Decoupling capacitors and perceived wisdom

[Howardlong]

This video questions the legitimacy and perceived wisdom of using parallel decoupling caps of order-of-magnitude different values to deal with their intrinsic resonances turned up on my feed a couple of days ago.



The TL;DR is to forget using, say, a 1nF plus 100nF, and just use two 100nF (with several caveats, of course).

Of course, it's horses for courses, but  my takeaway was... even more engineering work!

Meanwhile, I'll continue to use, as a starting point, the datasheet recommendations and vendor reference designs.

It reminded me of the perceived wisdom 15 years or so ago of using split ground planes to separate analogue and digital grounds in mixed signal designs, covered here

[iMo]

Afaik ADI, for example, since ever recommended 3 capacitors with 2 orders of magnitude apart (like 1n, 100n, 10u)..
PS: Btw., I have never ever heard about the wisdom on using parallel decoupling caps of "order-of-magnitude different values"..

This video questions the legitimacy and perceived wisdom of using parallel decoupling caps of order-of-magnitude different values to deal with their intrinsic resonances turned up on my feed a couple of days ago.

[ConKbot]

Working with SMT caps it's pretty cut and dry, your package size and layout sets parasitic introduce, so what does a 10n or 1n 0201 get me that a 100n 0201 doesn't? On the far side of the resonance they all converge on the same upwards impedance slope. So unless you're putting the resonance right on a specific frequency to eat it, the high frequency performance is similar.

[tom66]

There are some small changes in impedance for 100nF vs 1uF 0402 ceramic capacitor. This is because the 1uF capacitor is constructed differently.  But in general I do not see much need for going with, say, 10nF in the same package.  Best to stick with nearly or actually the largest capacitor value you can get in the package and then maybe a 100n-470nF in parallel in case that one has too high ESL.  If you look at the charts from cap manufacturers like Murata there is a small difference in high frequency impedance between caps of different value in the same package, although usually it is quite subtle.  Even better might be going 0402 1uF plus 0201 100nF, but not everyone wants to use 0201.

[Howardlong]

Afaik ADI, for example, since ever recommended 3 capacitors with 2 orders of magnitude apart (like 1n, 100n, 10u)..
PS: Btw., I have never ever heard about the wisdom on using parallel decoupling caps of "order-of-magnitude different values"..

This video questions the legitimacy and perceived wisdom of using parallel decoupling caps of order-of-magnitude different values to deal with their intrinsic resonances turned up on my feed a couple of days ago.

OK, then "orders-of-magnitude". Happy now?

I think you're missing the point, the video is suggesting NOT doing that.

[Howardlong]

Working with SMT caps it's pretty cut and dry, your package size and layout sets parasitic introduce, so what does a 10n or 1n 0201 get me that a 100n 0201 doesn't? On the far side of the resonance they all converge on the same upwards impedance slope. So unless you're putting the resonance right on a specific frequency to eat it, the high frequency performance is similar.

Yes, so the point is that blindly using, for example, a 1nF || 100nF for decoupling was probably often just a rule-of-thumb box ticking exercise.

That won't stop me using the vendor-recommended suggestions as a starting point, though.

[nctnico]

Afaik ADI, for example, since ever recommended 3 capacitors with 2 orders of magnitude apart (like 1n, 100n, 10u)..
PS: Btw., I have never ever heard about the wisdom on using parallel decoupling caps of "order-of-magnitude different values"..

This video questions the legitimacy and perceived wisdom of using parallel decoupling caps of order-of-magnitude different values to deal with their intrinsic resonances turned up on my feed a couple of days ago.

OK, then "orders-of-magnitude". Happy now?

I think you're missing the point, the video is suggesting NOT doing that.

The measurement method in his video is wrong because it doesn't represent a real PCB layout with all it's additional parasitics and package interconnects. Look for PDN impedance matching. The last time this topic came up I dug a little bit deeper and came across a much better explaination as to why using different values IS better in the end. Many people look at the real impedance only. But what about the imaginary axis? What happens at a reasonance peak in terms of phase shifts? Especially when the resonance frequency is concentrated at one frequency. You are basically dealing with HF here. For best RF performance, you want to keep the PDN impedance flat and hence you'll (very likely) need multiple different values in a real world design. Chip manufacturers know this and that is why you'll see that all of them provide lists with values to use for decoupling capacitors for high speed digital parts so designers without simulation software at least have a decent chance of getting the PDN impedance right.

This article shows the basics of having a flat PDN impedance: https://www.signalintegrityjournal.com/articles/1715-optimizing-power-distribution-networks-for-flat-impedance However, I think it will be extremely hard to apply this concept to a more complex circuit where a single supply rail powers many chips each with their own distribution network and internal decoupling.

[Gerhard_dk4xp]

http://www.hoffmann-hochfrequenz.de/downloads/experiments_with_decoupling_capacitors.pdf

OMG. I wrote that in 2007.

Where is my bottle of Gerontol-forte??

