[SI-LIST] Re: Antwort: Re: Questions about interplane capacitance
- From: Istvan Novak <istvan.novak@xxxxxxxxxxx>
- To: Joel Brown <joel@xxxxxxxxxx>
- Date: Fri, 14 Mar 2008 08:57:57 -0400
Joel,
Imagine the following. You establish your impedance target. For sake
of simplicity lets
assume you have a single load (a single pin) to feed with clean power.
We know from
previous analysis that minimizing the worst-case transient peak-to-peak
noise can be
done by setting the impedance target to be flat within the frequency
range of interest.
For sake of simplicity lets assume the frequency range of interest is
from DC to Fmax.
On a circuit schematics your ideal PDN solution is a series R-L source
impedance,
where R is your impedance target, and L is such that Fmax=1/(2*PI*L/R).
If the
PDN design is done properly, the load-current fluctuations will create
only acceptably
small voltage noise on the supply rail, in other words the load
impedance is much greater
than R, so your source operates in the 'unloaded' mode.
To turn the ideal schematics into an actual circuit, you can choose from
a large number of
methods to synthesize the source impedance. One attractive option is to
take a transmission
line of characteristic impedance of R, match it at both ends, and to
connect one end to
the DC source, the other end to your load. The required PDN impedance
(R) is usually
way lower than 50 ohms, so this is not your typical coax cable or signal
trace, but the
physics is the same regardless of the characteristic impedance: a
terminated transmission
line shows its characteristic impedance regardless of its length and
frequency. For PDN
analysis, the characteristic impedance can be considered constant even
if we have losses
and dispersion. The transmission line is a distributed R-L-G-C network,
where even
if we neglect R and G, the signal propagates through the series of
capacitive and inductive
contributors. The approximate characteristic impedance is sqrt(L/C) and
the approximate
propagation delay is sqrt(L*C). If you take this terminated
transmission line, it will show
R/2 impedance at the load point, regardless of its length and delay. If
the load current
fluctuates (within the bandwidth of Fmax), the voltage fluctuation will
be R/2*I(t).
(note: when you consider a one-dimensional transmission line, matched at
both ends,
your actual impedance shown to the load is R/2, unless you use source
termination only,
which gives you an impedance of R).
In the above circuit, the charge flows continuously from source to load,
and delay does
not matter.
The key to the above simplistic picture is the matching: if we consider
bypass capacitors,
those must have resistances properly matching the plane structure's
characteristic impedance.
I call them Bypass Resistors, because it is really their resistance what
matters: C and L
are less important. C is there only to make sure we do not draw DC
current with the
termination (so it should be as high value as possible), and L is there
as physical reality, so
it should be as low as conveniently possible.
Regards,
Istvan Novak
SUN Microsystems
Joel Brown wrote:
> I hate to prolong this but...
> When I think about a transmission line in the normal sense (signal
> propagation), A driver switches at one end but initially all it sees is Zo
> which looks like a resistor maybe 50 ohms and it does not even see the load
> until wave propagates the length of the line. So there is a time delay. Now
> think of the driver being the power pin of the IC and the load being the
> bypass cap on the corner of the board or visa versa and I see a propagation
> delay, so what am I missing?
>
> Joel
>
>
> -----Original Message-----
> From: SILR [mailto:silr@xxxxxxxxxxxx]
> Sent: Thursday, March 13, 2008 6:27 PM
> To: 'steve weir'; 'Joel Brown'
> Cc: 'Istvan Novak'; si-list@xxxxxxxxxxxxx
> Subject: RE: [SI-LIST] Re: Antwort: Re: Questions about interplane
> capacitance
>
> Joel asked:
> <<< ...why is charge propagation velocity not a factor when the PDN is
> purely resistive? >>>
>
>
> Well, here's a crazy way to think about this (which I'm sure not everyone
> will agree)... I guess this would be like how Eric B. might put it... "Be
> the Signal"
>
>
> I'm the charge on some corner of a board...there's an IC in the middle of
> the board...
