Twitter   Patreon WIGGLING 'LITE' IN GUEST MODE

Author Module Power Inlet
555x555
 OK, deep breath. At the risk of adding on to a topic which has been discussed a lot, I've been looking at various suggestions, examinations of others solutions, etc., for the inlet to a module from the power bus. Basically, a good bit of what I've read, and even more of what I've seen implemented in schematics online seems fishy, at best. I'm sure there's some mixture of me not understanding things and looking at explanations / schematics which are themselves confused. Hopefully some of you with more experience can clear things up. E.g. I gather that there's a lot about the nonideal nature of the components we choose that guides significant choices wrt the components, in this case. That hasn't totally settled in my brain. Anyway, without further ado, I figured let's try simulating this. So I plopped in this circuit: And then this is the result for each component type: This is the result if instead the load is constant: And this is the voltage (with the load varying, as before) on the voltage the simulated unfiltered constant-load module sees: Obviously there's a bit of arbitrariness to my choices of values here, but I don't think they're unrealistic. For the components going into the module, plus the capacitor, I picked a model ltspice had laying around, so as to try and simulate the nonideal characteristics. It's totally possible I'm off-base with some of these values, in which case, let me know. What I think I would conclude is: (1) The resistor is by far the best filter of external fluctuations. But it is still not very good at that. (2) A ferrite bead is the same as nothing at these frequencies. (3) The resistor is the worst for changing internal current fluctuations into voltage fluctuations. (4) Diodes are an OK compromise, and serve another purpose as well, obviously. (5) It seems like the degree to which an inductor will shut out power supply hum is also the degree to which it will behave like a resistor to internal current fluctuations and thereby lower the rails. No free lunch. If we are worried about total power supply fluctuation, it seems like the best solution is to add nothing in series with the module. If we are worried about just the fluctuation from other modules, a resistor is a good idea. But I'm not really sure why we wouldn't be worried about voltage self-fluctuations. I mean, I don't think we can take it as a guarantee that the effect of the current being pulled from the rails will simply add a tiny version of the same signal to itself. What about digital modules? Or modules processing multiple signals? And of course I doubt the PSRR of most things is nicely linear. Of course practice is worth a million simulations so tell me about that. Now, I've read elsewhere that the caps are not meant to filter, but to decouple, or to provide a charge reservoir. This makes no sense to me. Or rather, it makes perfect sense, but it seems like what a capacitor does in this context is to decouple a local rail from the power supply by filtering it, via the inherent ability of a capacitor to hold charge. Right? So isn't this all one and the same thing? Also read that the resistors are really fuses. So...is there a good reason I wouldn't replace them with a fuse, and save all the voltage fluctuation from the resistance? I mean, cost/convenience I guess? Also, how much does any of this matter if you have a well-designed module with voltage references where it's important and op-amps as the only thing on the rails?[/img]
mskala
 Ferrite beads aren't really meant for filtering out power supply hum, nor audio frequencies - their purpose is to limit radio-frequency interference. Similarly, decoupling capacitors' main function is to smooth out ultrasonic spikes, from logic chips, 555s, comparators, and other things that switch quickly. You might get a better picture of these components' strengths and weaknesses by looking at frequencies in the multi-MHz range, or sharp pulse waveforms. Of course that doesn't negate that many designers throw them in cargo-cult style where they aren't necessary or don't help.
555x555
