Well, gee, if you do go ahead and do a nodal analysis of an opamp like circuit, won't you find that the power supply circuitry and the bypass caps are also directly in series with the negative feedback network? Doesn't this have some effect?
I keep looking for that opamp ground pin. Apparently I'm not alone: http://www.analog.com/static/imported-files/application_notes/135208865AN-202.pdf
Demian's suggestion of modeling the circuit is a great one. Be sure to include the wiring, the parasitic elements of the components (like ESR, ESL, etc.), and so on. Even the switching action of the power supply diodes and the effects of any parasitic elements in them should be considered. A transient analysis would also be helpful, as would a 3D electromagnetic simulation.
I keep looking for that opamp ground pin. Apparently I'm not alone: http://www.analog.com/static/imported-files/application_notes/135208865AN-202.pdf
Demian's suggestion of modeling the circuit is a great one. Be sure to include the wiring, the parasitic elements of the components (like ESR, ESL, etc.), and so on. Even the switching action of the power supply diodes and the effects of any parasitic elements in them should be considered. A transient analysis would also be helpful, as would a 3D electromagnetic simulation.
I got this tought after reading AN-1135 from IR. It is not specifically about decoupling, but within the explenation lies why decoupling is mandatory.
Oscillation is only an effect of with improper decoupling. The cause is something more important (and far more simple).
The main cause is the law of current eternity. "Current comes out+current that comes back must be =0". However small the current is, it must obey this law.
Decoupling cap give path for this returning current. Consider the cap is the "battery" that powers the IC. Every current that leaves the +terminal must be replaced by exactly the same ammount entering it's -terminal.
For positive output, the flows comes from this cap's +terminal, entering +supply of opamp, leaving the output node of the opamp, goes to the load, and if this load is grounded, it travels through ground track, then returning to the -terminal of the cap.
Why we need cannot only put elko? Because elko only works for lower frequencies (working below 100khz), for signal >100khz we need cap that can pass this HF, like ceramic or tantalum. Only with elko the opamp will oscillate at HF, because there is no return path (of whatever the opamp has deliver via it's output node) for HF signal (elko cannot provide path for HF, ceramic/tantalum can). The opamp won't care, from DC to its Mhz maximum working frequency there should be path for this returning current (of whatever it has deliver through it's output node) back to opamp's +and- supply rails. And if the return path from the load is coming from ground path, there should be cap (of all frequencies) from ground path to +and- supply rails of the opamp.
But this makes me more confused. If we follow this logic, then the "law" to "separate dirty ground and clean ground" is wrong.
Separating dirty ground (supply decoupling) from clean ground (signal gnd), cannot be done just like that, we must study the loop first.
If we blindly follow that law of "dirty" and "clean" grounds, it just make the path much longer (and more inductive with longer PCB track).
Oscillation is only an effect of with improper decoupling. The cause is something more important (and far more simple).
The main cause is the law of current eternity. "Current comes out+current that comes back must be =0". However small the current is, it must obey this law.
Decoupling cap give path for this returning current. Consider the cap is the "battery" that powers the IC. Every current that leaves the +terminal must be replaced by exactly the same ammount entering it's -terminal.
For positive output, the flows comes from this cap's +terminal, entering +supply of opamp, leaving the output node of the opamp, goes to the load, and if this load is grounded, it travels through ground track, then returning to the -terminal of the cap.
Why we need cannot only put elko? Because elko only works for lower frequencies (working below 100khz), for signal >100khz we need cap that can pass this HF, like ceramic or tantalum. Only with elko the opamp will oscillate at HF, because there is no return path (of whatever the opamp has deliver via it's output node) for HF signal (elko cannot provide path for HF, ceramic/tantalum can). The opamp won't care, from DC to its Mhz maximum working frequency there should be path for this returning current (of whatever it has deliver through it's output node) back to opamp's +and- supply rails. And if the return path from the load is coming from ground path, there should be cap (of all frequencies) from ground path to +and- supply rails of the opamp.
But this makes me more confused. If we follow this logic, then the "law" to "separate dirty ground and clean ground" is wrong.
Separating dirty ground (supply decoupling) from clean ground (signal gnd), cannot be done just like that, we must study the loop first.
If we blindly follow that law of "dirty" and "clean" grounds, it just make the path much longer (and more inductive with longer PCB track).
Attachments
I would choose the "big fat" polyprop 100nF cap for decoupling instead of the ceramic. I would try it and see if the OPamp in its application is stable and if it is, then do listening tests between different big fat PP caps, and I might even try a SMD PPS cap and some polystyrene caps.
100nF is the rule of thumb it seems, but rule of thumbs might not be the optimal value. I sometimes have used 10n polystyrenes and sometimes 0,47uF BG HiQ caps.
Sigurd
100nF is the rule of thumb it seems, but rule of thumbs might not be the optimal value. I sometimes have used 10n polystyrenes and sometimes 0,47uF BG HiQ caps.