>
> I have to get to that IC using this transmission line called the PDN. If
> this Transmission Line had the classical Ls and Cs (and Rs) to describe its
> characteristics, then that means I have to deal with changing EM fields to
> get to that IC... but these time varying EM fields always cause some delay
> in my charge getting to that IC in a timely manner...
>
> BUT!!! ...if this Transmission Line had nothing but Rs (no Ls and Cs) to
> describe its characteristics then that means I don't have to deal with time
> varying EM fields any longer... I just have to deal with resistive losses
> only but this only equates to a drop in Static Voltage Potential and nothing
> more... (so keep the Z low!!!) And I can get there (to the IC) in no time
> and the IC gets the charges that it needs and it wouldn't even know anything
> is different...
>
> Again, this is a crazy way to look at this... but it works for me, for
> now... until I get grilled by the senior members... :-)
>
> Silvester
>
> -----Original Message-----
> From: si-list-bounce@xxxxxxxxxxxxx [mailto:si-list-bounce@xxxxxxxxxxxxx] On
> Behalf Of steve weir
> Sent: Thursday, March 13, 2008 5:57 PM
> To: Joel Brown
> Cc: 'Istvan Novak'; si-list@xxxxxxxxxxxxx
> Subject: [SI-LIST] Re: Antwort: Re: Questions about interplane capacitance
>
> Joel, it is transmission line theory. Think about an infinitely long
> signal transmission line for a moment. At any instant a driver sees a
> constant impedance across frequency. The voltage to current relation is
> independent of prior history. What Istvan describes is a very low impedance
> transmission structure for power.
>
> Sadly, a lot of IC vendors are still playing catch-up in terms of power
> delivery. If they had their game together, they would be telling you:
> The actual voltage tolerances at the die, the current spectrum at the die,
> and the parasitics of the die and package. From that you could engineer
> your PDN. Instead often what we see are partial recipes like you describe.
> On a good day the recipes cost extra money. On a bad day, they result in
> failures.
>
> One can build a resistive networks several ways. One is to add discrete
> resistance by any number of methods, another is to use the FDTIM method that
> Larry Smith has long championed. And even though resistive networks have
> attractive qualities, reactive networks may still be cheaper and equally
> effective. It all depends on the circumstances. It is somewhat akin to
> surface finishes: there is no ideal one. But there are several that when
> used properly work very well. One just needs to respect the limitations of
> each method.
>
> Part of the problem figuring this stuff out is having sufficient experience
> to judge trade-offs early in the design process. That's just something that
> has to be learned. As more training materials come out on power delivery,
> it will probably get easier. In the shameless plug department, we (
> Teraspeed ) do very nice jobs optimizing power delivery systems for
> customers. If you've got a design you want advice on we can get you
> squared-up pretty quickly.
>
> Best Regards,
>
>
> Steve.
> Joel Brown wrote:
>
>> Steve,
>>
>> So even though each cap has a relatively high ESR (1.6 Ohms) the PDN
>> as a whole has a relatively low impedance which will result in low
>> noise on the PDN. This goes against intuition and previous thinking
>> that an IC needs a local bypass with low ESR and ESL to supply the
>> needed charge during switching transients. I am starting to see that
>> mathematically a resistive PDN lowers noise compared to one that is
>> the inductive (you did a good job explaining that). The thing that I
>> am having trouble grasping or
>>
> visualizing
>
>> is that why is charge propagation velocity not a factor when the PDN
>> is purely resistive? Is the PDN model simply a resistor in series with
>> the
>>
> load
>
>> and distance has no effect? Perhaps there is an analogy that would
>> make
>>
> this
>
>> concept easier to understand? Also, when I see app notes from IC
>> vendors that recommend using a 0.1uF and 1000pF cap on each supply pin
>> and then instead I use distributed high ESR capacitors I feel like I
>> am doing something quite different and contrary from the recommended
>> and when I
>>
> query
>
>> the vendors on how they arrived at recommendations in the app note the
>> answer I get is "we recommend that you do it exactly as shown in the
>> app note, we know it works that way and if you don't follow it it may
>> not
>>
> work".
