The single most important advantage of a chopper-stabilized, or auto-zero, amplifier is the fact that it has extremely low offset voltage. Unfortunately, there is some hesitation when it comes to using auto-zero amplifiers as most engineers associate them with the original chopper amplifier designs. This stigma has been perpetuated by either engineers who worked with these older chopper amplifiers and remember the difficulties they had with them, or younger engineers who learned about chopper amplifiers in school but probably did not understand them too well. I hope to lift this stigma from modern auto-zero amplifiers and if I've done my job well then by the end of this presentation, you should be quite comfortable using and designing with these high performance op amps.
The early concept of a chopper-stabilized amplifier involved dividing the input signal into two frequency bands and processing them separately. The high-frequency content was passed directly to and amplified by an op amp. The low-frequency and dc content was modulated into an ac signal with a switch, hence the term "chopper." This new signal was then ac coupled to remove dc errors, amplified by a separate amplifier, then synchronously demodulated at the output. This ac output signal was then filtered to remove switching spikes and noise, though not entirely successfully, and summed back into the high-frequency amplifier, combining both amplified high and low-frequency components back into a complete signal.
As you might imagine, there were a number of difficulties implementing this architecture. The modulation/demodulation switches had to be designed to not contribute any dc offsets themselves. It was also a challenge to remove the switching spikes and noise from the chopped signal which caused modulation clock feedthrough and intermodulation distortion at the output. Filtering the demodulated low-frequency signal helped reduced switching feedthrough, but often times there still remained a considerable amount of voltage spikes and noise on the output. This made chopper amplifiers suitable for only very low frequency or dc applications.
Modern auto-zero amplifiers achieve the same level of precision but are not nearly as troublesome to use. Throughout the text, I refer to these amplifiers as auto-zero or auto-correction amplifiers. I try to stay away from the word "chopper" not only to avoid the instinctive reaction from most designers, but also because the modern implementations of this architecture are quite different from their predecessors. The term "auto-zero" is really a more accurate description of how the amplifier works; the input signal waveform is certainly not chopped up anymore. First, we'll examine the theory behind auto-zero amplifiers to better understand exactly how these op amps work.
Basic auto-zero theory
Modern auto-zero amplifiers can be described using a model that consists of two amplifiers: A main amplifier to provide the primary signal amplification, and a nulling amplifier to correct offset error voltages. Figure 1 shows the simplified schematic. Keep in mind this is simply a model used to describe the basic technique used in auto-zero correction. Each manufacturer has their own implementation of the architecture.
Both the main and the nulling amplifiers have an associated input offset voltage, which can be modeled as a dc voltage source in series with the non-inverting input. These are labeled as VOSX, where x denotes the amplifier associated with the offset; A for the nulling amplifier, B for the primary amplifier. The open loop gain of each amplifier, the gain for the +IN and --IN inputs, is given as AX. Both amplifiers also have a third voltage input with an associated open loop gain of BX.
There are two modes of operation set by two sets of switches in the overall amplifier: An auto-zero phase and an amplification phase.
Figure 1: Auto-zero phase of the amplifier
In this phase, all phase-A switches are closed and all phase-B switches are opened. Here, the nulling amplifier is taken out of the gain loop by shorting its two inputs together. Of course, there is a degree of offset voltage, shown as VOSA, inherent in the nulling amplifier that maintains a potential difference between the +IN and -IN inputs. The nulling amplifier feedback loop is closed through phase-A2, and VOSA appears at the output of the nulling amp and on CM1. This capacitor can either be an external capacitor, on the order of 1uF, or an internal capacitor. Most recent auto-zero amplifiers do not require external capacitors as they can achieve effective CM values through Miller equivalence.
Mathematically, we can express the output of the nulling amplifier in the time domain as:
which by simple rearrangement is:
This shows us that the offset voltage of the nulling amplifier times a gain factor appears at the output of the nulling amplifier, and thus on the CM1 capacitor.
