[SI-LIST] Re: Decoupling capacitors
- From: steve weir <weirsi@xxxxxxxxxx>
- To: Stefan Ludwig <ludwig@xxxxxxxxxxxxxxxxx>
- Date: Tue, 17 May 2005 09:37:45 -0700
Stefan exactly! The oppressively competitive computer market has forced
Intel to design for total system cost. It's a lesson that many other chip
builders need to learn well if they expect to succeed in a market that is
becoming more consumer electronic, sic low total cost, oriented.
Regards,
Steve.
At 06:27 PM 5/17/2005 +0200, Stefan Ludwig wrote:
>A slightly different angle on this is the following:
>
>Intel's 90nm processors consume 130W these days (desktop and server
>parts). That can be more than 100 Amps! You simply can't afford a "cheap",
>"un-bypassed" package with this sort of power and current requirements.
>Once you bypass to the max on the die itself and in the package, you end
>up with a lot less stringent requirements on the PCB.
>
>A nice side effect of the above and a requirement from Intel's customers
>is that their processors need to run with the cheapest possible boards
>produced by the customers' contract manufacturers. Intel must engineer a
>rock solid chip/package solution, which then allows their customers to
>have a simpler board structure and cheaper bypassing. For a while, Intel
>delivered their processors in the form of cartridges, so that even the
>least capable customer can build a PC board that still works. It had the
>processor die and cache SRAMs on it, as well as lots of bypass caps.
>
>Nowadays, they sell you boxed chips, with the heat sink and cooling fan,
>so you don't blow up "your precious!"
>
>Stefan
>
>"May you live in interesting times!" - Chinese curse (<- NOT, as it turns
>out if you google for it.)
>
>steve weir wrote:
>
>>Ken,
>>
>>Thanks.
>>
>>It is precisely because Lindenhurst has a relatively high spreading
>>inductance for power that I think people need to pay close attention to
>>it. Carpet bombing that sort of board with bypass caps is futile, and
>>because of the amount of bypass included in the processors,
>>unnecessary. Now, you may have other chips on your board that impose
>>that cost of extra and thin layers close to the IC mounting surface, but
>>as Lindenhurst demonstrates, with proper engineering of the chip /
>>package, that expense is not mandatory.
>>
>>Regards,
>>
>>Steve.
>>
>>At 09:25 AM 5/17/2005 -0600, Ken Cantrell wrote:
>>
>>
>>>Steve,
>>>"do you develop it with geographically specific transfer impedances to each
>>>IC
>>>attachment or do you globalize?" - geographically specific transfer
>>>impedances.
>>>"Do you develop an independent impedance versus frequency for each big chip,
>>>or do you use the worst case device?" - independent impedance versus
>>>frequency.
>>>However, I sometimes globalize in the sense that I place some of the
>>>capacitance(mid-to-high)near the chips and some of the capacitance in an
>>>inter-chip grid,being careful about lambda/4 placement. It just depends on
>>>the situation.
>>>
>>>I had looked at the Lindenhurst material some time ago, and frankly, didn't
>>>get as much out of it as you. I thought that the basic stack-up was OK, but
>>>could be improved on for my situation. I prefer to have tight power/ground
>>>coupling, as close to the chip as possible, even if it means a two layer
>>>hit. But that's the nice thing about the list, exposure to other view
>>>points; someone else seeing something that you missed. I will review.
>>>
>>>Thanks,
>>>Ken
>>>
>>>
>>>-----Original Message-----
>>>From: steve weir [mailto:weirsi@xxxxxxxxxx]
>>>Sent: Monday, May 16, 2005 5:34 PM
>>>To: Ken Cantrell; Ken Cantrell; Larry.Smith@xxxxxxx
>>>Cc: joepaul@xxxxxxxxxxx; si-list@xxxxxxxxxxxxx
>>>Subject: RE: [SI-LIST] Re: Decoupling capacitors
>>>
>>>
>>>Ken, if Larry's F^N method meets your specifications then you are all
>>>set. There is no question that it does what is claimed which is yield a
>>>relatively flat impedance profile with a higher ESR than a big "V" with the
>>>same reject band end-point insertion loss.
>>>
>>>The real key in all of this is as Istvan noted the specifications. It also
>>>turns-out that the inductance in both the components and the mounting
>>>structure remains the enemy. Even more so for "capacitors by the decade"
>>>than the other two methods. This raises an interesting issue, because if
>>>one wants to drive down the real-estate as a primary goal, the two
>>>available choices are IDC and X2Y
>>>
>>>If you don't mind my asking, when you develop you specification, do you
>>>develop it with geographically specific transfer impedances to each IC
>>>attachment or do you globalize? Do you develop an independent impedance
>>>versus frequency for each big chip, or do you use the worst case device?
