DC power filter

We sometimes have a "quiet" power pour on a pc board, like +3.3Q. Ideally is has its own LDO from +5 or something.

But sometimes we just add a lowpass filter from, say, +3.3V to +3.3Q, and that tends to be a ferrite bead and a biggish cap. That's not very scientific, and the bead inductance can resonate with downstream capacitance if the load current wiggles.

Here's one experiment:

formatting link

The bead is VISHAY ILHB1206ER601V MOUSER 70-ILHB1206ER601V

rated 600r, 2.5 amps. Measures 3uH and 70 mohms.

VISHAY ILHB1206ER601V MOUSER 70-ILHB1206ER601V

Q is OK, but low frequency filtering looks mediocre. The 56u polymer cap catches fast spikes pretty well, although we'll also have a lot of ceramic bypass caps everywhere too. Power pours themselves filter the really fast stuff.

Supply ripple changes gate and FPGA prop delay, which is jitter in our world.

Reply to
John Larkin
Loading thread data ...

I was wondering the intent behind your test config for this RLC filter.

  1. What does the R=1 simulate? Doesn't that unload the LC resonance of ~10kHz

  1. Why drive with 0.1 ? voltage source ~ 1kHZ square wave then capture with 1M? DSO? Wouldn't it better to use actual ESR of regulator for DC source and use a current pulse pump to simulate dynamic CMOS loads.

  2. Would it be better to use an AC coupled 50? termination in the DSO to look for ripple above the SMPS switch rate or LC resonant frequency since this is where load regulated noise comes from in an LDO?

Tony Stewart Test Engineer EE'75

Reply to
Tony Stewart

Well, easy enough. Solve the network and add ESR until the poles become real.

Yeh, at what frequency? Ferrite beads have a diffusive characteristic, i.e., Z ~ sqrt(F) over a substantial range. By Kramers-Kronig, that's equivalent to saying X_L = R, or Q = 1, over the same range.

Hence, it might measure "3uH" at 1MHz, "1uH" at 10MHz, etc.

Real parts vary above and below that curve, of course, and so the Q follows the slope. Q generally stays low, so it doesn't matter much. It's not like you can make them resonate any sharper than the Q.

Q generally peaks at low frequencies, above the L/R time constant set by DCR

  • asymptotic L, and below the core loss curve.

And, at what bias current? Ferrite beads saturate typically by 1/20th or so of their DC current rating. Laird is one of the few that provides bias curves; I don't bother with anyone else.

Well, yeah. The FB looks like R || L. The R adds a zero, giving a shitty asymptote.

Easy solution: make a nice, clean CLCLC, and put damping on either end to keep it well-behaved under load conditions. At each end, an R+C provides parallel termination resistance, and a series L (optionally L || R) isolates that termination from the source and load. So it overall looks like: L(R+C)CLCLC(R+C)L

With good layout and component choice, and shielding to exclude radiating fields, this many stages should be good enough for practically unlimited attenuation (>100dB) at high frequencies, and can always be extended with larger values and more stages for lower frequencies.

I made a particularly nice CDN, LISN, bias tee, whatever you want to call it, with this scheme. The crossover frequency varies a bit with load impedance (say in the 10-30MHz range); the Q is never very high, for any point on the Smith chart.

formatting link
EUT side has a coupling cap to 50 ohms; R+C in the middle of two chokes; DC side stays nice and quiet.

Mind the strays of larger components, which will introduce zeroes in the transfer function. No, this does not happen randomly or unpredictably, they're quite predictable. Just need suitable models. The real problem is a lot of models are either overly simple (e.g. RLC parallel or series), hard to use, or don't model modes beyond the first SRF. Consequently, you have to either assume, measure, or guard-band those modes.

Coilcraft provide extensive models, but they use nonphysical elements which makes simulations difficult or impossible. Luckily, I'd found a converter which makes them work in any analysis:

formatting link

Tim

--
Seven Transistor Labs, LLC 
Electrical Engineering Consultation and Design 
 Click to see the full signature
Reply to
Tim Williams

R=1 kills the resonant Q a little. It's not simulated.

That shouldn't matter too much. My big concern was resonance, and this combination looks pretty good. The regulators and loads could vary a lot.

I'm mostly concerned with mid-frequency ripple, so the scope termination doesn't matter.

Regulators generally have mediocre high-frequency loops, so we rely on output capacitance to keep things stiff at all but low frequencies.

A more compulsive test would combine a specific regulator (switch or linear) with a filter, and test the load transient response at various base loads. That could become a project.

I wonder if just a regulator and its output caps ever parallel resonate. Probably so, maybe often.

I've done some load transient testing with LM317 types with ceramic output caps, and they do ring pretty good after a load step. I posted that a while ago, with a fix that seems to work.

Reply to
John Larkin

I did this experimentally because I don't have a good model of the bead. And I'd prefer to not add resistance in series with my big cap; its ESR is nice, around 20 mOhms. To kill Q, I prefer to add a resistor across the inductor. The bead Q isn't high, which is good.

On my AADE meter, which runs it in the very roughly 500 KHz range, I think. Pretty low. Inductance is a fuzzy concept.

That's why we buy them.

Beads seem to have pretty good Qs at low frequencies. They behave like inductors in the roughly 1 uH range. LT Spice has sims of the Wurth parts, but we don't use them. The correlation of Z-at-100-MHz and inductance is weak.

I'm not counting on that.

I'm expecting 100 mA, maybe 200. This 1206 bead is rated for 2.5 amps.

It cuts the ringing about in half. Hardly worth it, but the cost of the resistor is maybe a penny.

The problem is that the supply current can depend on things out of my control, namely the customer's timing pattern. If I have any supply resonance, somebody will find it. And I want to keep it simple. There are something like 16 supply rails on this board.

The real concern is: does a bead and a cap have low frequency resonances? Sometimes they do. This combo ain't bad.

Reply to
John Larkin

ElectronDepot website is not affiliated with any of the manufacturers or service providers discussed here. All logos and trade names are the property of their respective owners.