Gerhard

[NE666]

I'm not qualified to wade in on the actual video and question at hand however, I did have this YT contributor's "A Critical Resistor Pitfall" video amongst my suggestions only a week or so ago, and it was clear even to me that his presented theory of ohmic heating causing the noise he was seeing wasn't going to stand up to scrutiny.

[wraper]

Different capacitances are only worth it if they are of different dimensions. Otherwise there is virtually no benefit from using smaller capacitance as they have roughly the same inductance for particular package size regardless of capacitance.

[Howardlong]

http://www.hoffmann-hochfrequenz.de/downloads/experiments_with_decoupling_capacitors.pdf

OMG. I wrote that in 2007.

Where is my bottle of Gerontol-forte??

You were ahead of your time ;-)

How much was a 100 uF 6.3V ceramic multilayer 1208 in 2007 I wonder? Bleeding edge stuff!

Tantalums seem to have gone the way of the dodo since then, presumably down to a combination of MTBF, failure mode, cost, and ecological impact.

[tom66]

http://www.hoffmann-hochfrequenz.de/downloads/experiments_with_decoupling_capacitors.pdf

OMG. I wrote that in 2007.

Where is my bottle of Gerontol-forte??

You were ahead of your time ;-)

How much was a 100 uF 6.3V ceramic multilayer 1208 in 2007 I wonder? Bleeding edge stuff!

Tantalums seem to have gone the way of the dodo since then, presumably down to a combination of MTBF, failure mode, cost, and ecological impact.

It's funny looking at boards made, say, just 15 years ago and seeing them littered with tantalum and electrolytic capacitors now, and looking at boards made today and seeing just MLCC.  There are a few factors that made it possible.  One was ceramic capacitors getting better (higher energy density, lower cost), sure, but another was that you just need less capacitance now, with faster switch mode power supply controllers for instance, with fast switching rates meaning that less energy is stored per cycle and faster control loops meaning they can respond more quickly to a load or input change.  Tantalums were really compromised in multiple ways, having reliability issues at high temperature and voltage, a propensity to combust if abused, and limited temperature range unless you paid for premium devices.  MLCC has its own issues, DC bias being a particular pain for some larger caps, but they are a lot better than tantalums for 99% of applications.

[Howardlong]

MLCC has its own issues, DC bias being a particular pain for some larger caps

Add aging to the list for class II and III MLCCs too.

[David Hess]

Many years ago I had a wideband, 144 MHz to 1.2 GHz, power coupling capacitor application with 12 x 1206 size SMD capacitors in parallel in a transmission line environment, so there was zero lead length.  (1) 12 x 0.1 microfarad capacitors failed with excessive loss at high frequencies, while 4 x 0.1, 4 x 0.01, and 4 x 0.001 worked.

I do not know how to reconcile this with the advice about not staggering the capacitance values.  It seem like having a 50 ohm impedance would be a better situation for parallel capacitors of the same value than the much lower impedance of a PDN.

(1) Axial leaded parts could have been used since the leads would form the transmission line, but surface mount parts were actually easier for assembly.

[Howardlong]

It seem like having a 50 ohm impedance

That was my first thought regarding the video's practical demonstrations: there was an assumption made that the power rails are 50 ohm terminated, when in fact it's far more complex than that, including non-linearities and complex impedances. I'm not discounting everything he's saying, far from it: I found it interesting, but more as a food-for-thought than encouraging me to fundamentally re-think my strategy of seeing what the data sheets suggest, and how the vendors' reference designs do it as starting points.

[PlainName]

Different capacitances are only worth it if they are of different dimensions. Otherwise there is virtually no benefit from using smaller capacitance as they have roughly the same inductance for particular package size regardless of capacitance.

The chart shows the different capacitances give lowest impedance at different frequencies. Isn't that what you want?

Edit: I misunderstood the thrust of the charts, but am now enlightened.

[Nominal Animal]

What I'd like to know, is whether I should try to minimize supply pin to capacitor distance, or say the loop length/area from supply pin to capacitor to ground pin, and why.

Simple theory says that supply pin to capacitor distance, because the ground end of the capacitor only needs the ground potential reference to work, with theoretically no current at all.  But what say other practical considerations?

The above diagrams showing how package size affects the minimum impedance point is just such a practical consideration.  Similarly, when one uses two or more capacitors of different sizes, constructions, or dielectrics, their interaction (and nonzero ESR and ESL) often creates a new impedance maximum at high frequencies (above where the two minima would be separately).  In the case of current microcontrollers running at tens or hundreds of MHz, that new maximum can hit the internal switching rate (clock frequency), making the pair less effective than the bigger capacitance one would be alone.  Many of you know these effects much better than I do; I only know it can happen.  But it does show that one should never trust theory blindly, always check.