>>>
>>>Lindenhurst is the current example that I am referring to. Search Intel
>>>for DDR2 and it should come up. If you can't find it, I will send you a
>>>copy. Lindenhurst is a great example of how good engineering was used to
>>>yield high quality, and low-cost:
>>>
>>>8 layers, 2 power, 2 logic common and 4 signal. All signal layers strongly
>>>reference logic common. Cavity resonance and issues of switching return
>>>layers are minimized. The two power layers are relatively far from the
>>>logic common layers, about 15 mils for a net effective of about 7mils, but
>>>with the nearest power layer to the IC over 25 mils down the stack.
>>>
>>>What, 7 mils effective. that's 35pH, or some 22mohms at 100MHz? How can
>>>that be? What happened to a milliohm at 100MHz? How can Intel be off by
>>>more than an order of magnitude and still make servers that run
>>>non-stop? Hasn't it been drilled into everyone's head that we are supposed
>>>to need multiple pairs of power / common with thin dielectric right up
>>>against the IC to meet these "modern requirements"?
>>>
>>>Well, those smart people in Santa Clara apparently missed the class where
>>>we were all "taught" these truisms. Instead they got the funny idea that
>>>big structures make for big HF losses and they should move the HF storage
>>>into the package and drop the definition of HF for power delivery, and
>>>therefore the package / internal storage cut-off frequency way, way
>>>down. The result is dramatic system savings over the ever more futile
>>>efforts to try and jam many amperes at high frequency through local board
>>>interconnect and into a package. The only way to save even more ( and it
>>>is MANY millions more each year ) money is to make the in-package
>>>capacitors X2Ys. I believe it was Scott in a comment to Chris who once
>>>noted that a well designed package does not demand much from the PCB.
>>>
>>>There is a lot to be learned from these very shrewd decisions by
>>>Intel. Alas for the other chips we buy, our other vendors may not be so
>>>forward thinking and impose on us expensive PCB level solutions.
>>>
>>>Regards,
>>>
>>>
>>>Steve
>>>
>>>
>>>At 04:27 PM 5/16/2005 -0600, Ken Cantrell wrote:
>>>
>>>
>>>>Steve,
>>>>I guess I'm partial to Larry's method because I'm in HPC and have "insanely
>>>>low PDS requirements" also. I could argue about "higher inductance with a
>>>>reduced count" between big V and fine multi-pole being as big of an issue
>>>>
>>>as
>>>
>>>
>>>>you are implying, but that is a topic in itself. I have found that a good
>>>>board layout eliminates most of the issues (whatever they are), just as
>>>>theory overestimates the severity of most problems. I will review your
>>>>material (didn't mean to make it seem so black and white, know it isn't),
>>>>and am looking forward to your paper as stated in my response to Istvan.
>>>>On the Intel 8 layer server stack-up, are you talking about the Lindenhurst
>>>>board stack-up? Do you have a pointer handy?
>>>>Thanks,
>>>>Ken
>>>>
>>>>-----Original Message-----
>>>>From: steve weir [mailto:weirsi@xxxxxxxxxx]
>>>>Sent: Monday, May 16, 2005 3:45 PM
>>>>To: Ken Cantrell; Larry.Smith@xxxxxxx
>>>>Cc: joepaul@xxxxxxxxxxx; si-list@xxxxxxxxxxxxx
>>>>Subject: RE: [SI-LIST] Re: Decoupling capacitors
>>>>
>>>>
>>>>Ken, please see also my comments to Larry. Please do review, because it is
>>>>not as nearly black and white as you think. The price for reduced
>>>>capacitor count is a higher inductance in the capacitor network. Do we
>>>>care? Well that all depends on how sharp and uniform a cut-off one
>>>>presumes for the IC packages. ( Assuming of course we are using bypass
>>>>capacitors to bypass power and not signals. ) If that cut-off is not sharp
>>>>or uniform, then at the end of the day equal inductance is what we seek and
>>>>the quantity advantage of the fine multipole evaporates. If one makes what
>>>>I consider a mistake, but is still done, of relying on the power bypass
>>>>network for signal return, then the bypass network is further enlisted to
>>>>support much higher frequencies. This is I feel very good at one
>>>>thing: making assemblies more expensive.
>>>>
>>>>In my highly opinionated view, the Intel 8 layer server stack-up should be
>>>>the basis of a required one or two week course for all SI and PI
>>>>engineers. Big "V" works great on those boards designed and built to very,
>>>>very demanding cost targets, while supporting behemoth processors and big
>>>>fast front-side busses. It is a shining example of using the principles
>>>>that experts like Chris have been driving home here for a long time.
>>>>
>>>>Big "V" is not without its limitations. Neither is fine multipole. In the
>>>>hands of an experienced engineer, each can be used effectively. Fine
>>>>multipole or what I prefer to call F^N was developed by smart people at Sun
>>>>to solve a problem and they succeeded. It just doesn't make it the only
>>>>viable solution.