>
>> Also I am wondering how all this relates to X2Y caps, I suppose they
>> could be used with series resistors but that would somewhat defeat the
>> purpose
>>
> by
>
>> adding inductance. Its not clear to me what approaches I should
>> attempt on future designs.
>>
>> Joel
>>
>>
>> -----Original Message-----
>> From: si-list-bounce@xxxxxxxxxxxxx
>> [mailto:si-list-bounce@xxxxxxxxxxxxx]
>>
> On
>
>> Behalf Of steve weir
>> Sent: Wednesday, March 12, 2008 11:26 PM
>> To: Doug Brooks
>> Cc: Istvan Novak; si-list@xxxxxxxxxxxxx
>> Subject: [SI-LIST] Re: Antwort: Re: Questions about interplane
>> capacitance
>>
>> Doug, Istvan's representations are analytically exact. When the
>> characteristic impedance of the transmission structure is high, and/or
>> the rise times are slow then capacitors can be placed in close enough
>>
> proximity
>
>> that they load the transmission structure so as to make it appear a
>> lower impedance with some little bumps. For example if one had a 10"
>> x 10" 4
>>
> mil
>
>> er4.0 board = 22.5nF and loaded it with bypass every square inch of
>> 150pH, then for slow enough signals, the distributed impedance would
>> look like 80mOhms. If the network were constructed from 200
>> capacitors, an ESR
>>
> value
>
>> of 1.6Ohms / cap would make that impedance uniform down to the RC knee
>> of the parts. Assume that were matched by an AVP regulator of 80mOhms
>> and
>>
> the
>
>> entire thing looks like 80mOhms from DC to over 1GHz. 80mOhms would
>>
> support
>
>> a 32 bit transition w/ about 5% droop. The impedance scales with
>>
> dielectric
>
>> height and the inverse square root of eR. Scale the dielectric down
>> to 0.1mils and now 320 lines can switch simultaneously in one
>> direction with
>>
> 5%
>
>> droop, and at arbitrary edge rates.
>>
>> Istvan has shown using analysis with the reverse pulse technique an
>> inductive PDN shunt impedance acts like a noise high pass filter (
>> See DC papers from at least as far back as DC East 2005 ). Put in a
>> square wave noise pulse ( load current ) and the leading edge changes
>> by Vdelta = -L*di/dt below the baseline. Allow the pulse to persist
>> long enough and
>>
> the
>
>> system recovers back to the baseline which would be -I*Rpdn.
>> Return the load current to zero, and now the energy stored in the
>>
> inductance
>
>> kicks back -L*di/dt. The p-p noise is then 2*Ldi/dt - (Imax-Imin)*Rpdn.
>>
> If
>
>> Rpdn is very small then it approximates 2*Ldi/dt.
>> This behavior is apparent in the transient response plots of virtually
>> any non-AVP VRM.
>>
>> Now, suppose that the VRM and PDN can be made to appear resistive
>> right through Fknee. Then the response to a current pulse of I is
>> simply Vdelta
>>
> =
>
>> -I*Rpdn. There is no component of di/dt, and so the total p-p noise
>> is
>>
> just
>
>> (Imax - Imin)*Rpdn. AVP schemes position the DC operating point
>> intentionally high so that at Imin they are at their margined high
>> limits and at Imax they are at their low limits. This allows
>> increasing Rpdn
>>
> while
>
>> still meeting the same current and voltage specifications.
>>
>> Best Regards,
>>
>>
>> Steve.
>>
>>
>> Doug Brooks wrote:
>>
>>
>>> Istvan,
>>>
>>> With all due respect, I would modify your argument a little bit. In
>>> very simplistic terms, suppose we need x amount of charge to
>>> transition from a zero to a one in one ns. That amount of charge (I
>>> suggest) must be within
>>> 6 inches of the need (what I think we are referring to as the service
>>> radius). If not, it takes a little longer to reach the logical one state.