 OK. I set the noise-injecting module to give 200ns current draw pulses every 1us, with 2ns rise times on each side. Still this doesn't seem to have a huge effect in simulation. I feel like I'm missing some parasitic something that ought to be there, or the modules aren't modeled well as pure current draws? Anyway, here are the results: Zoomed out, it seems like the diodes are actually the best filters, but zoomed in everything looks the same. And the ferrite bead still has very little effect.
mskala
 What do these curve signify? They seem to be voltages, but measured *where*?
555x555
 Sorry for the ambiguity. These are voltages measured at what would end up being "V+" of the module, i.e. at the current draw inlet for how I'm modeling the modules. I18 in the diagram. The schematic I posted is actually one of six, so I can test with each of the components. However, I realized I effed it up and had M+ and MGND connected everywhere on the most recent test, instead of one for each copy. Hence why everything was the same but the levels. Oops. The first test didn't have this problem. Here's the curves again for the high frequency spike test, correctly this time: So this is closer to what I'd expect. The hf transients are suppressed by the ferrite. But the overall amplitude is higher. The inductor does very little to get those hf problems; not sure why. I did "pick an inductor" rather than use an ideal one. Either I'm missing something (likely), or the way various analyses leave off cables and bus resistance and inductance makes the traditional analyses problematic (also possible).
mskala
 Are you really measuring at I18, which would be across the capacitor C11? If I'm reading your diagram correctly, the module is drawing something like 100mA (which is a LOT, but let's go with that...) and connected directly across a 10uF capacitor. If the power supply were to suddenly vanish, it would discharge that capacitor at a rate of 10000 volts per second (100mA divided by 10uF). Even with whatever other stuff is going on, it shouldn't be possible for the capacitor voltage to change much faster than that. And yet your curves are showing the voltage changing at billions of volts per second. That doesn't add up for me unless you're really measuring somewhere else, not directly across the capacitor. For instance, might you be measuring at "M+"? That doesn't make a lot of sense to me either, though, because C11 should still be having an effect and it doesn't seem to be.
slow_riot
 Firstly, I would recommend using AC sweep mode to get a graph of amplitude versus frequency, this is standard practice and will make it easier to reference your measurements against benchmarks. Tying into the above point, I would design your RC or LC on the inputs in reference to the PSRR curves for opamps. You are asking too much from a passive filter to attenuate all the way down to DC, as resistance, inductance and capacitance do bring in their own issues. Luckily, opamps have excellent PSRR except at high frequencies. For protection, there are 2 main methods, one is series diodes (pref. Schottky), the other is shunt. Series is the easiest, but it lowers the voltage the module sees by one diode drop. Shunt, where the diodes are connected from a +/- polarity to 0V, does not have the diode drop. It may also give better protection for misplaced header where + or - is connected to ground (it may also not). The diode generally needs to be protected by a current limiting device such as a polyfuse such that current surge during the reversed polarity condition does not blow the diode, and then blow the circuit. The limitation of polyfuses is that they can have high parasitic resistance, 3-5 ohms typically. For some circuits where you don't want common impedance coupling this may be a problem.
555x555
Thanks guys. The schematic doesn't represent a design in any way at this point; I'm just trying to understand everything first. That's partly why the huge current draws, to exaggerate the effects and see what can be done. Also this is the handy thing about a simulation: if you don't model it, it isn't there. So if I'm not understanding the problem, this will give results that don't look right.