Figure 2: Output phase of the amplifier
When the phase-B switches close and the phase-A switches open for the amplification phase, this offset voltage remains on CM1 and essentially corrects any error from the nulling amplifier. The voltage across CM1 is designated as VNA. Let us also designate VIN as the potential difference between the two inputs to the primary amplifier, or VIN = (VIN+ minus VIN-). Now the output of the nulling amplifier can be expressed as:
Because phase-A is now open and there is no place for CM1 to theoretically discharge, the voltage VNA at the present time t is equal to the voltage at the output of the nulling amp VOA at the time when phase-A was closed. If we call the period of the auto-correction switching frequency TS, then the amplifier switches between phases every 0.5 x TS. Therefore, in the amplification phase:
And substituting Equation 4 and Equation 2 into Equation 3 yields:
For the sake of simplification, let us assume that the auto-correction frequency is much faster than any potential change in VOSA or VOSB. This is a good assumption since changes in dc offset voltage are a function of temperature variation or long-term wear time. Given this assumption, VOS is effectively made time invariant, and we can remove the time dependence of the VOSA variable. Equation 5 can then be rewritten as:
We can already get a feel for the corrective nature of the auto-zeroing. Note the VOS term has been reduced by a factor of 1+BA. This shows how the nulling amplifier reduces its own offset voltage error even before correcting the primary amplifier. Continuing on, the primary amplifier output voltage is the voltage at the output of the complete auto-zero amplifier. It is equal to:
In the amplification phase, VOA = VNB, so this can be rewritten as:
The auto-zero architectures are optimized in such a way that AA = AB and BA = BB, and BA >> 1. In addition, the gain product of AABB is much greater than AB. This allows Equation 10 to be simplified to:
Most obvious is the gain product of both the primary and nulling amplifiers. This AABA term explains why auto-zero amplifiers have extremely high open loop gain. To understand how VOSA and VOSB relate to the overall effective input offset voltage of the complete amplifier, we should set up the generic amplifier equation of:
Where k is the open loop gain of the amplifier and VOS,EFF is its effective offset voltage. Putting Equation 12 into the form of Equation 11 gives us:
And from here, it is easy to see that:
Thus, the offset voltages of both the primary and nulling amplifiers are reduced by the gain factor BA. Considering the open-loop gains of the local amplifiers, AA and BA, could be on the order of 10,000 or higher, it quickly becomes evident that even an inherent offset voltage of millivolts is reduced to an effective input offset voltage of the complete auto-zero amplifier of microvolts.
As a practical example to show some numbers, the AD8552 has a typical input offset voltage of 1uV operating from a 5V supply. It is often easy to overlook the practical significance of numbers, particularly numbers which are either extremely small or large.
Consider the following analogy: Imagine this amplifier is being used in an interstate weigh scale with 5V corresponding to the full scale value of a fully loaded 18-wheel truck. Such a truck could weigh on the order of 25 tons, which is 50,000 pounds, or 28,800 kilograms. One microvolt of precision would tell you if the keys were left in the ignition.
Now, there are practical limitations to going around and measuring the weight of drivers' keys in their trucks, namely other error sources in the circuit itself outside of the amplifier. Thermocouple voltages between the solder and the copper traces on the board may be larger than the offset voltage of the auto-zero amp. We'll consider these effects later, but the point here is to show that the amplifier is no longer the dominant source of error in the system.
Another point to consider is how to test and guarantee the performance of these devices, when the amplifiers are more accurate than the systems they are tested on! The best solution is to simply characterize the device to show what the typical specs are, but use wider limits for guaranteed minimums and maximums. This prevents raising the cost of these devices due to extensive testing. Most manufacturers test their amplifiers as well and as quickly as they can. Even if that means the guaranteed limits are well outside of what the device can reliably perform, most of these limits are still better than most standard op amps. The 5uV maximum offset voltage of the AD8552 is still better than the highest grade of the OP177, even though the offset voltage of the AD8552 is really around 1uV or less.
Offset errors from all sources are corrected
There are a number of benefits in using auto-zero correction in addition to the obvious effect of low offset voltage. From a general point of view we are looking at an error correcting amplifier, an amplifier that corrects itself from any source of offset voltage error. This error correction occurs during every period of the auto-zero clock so the amplifier provides a continuous error correction. So as the inherent offset voltage of both the main and the nulling amplifiers shift with a change in temperature, as is common with any op amp, the overall low offset voltage of the total auto-zero amplifier is maintained. It also corrects for any long term drift or aging effects that could change offset voltage over time.
