On Wed, Jun 04, 2003 at 09:37:19PM +1000, Ian Wilson wrote:

Thanks for the well reasoned response Ian. 
I went through this a few months ago, but needed 0402 package
(was a dense board) and a similar line of research showed 100n 
was the best choice there.  But it depends on the particular 
caps chosen. 

Sun publishes some of the best SI reports in this area I have 
found particularly covering the concept of (maesured) low plane 
impedance over frequency. see for example
http://groups.yahoo.com/group/si-list/files/Signal%20Integrity%20Documents/Published%20SI%20Papers%20from%20Sun/

Minor quibbles with some of the details inline 



> On 07:14 PM 4/06/2003, John Sheahan said:
> >On Wed, Jun 04, 2003 at 09:34:18AM +0200, Norbert Hoppe wrote:
> >> When selecting parallel caps, it is important to remember that as the 
> >larger
> >> value capacitor goes inductive, the smaller value cap is still 
> >capacitive.
> >> At a particular frequency, a LC circuit is developed between the 2 caps.
> >> An infinite impedance could be generated with no decoupling benefit 
> >provided.
> >> When this occurs, single-capacitor decoupling is all that one can use
> >> for this application.
> >
> >actually - if you look at ESL graphs for multilayer SMD caps - you will
> >see it depends much more on case size than on capacitance.
> >So the 100n tends to win.
> >This quote I think may be  older wisdom for thruhole components.
> >john
> 
> The first resonance (at least) of a cap is series, so looks like a short 
> circuit. By adding a number of different valued caps you can scatter a 
> number of these nice AC shorts around your board and around your 
> frequencies of interest.

unfortunatley, these resonant circuits can sometimes clobber each other when 
there are different values involved.  But yes - that series resonance
helps a lot.

> 
> Above resonance the reactive impedance starts to rise as the impedance 
> characteristic is now inductive.  In many cases this is not an issue, as 
> the effective reactance is still low in the frequencies of interest.  In 
> other situations, though, it is a critical issue and hence designers have 
> used, and will continue to use, a variety of values in parallel - very 
> common in RF environments.
> 

sure, we want nice low ESL and ESR values here. 
The capacitance value (decoupler value) is really 
useful only below the series resonance point, which 
also may matter. And that is a big chunk of spectrun from the bulk cap
and maybe PSU up to say 100KHz region, thru to the 100M range.


> However, big small caps, or is that small big caps, you know ... large 
> capacitance in small volume, have pretty cruddy material, X7R if you are 
> lucky or Z5U if capacitance is big.  These materials have pretty poor, and 
> frequency dependent, ESR which decrease their value as decouplers.  Due 
> largely to the effects of the lossy material, a Kemet, for example, X7R 
> shows a sloppy self resonance and the following series impedance at 100 MHz:
> 
> Value   Size  Impedance
> 103      0603    ~1 ohm
> 103     0805    ~0.5 ohm
> 103     1206    ~0.3 ohm
> 104     0805    ~1 ohm
> 104     1206    ~1 ohm
> 
> So if you spec a 10nF 0603 you have a resistor, not a decoupler, at 100 
> MHz. 

probably only a NPO lower than 1n and the interplane capacitance is
working here.  The 100n/10n argument is null as we are essentially
at/past series resonance.   Ovviously here we have an L or an R , not
a C. But we do have a low value L or R so we are still decoupling,
- the point of the dcoupler is to attenuate noise. C is just the way
we often tend to do this.
A sprinkling of 10-100pF value caps appeal to some in the >100Mhz range. 

> According to Kemet, the 0603 only performs better than the other 
> sizes at over the narrow freq range of about 10 to 30 MHz.  The lossy 
> dielectric, and the need to use thinner metal in the large capacitances (to 
> keep the pkg the same), is killing the performance.  Maybe that is an 
> overstatement - but compare the self resonance curves of a COG/NPO material 
> to that of a X7R, or worse Z5U,  you can see the dramatic effect the losses 
> have on the resonance shape.  COG/NPO has dissipation factors in the order 
> or 0.1% while the other materials are between about 2.5 to 5% or more.  (In 
> fact, the lower Q of the high capaciatnce devices is partially a good 
> thing.  Having high Q resonances around a board is a shocker when you find 
> you fail EMC.)
> 

agree

> Note also that the 100n 0805 has roughly *twice* the impedance @ 100 MHz 
> than the cap *one tenth* the value in the same pkg!  In this case, if you 
> are operating above 30 MHz the 1206 10n wins, followed closely by the 0805 
> 10n. 100n in any pkg and 10n in 0603 have about twice the impedance.


here you would do better with 1n NPO however..

> 
> See figures 4, 5 & 6 of:
> http://www.kemet.com/kemet/web/homepage/kechome.nsf/vapubfiles/F3102Gce/$file/F3102GCe.pdf
> 
> There is always progress in material science so the small-packaged, larger 
> capacitance devices get better over time.
> 
> For modern high speed decoupling - I use lots of 10n devices, a few bulk 
> devices and good (hopefully) layer stackup and split plane 
> arrangement.  Specific devices operating at speed will have special 
> treatment.  Currently, I side with the get to the plane fast crowd and have 
> my supply via close to the power pads and then decoupling caps strung to 
> these same vias with nice fat tracks.  I don't, usually, have a via, track, 
> cap, then component pad arrangement - though my guess is, with the right 
> sort of component selection either arrangement can be done well.
> 

I'd perhaps also suggest power plane as near to the component side as 
practical, as
via inductance is more significant than ESL in things like 0603. And
prefer the decoupler on the component side (appropriate component that
is) where this can be done close enough. 
I'd side with the stackups that have power plane and ground plane
together,  rather than every-other as has been mentioned.

> Ian Wilson
> 
> 


* * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
* To post a message: mailto:[EMAIL PROTECTED]
*
* To leave this list visit:
* http://www.techservinc.com/protelusers/leave.html
*
* Contact the list manager:
* mailto:[EMAIL PROTECTED]
*
* Forum Guidelines Rules:
* http://www.techservinc.com/protelusers/forumrules.html
*
* Browse or Search previous postings:
* http://www.mail-archive.com/[EMAIL PROTECTED]
* * * * * * * * * * * * * * * * * * * * * * * * * * * * * *

Reply via email to