[DimitriP]

It seem like having a 50 ohm impedance

That was my first thought regarding the video's practical demonstrations: there was an assumption made that the power rails are 50 ohm terminated, when in fact it's far more complex than that, including non-linearities and complex impedances. I'm not discounting everything he's saying, far from it: I found it interesting, but more as a food-for-thought than encouraging me to fundamentally re-think my strategy of seeing what the data sheets suggest, and how the vendors' reference designs do it as starting points.

To paraphrase ...."it's 50 ohms somewhere"

[Someone]

Many years ago I had a wideband, 144 MHz to 1.2 GHz, power coupling capacitor application with 12 x 1206 size SMD capacitors in parallel in a transmission line environment, so there was zero lead length.  (1) 12 x 0.1 microfarad capacitors failed with excessive loss at high frequencies, while 4 x 0.1, 4 x 0.01, and 4 x 0.001 worked.

I do not know how to reconcile this with the advice about not staggering the capacitance values.  It seem like having a 50 ohm impedance would be a better situation for parallel capacitors of the same value than the much lower impedance of a PDN.

You have the explanation right in the post, DF vs frequency. Were the 1n parts the same material (and voltage rating) as the 100n?

[David Hess]

Many years ago I had a wideband, 144 MHz to 1.2 GHz, power coupling capacitor application with 12 x 1206 size SMD capacitors in parallel in a transmission line environment, so there was zero lead length.  (1) 12 x 0.1 microfarad capacitors failed with excessive loss at high frequencies, while 4 x 0.1, 4 x 0.01, and 4 x 0.001 worked.

I do not know how to reconcile this with the advice about not staggering the capacitance values.  It seem like having a 50 ohm impedance would be a better situation for parallel capacitors of the same value than the much lower impedance of a PDN.

You have the explanation right in the post, DF vs frequency. Were the 1n parts the same material (and voltage rating) as the 100n?

This was a power application, up to 50 watts, so I started with a design which was a best effort short of using RF rated capacitors within the mechanical considerations.  The capacitors were all common C0G parts of the same voltage rating, probably 100 volts, and the same series and package size.

I started with 12 x 0.1 microfarads because it did seem like the low inductance, especially with transmission line construction, would work fine at the highest frequencies.  It did not and there was a unacceptable SWR mismatch.  Dividing the capacitors between 0.1, 0.01, and 0.001 microfarads lowered losses to below 0.5 dB.  If that had not worked, then I would have used surface mount RF capacitors.

[iMo]

We want this, afaik. Also the slopes in the picture differ from the ideal situation, capacitors have self resonance peaks there, etc. But still the recommendation to have 3 caps 2 decades apart (starting with the lowest capacitance closest to the device you decouple) when dealing with fast or RF circuits is worth the effort, imho.

[Siwastaja]

We want this, afaik.

But that is what you don't get when paralleling actual capacitors of different values. It looks like that if you look at their curves in isolation and overlap them mentally, but that's only because they have different absolute Y axis. What you actually get it as shown by wraper.

[Someone]

Many years ago I had a wideband, 144 MHz to 1.2 GHz, power coupling capacitor application with 12 x 1206 size SMD capacitors in parallel in a transmission line environment, so there was zero lead length.  (1) 12 x 0.1 microfarad capacitors failed with excessive loss at high frequencies, while 4 x 0.1, 4 x 0.01, and 4 x 0.001 worked.

I do not know how to reconcile this with the advice about not staggering the capacitance values.  It seem like having a 50 ohm impedance would be a better situation for parallel capacitors of the same value than the much lower impedance of a PDN.

You have the explanation right in the post, DF vs frequency. Were the 1n parts the same material (and voltage rating) as the 100n?

This was a power application, up to 50 watts, so I started with a design which was a best effort short of using RF rated capacitors within the mechanical considerations.  The capacitors were all common C0G parts of the same voltage rating, probably 100 volts, and the same series and package size.

I started with 12 x 0.1 microfarads because it did seem like the low inductance, especially with transmission line construction, would work fine at the highest frequencies.  It did not and there was a unacceptable SWR mismatch.  Dividing the capacitors between 0.1, 0.01, and 0.001 microfarads lowered losses to below 0.5 dB.  If that had not worked, then I would have used surface mount RF capacitors.

I just pulled up some example series of 1206 C0G parts across that range. The parasitic inductances (package dominated) were all consistent with minimal resonance, but as predicted the DF is found to be lower for the smaller parts. So quite a different mechanism than the people chasing lowest |Z| by using diversity in parallel.

[Siwastaja]

Also see:
https://www.eevblog.com/forum/projects/good-practice-for-mlcc-power-by-pass-capacitor/msg5588349/#msg5588349
https://www.eevblog.com/forum/beginners/trouble-finding-100uf-ceramic-capacitors-do-they-exist/msg5582751/#msg5582751

[tggzzz]

Different capacitances are only worth it if they are of different dimensions. Otherwise there is virtually no benefit from using smaller capacitance as they have roughly the same inductance for particular package size regardless of capacitance.

Hence 0306 capacitors as well as 0603: shorter and wider both reduce inductance.

From https://datasheets.kyocera-avx.com/LICC.pdf