Amplifiers built using CMOS transistors can have offset voltage temperature drifts (TCVOS) of 2uV/C and precision bipolar amplifiers can have a TCVOS of 0.2uV/C. Auto-zero amplifiers can have typical TCVOS numbers down to 5nV/C. This means that the offset voltage will change only 0.5uV over a 100C change in temperature, which is still less than the original offset voltage of the device.
The open-loop gain of the auto-zero amplifier is extremely high, as explained in Equation 11. Most auto-zero amps have typical open-loop gain numbers on the order of 136dB (6,000,000x gain) or higher. Because of this, the fractional gain error of these amplifiers is extremely low. Fractional gain error is the differential error voltage between the inputs of an op amp. This is not an inherent amplifier offset voltage but an error source due to the finite gain of the op amp. Remembering back to our control theory classes, the gain of an amplifier can be expressed as:
Figure 3: Amplifier control loop diagram
Where A(s) is the open-loop gain of the amplifier and b is the feedback attenuation of the amplifier, which is based on the resistors placed around the amplifier. Ideally, we would like the closed-loop gain to equal 1/b, but this can only occur as A(s) approaches infinity. For finite open-loop gains the closed-loop gain will be less than the ideal gain, and this difference is known as fractional gain error. Clearly, this can be reduced by using amplifiers with high open-loop gains, and this is the advantage with auto-zero op amps.
The common-mode rejection ratio (CMRR) of an amplifier is a measure of how much the amplifier's output voltage changes with changes in the input common-mode voltage. This is essentially a change in input offset voltage with common-mode voltage, and ideally this should be zero, corresponding to an infinite CMRR. Because it is an input offset voltage change, the level of output voltage swing due to this error is proportional to closed-loop gain, making this a difficult problem for high gain or precision applications. But precisely because it is an input offset voltage, the error correction of the auto-zero cycling greatly reduces this effect. This yields common-mode rejections on the order of 135dB or higher.
Power supply rejection, or PSRR, is also greatly improved for exactly the same reasons. Changes in the power supply to the amplifier create small changes in input offset voltage, which are then error corrected through auto-zero correction. Although most traditional amplifiers have a PSRR of 80dB to 100dB, auto-zero op amps can easily achieve typical numbers of 135dB.
Another advantage not commonly known about auto-zero amplifiers is the fact that they exhibit no flicker noise. This is a unique and significant feature of these amplifiers as flicker noise is present in all traditional amplifiers. Flicker noise, also known as 1/f noise, is in fact present in any semiconductor as it is a function of the recombination of electrons and holes. Not to be confused with either shot noise or thermal noise, both of which are equal energy per frequency, flicker noise increases with a decrease in frequency at a rate of 3dB per octave. This means that at lower frequencies the flicker noise of an amplifier can become quite large, which can become a significant error source for sub-Hertz precision applications such as weigh scales or strain gauges.
However, flicker noise can be modeled as a low frequency random change in input offset voltage, and as we've already demonstrated, an auto-zero amplifier corrects these changes in offset voltage. This means the low frequency noise of an auto-zero amplifier can be lower than traditional amplifiers, which could have 1/f corner frequencies in the 1-10Hz range.
Up to now we've considered only the performance advantages to auto-zero amplifiers. With the more recent auto-zero amplifiers, there is price-to-performance ratio advantage as well. To understand why, we should consider the general advantage of these amplifiers: They are self-error correcting op amps. This means that these amplifiers do not necessarily need to be fabricated on a precision process, like bipolar, or require trimming resistors on the IC or on the PC board of the application circuit. From a philosophical standpoint, electrical complexity has replaced mechanical complexity. This is a common theme in engineering advancement.
One way of mechanically achieving amplifier precision is by using on-board resistors and trimming them with a laser. In production this requires lining the laser up with the resistor, trimming some of the resistor, measuring the output, and repeating this process until the required level of precision is reached. This requires a good deal of expertise and time, not to mention the cost of the laser machine. There is also a limit to how small you can make the size of the IC. The trimming resistor must be relatively large and cannot be reduced as it must stay well larger than the wavelength of the trimming laser. This is mechanical complexity.