>>>>
>>>>Regards,
>>>>
>>>>
>>>>Steve.
>>>>At 11:21 AM 5/16/2005 -0600, Ken Cantrell wrote:
>>>>
>>>>
>>>>>Gosh -
>>>>>"In many common cases, it is just a rote practice that provides no
>>>>>actual value. Life with SMT devices is much different today than when we
>>>>>had leaded devices and those practices were first adopted."
>>>>>I know this has been a long, ongoing issue, and I've read the majority of
>>>>>Steve's and Istvan's material. Maybe I need to review. However, if you
>>>>>take any Zo vs Frequency (100KHz to 100MHz)profile with a target
>>>>>
>>>impedance,
>>>
>>>
>>>>>and plot big V vs Larry's method, you always end up with fewer caps with
>>>>>Larry's method, and it's more like a 20% to 30% reduction in count over
>>>>>
>>>big
>>>
>>>
>>>>>V. And you don't have any abrupt impedance changes at the ends, just a
>>>>>
>>>>bunch
>>>>
>>>>
>>>>>of little cycloids that are near or at the target impedance line. At
>>>>>
>>>>least
>>>>
>>>>
>>>>>that's what my handy-dandy copy of D. Brooks' Bypass Impedance Calculator
>>>>>shows.
>>>>>The real issue, at least for me, is what Larry states below about low and
>>>>>high fequency anti-resonances. How do you, Steve/Howard/Istvan, address
>>>>>this?
>>>>>
>>>>>Not a big fan of big V,
>>>>>
>>>>>Ken
>>>>>
>>>>>
>>>>>-----Original Message-----
>>>>>From: si-list-bounce@xxxxxxxxxxxxx
>>>>>[mailto:si-list-bounce@xxxxxxxxxxxxx]On Behalf Of Larry SMITH
>>>>>Sent: Monday, May 16, 2005 10:14 AM
>>>>>To: weirsi@xxxxxxxxxx
>>>>>Cc: joepaul@xxxxxxxxxxx; si-list@xxxxxxxxxxxxx
>>>>>Subject: [SI-LIST] Re: Decoupling capacitors
>>>>>
>>>>>
>>>>>Steve - The "big V" approach makes a very low impedance at the series
>>>>>resonant frequency of the capacitor chosen. With this method, problems
>>>>>can arise about a decade above and below the SRF where the decoupling
>>>>>capacitors have to interface with other components in the system.
>>>>>
>>>>>On the low frequency side, the output of the VRM is inductive. It's
>>>>>impedance crosses the capacitive impedance of all the big V caps in
>>>>>parallel and usually makes a fairly high Q resonant peak in the kHz or
>>>>>low MHz region.
>>>>>
>>>>>On the other end of the spectrum, the inductance of all the caps in
>>>>>parallel usually resonates with the capacitance of the power planes and
>>>>>make a high Q resonant peak in the 100MHz region. These are the two
>>>>>"gotcha's" associated with the big V method. "Antiresonant" peaks at
>>>>>any frequecy can (and often does) cause trouble in power distribution
>>>>>systems.
>>>>>
>>>>>You can add additional capacitor values to address the newly created
>>>>>peaks, but then you end up with more peaks.. If you add enough
>>>>>capacitor values to eliminate all the peaks, you end up with a flat
>>>>>frequency profile. This is the one I like.
>>>>>
>>>>>regards,
>>>>>Larry Smith
>>>>>Sun Microsystems
>>>>>
>>>>>steve weir wrote:
>>>>>
>>>>>
>>>>>>Joe Paul, the ESL may be the same but the ESR of the 1uF in the same
>>>>>>chemistry and voltage will definitely be about 35-40% that of the
>>>>>>0.1uF. It is very unlikely that 0.1uF will provide any cost or
>>>>>>
>>>>>performance
>>>>>
>>>>>
>>>>>>advantage over 1uF in the same 0603 case from the same mfg at a low
>>>>>>
>>>>>voltage
>>>>>
>>>>>
>>>>>>rating. Due to cover layer considerations, depending on the voltage
>>>>>>
>>>and
>>>
>>>
>>>>>>chemistry, the 1uF may actually exhibit lower mounted inductance than
>>>>>>
>>>>the
>>>>
>>>>
>>>>>0.1uF
>>>>>
>>>>>
>>>>>>At 2ns/ 160MHz, both capacitors are fully inductive and cover layer
>>>>>>
>>>>issues
>>>>
>>>>
>>>>>>aside will have very similar performance. The 1.0uF capacitor has the
>>>>>>benefit of more capacitance which generally makes it easier to
>>>>>>
>>>stabilize
>>>
>>>
>>>>>>the transition from the bulk capacitors / VRM. Fans of the big "V"
>>>>>>
>>>like
>>>
>>>
>>>>>>Dr. Johnson, Istvan Novak, and myself will usually advise that at the
>>>>>>
>>>>>same:
>>>>>
>>>>>
>>>>>>cost, package size and chemistry, take the bigger capacitor. If you
>>>>>>
>>>>want
>>>>
>>>>
>>>>>>to find out why some people do things differently, take a look at
>>>>>>
>>>Larry
>>>
>>>
>>>>>>Smith and company's papers on multipole capacitor networks. If
>>>>>>
>>>nothing
>>>
>>>
>>>>>>else, those papers should help you better understand what you are
>>>>>>
>>>doing
>>>
>>>
>>>>>>whether or not you elect to follow the methods they describe.