So two updates:

First, if we look at what *another* module sees, it is clear that the ferrite beads and the inductor are huge wins at high frequency. If your module switches a lot of current and you do not add one of the two, then you are an asshole. Or my simulation is still wrong, which is very possible. Graphs:

Second, thinking through:

 mskala wrote: Are you really measuring at I18, which would be across the capacitor C11?...If the power supply were to suddenly vanish, it would discharge that capacitor at a rate of 10000 volts per second (100mA divided by 10uF). Even with whatever other stuff is going on, it shouldn't be possible for the capacitor voltage to change much faster than that. And yet your curves are showing the voltage changing at billions of volts per second.

I'm *not* measuring across the capacitor; I'm measuring from the capacitor to "true" ground, back at the power supply. But the module's ground itself is changing. Here are voltages across the capacitor:

And zoomed in:

This looks more like the filtering that other modules see.

It seems, given the discrepancies between the two, that ground ripple is a much bigger problem than power ripple. OK we already knew that. Since I'm only modeling one rail right now, it's worse than it would be in real life since at least some of the current draws would cancel. Also maybe the filtering of two power supply references wrt ground would hold ground in place a little better? I'll have to try that...
555x555
 OK, I think I'm getting the picture here. I tried this circuit, substituting various components in the per-module filters: I'm happy to go into details or post graphs, but I tried things a little more flexibly this time, so I don't quite have a graph that sums it all up. I measured the voltage across the current sources to get an idea of what voltages a module with a given filter "sees." This is what I came up with; please correct me if I'm wrong: (1) In a properly designed system, self-noise seems to always be the worst problem, in terms of absolute amplitude of voltage shift. This is minimized by minimizing the impedance to the module at the frequencies in which it is producing self-noise. Diodes don't just add a constant voltage drop, but also add some impedance. In my tests, diodes on the audio-rate module raised the self-induced voltage drops by about 3x for schottky, and 6x for silicon. Now, of course this "noise" corresponds with the signal, and so in practice might not be all that detrimental. That would depend on the particular circuit. Either way, it seems worthwhile to have a separate inductance-blocked path from a switching to an analogue section of a module, although that probably happens automatically a lot of the time through producing +5V or whatever is needed. I'm not talking about a dirty ground, but analogue vs. digital rails. (2) The noise from other modules, bleed, is almost completely a high-frequency problem, at least with the distribution system I hypothesized. Before compensation, it is actually worse than self-noise. But it can be compensated for with a little inductance on the power line. Ferrite beads seem to do pretty well, but an inductor is better. (3) You are going to be able to do very little about ripple, unless you first decide that you don't care about self-noise. Any filter that gets at ripple adds impedance and therefore voltage drops at audio rate current fluctuations. Ripple is best solved at the power supply. (4) It seems like bleed is better taken care of at the module causing it than at another module. Cooperation is necessary here. Which would then tell me that best practice is: (1) lower impedance however possible, (2) but add a little inductance to shut out high frequency content. Which mostly means that the problem is not solved at the per-module level, but at the level of power distribution. Although there's also (3) minimize current-draw fluctuations in the design, if possible, especially at high-frequency. That means wasting power, which is going to be the opposite of what most chip designers want. Does all of this sound about right? Questions: (1) So, is there anything realistically doable about ripple and audio rate noise, after it has already been injected into the rails? (2) Is there any real way to limit self-noise beyond what I've detailed above? Especially high-frequency self-noise?
mmagin
 I'll never understand the obsession with not reducing the power rail by a fraction of a volt. For reverse polarity protection, a schottky diode is a Vf of about 0.3. And if you're going to use a RC low-pass to filter noise, a 10 ohm resistor is just 0.2 volt at 20 mA. I don't like the shunt-style reverse polarity protection diodes. Polyfuses are finicky things that take a long time to return to their original value in some cases.
slow_riot
 555x555 wrote: Questions: (1) So, is there anything realistically doable about ripple and audio rate noise, after it has already been injected into the rails? (2) Is there any real way to limit self-noise beyond what I've detailed above? Especially high-frequency self-noise?

I personally look at achieving high PSRR, but I don't rely on it. All I can do is implore customers not to cheap out on a power supply. Some components don't have high PSRR (not specified for e.g. an SSM2164 unless I'm mistaken... edit, I am mistaken!).

The parasitic impedance of series devices like a diode or polyfuse attenuates noise on the supplies before the module, but devices upstream interact.

One interesting semi recent (?) component is a fusible resistor (e.g. Vishay NFR25), which is set at a "hold" resistance and tries to go open circuit in overcurrent situations. In some cases you could use a low ohmage (say 10) as the impedance on a supply inlet as well as overcurrent protection. Or alternatively to use a very low ohm like 0.33 just for OCP and minimum parasitics. Compared to polyfuses they are slower responding but they work for me.

I don't think a discussion about power supply noise is complete without looking at ground/earth. As you can't do anything about noise on that.

There was a good thread called "sacred rails/holy ground" with info about managing current flow. It's always worth looking exactly where is going where, minimising it, trying to route it across the rails.
555x555
 mmagin wrote: I'll never understand the obsession with not reducing the power rail by a fraction of a volt. For reverse polarity protection, a schottky diode is a Vf of about 0.3. And if you're going to use a RC low-pass to filter noise, a 10 ohm resistor is just 0.2 volt at 20 mA.

For the majority of applications, I'd agree: 0.3V reduction is completely acceptable. The problem I'm pointing to is that, at least in my tests, diodes do not simply have a current-independent voltage drop, but also a current-dependent voltage drop, i.e. an impedance. In my tests, the ripple in the voltage drop created by a 7mA change in current at 640 Hz is about 2mV p-p with no protection diode, 20mV p-p with a silicon diode (1N4148) and 11mV p-p with a Schottky diode (1N5817). That's compared to 128mV p-p with a 10 Ohm resistor, so if you're putting in a resistor but don't want to put in a diode...well that is ideology at it's purest .

 Quote: I don't like the shunt-style reverse polarity protection diodes.

Why not?

 slow_riot wrote: The parasitic impedance of series devices like a diode or polyfuse attenuates noise on the supplies before the module, but devices upstream interact.

Exactly what I was noticing. There is a tradeoff between attenuating noise from upstream sources and creating a varying voltage drop on the module. And as far as I can tell, it is probably not often a good tradeoff; the voltage fluctuations are much worse from the varying current draw of one's own module.

 Quote: I don't think a discussion about power supply noise is complete without looking at ground/earth. As you can't do anything about noise on that. There was a good thread called "sacred rails/holy ground" with info about managing current flow. It's always worth looking exactly where is going where, minimising it, trying to route it across the rails.

Definitely

I read through that thread and a few others back before I built my case. Homemade busbars are the result. Now, however, I'm wondering whether I should've gone with more robust wiring from bar to module.

There's a way where of course the distribution/ground shouldn't be a different topic than on-module regulation, but it is, because as a designer, one can't do much more than recommend distribution and ground. And one has to figure out how to work with whatever's out there, as best as possible anyway.
cygmu
 slow_riot wrote: There was a good thread called "sacred rails/holy ground" with info about managing current flow. It's always worth looking exactly where is going where, minimising it, trying to route it across the rails.

Here's that thread, for reference
https://www.muffwiggler.com/forum/viewtopic.php?t=110012&start=all&pos tdays=0&postorder=asc
BugBrand
 Why is the value of the electro (10u C11) never questioned?
555x555
 BugBrand wrote: Why is the value of the electro (10u C11) never questioned?

I'm open to questioning it . I read in a few places that some people want to use 470uF w/ a 10 Ohm, thereby lowering the frequency of the lowpass to 33Hz. I just don't know if I like the idea of turning on 20-30 modules = 9.4mF-14.1mF instead of = 200-300uF.
555x555
 Ok, in the simulation anyway, a 10 ohm resistor plus a 470uF cap does indeed cut the hum in half, as expected. Plus, although to a lesser extent, the cap/resistor also successfully cuts the amplitude of the self-noise for the audio-rate module from 128mV to 25mV. However, there seems to be diminishing returns here, as even increasing it to ridiculously high values wouldn't push it past 25mV. And with no resistor, that's 2mV, so... Without the resistor, increasing the 10uF doesn't have much effect. In fact, one could probably decrease it for some modules, though I haven't tried that.
555x555
 One more update: I actually did the calculation of the crossover for 33u inductor and a 10u cap: 8.7kHz. If we reduce the inductor to 2.2u, that pushes it out of the audio range, at 33kHz. And measuring the voltage in the audio module, the reduced impedance moves the self-noise from 2mV to 500uV. Again, it seems like there is a tradeoff between high impedance blocking external noise and low impedance avoiding internal noise. It seems to me that best practices ought to be (1) make the distro impedances as low as possible; (2) take care of hum at the power supply; (3) on the power inlet, use an LC filter with a crossover out of the audio range; and (4) separate high-frequency switching circuits with a separate power supply and a high-value inductor. But I'm not sure what to do with something like a sample-and-hold or a very fast envelope generator, which bridges slow and fast in weird ways. Question: *can* you take care of hum at the power supply, or is there an interaction with earth, or high-power devices like incandescent light bulbs, that makes hum more of a local problem?
555x555
 On second thought, this probably isn't about the LC frequency, but about the inductive reactance vs. the parasitic resistance on the power inlet. At 25mOhms, 33uH gives the same reactance at 120Hz, 2.2uH at 1800Hz, which is about twice the 640Hz frequency I'm testing with. To really push that out of the audio spectrum, you need about 120nH, which is starting to look on the order of the parasitic inductance from the wiring to the bus. So a ferrite bead is looking to actually be the better solution. /thinking aloud
slow_riot
 Lots of research combined with experimentation, forgery etc likely to give best results. LC (pi network) can be dangerous as without a damping resistance you have created a highly resonant filter. You didn't follow my suggestion about AC sweep mode, which will show the whole slope including Q point. With grounding, never have such neglible differences in values stirred so much debate! Neil Muncy is seminal: http://www.aes.org/e-lib/browse.cfm?elib=7945 Bill Whitlock too: http://web.mit.edu/~jhawk/tmp/p/EST016_Ground_Loops_handout.pdf My biggest influence is actually someone from the industrial sector, which in general has moved towards meshed multi point grounding rather than single point. http://www.cherryclough.com/PCB-design-and-layout-techniques-for-EMC http://www.cherryclough.com/Electronic-Design-Techniques-for-EMC
mmagin
555x555 wrote:

 Quote: I don't like the shunt-style reverse polarity protection diodes.

Why not?

Partially it seems like an inelegant design to me.
More practically, you absolutely have to have some kind of overcurrent protection in the module then, you cannot expect a particular current limit from the power bus. And if it's not self-resetting, that's probably impractical for a commercial product. And polyfuses are kind of finicky and unideal in their behavior.

On the other hand I suppose the added resistance of a polyfuse designed for the low currents we're dealing with could be substituted for the series resistor discussed elsewhere in this thread
555x555
 @slow_riot: those are some excellent resources. Thank you! I guess I'm diving back into the info-hole! @mmagin: Agreed, it's inelegant. However, the elegant design would have been to have somehow made it impossible to plug things in wrong in the first place with some sort of keyed connector. After that's been botched, I'm not sure I like any of the electrical "solutions."
555x555
 OK, I figured out how to do the math on this. So taking this circuit: I calculated the laplace transforms v[s] and il[s], the voltage the module sees and the current it pulls from the bus. First, v[s], depending on vp[s] and i[s]: Then, il[s], depending on vp[s] and i[s]: As you can see, after massaging into the typical forms, v[s] ends up being a low pass filtered vp[s] plus a low pass plus band pass filtered i[s]. il[s] ends up being a band pass filtered vp[s] plus a low pass filtered i[s]. A couple things about interpretation here: First, we can see that v[s] is basically a function of vp[s]--the voltage noise as seen at the bus--and il[s]---the self-noise induced by the current draw across the impedance to the bus. And further, these appear to be independent, and can be considered and analyzed independently. (Of course, component choices won't affect them independently.) Similarly, il[s] is a function of the fluctuating power charging the capacitor across the impedance of the lines plus input network, and a function of the current draw of the module. Again these can be considered independently. So really we have four transfer functions: v[s]/vp[s], the filtering of bus voltage fluctuations w.r.t. the module, v[s]/i[s], the self-noise resulting from the current draw of the module, il[s]/vp[s], the current resulting from the voltage fluctuations of the bus, and il[s]/i[s], the filtering of current fluctuations from the module w.r.t. the bus. il[s]/vp[s] is a tricky one to interpret. Since by hypothesis i[s] is independent of the power supply (of course it isn't independent in practice or none of this would matter), that means that the current resulting from il[s]/vp[s] can only go to the capacitor, and must be the current required to charge the capacitor, given a certain fluctuation of the bus power. Based on the relative phase of this current and the current on the bus causing the voltage fluctuations, this either dampens the effect by drawing some of the fluctuating current out of the bus / supplying the bus with current from our capacitors (bypassing the power supply return), or it draws additional current from the bus, thus amplifying the effect for the next module. Which it is probably depends on the power distribution topology; I'm not sure. Next, apart from the standard band/low pass equations, there are a few scalar terms. Notably R, as expected, affects the low pass part of the v[s]/i[s] transfer function. Additionally, the ratio between L and C affects the band pass portion of v[s]/i[s], and its reciprocal affects the il[s]/vp[s] transfer function. Note that, apart from its inherent role in determining Q, R does not appear elsewhere in these equations. Unless the inductance is negligible, which at high frequencies even the parasitic inductance of the cable harness starts to look non-negligible, R only affects the circuit by inducing a self-noise proportional to R multiplied by a lowpass of our currents. Shifting the filter to a lower frequency filters both the power bus noise and the self-noise, contrary to what I thought looking at the simulations. The problem is that the pass band of the current (and the stop band, but to a lesser result) will be multiplied by R. The band pass component of the self-induced noise is pernicious, unless we know the frequency spectrum of our module's current draw very well. This suggests that it is best to increase C rather than L in order to get the characteristic frequency down. When it comes to the current draw from the power bus, these equations tell us little except that the low pass filter is perfectly glad to filter our module's current draw, bidirectionally, from the bus. il[s]/vp[s] is the current we pull in order to maintain the more steady voltage of the filter. Whatever you end up doing to the ratio of C to L to limit the band pass of the current showing up as self-noise, will end up increasing the current pulled by the module to maintain the voltage. Both of these follow a band-pass pattern. *Whew!* Let me know if I missed anything .
 Page 1 of 1
Powered by phpBB © phpBB Group