Sigurd
syn08 said:
Sigurd wrote
This 'rule of thumb' is an excellent example of very tired thinking. The 'rule' was perfectly adequate when applied to slow TTL logic. It has become increasing untrue as logic switching speeds (and opamp gain bandwidth products) increase.
Reading a little more widely will reveal recommendations from the likes of Analog Devices for decoupling capacitors in the range of 220-330nF for their instrumentation amps which have applications not unlike audio - high gain amplifiers acting on low level signals in environments that are not under the control of the designer.
A little light reading from TI - and certainly not the only example ( but this is quite digestible)
How (Not) to Decouple High-Speed Operational Amplifiers by Bruce Carter
This 'rule of thumb' is an excellent example of very tired thinking. The 'rule' was perfectly adequate when applied to slow TTL logic. It has become increasing untrue as logic switching speeds (and opamp gain bandwidth products) increase.
Reading a little more widely will reveal recommendations from the likes of Analog Devices for decoupling capacitors in the range of 220-330nF for their instrumentation amps which have applications not unlike audio - high gain amplifiers acting on low level signals in environments that are not under the control of the designer.
A little light reading from TI - and certainly not the only example ( but this is quite digestible)
How (Not) to Decouple High-Speed Operational Amplifiers by Bruce Carter
VivaVee said:
A practice that is often seen is to parallel a bunch of capacitors (like an electrolytic, a stacked metal film, and a polystyrene) with the hope of attaining a low impedance across a very wide band of frequencies. Some time, for fun, either measure the impedance of what you get by doing that or do a complete simulation.
Parrareling different cap types and values is not as easy as it seems. EVA show us something
http://www.diyaudio.com/forums/showthread.php?postid=1566979#post1566979
http://www.diyaudio.com/forums/showthread.php?postid=1566979#post1566979
Good tips, PMA 
This diagram is about bootstrap, but maybe we can learn something about decoupling cap. There are 3 way to put R. The dioda makes 1 direction flow, the energy from the cap cannot flow back to the rail.
The first one is R to the opamp's supply pin. It helps for current flowing to the opamp, but don't help for rail current charging the cap.
The second one is R to both the opamp's supply pin and decoupling cap. It helps for rail powering the opamp, but don't help for cap powering the opamp
The third one is R only to decoupling cap. It helps for cap powering the opamp and rail charging the cap.
This diagram is about bootstrap, but maybe we can learn something about decoupling cap. There are 3 way to put R. The dioda makes 1 direction flow, the energy from the cap cannot flow back to the rail.
The first one is R to the opamp's supply pin. It helps for current flowing to the opamp, but don't help for rail current charging the cap.
The second one is R to both the opamp's supply pin and decoupling cap. It helps for rail powering the opamp, but don't help for cap powering the opamp
The third one is R only to decoupling cap. It helps for cap powering the opamp and rail charging the cap.
Attachments
CG wrote:
The simulation, if you include a reasonably complete and accurate model of the components you are using, will give you an idea of what can go wrong. But only by measuring the circuit (PCB + components +/- wiring) will you see what you are getting. This is a perfect example of where simulations will give you a hint but physical measurement is needed to see what is really happening.
This requires something like a network analyser which is not exactly lurking on everyones test bench.
Although a quick look at Ebay found a HP 8754A 4-1300MHz Network Analyzer for USD999. This is still quite a hurdle for many (most?).
The simulation, if you include a reasonably complete and accurate model of the components you are using, will give you an idea of what can go wrong. But only by measuring the circuit (PCB + components +/- wiring) will you see what you are getting. This is a perfect example of where simulations will give you a hint but physical measurement is needed to see what is really happening.
This requires something like a network analyser which is not exactly lurking on everyones test bench.
Although a quick look at Ebay found a HP 8754A 4-1300MHz Network Analyzer for USD999. This is still quite a hurdle for many (most?).
VivaVee, your link provided the question that I have at this time. E.G. is the .1uF default bypass cap obsolete? I certainly would wish it to be true. It would depend on the real contribution that the bypass does and at what frequency. I hope that Scott Wurcer will give us more complete input on this. I know that he has greater understanding than most, because he helped me 24 years ago with a bypassing problem with an RF amp. He has done significant research on the subject.
VivaVee said:
Actually, if you include all the items you mention you can get a very good idea of the impedances and all through simulation. But, knowing what to put into the simulation is the key. Few people bother with that. Plus it really takes good models to make that all work. You pretty much need a network analyzer to make your own models or at least verify the ones you "invent."
The kinds of simulation tools that are good to use for this are not so inexpensive and are primarily used for RF type work.
However, this one is free and is pretty good. Especially for the price.
http://www.practicalrf.com/$Newsletter/e-letters/February2006/RFSim99.htm
There's a couple of PC based network analyzers that work pretty well down to a few KHz, and are not stupid amounts of money. This one is pretty good:
http://n2pk.com/
John Curl wrote
Being perverse, I would say it depends.
YES: If you are talking about a 100nF ceramic in a leaded package with a Y5U dielectric, then definitely obsolete unless you are happy
building circuits with 741 opamps. Which might be all you need in some instances.
NO: If you are talking about a 100nF X7R ceramic in a surface mount package (metric 1608 or smaller) properly located adjacent to the supply pin and routed to the ground plane on your multilayer (4 layers if you are cheap, 6 layers acceptable at a pinch) PCB.
Put the first tranche of caps alongside each and every supply pin and then add more to decouple the power and ground planes. Spread harmonically across the entire PCB, this last set will push the GND-Vcc plane resonances high enough in frequency to not be any trouble to your active circuitry.
The tolerances of the caps will lead to a reasonably broad trough in the power supply impedance vs frequency curve that should see you right in the low hundreds of MHz for practical purposes.
But then, someone will try to operate a 5W 900MHz transceiver next to your cabling, and all heck breaks loose. Then my default answer breaks down and you have to do some real engineering - cookbook stuff is no longer enough.
But if RF immunity is not a big issue (and you may have more effective control over what the end user does or expects than I do) then the qualified default works well. Just note that the effective capacitance at the supply pins is in the ball park of the 220n-330n recommendation from ADI for sensitive analog instrumentation I mentioned before. But cheaper, much cheaper.
Caveat: no-one listens to my industrial electronics, unless you count the 'scream' of an ADC hitting FFFF. Or was that just me?... for audio use, I stick to the same strategy but use COG/NPO versions of the 100nF cap.
P.S. I have some further references which I trying to track done which would help you, I think. Just trying to find a URL to the papers.
Being perverse, I would say it depends.
YES: If you are talking about a 100nF ceramic in a leaded package with a Y5U dielectric, then definitely obsolete unless you are happy
building circuits with 741 opamps. Which might be all you need in some instances.
NO: If you are talking about a 100nF X7R ceramic in a surface mount package (metric 1608 or smaller) properly located adjacent to the supply pin and routed to the ground plane on your multilayer (4 layers if you are cheap, 6 layers acceptable at a pinch) PCB.
Put the first tranche of caps alongside each and every supply pin and then add more to decouple the power and ground planes. Spread harmonically across the entire PCB, this last set will push the GND-Vcc plane resonances high enough in frequency to not be any trouble to your active circuitry.
The tolerances of the caps will lead to a reasonably broad trough in the power supply impedance vs frequency curve that should see you right in the low hundreds of MHz for practical purposes.
But then, someone will try to operate a 5W 900MHz transceiver next to your cabling, and all heck breaks loose. Then my default answer breaks down and you have to do some real engineering - cookbook stuff is no longer enough.
But if RF immunity is not a big issue (and you may have more effective control over what the end user does or expects than I do) then the qualified default works well. Just note that the effective capacitance at the supply pins is in the ball park of the 220n-330n recommendation from ADI for sensitive analog instrumentation I mentioned before. But cheaper, much cheaper.
Caveat: no-one listens to my industrial electronics, unless you count the 'scream' of an ADC hitting FFFF. Or was that just me?... for audio use, I stick to the same strategy but use COG/NPO versions of the 100nF cap.
P.S. I have some further references which I trying to track done which would help you, I think. Just trying to find a URL to the papers.
Could decoupling capacitor size be approximated by opamp's output capability at a certain high frequency? The current that must pass decoupling cap can be approximated by the output current that flow from output pin at any high frequency. Maybe with help of some safety factor (x5, x10) from that current.
If opamp(1) can put more current at a certain Mhz frequency than opamp(2), then opamp(1) will need larger decoupling size than opamp(2).
If opamp(1) can put more current at a certain Mhz frequency than opamp(2), then opamp(1) will need larger decoupling size than opamp(2).
The caps esr at the frequency of interest will impact the peak available current. Bigger caps may be necessary but they won't work because of the ESR or self resonance issues. A network of smaller caps may well work better. The info in the app note was very good at explaining the issue. I'm not sure how my assembly house will react when I tell them to put the caps on edge. . .
The decoupling caps are much more involved with the capacitive component of the load. Both loads may still have 10,000 pF of cable attached to them and if the opamp is trying to respond to a click the dv-dt may require a lot of current to fill the cap. A 10,000 pF cable (a long one) would be 800 Ohms or so at 20KHz. The 10K input Z of the power amp is at the other end of the cap.
Its easy to see how a high GBW opamp a long cable and less that optimized bypassing could lead to oscillation. Take this problem to a power amp with wide bandwidth and a long high capacitance cable and it can lead to a lot of smoke.
Its easy to see how a high GBW opamp a long cable and less that optimized bypassing could lead to oscillation. Take this problem to a power amp with wide bandwidth and a long high capacitance cable and it can lead to a lot of smoke.
Lumanauw, bypassing has many factors, but the most important here is the maintaining of a reasonably low impedance over a working frequency range. Usually, ANY cap over .1uf is useless for this job, because the capacitor self resonates and goes inductive above a certain frequency. The real question here is whether we can or should use even less capacitance, in order to: 1 Have better stablity, 2 Have a better fit, 3 to save money or invest in a higher quality cap with a smaller value. Usually, peak current with frequency has very little to do with audio, just because the power supply normally handles it OK.
- Status
- Not open for further replies.