>>>>>>
>>>>>>If you really want to see how your capacitors perform and have access
>>>>>>
>>>to
>>>
>>>
>>>>a
>>>>
>>>>
>>>>>>VNA, I suggest building a test board. You can get details on such a
>>>>>>
>>>>board
>>>>
>>>>
>>>>>>from Istvan's papers on his web site, or from mine on the X2Y web
>>>>>>site. You can put together a decent set of CPW test fixtures for
>>>>>>
>>>under
>>>
>>>
>>>>>>$200. cash and some time. Your biggest expense will be a pair of SMA
>>>>>>connectors per fixture.
>>>>>>
>>>>>>If you want more information on bypass network design, Istvan's web
>>>>>>
>>>>site,
>>>>
>>>>
>>>>>>the Teraspeed web site, and the X2Y web site all have papers on the
>>>>>>
>>>>>subject.
>>>>>
>>>>>
>>>>>>Steve.
>>>>>>At 04:58 PM 5/16/2005 +0530, Joe Paul M wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>>I have a doubt regarding decoupling capacitors.
>>>>>>>
>>>>>>>I have the option for using 1 uf9AVX 06036D105KAT2A) or 0.1uF (AVX
>>>>>>>0603ZC104KAT2A) at same cost.
>>>>>>>
>>>>>>>Concerned rise time is about 2nS.
>>>>>>>
>>>>>>>Is there any issue in using 1uF caps, if it has same ESR and ESL and
>>>>>>>package (0603) as 0.1uF.
>>>>>>>
>>>>>>>Thanks all
>>>>>>>Joe Paul
>>>>>>>
>>>>>>>------------------------------------------------------------------
>>>>>>>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 FAQ wiki page is located at:
>>>>>>> http://si-list.org/wiki/wiki.pl?Si-List_FAQ
>>>>>>>
>>>>>>>List technical documents are available at:
>>>>>>> http://www.si-list.org
>>>>>>>
>>>>>>>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
>>>>>>>
>>>>>>>
>>>>>>
>>>>>>------------------------------------------------------------------
>>>>>>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 FAQ wiki page is located at:
>>>>>> http://si-list.org/wiki/wiki.pl?Si-List_FAQ
>>>>>>
>>>>>>List technical documents are available at:
>>>>>> http://www.si-list.org
>>>>>>
>>>>>>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
>>>>>>
>>>>>>
>>>>>>
>>>>>------------------------------------------------------------------
>>>>>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 FAQ wiki page is located at:
>>>>> http://si-list.org/wiki/wiki.pl?Si-List_FAQ
>>>>>
>>>>>List technical documents are available at:
>>>>> http://www.si-list.org
>>>>>
>>>>>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
>>>>>
>>>>>
>>>>>
>>>>>______________________________________________________________________
>>>>>All email being sent to or from SRC Computers, Inc. will be scanned by a
>>>>>third party scanning service.
>>>>>______________________________________________________________________
>>>>>
>>>>
>>>>______________________________________________________________________
>>>>All email being sent to or from SRC Computers, Inc. will be scanned by a
>>>>third party scanning service.
>>>>______________________________________________________________________
>>>>
>>>
>>>______________________________________________________________________
>>>All email being sent to or from SRC Computers, Inc. will be scanned by a
>>>third party scanning service.
>>>______________________________________________________________________
>>>
>>
>>
>>------------------------------------------------------------------
>>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 FAQ wiki page is located at:
>> http://si-list.org/wiki/wiki.pl?Si-List_FAQ
>>
>>List technical documents are available at:
>> http://www.si-list.org
>>
>>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
>>
>>
>>
>>
>
>--
>
>Ludwig Systems Engineering Consulting - Design - Implementation
>WWW: www.ludwigsystems.com System Architectures - FPGAs - PCBs
>Ph/Fx: +41-43-355-58-73/74 Hardware - Firmware - Software
>
>
------------------------------------------------------------------
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 FAQ wiki page is located at:
http://si-list.org/wiki/wiki.pl?Si-List_FAQ
List technical documents are available at:
http://www.si-list.org
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: