Best output Capacitor

[gearheadgene]

What makes a good output capacitor for class D? What are the tradeoffs when selecting it? I've seen polypropylene, or other polyester caps used, but can't figure why. Is there any audible difference between these or any other types? Maybe this is one of those religeous questions

Seems to me that a good quality ceramic cap, with low ESR would work out nicely. Ripple current is surely going to play into the decision - which makes low ESR types ideal.

I'm inclined to use surface mount, if possiblet and ceramic chip caps are good for that. But if they are not the best choice, for whatever reason, through hole is acceptable.

[Pafi]

Gearheadgene!

a good quality ceramic cap

Do they exist in high capacitance and high voltage? High value, high voltage ceramic caps are tipically made of Y5V, wich is not a capacitor material in reality:
https://avxcorp.com/docs/catalogs/cy5v.pdf
As you see from first 3 diagrams, it's good for temperature sensor, distorter, resonator, or something else, but absolutely unusable for ClassD audio.

NP0 material is quite good, but there is no bigger cap then ~10 nF made of this material. (edit: I was wrong, 100 nF exists, but it's big, very expensive, and presumably not as good as smaller ones)

X7R is more or less usable, however they also exhibit piezo effect, resonates at some MHz, and they are also temperature- and voltage-dependent:
http://www.avx.com/docs/Catalogs/cx7r.pdf
Voltage dependance is not specified here, but I found it's more then 15 % at nominal voltage. Maybe if you connect 5 pieces of 100 nF paralell, you could get a reasonable alternative for small power, in a post-filter fed back amp. In pre-filter fed back amp it has approximately "only" 1 percent of distortion at high freq.

Polypropylene and polyester caps are practically perfect compared to ceramical caps in these aspect. The only problem that it's almost impossible to get SMD version.

[gearheadgene]

I'm not following your reasons. All that NPO Z5U and X7R stuff refers to temperature tolerance. For example, NPO has a very tight tolerance over temperature, while Z5U has really lousy temperature coefficient. Maybe that's what your getting at? The poor tempco caps vary drastically vs temperature. So I suppose they don't work well for class D because the roll off frequency is going to change all over the place as the part heats up. The opposite must be true of an NPO since it is very stable over temperature.

Piezo effect of X7R? Really? I never heard of that. I'm more inclined to believe it's due to ceramic material instead of the tempco.

I wasn't aware of the voltage dependancy either. So you are saying the poly caps are stable over temp, and voltage - ok. So it's not an audible subjective issue, it's purely the ability of the device to remain withing spec - that's the issue.

gene

[Pafi]

All that NPO Z5U and X7R stuff refers to temperature tolerance.

They are material classes, characterised by temperature coefficient and many other features. Check out linked datasheets!

Piezo effect of X7R? Really? I never heard of that.

I even heard it. It plays speaker. (And microphone, when its on the input.)
http://en.wikipedia.org/wiki/EIA_Class_2_dielectric

The strongest problems are resonance and/or voltage-dependence.

So it's not an audible subjective issue

It can be audible also! Voltage dependence causes distortion, resonance can cause false comparation.

[gearheadgene]

Interesting. That link concurs that it's the ceramics that are piezo electric. It also agrees with you about voltage coefficient - which is something that never occurred to me as a possible side affect.

Looking closer at these polyester and polypropylene caps - I see that they are VERY inductive. I'm still looking around, but most of what I'm finding is that they have self-resonant frequency either too high, or too low, and the impedance (i.e. Q factor is high) is sharp. Good luck finding a low-ESR type. I have a prototype amp running that uses a metal polypropylene type, 0.47uF cap, switching frequency is around 450 kHz. I'm inclined to think that that cap just doesn't look like 0.47uF at that frequency - in fact it is probably an order of magnitude lower, perhaps. Also the output impedance of the amp is going to be much higher than I expected. This is going to make speaker impedance an issue. My prototype doesn't have feedback at the speaker cocnector which should help drop the output impedance - but still, this is a bad starting point.

I just happened upon these Kemet MDK parts (http://www.evoxrifa.com/smd_catalog/dil_caps/mdk.pdf). Interesting data sheet - I recommend taking a look. They spec the ESR at 500 kHz, claim no voltage coefficient nor dissipation factor. Tempco is low, looks like around 1% from 0-70C.

These parts look outstanding, but I can't find any distribution.

[Pafi]

it is probably an order of magnitude lower, perhaps.

Why do you think this?

Thru-hole capacitors all have a significant ESL, simply because they have too long leg. You can do several things to reduce this effect. But even if you have the best cap, a wrong PCB layout can ruin everything!

the output impedance of the amp is going to be much higher than I expected

For example almost infinite at resonance frequency? This is natural. Zout=1/(1/(j*omega*L+ESR) + j*omega*C). Or is it higher then this?

They spec the ESR at 500 kHz, claim no voltage coefficient nor dissipation factor.

DF can be calculated from ESR. DF=ESR*omega*C. It's quite high, but this is not a big problem in an audio filter.

[gearheadgene]

Sorry for the delay, but I've been doing a lot of thinking about this.

Let me start with the output impedance. You need the LC filter to pass the audio while removing the switching information. It's really enlightening to look at the filter response, AND it's impedance looking back into the amp.

NOTE: I like laPlace notation, so s = j*omega

Neglecting parasitics, the filter response is
1 / (S^2 * L * C + 1)

This transfer function is 2nd order pole at s = sqrt (1/LC), with a 40 dB per decade roll off. That's the filter.

But, the impedance looking into the amp is much different - basically the parallel combo of the L and C. First, neglecting any parasitics,

Zin = L || C
Zin = [ s * L ] / [ s^2 * L * C + 1]
There's a zero at s=0, 2nd order pole at s = sqrt(1/LC).

Which means, the impdeance starts at 0-Ohms at DC, and increases to a peak at the resonant frequency, then falls back to zero at infinite frequency. You can run this in LTSpice and see it. I used some reasonable L and C values, and see that the output impedance rises from near zero, to several Ohms by 20 kHz. Bummer, since it's really important for the amp to have low output Z, maybe under 0.5 Ohm throughout the audio spectrum.

Now, try adding some parasitics, for example adding ESL to the cap.
Let L -> L1
C -> C + L2 where L2 is ESL of the cap.
Zin = [ (s^2 * L2 * C + 1)*(s * L1) ] / [ s^2 * C * (L1 + L2) + 1]
If you plot this, or simulate in spice, the output impedance has the same properties as before PLUS an additional 2nd order zero. The output impedance actually approaches infinite at high frequency. Maybe this is not a problem for audio, but it could be for EMI and ESD. So, this seems to imply the need for an additional high frequency pole, above the switching frequency, using a very low ESL cap (like perhaps 0.047 uF, small body smt). At least that pushes the pole way, way up in frequency where the energy is hopefully very very small.

I'm afraid I don't see a way around the rising output impedance throughout the audio band with this topology. You'd need a low frequency pole to negate the low frequency zero, but also not alter roll-off / filter affect required to remove the switching energy. I'm not a filter expert, and so far have come up with no ideas. I do think, however, that if the filter is part of the amplifier feedback loop, then the output impdance will be reduced, hopefully. I still need to verify that.

Anyway, more parasitics, more poles and zeroes, more complicated transfer functions. But you get the point.

[gearheadgene]

Regarding capacitor choice, you are right about voltage depedency and also piezo effect.

Here's some usefull comparizon info:
http://www.wima.com/EN/characteristics.htm

Wima claims to have low ESL
http://www.wima.com/EN/selfinductance.htm

But, if you go through the math you will find the parts have ESL on the order of 40 nH and some had around 100 nH.

Here's a very interesting SMT cap from AVX that looks promising:
http://www.avx.com/docs/Catalogs/cb-petht.pdf

regards

gene

[Pafi]

gene!

to several Ohms by 20 kHz

In a purely reactant (no resistor) network there are only two kind of limit value: zero, and infinite. Put the exact resonance freq to the equation!

You can lower this impedance with feedback significantly. Of course, when you do this, knowing of parasitic effects is important!

But, if you go through the math you will find the parts have ESL on the order of 40 nH and some had around 100 nH.

Why do you think this? I measured 20 nH with uncut legs on a 4,7uF 250V Wima capacitor.

Here's a very interesting SMT cap from AVX that looks promising:

Yes, it's good indeed, I've been using it already (1 uF 100 V):
http://users.hszk.bme.hu/~sp215/PWM_1kW/1.jpg
but unfortunately I haven't find any public source for them. I get some of them from a friend, but I'm going to run out of them soon. Have you find any source?

[gearheadgene]

I kept the resistor out of the math just to simplify things a little. My point is that the impedance starts out at F=0, with a value of R (zero in my example) and rises at 20dB per decade. So, the rise in output impedance is unavoidable - and unwanted. Yes, at resonant frequency, impedance is quite high, and very peaky with large values of R (infinite in my ex
ample).


Why do I think the ESL is 40nH or maybe 100nH? Hey, you know what? I made a math error

w = 1 / sqrt(LC) or
w^2 = 1/LC or
L = 1/ [C* w^2]
For example, the parts here:
http://www.wima.com/EN/WIMA_MKS_02.pdf
see graph Z vs. F,
1uF cap resonant at 2 MHz , so
L = 1 / (1e-6)(6.28 * 2e6)^2
L = 6.3 nH

So I was wrong - happily! Thanks for calling me on that. If you measured 20 nH, maybe that's a tad high or maybe just a different package configuration - but it is still pretty good.

Digikey carries some of the AVX part, for example CB052E0105JBC, CAP FILM 1.0UF 100VDC 5% 2824 ($1.80 each,min qty = 1) . I thought I saw a few on mouser. Lastly, maybe TTI - but those guys make you buy in bulk.

gene