high frequency harmonic content)... with a corner somewhere around the loop frequency, and thus post-filtering components become smaller as this corner goes up, as one needs to get rid of switching content and extend the HPF as far as possible. A linear supply to me just seems to be accepting that this filter needs good rejection down to 50/60 Hz while maintaining low impedance. Furthermore, decoupling I've always seen as stiffening up this filter at higher frequencies... basically starting at a frequency corresponding to the capacitor and the source impedance (?) and ending at the self-resonant frequency of the cap and other parasitics. Is this the right picture?

E

[RandallMcRee]

No need to create the wheel again--there are many long threads about low-noise power supply design over at diyaudio.

The one that stands out as being closest to your question/quest is this one:

https://www.diyaudio.com/forums/vendor-s-bazaar/297147-silentswitcher-mains-free-15v-6-5-3-3v-power.html

Jan Didden created the silent switcher project more or less in answer to, methinks, just the question you are asking.

Randall

[ezalys]

I don't really see any schematics in the linked thread or any particular overarching comprehensive explanation about why designing this switcher with a switching frequency of 1 MHz and then post-regulating is a much better way to go than using passives to do the same... or even if it's possible with passives or WHY it's not possible with passives. Or using passives and then post-regulating.

Also, does this switching frequency of 1 MHz imply that the noise that's below that is cleaned up?

I've also encountered these resonant converters... which I understand as a narrow BPF driven at its center frequency. Does this filter noise above and below the center frequency? Do you now only have to deal with the noise you've introduced with the switching in addition to what's within the BPF? What I'm really lacking is an understanding of power supplies in terms of filtering and impedance.

[dmills]

One NICE trick is to lock the switchers to a (harmonic) of your sample rate, that way any switcher residual aliases down to DC and just appears as a small DC offset.

Apart from the usual low noise things (pay attention to current loops) low noise with switchers really amounts to understanding that LDOs have PSRR that falls with frequency, so you really want to dump the switcher ringing before you hit the LDO.
Fortunately the carefully dimensioned L/C LPF is made for this, but watch the parasitics and you are usually better with a couple of low Q stages instead of one high Q one (that WILL in production have the series resonance right where you don't want it), typically a few ohms in parallel with the filter inductor will help to kill the Q.

One other thing about LDOs, they sometimes have horrible low frequency 1/f noise, this is worth being a little careful of.

Johnson has useful notes about LC power filtering in "High speed digital design".

[RandallMcRee]

Hey all,

I'm interested in low-frequency low noise low distortion signal acquision and generation, mostly for audio and control loop characterization. 130 dB of dynamic range up to 400 kHz let's say on a +/- 1 volt signal.
E

Taking a step up in abstraction, I would say that the solution to your problem of measurement has less to do with power supply design and more to do with a low-noise voltage reference for the A/D converter. At least that has been my approach, in trying to measure signals in your voltage ballpark but at the DC limit. Inspecting data sheets for A/D converters supports this as well as app notes such as DN568, Reference Filter Increases 32-Bit ADC SNR by 6dB:
https://www.analog.com/media/en/reference-design-documentation/design-notes/dn568f_web.pdf

What do you think?

[ezalys]

I think you're right. The PSRR of the ADC is 70 dB... for 130 dB of dynamic range that's 60 dB on 5 volts that's 5 mV... which doesn't seem super super difficult. I worry that this PSRR figure drops off at higher frequency though. I also worry about other components that might be more susceptible.

[David Hess]

Just like CMRR and open loop gain, PSRR falls with increasing frequency and generally at the same rate of 6dB/octave or 20dB/decade starting at a relatively low frequency.

The underappreciated problem with switching regulators is that they can inject considerable common mode noise into the ground or common reference.  Linear regulators can also but there is more difficulty at higher frequencies where magnetic and capacitive coupling become more significant.

[T3sl4co1l]

*Low pass, yes.

That is the only way to think about them, in a certain respect.  Namely, in terms of immunity from input/mains variations, and stability in the face of load variations.