>>> I look at it, not from the standpoint of a dip in the rail, as much
>>> as the ability to satisfy the rise time requirement (unless you are
>>> referring to a dip in the rail that occurs during the rise time
>>> itself.) In the slightly longer term, the charge will replenish
>>> fairly quickly, but not, perhaps fast enough to meet the rise time
>>>
> requirement.
>
>>> Doug Brooks
>>>
>>>
>>>
>>>
>>>
>>>
>>>> Andreas,
>>>>
>>>> Yes and no. It is true that charge moves with finite speed, so for
>>>> any given time duration the charge has to come from locations closer
>>>> than the ratio of distance over speed. BUT the whole notion of
>>>> service radius is based on the assumption that as you deplete the
>>>> charge available in the immediate vicinity of the active device, you
>>>> have to wait for replenishment, otherwise you get a big dip on the
>>>> supply rail.
>>>>
>>>> Having a matched
>>>> transmission medium to deliver power to the active device, the
>>>> charge moves without interruption, and as you deplete the planes
>>>> close to the device, it gets replenished on the fly from areas
>>>> further away, so the service area concept is pretty much meaningless
>>>> in this scenario. Current flows without interruption.
>>>> The bucket brigade of infinitesimally small inductive and capacitive
>>>> elements of the transmission line transmits the power continuously.
>>>> If the load current changes, for any I(t) time function of load
>>>> current, the transient noise at the load point will be I(t)*Zo,
>>>> where we assume that Zo is the resistive and frequency independent
>>>> characteristic impedance of the transmission medium. This is a very
>>>> simplistic one-dimensional model, but it gives a good insight of why
>>>> the service radius matters only on PDNs where the network is not
>>>> matched.
>>>>
>>>> Regards,
>>>>
>>>> Istvan Novak
>>>> SUN Microsystems
>>>>
>>>>
>>>>
>>>>
>>>>
>>>> Andreas.Lenkisch@xxxxxxxxxx wrote:
>>>>
>>>>
>>>>
>>>>> Istvan,
>>>>> I'm wodering a little about your comments to the service radius.
>>>>> Independant if the impedance is resistive, we have still a
>>>>> propagation time which would limit the service radius from my
>>>>>
>>>>>
>> understanding.
>>
>>
>>>>> Do I'm wrong?
>>>>>
>>>>> regards
>>>>> Andreas
>>>>>
>>>>>
>>>>>
>>>>> Istvan Novak <istvan.novak@xxxxxxxxxxx> Gesendet von:
>>>>> si-list-bounce@xxxxxxxxxxxxx
>>>>> 11.03.2008 13:14
>>>>>
>>>>> An
>>>>> Joel Brown <joel@xxxxxxxxxx>
>>>>> Kopie
>>>>> si-list@xxxxxxxxxxxxx
>>>>> Thema
>>>>> [SI-LIST] Re: Questions about interplane capacitance
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> Joel,
>>>>>
>>>>> Just one quick comments to the good summary from Steve:
>>>>>
>>>>> While considering planes and bypass capacitors in terms of
>>>>> effective capacitances and inductances is a valid approach, we need
>>>>> to keep in mind that focusing on the capacitive or inductive nature
>>>>> of parts without looking at the wider picture misses a very
>>>>> important and useful class of solutions, namely that of matched
>>>>> transmission lines. As it was pointed out earlier several times on
>>>>> the SI list, the best
>>>>> (self) impedance for a power distribution network is a resistive
>>>>> one, neither capacitive, nor inductive.
>>>>> We can get resistive impedance from a matched transmission line,
>>>>> regardless of its capacitance and inductance, and in such cases the
>>>>> notion of 'service area' of parts become meaningless: you can put
>>>>> bypass components further away from the active devices without
>>>>> sacrificing performance.