With an error correcting auto-zero amplifier, we are not interested in the initial offset voltage error of the internal amplifiers; we know the auto-zero architecture will correct any offset voltages. This allows these amplifiers to be constructed on a less precise, less expensive process like CMOS. It also saves the die area from relatively large trimming resistors and the cost of purchasing laser trimming machines. And as process continue to shrink, from 0.6m CMOS to 0.18m and beyond, the die size scales with it. Auto-zero amplifiers have replaced mechanical complexity with electrical complexity, a clear advantage.
Because the auto-zeroing operates in discrete time, its correction capability is limited by the period of the auto-zero clock. This means that as a change in offset voltage frequency increases, the amplifier is not as able to correct itself. For most of the error sources considered here, this is not a problem. Temperature and aging effects certainly do not change several hundred times a second and flicker noise decreases with an increase in frequency. Raising the frequency of the auto-zero clock can improve its high frequency accuracy performance somewhat, but there are practical limitations to doing this. In practice, auto-zero correction occurs anywhere from 100Hz to 10kHz or higher depending on the amplifier. This number is given in the manufacturer's datasheet.
Offset voltage errors are corrected through an iterative process; each clock cycle further reduces the effective offset voltage up to its mathematical limit. From a practical standpoint, offset voltage is approximately corrected within several clock cycles, but as the number of iterations increases, the closer the offset voltage will be corrected to zero. The longer you wait, the more accurate the answer is.
This brings up another practical limitation to specifying minimum limits to open-loop gain, CMRR, and PSRR. In production, these parameters must be measured on each op amp to guarantee the minimum performance. Although the minimum limits could be raised, assuming the resolution issues can be taken care of in the testers, it is not practical to wait for longer periods of time to get the "right" answer, as this increases the cost of testing the op amp, which in turn increases the cost of these devices to customers.
The discussion of auto-zero amplifiers has focused primarily on the auto-zero correction itself, particularly with respect to low frequency or dc performance, frequencies significantly lower than the auto-correction clock. The ac characteristics of an auto-zero is driven by the main amplifier, which can achieve whatever performance the manufacturer desires.
The single most fundamental misunderstanding about auto-zero amplifiers is the relationship between the auto-zero frequency and the usable frequency range of the amplifier. The bandwidth of the amplifier is independent of the correction frequency of the auto-zero circuitry. This is an important point that deserves mentioning again in a separate paragraph:
The gain-bandwidth product of an auto-zero amplifier is independent of its auto-zero correction frequency.
This can be easily understood by looking back to Figure 1. Notice the switching circuitry affects only the nulling amplifier, which is only used to correct the offset voltage of the device. The main amplifier is never switched out of the loop. At frequencies much higher than the clock frequency for the switches, the model simply looks like the main amplifier. The gain-bandwidth product of the auto-zero amplifier is the gain-bandwidth product of the main amplifier.
Another way to think about this effect is to consider the nulling amplifier as simply increasing the amplifier's dc open-loop gain while lowering the dominant pole by an equal amount, thus keeping the gain-bandwidth product the same. I find this a good way to think about the frequency response of auto-zero op amps.
Most precision op amps have a dominant pole anywhere from several Hertz to several hundred Hertz. Auto-zero op amps have an effective dominant pole on the order of milliHertz to sub-milliHertz. This means you only see the maximum possible open-loop gain of the amplifier with signals whose periods are on the order of 100 seconds or more. Again, the longer you wait, the more accurate the answer is.
The downside to using auto-correction circuitry is feedthrough of the switching clock. Small amounts of this clock can appear at the output of the amplifier mostly due to charge injection through the switches. Charge injection occurs due to gate-to-source and gate-to-drain parasitic capacitance in the MOSFET switches. This charge injection error is the primary limitation to increasing the auto-zero clock frequency to achieve higher levels of performance, as higher clock frequencies increase the amount of charge injection.
Clock feedthrough can be modeled as an input error signal source, so the amount of feedthrough to the output is proportional to the closed-loop gain of the amplifier's configuration. Unfortunately, this error source cannot be corrected as it is due to the correction circuitry itself.
Figure 4: Spectral analysis of output with +60dB of gain
Any amount of clock feedthrough will also cause a degree of intermodulation distortion in the presence of input signals. Intermodulation distortion (IMD) shows up on the output as spurious tones at both the sum and difference frequencies between the input and auto-zero clock. Similar to clock feedthrough and input noise, the level of IMD is proportional to closed-loop gain.
Figure 5: Example of intermodulation distortion with +60dB gain
The output will also contain energy at the harmonics of the auto-zero clock as well as IMD products from the input signal and these harmonics. The absolute level of IMD and clock-feedthrough is difficult to predict and could vary from device to device, even within the same manufacturer. Again, it is best to pay close attention to the datasheet to see the level of these effects.
There are several ways to deal with IMD in auto-zero op amps. The first solution is to do nothing at all. If input range of interest is only low frequencies and the clock feedthrough is sufficiently low, there may be no need to regard IMD as a serious source of error, as its products will be very close to the auto-zero clock, well outside of the application's bandwidth.
The second method is to use a feedback capacitor around the amplifier, creating a low pass filter as shown in Figure 6. The cutoff frequency (fc) of this filter is:
Of course, this cutoff frequency should be higher than the application's bandwidth as it will filter input signal as well. Figure 7 shows the performance improvement using this technique.
Figure 6: Using a feedback capacitor to reduce IMD
Figure 7: Spectral output using feedback capacitor to reduce IMD
Another technique for improving IMD involves modifying the auto-zero clock. By using a randomized clock frequency, as opposed to a fixed clock frequency, the clock feedthrough and IMD energy is randomized as well. If the auto-zero clock is optimized with a uniform probability density function, then IMD disappears. The noise floor of the amplifier will increase slightly, however. Conceptually, the energy in the IMD and clock feedthrough spurs is redistributed, smeared across all frequencies and showing up as broadband noise. Figure 8 shows the spectral output of the AD8572, which uses such a technique to eliminate IMD.
Figure 8: The AD8572 is optimized to eliminate IMD products
Intermodulation distortion and clock feedthrough are not the only side effects to using auto-zero correction. The charge injection from the auto-zero clock can create current spikes on the input bias current. The current by itself is not a major issue, but it does create a voltage spike at the input that is proportional to the size of the resistors used around the amplifier. For this reason, excessively large resistors are not recommended as it could cause excessive offset voltage spikes.
Applications for Auto-Zero Amplifiers
We've talked a bit on auto-zero technology, now we'll shift our focus on applications for these devices. The most obvious is any kind of precision or high gain application. Weigh scales have already been mentioned, but also strain gauges, pressure measurement, and other instrumentation apps.
Other applications that are often overlooked are ones that may not necessarily require absolute precision, but cannot tolerate any drifts over temperature. This is found in many automotive applications where there may be wide swings in operating temperature. The continuous correction of auto-zero amplifiers makes them a natural fit for such applications with the added benefit of having high precision as well. This same principle hold true for zero drift over lifetime circuits, such as remote sensors, where routine calibration is not practical.
Some applications require good ac performance, but also need low dc offset voltage. A precision current sensor is an example of one such application, as shown in Figure 9. Auto-zero amplifiers lend themselves well to such apps, but the intermodulation distortion of the op amp must be considered. Non-fixed correction frequency auto-zero op amps are the best choice here to minimize IMD. Again, the penalty for lowering IMD is an increase in the noise floor, lowering the overall resolution slightly.
Despite their internal complexity, auto-zero amplifiers can be configured and used in a similar fashion to traditional op amps. Older devices require the addition of one or two external capacitors which are used to store the offset voltage correction charge, shown as CM1 and CM2 in Figure 1. These capacitors are usually around 1uF, the optimum value should be determined from their datasheets. Low leakage capacitors should be used here as leakage current affects the op amp's correction ability and increases error.
More recent auto-zero amplifiers have these storage capacitors internal in the package, eliminating the need for external support components. From the user's perspective, they are configured identical to a standard op amp, making them pin-for-pin compatible replacements.
Figure 9: A precision current detector should use a randomized auto-zero op amp
Maximizing performance in high precision applications
Although you may have the highest precision amplifier in your circuit, you may not get the highest precision from your application. By using an auto-zero amplifier, the upper limit of precision may quickly become limited by the circuit design itself or even the layout used. Maximizing performance for a high precision application is an entire course in itself, but we will introduce a few ideas here to think about.
One potential source of offset voltage errors are thermoelectric voltages on the circuit board. This voltage, known as Seebeck voltage, occurs at the junction of two dissimilar metals and is proportional to the temperature of the junction. Seebeck voltage varies greatly depending on the two metals used, but it can be on the order of microvolts to millivolts, and can have a typical temperature variation from a few microvolts to tens of microvolts per degree Celsius.
The most common metallic junctions on a circuit board are solder-to-board traces and solder-to-component leads. If the temperature of the PC board at one end of a component is different from the temperature of the other end, the Seebeck voltages will not be equal, resulting in a thermal voltage error. Figure 10 shows an example.
In a high gain configuration, RF is much greater than R1 and the closed-loop gain is simply 1 + (RF / R1). A Seebeck difference across R1 appears as an offset voltage directly on the inverting input of the amplifier. This offset in turn gets amplified by the closed-loop gain, resulting in a substantially larger voltage at the output.
Figure 10: Using dummy components can minimize thermoelectric voltage errors
This thermocouple error can be reduced by using a dummy component to match the thermoelectric error source. In this example, we add a resistance in series with the non-inverting input, RS. The addition of RS does not affect the ac performance of the amplifier. Placing the dummy component physically close to its counterpart, R1, will ensure their Seebeck potentials are near equal, thus minimizing the thermocouple error. By setting RS equal to the parallel combination of RF and R1, error caused by input bias current into the amplifier inputs is minimized as well. Seebeck voltage differences across the feedback resistor RF are not as critical as this voltage error appears directly on the output and is not amplified by the closed-loop gain.
Another method for minimizing Seebeck error is to maintain a constant ambient temperature throughout the circuit board. This could become rather costly as it may require fans, temperature monitoring, or thermal insulation on the board. Using a ground plane on the PC board will help distribute heat throughout the board and has the added benefit of reducing EMI noise pickup as well.
Minimizing errors due to input bias and leakage currents is also key to achieving the highest degree of performance. Setting RS equal to the parallel combination of RF and R1 was already mentioned as a way to minimize error due to input bias current, but it does not address input offset current error. Older auto-zero amplifiers were constructed on bipolar processes, which have bias currents on the order of hundreds of nanoamps and offset current errors on the order of a few nanoamps. This translates into several microvolts of error if resistances larger than 1kW are used around the amplifier. More recent auto-zero amplifiers are designed using CMOS, which minimizes input bias and offset currents. However, the maximum supply voltage of these devices are somewhat lower than their bipolar counterparts. For higher supply voltages and minimum input bias current, JFET input auto-zero op amps should be considered.
The PC board surface should remain clean and free of moisture to avoid leakage currents between adjacent traces. Surface coating of the circuit board will reduce surface moisture and provide a humidity barrier, thus reducing parasitic resistance on the board. Using guard rings around the amplifier inputs can further reduce leakage currents. Figure 11 shows a example layout using a surface mount dual amplifier.
Figure 11: Using guard rings to minimize leakage current
The guard ring does not need to be a specific width, but it should form a continuous loop around both inputs. By setting the guard ring voltage equal to the voltage at the non-inverting input, both parasitic resistance and capacitance are reduced. For further reduction of leakage currents, components can be mounted to the PC board using Teflon standoff insulators.
The future of auto-zero technology
Modern auto-zero amplifiers are more sophisticated than their older cousins of even a few years ago, benefiting from design advances as well as improvements in process technology. Although the die size for these complex amplifiers is larger than for an equivalent standard amplifier, the design is scalable and can be reduced as process feature sizes shrink. This not only makes these precision amplifiers more affordable, it also allows them to become available in smaller packages, including the SOT23.
Design techniques have already improved to the point of not requiring external capacitors to store the correction voltage for the auto-zero amplifier. As designs continue to improve, clock feedthrough and intermodulation distortion will be reduced. The ultimate goal is to reduce all clock feedthrough energy below the level of the amplifier's noise floor, thus making these amplifiers appear in every sense equal to any traditional amplifier.
The concept of randomizing the auto-zero clock is still in its infancy, and there are many ideas that can be expanded and implemented in future designs. One idea is to shape the noise floor of the amplifier, lowing the noise floor at lower frequencies by shifting energy up to higher, out-of-band frequencies. This could potentially yield low noise, 1uV offset amplifiers that also have no 1/f noise. Throw in an extremely high gain-bandwidth product and slew rate and you start to approach the ideal op amp! An interesting concept that may not be that far in the future.