A switcher does indeed add its own noise, so you need a filter to take that back down.

It is impossible to make a filter with zero impedance, so you must make do with nonzero.  The maximum impedance is deltaV / deltaI, for some maximum ripple deltaV at the maximum expected change in load current deltaI.

Note that, in current mode converters (the one true control scheme), the first inductor and capacitor are in the loop, so their L and C values are somewhat irrelevant to the filter (if you know the control loop transfer function, you can solve for the equivalent output impedance; or simulate it).  Everything afterwards is normal passive stuff (mind, making sure to include the supply's actual output impedance in the calculation).

Tim

[ezalys]

Common mode noise is something I still don’t really understand very well. I understand that the power supply is generating large current spikes, but how does this turn into common mode noise? How does the power supply manage to raise both ground and the supply line together? Doesn’t it need to “push against” a third point to do this?

[T3sl4co1l]

Consider two cables attached to either end of a PS module.  Consider that the switching loop is switching current into its local ground path.  Even a solid ground plane has nonzero impedance, and therefore voltage drop.  The worst case condition is with cables connecting at either end, which will transmit the full voltage drop across the board.

Because it's a ground plane, the drop does depend on the distance between cables, the width of the board, and the size of the switching loop (between input/output bypass cap, diode and switch).  The voltage drops off inversely with distance.  Consider a very small (say 1mm) loop in the middle of a 1 x 1m board: we expect very little voltage in this case; in contrast, consider a 10mm loop in a 100 x 100mm board: this will have only, say, 1/10th the voltage between connectors as ground loop in the switching loop itself.

Note that EMC is concerned with sub-mV levels of emissions, so it doesn't take much to be a problem!

It is sufficient to consider the voltages or currents between different points on the ground node itself, alone, but it is also important to note what happens to other wires/signals.  In the most common case -- power supplies -- large capacitors bypass the +V and GND wires together, so that they act as a supernode, i.e., effectively in parallel for RF purposes.

Therefore, when we draw the common mode equivalent circuit for a system, we often simplify cables into single wires, treating them as the same (common mode) voltage and current.

We treat signals separately, when they can't be justifiably connected together.  Say, because the impedance or voltage or current is very different from others in the group.  Coincidentally, this condition is likely to identify possible offenders for us.

Or we simply ignore signals altogether.  A signal trapped within a shield, is as good as no signal at all -- in the common mode, it's all shield!  So, we don't worry about what's inside a coax cable, say.

We are, of course, still concerned with what that signal might do as it goes through imperfectly shielded areas!  Say we have a coax cable going into one of those awful two-pin plastic connectors.  Like one of these,
https://www.digikey.com/product-detail/en/te-connectivity-amp-connectors/1-1337543-0/A97553-ND/1755940
We can model the cable as a continuous single conductor (the shield), except for where it ties into circuit ground, where that pin length corresponds to about 10nH of stray inductance.  In that case, we get (differential to common) mode conversion, as the two pins act like a small (~10nH), weakly coupled (k ~ 0.5?) transformer, and some of the signal current is coupled directly in series with the shield.

Whereas a connector like this,
https://www.digikey.com/product-detail/en/te-connectivity-amp-connectors/5-1634556-0/A97570-ND/1755957
has a clean ground path surrounding the signal in all places, except for a very modest gap between the ground pins themselves.  Which will let some signal through, but very little over the useful bandwidth of a BNC connector.

Tim

[ezalys]

Got it! So it's mostly IR drops in the ground plane. That makes sense.

So someone on the diyaudio forums suggested floating the converter and analog frontend by using an isolated power supply on the board itself and then use optoisolators and such to transfer clock and data lines over. Is this a good idea?

[T3sl4co1l]

Not just IR drops, but L*dI/dt drops too.  Which in copper, is dominant over about ~10s kHz.

Optos are pretty nice.

Beware that an isolated supply is yet another cut between planes, except even worse than that, you have AC noise directly across it.  And unless it's a special design, there will be significant capacitance along that cut, too.

Tim