>>>>>
>>>>> Regards,
>>>>>
>>>>> Istvan Novak
>>>>> SUN Microsystems
>>>>>
>>>>> Joel Brown wrote:
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> Interplane capacitance is frequently cited as the only effective
>>>>>> bypass capacitance on a PCB at frequencies above 200 MHz.
>>>>>> I am currently working on a design which brings up some questions
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> regarding
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> interplane capacitance.
>>>>>>
>>>>>> 1. Power planes normally carry "standard" voltage rails that are
>>>>>> used throughout a board such as +5V and +3.3V.
>>>>>> High speed ICs usually have core voltages that are local to the IC
>>>>>> and
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> are
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> provided by a local regulator which converts the standard rail to
>>>>>> the
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> core
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> voltage (example 3.3 to 1.8V).
>>>>>> The local core voltage is distributed on a plane area that is
>>>>>> local to
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> the
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> IC and therefore is small in area (0.25 sq in or less) which
>>>>>> results in
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> a
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> very small amount of interplane capacitance.
>>>>>> Is this very small amount of capicitance effective for bypassing
>>>>>> the IC?
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> I
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> am sure it depends somewhat on the current waveform being drawn by
>>>>>> the
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> IC
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> but this can only be estimated because semiconductor manufacturers
>>>>>> do
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> not
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> provide current consumption profile as a function of frequency. To
>>>>>> make matters worse, some ICs have several different VCC pins which
>>>>>> the manufacturer recommends connecting to separate networks of
>>>>>> bypass caps
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> and
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> ferrite beads. This cuts the power distributuion up even more
>>>>>> resulting
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> in
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> less (practically zero) interplane capacitance. It is somewhat
>>>>>> ironic
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> that
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> the the voltages such as +5V and +3.3V which are required at
>>>>>> points
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> across
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> the whole board and therefore have the most interplane capacitance
>>>>>> are
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> also
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> the voltages which have least requirement for interplane
>>>>>> capacitance
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> because
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> they do not directly supply high speed rails.
>>>>>>
>>>>>> 2. There has been a lot of emphasis on reducing the mounted
>>>>>> inductance
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> of
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> bypass capacitors. Even with this reduced inductance they are
>>>>>> still only effective up to several hundereds of MHz at which point
>>>>>> the interplane capacitance becomes the only bypass capacitance
>>>>>> mechanism. However there
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> is
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> inductance between the connection of the IC to the planes. This
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> inductance
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> consists of vias and package inductance. I did look for some
>>>>>> numbers for package inductance and did not find much, it seems to
>>>>>> be a closely held secret. Also it is unknown how much bypass
>>>>>> capacitnace is internal to
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> the IC
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> package. Just for example if we assume 250pH for the vias and 500
>>>>>> pH for
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> the
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> package, then the impedance at 500 MHz would be 2.36 Ohms. This
>>>>>> seems
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> rather
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> high for the interplane capacitance to be of much benefit.
>>>>>>
>>>>>> In summary how much interplane capacitance is needed to be
>>>>>> beneficial,
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>> and
>>>>>
>>>>>
>>>>>
>>>>>
>>>>>> why is it beneficial given the inductance in the vias and package?
>>>>>>
>>>>>> Thanks - Joel
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
------------------------------------------------------------------
To unsubscribe from si-list:
si-list-request@xxxxxxxxxxxxx with 'unsubscribe' in the Subject field
or to administer your membership from a web page, go to:
//www.freelists.org/webpage/si-list
For help:
si-list-request@xxxxxxxxxxxxx with 'help' in the Subject field
List technical documents are available at:
http://www.si-list.net
List archives are viewable at:
//www.freelists.org/archives/si-list
or at our remote archives:
http://groups.yahoo.com/group/si-list/messages
Old (prior to June 6, 2001) list archives are viewable at:
http://www.qsl.net/wb6tpu
Other related posts:
