Download Steering in the right direction

Page 1 of 13
Yet More On Decoupling, Part 6 – steering in the right direction
Kendall Castor-Perry
This article was published on EDN: http://www.edn.com/design/powermanagement/4416874/Why-bypass-caps-make-a-difference---Part-6--Simulating-thecomplete-op-amp-PS-circuitThis is the last of the “Yet more...” series. One final plot twist, a moderately satisfying
ending, and my work here will be done. Until the sequel, that is...
Let the op-amp do the pinging!
Instead of exciting the supply rail impedance with a test current source, let’s drive a
signal into the amplifier, so that the load current itself causes the excitation on the supply
rails. There must still be supply voltage variations resulting from this, and we know they
can affect the amplifier output. This has profound implications for sampled data systems
and would explain why, so often, the fine tuning of such a high precision system involves
fiddling about with not only amplifier choice but also supply decoupling.
The test fixture stays the same but now we set up the input voltage source to provide a
+/-1 V signal. Because the amplifier is configured with a gain of -1, we’ll invert the
input signal when plotting so that we can overlay it on the output and examine the settling
behaviour. This signal causes a +/-10 mA current to flow in the load, which is a
combination of the 111.1 ohm load resistor and the 1 kohm feedback resistor. Now, that
feedback resistor doesn’t return to ‘real’ ground, only the nearly-at-ground input of the
amplifier, but the error introduced by this is going to be a small fraction of the other
errors we’re looking for.
To add a pinch more realism, an additional capacitance of 22 pF was added across the
load. This is typical of the load such an amplifier might have to contend with – ADC
input stages, other buffer amps, trace capacitance, and the occasional ‘scope probe.
There’s no little capacitor across the feedback resistor, for two reasons. Firstly, the best
value would depend on the amplifier, and secondly I wanted to look at CFAs as well –
just out of interest, at this stage.
Now, it’s clear that the very simple macromodel introduced in part 3 isn’t going to work
in this circuit; the output current isn’t supplied from the power pins. In any normal
amplifier, when +10 mA is being delivered to the load, it is flowing out of the output pin
and must be flowing into the +ve supply pin. Conversely, when the load current is -10
mA, this current flows into the output pin and comes out of the –ve supply pin. Each pin
therefore should inject a current demand of 0 to 10 mA on its supply rail, phased like the
way we tested the circuit earlier. So, if the amplifier model obeys these rules, the
excitation of the supply impedance should be comparable to the tests in part 5. Let’s try
some of our models and look at the waveforms. In the busy figure 6.1, the most
important traces are the grey and pink traces showing that this model is indeed taking an
extra 10 mA from the +ve rail when the output voltage is +1 V, and -10 mA from the
Page 2 of 13
negative rail when at -1 V. On this scale the supplies aren’t moving much, but when we
zoom in (figure 6.2) and apply some offset to get both rails on the same scale (400 mV
top to bottom), we see the expected voltage step and frizz on the supply rails, together
with the beginnings of some activity on the amplifier output (green trace).
6V
V (o u t )
V (i n )
V (p l u s )
V (m i n u s )
I x (U 2 : V + )
I x (U 2 : V - )
1 8m A
5V
1 5m A
4V
1 2m A
3V
9m A
2V
6m A
1V
3m A
0V
0m A
-1 V
-3 m A
-2 V
-6 m A
-3 V
-9 m A
-4 V
-1 2 m A
-5 V
-1 5 m A
-6 V
0 µs
5 µs
1 0µs
1 5µs
2 0µs
2 5µs
3 0µs
3 5µs
4 0µs
4 5µs
-1 8 m A
5 0µs
Figure 6.1: waveforms from the test rig driving a load, showing op-amp supply currents
(grey and pink traces).
Page 3 of 13
1 .2 V
V (o u t )
V (i n )
V (p l u s ) -3 . 9
V (m i n u s ) + 5 . 9
1 .1 V
1 .0 V
0 .9 V
0 .8 V
1 4µs
1 5µs
1 6µs
1 7µs
1 8µs
1 9µs
2 0µs
2 1µs
2 2µs
2 3µs
2 4µs
2 5µs
2 6µs
Figure 6.2: zooming in on figure 6.1 to show that the supply voltages are moving.
This is with the ceramic caps on the regulators. By the way, if you probe the current in
the printed circuit trace, you’ll see significant ringing rather than the fairly clean square
wave. In other words, the sum of the currents in the trackwork and the local decoupling
capacitor is clean, but the transient current demand from the amplifier provokes high
frequency circulating currents that are flowing in the power feed, and that could get all
over your circuit board., without an obvious voltage cause.
The final plot twist
It’s rogue’s gallery time again because, you guessed it, many models out there don’t let
you do this; they don’t steer load current from the power pins, but from a magical,
everlasting internal battery supply.
Page 4 of 13
Figure 6.3: LT LT1354 model, taking a constant current from each rail (but not the same
value).
Figure 6.4: LT LT1055 model; some little high frequency squeaks modulating the supply
current.
Figure 6.3 shows common behaviour, from one of the macromodels that didn’t model
supply rejection either – I only include it because it’s a good amplifier in the real world!
Although the amplifier is certainly delivering current into the load, the supply pins only
consume a constant, quiescent current. And in the case of the LT1354, the 1 mA current
Page 5 of 13
that goes into the +ve supply pin doesn’t all come out of the –ve supply pin, about 100
uA goes missing. Mr. Kirchhoff will be upset... This behaviour was quite common
among models from several makers. The LT1055 (figure 6.4) offered a variant: a little
dynamic twiddle of a current on the signal edges, superimposed on a static current drain.
Several TI models failed this test. Some models, including the TLE2082 (clear winner of
the “impossible PSG” competition earlier in the series), sat there analysing so slowly that
it was going to take almost a week to complete (I saw that behaviour with some exElantec devices also). The OPA350 gave very strange results for 22 nF decoupler, with
the +5 V supply voltage collapsing, despite no unusual current being taken from the
regulator. At longer timescales, LTspice wasn’t able to complete the analysis. Perhaps
using excessive supply voltage has an impact in transient SPICE simulations.
Winner of the “does it do that?” competition this time was TI’s OPA2604E model, for a
device beloved of audio designers. For some decoupler values, it analysed fine for a few
microseconds and then spontaneously burst into oscillation on the ‘screen’ in front of me,
as if I had touched it somewhere sensitive with a high capacitance probe! Next thing we
know, we’ll have thermal feedback from the simulator to the keyboard so you really do
burn your fingers when the circuit goes wrong! But I digress..
Whoa, yeah! Change the decoupling, change the settling time
To see what’s actually happening at the op-amp output we need to zoom in further. Let’s
do that and try some other amplifiers to see what they look like, still with the livelier
ceramic regulator output cap. In Figure 6.5, the LT1723 shows some evidence of the
ringing with minimum decoupler coming through, at about 1 mV p-p. The time delay
between the negative-going edge of the blue trace (input) and the bouncing up of the
green trace (output) indicates how long it took the device to respond. This is affected by
both slew rate and closed-loop bandwidth, which in this configuration we’d expect to be
half the GBW. The LT1812 in figure 6.6 has a lower bandwidth but higher slew rate, and
the green trace “gets there” more quickly, indicating that at this signal level the slew rate
is dominant. Ringing with 22 nF is 2 mV p-p, which is 0.1% of full scale.
Page 6 of 13
1 .0 1 0 V
V (o u t )
V (i n )
V (p l u s ) -3 . 9
V (m i n u s ) + 5 . 9
1 .0 0 8 V
1 .0 0 6 V
1 .0 0 4 V
1 .0 0 2 V
1 .0 0 0 V
0 .9 9 8 V
0 .9 9 6 V
0 .9 9 4 V
0 .9 9 2 V
0 .9 9 0 V
1 4 . 6 µ s1 4 . 8 µ s1 5 . 0 µ s
1 5 . 4 µ s1 5 . 6 µ s1 5 . 8 µ s
1 6 . 2 µ s1 6 . 4 µ s1 6 . 6 µ s
1 7 . 0 µ s1 7 . 2 µ s
Figure 6.5: zooming in on the output of the LT1723 to show the ripple, which is worst for
the 22 nF decoupler.
1 .0 1 0 V
V (o u t )
V (i n )
V (p l u s ) -3 . 9
V (m i n u s ) + 5 . 9
1 .0 0 8 V
1 .0 0 6 V
1 .0 0 4 V
1 .0 0 2 V
1 .0 0 0 V
0 .9 9 8 V
0 .9 9 6 V
0 .9 9 4 V
0 .9 9 2 V
0 .9 9 0 V
1 4 . 6 µ s1 4 . 8 µ s1 5 . 0 µ s
1 5 . 4 µ s1 5 . 6 µ s1 5 . 8 µ s
1 6 . 2 µ s1 6 . 4 µ s1 6 . 6 µ s
1 7 . 0 µ s1 7 . 2 µ s
Figure 6.6: looking at the higher slew rate but lower GBW LT1812 on the same scale as
figure 6.5.
Page 7 of 13
What about slower amplifiers? Figure 6.7 shows the LT1632. Note the significantly
higher level of ringing now. In fact, if you look at a PSG plot for this model, you can see
3 dB peaking before it settles down. If this peaking coincides with a supply impedance
peak it will make the ringing look worse than it would do if the model were correct.
Figure 6.8 shows the even slower LT1630. It’s the slowest one here, yet is not showing
quite so much ripple. Why is that?
1 .0 1 0 V
V (o u t )
V (i n )
V (p l u s ) -3 . 9
V (m i n u s ) + 5 . 9
1 .0 0 8 V
1 .0 0 6 V
1 .0 0 4 V
1 .0 0 2 V
1 .0 0 0 V
0 .9 9 8 V
0 .9 9 6 V
0 .9 9 4 V
0 .9 9 2 V
0 .9 9 0 V
1 4 . 6 µ s1 4 . 8 µ s1 5 . 0 µ s
1 5 . 4 µ s1 5 . 6 µ s1 5 . 8 µ s
1 6 . 2 µ s1 6 . 4 µ s1 6 . 6 µ s
1 7 . 0 µ s1 7 . 2 µ s
Figure 6.7: the output of the LT1632 model, whose slightly peaky PSG may be
exaggerating the ringing a little.
Page 8 of 13
1 .0 1 0 V
V (o u t )
V (i n )
V (p l u s ) -3 . 9
V (m i n u s ) + 5 . 9
1 .0 0 8 V
1 .0 0 6 V
1 .0 0 4 V
1 .0 0 2 V
1 .0 0 0 V
0 .9 9 8 V
0 .9 9 6 V
0 .9 9 4 V
0 .9 9 2 V
0 .9 9 0 V
1 4 . 6 µ s1 4 . 8 µ s1 5 . 0 µ s
1 5 . 4 µ s1 5 . 6 µ s1 5 . 8 µ s
1 6 . 2 µ s1 6 . 4 µ s1 6 . 6 µ s
1 7 . 0 µ s1 7 . 2 µ s
Figure 6.8: the output of the LT1630 model. The current demand is so sluggish that the
ringing is reduced.
Now we start seeing something interesting The LT1630 takes so long to settle down,
and the current excitation’s edges are so slowed down, that despite its poor bandwidth,
and hence supply rejection, the output ringing is not that severe (once the output has
actually settled). The closed-loop small signal response and the large-signal slewing
behaviour can combine to ‘take the edge off” (literally!) the output voltage, hence the
load current, hence the supply excitation. The longer it takes the output voltage to reach
the final value, the gentler will be the excitation of the supply resonances.
What about our secret weapon for reducing all this ringing, the tantalum output capacitor
for the regulators? Well, it’s hardly surprising that it has a marvellous effect here; figure
6.9 shows what’s left on the worst-case LT1632 output. Now, the value of the
decoupling capacitor really doesn’t seem to be having much important effect. The
equivalent trace for the LT1723 is now so clean and fast that it’s not worth printing;
we’re going to have to zoom even further in on both amplitude and time.
Page 9 of 13
1 .0 1 0 V
V (o u t )
V (i n )
V (p l u s ) -3 . 9
V (m i n u s ) + 5 . 9
1 .0 0 8 V
1 .0 0 6 V
1 .0 0 4 V
1 .0 0 2 V
1 .0 0 0 V
0 .9 9 8 V
0 .9 9 6 V
0 .9 9 4 V
0 .9 9 2 V
0 .9 9 0 V
1 4 . 6 µ s1 4 . 8 µ s1 5 . 0 µ s
1 5 . 4 µ s1 5 . 6 µ s1 5 . 8 µ s
1 6 . 2 µ s1 6 . 4 µ s1 6 . 6 µ s
1 7 . 0 µ s1 7 . 2 µ s
Figure 6.9: LT LT1632 settling cleanly with tantalum regulator cap. Compare to figure
6.5.
1 0 0 1 .0 m V
V (o u t )
V (i n )
V (p l u s ) -3 . 9
V (m i n u s ) + 5 . 9
1 0 0 0 .9 m V
1 0 0 0 .8 m V
1 0 0 0 .7 m V
1 0 0 0 .6 m V
1 0 0 0 .5 m V
1 0 0 0 .4 m V
1 0 0 0 .3 m V
1 0 0 0 .2 m V
1 0 0 0 .1 m V
1 0 0 0 .0 m V
9 9 9 .9 m V
9 9 9 .8 m V
9 9 9 .7 m V
9 9 9 .6 m V
9 9 9 .5 m V
9 9 9 .4 m V
9 9 9 .3 m V
9 9 9 .2 m V
9 9 9 .1 m V
9 9 9 .0 m V
0 . 9 0 µ s1 . 0 0 µ s1 . 1 0 µ s1 . 2 0 µ s1 . 3 0 µ s1 . 4 0 µ s1 . 5 0 µ s1 . 6 0 µ s1 . 7 0 µ s1 . 8 0 µ s1 . 9 0 µ s2 . 0 0 µ s2 . 1 0 µ s
Figure 6.10: LT LT1723 settling with tantalum regulator cap, 500 kHz squarewave, 100
uV/div.
Page 10 of 13
Figure 6.10 shows the step response of the LT1723 model with a full scale of +/-0.1%;
figure 6.11 shows the LT1812 on the same scale. Now, the LT1812 is an amplifier with a
much higher slew-rate than the LT1723 (750 V/us versus 7 0V/us), but a lower gainbandwidth product, than the LT1723 (100 MHz versus 200 MHz, so the data sheet says,
and the PSG sweeps bear this out). So the LT1723 takes longer to acquire the input
signal; it also has a lower initial response ring frequency because on the first few parts of
the ring it’s still in slew limiting except at the turning points. But the combination of less
rail excitation (probe the fixtures yourself!) and more power supply rejection means that
the differences between the various decoupling capacitors (swept 22 nF to 470 nF as
always) are much less pronounced than on the LT1812.
1 0 0 1 .0 m V
V (o u t )
V (i n )
V (p l u s ) -3 . 9
V (m i n u s ) + 5 . 9
1 0 0 0 .9 m V
1 0 0 0 .8 m V
1 0 0 0 .7 m V
1 0 0 0 .6 m V
1 0 0 0 .5 m V
1 0 0 0 .4 m V
1 0 0 0 .3 m V
1 0 0 0 .2 m V
1 0 0 0 .1 m V
1 0 0 0 .0 m V
9 9 9 .9 m V
9 9 9 .8 m V
9 9 9 .7 m V
9 9 9 .6 m V
9 9 9 .5 m V
9 9 9 .4 m V
9 9 9 .3 m V
9 9 9 .2 m V
9 9 9 .1 m V
9 9 9 .0 m V
0 . 9 0 µ s1 . 0 0 µ s1 . 1 0 µ s1 . 2 0 µ s1 . 3 0 µ s1 . 4 0 µ s1 . 5 0 µ s1 . 6 0 µ s1 . 7 0 µ s1 . 8 0 µ s1 . 9 0 µ s2 . 0 0 µ s2 . 1 0 µ s
Figure 6.11: LT LT1812 settling with tantalum regulator cap, 500 kHz squarewave, 100
uV/div.
Page 11 of 13
1 0 0 1 .0 m V
V (o u t )
V (i n )
V (p l u s ) -3 . 9
V (m i n u s ) + 5 . 9
1 0 0 0 .9 m V
1 0 0 0 .8 m V
1 0 0 0 .7 m V
1 0 0 0 .6 m V
1 0 0 0 .5 m V
1 0 0 0 .4 m V
1 0 0 0 .3 m V
1 0 0 0 .2 m V
1 0 0 0 .1 m V
1 0 0 0 .0 m V
9 9 9 .9 m V
9 9 9 .8 m V
9 9 9 .7 m V
9 9 9 .6 m V
9 9 9 .5 m V
9 9 9 .4 m V
9 9 9 .3 m V
9 9 9 .2 m V
9 9 9 .1 m V
9 9 9 .0 m V
0 . 9 0 µ s1 . 0 0 µ s1 . 1 0 µ s1 . 2 0 µ s1 . 3 0 µ s1 . 4 0 µ s1 . 5 0 µ s1 . 6 0 µ s1 . 7 0 µ s1 . 8 0 µ s1 . 9 0 µ s2 . 0 0 µ s2 . 1 0 µ s
Figure 6.12: as figure 6.10 but with an extra 4.7 nF capacitor as decoupling.
Did you also notice that the actual settled output value is lower with the LT1812 than
with the LT1723? The LT1812 has rather a low open-loop gain in comparison to the
LT1723, and here’s an example of where this is relevant. The settled value of the output
step is dependent on the low frequency error remaining; with a low-frequency open loop
gain of only ~66 dB, we pick up an extra ~0.05% of error, which shows in the plots. It’s
not quite symmetrical when we look at the –ve excursion, probably due to a little bit of
input bias current and offset voltage in the model.
The settling behaviour is always better with the larger values of decoupling, up to 470 nF.
If we’re going to be using such large decoupling capacitors, might it be sensible to put an
additional small decoupling capacitor in parallel, to take us up to the three capacitor
situation discussed in “Know the...”? Figure 6.12 shows the result of adding a 4.7 nF
ceramic capacitor (with 0.7 nH inductance and 80 milliohms of ESR) in parallel with the
existing swept decoupling capacitor. The result is a new, undamped resonance at an even
higher frequency, as we’d expect. The extra cap appears to add no value at longer
timescales, and just additional ringing in the first few hundred ns. This points to a
preference for a single, well-chosen decoupling capacitor rather than several in parallel.
The Big Finish
Finally, therefore, we have arrived at where I’ve been trying to get to all along: the
accuracy of the op-amp’s output is unmistakeably affected by the choices you make on
power supply regulator and decoupling components. We’ve seen that if there’s no
damping in the regulator output capacitance, an op-amp can self-harm through its
Page 12 of 13
interaction with the supplies. Even with damping from tantalum capacitors, the value of
local supply decoupling affects the step output of fast amplifiers, delivering an extra error
term in even a moderately accurate system. As the decoupling capacitor value is
increased, the frequency and amplitude of this resonant tail fall off rapidly. Clearly, for
accuracy to be maintained here, we need as large a decoupling capacitor as possible, with
470 nF being clearly superior to the commonplace 100 nF in these simulations.
Remember, all this work has been done for a gain of -1, i.e. a noise gain of 2. The effect
of increasing the noise gain will be the same as choosing a lower GBW amplifier. High
gain systems are even more likely to be impacted by these supply disturbances.
Takeaways and conclusions for the series







Signals on the supply pins of an op-amp will find their way to the output
especially at higher frequencies, but also at LF if the amplifier has poor low
frequency open loop gain.
Self-inflicted supply variations due to varying load current are always present
and can never be completely rejected by the amplifier. They will affect the
amplifier output to some accuracy level, and you can investigate this to a
surprising level of accuracy in simulation.
Op-amp vendors put many obstacles in your way to simulating this effectively
– press them (and your regulator and passives suppliers) for proper SPICE
models that work! Or just use op-amps with working models, by testing out
the models with the test fixtures from this series.
Don’t use low ESR regulator capacitors just because the regulator vendor says
they are OK for ‘stability’. Your system performance is likely to be
significantly better with the additional losses that tantalum capacitors can
provide in this role.
A bit of slew-rate limiting in the op-amp isn’t necessarily bad; the resonances
on the supplies are excited less strongly the slower the amplifier output is
slewing. But if there are some fast-slewing, wide-bandwidth amplifiers
affecting the supply rails, then low-bandwidth amplifiers, with their
consequently poor supply rejection, will be particularly sensitive to the
resulting noise. Careful in a mixed system!
Use the highest value local ceramic decouplers you can get, in the largest
package that will fit, and don’t use Y5V dielectric material unless you’ve fully
allowed for the value change with bias voltage. But the volumetric advantage
of the high-k dielectric may be completely lost, with this allowance.
Adding another small ceramic capacitor in parallel with the main decoupler
has a small effect on the amplifier output – but it’s often a negative one, so
test carefully; most times, it may be unnecessary.
2013 Postscript
Just as in part 4, this final part required the op-amp macromodels to exhibit behavior that
resembled what the actual parts would do. In this case, what the amplifiers needed to do
Page 13 of 13
was use their output stages to steer currents flowing into the supply pins of the device.
Somewhat better news to report here. All the LT parts that I had grumbled about in the
original article have now acquired quite reasonable-looking current steering, and look
much more plausible to me. Redoing the plot for the LT1723 showed that the apparent
static current consumption had fallen somewhat, possibly due to correct values of the
‘off’ state current in the output stage. The LT1354, LT1055, LT1630 and LT1813
(standing in for the LT1812) all show viable behaviour when these plots are run.
I did find, by looking at the simulation files with the same time-stamps as the
corresponding images, that I had slightly misrepresented the sim in the original text. My
additional load cap of 22 pF (scope probe!) didn’t have a 100 ohm resistor in series as I
had claimed. So I tweaked the text of this version to fix that.
My image backup system unaccountably had different-looking files for figures 3 and 4 of
this part, so I extracted them from the original .doc file rather than using the original
.wmf versions. That’s why they look different (if they look different).
One interesting feature is particularly clear when you look at the supply pin currents of
the LT1354. The bandwidth of the part clearly limits the risetime of the output
squarewave seen on the load resistance. When the voltage ramps down to the opposite
extreme and the current in the load changes direction, the current in the previously active
pin falls to zero quite abruptly, and is not shaped by the bandwidth of the whole
amplifier. The implication here is that a bandlimited signal driven into a resistive load
can cause high frequency components to appear in the currents taken by the amplifier
power pins. Obvious when you think about it, but perhaps not obvious enough to be
considered as a source of EMI. Keep an eye out for that even in low bandwidth circuits!
The TI models haven’t changed, as I mentioned in part 4, so I didn’t redo the analyses.
LTspice has got faster, stabler and more convergent in the intervening five years, so it’s
possible that it will give a different account of the inaccurate behaviour of the model. As
I’ve observed elsewhere, what’s the point in having a more precise version of the wrong
answer?
Anyway, spoiler alert, we finally found out that bypass capacitor choice can definitely
have an effect on the signal integrity at your amplifier outputs. Who would have thought
it? The base rules I give at the end are as valid now as they were then. Don’t treat
bypass capacitors as an unavoidable expense imposed upon you by IC manufacturers.
And don’t assume that if an amplifier appears not to be oscillating, that you have done all
you need to with its supply impedance!