How important are the differences between sigmadelta and successiveapproximation architectures in choosing an analogtodigital (A/D) converter? They can often be an important factor in initiating the selection of a converter for a specific application. We describe here four major circuit architectures used in A/D converter (ADC) design and outline the role they play in converter choice for various kinds of applications. The descriptions are augmented by three examples that illustrate tradeoffs and issues associated with architectural considerations.
Though not detailed or exhaustive, this overview is intended to raise issues that should be understood when considering converters of different architectures. Sources of moredetailed information on converter architectures can be found in the References and will be made readily accessible by direct links at appropriate points. As one might expect in a survey of this kind, these descriptions are not comprehensive; and variations within each of the architecture families make generalizations less than fully accurate. Nevertheless, such generalizations are useful for the system designer to keep in mind when conducting a high level overview of a proposed system's requirements.
Converter Architectures
An overwhelming variety of ADCs exist on the market today, with differing resolutions, bandwidths, accuracies, architectures, packaging, power requirements, and temperature ranges, as well as hosts of specifications, covering a broad range of performance needs. And indeed, there exists a variety of applications in dataacquisition, communications, instrumentation, and interfacing for signal processing, all having a host of differing requirements.
Considering architectures, for some applications just about any architecture could work well; for others, there is a "best choice". In some cases the choice is simple because there is a clearcut advantage to using one architecture over another. For example, pipelined converters are most popular for applications requiring a throughput rate of more than 5 MSPS with good resolution. Sigmadelta converters are usually the best choice when very high resolution (20 bits or more) is needed. But in some cases the choice is more subtle. For example, the sigmadelta AD7722 and the successiveapproximations AD974 have similar resolution (16 bits) and throughput performance (200 ksps). Yet the differences in their underlying architectures make one or the other a better choice, depending on the application.
The most popular ADC architectures available today are successive approximations (sometimes called SAR because a successiveapproximations (shift) register is the key defining element), flash (all decisions made simultaneously), pipelined (with multiple flash stages), and sigmadelta (SD), a chargebalancing type. All A/D converters require one or more steps involving comparison of an input signal with a reference. Figure 1 shows qualitatively how flash, pipelined, and SAR architectures differ with respect to the number of comparators used vs. the number of comparison cycles needed to perform a conversion.
Flash Converters
Conceptually, the flash architecture (illustrated in Figure 2) is quite straightforward: a set of 2^{n}1 comparators are used to directly measure an analog signal to a resolution of n bits. For a 4bit flash ADC, the analog input is fed into 15 comparators , each of which is biased to compare the input to a discrete transition value. These values are spaced one leastsignificant bit (LSB=FS/2^{n}) apart. The comparator outputs simultaneously present 2^{n}1 discrete digital output states. If for example the input is just above ¼ of full scale, all comparators biased to less than ¼ full scale will output a digital ‘1’, and the others will output a digital ‘0’. Together, these outputs can be read much like a thermometer. The final step is to leveldecode the result into binary form.
Design Considerations and Implications: The flash architecture has the advantage of being very fast, because the conversion occurs in a single ADC cycle. The disadvantage of this approach is that it requires a large number of comparators that are carefully matched and properly biased to ensure that the results are linear. Since the number of comparators needed for an ^{n}bit resolution ADC is equal to 2^{n}1, limits of physical integration and input loading keep the maximum resolution fairly low. For example, a 4bit ADC requires 15 comparators, an 8bit ADC requires 255 comparators, and a 16bit ADC would require 65,535 comparators!
Pipelined Architecture
The pipelined (or pipelinedflash) architecture effectively overcomes the limitations of the flash architecture. A pipelined converter divides the conversion task into several consecutive stages. Each of these stages, as shown in figure 3, consists of a sample and hold circuit, an mbit ADC (e.g. a flash converter), and an mbit D/A converter (DAC). First the sample and hold circuit of the first stage acquires the signal. The mbit flash converter then converts the sampled signal to digital data. The conversion result forms the most significant bits of the digital output. This same digital output is fed into an mbit digitaltoanalog converter, and its output is subtracted from the original sampled signal. The residual analog signal is then amplified and sent on to the next stage in the pipeline to be sampled and converted as it was in the first stage. This process is repeated through as many stages as are necessary to achieve the desired resolution. In principle, a pipelined converter with p pipelined stages, each with an mbit flash converter, can produce a highspeed ADC with a resolution of n = p × m bits using p × (2^{m}1) comparators. For example, a 2stage pipelined converter with 8bit resolution requires 30 comparators, and a 4stage 16bit ADC requires only 60 comparators. In practice, however, a few additional bits are generated to provide for error correction.
Design Considerations and Implications: Pipelined converters achieve higher resolutions than flash converters containing a similar number of comparators. This comes at the price of increasing the total conversion time from one cycle to p cycles. But since each stage samples and holds its input, p conversions can be underway simultaneously. The total throughput can therefore be equal to the throughput of a flash converter, i.e. one conversion per cycle. The difference is that for the pipelined converter, we have now introduced latency equal to p cycles. Another limitation of the pipelined architecture is that the conversion process generally requires a clock with a fixed period. Converting rapidly varying nonperiodic signals on a traditional pipelined converter can be difficult because the pipeline typically runs at a periodic rate.
Successive Approximations
The successiveapproximations architecture can be thought of as being at the other end of the spectrum from the flash architecture. While a flash converter uses many comparators to convert in a single cycle; a SAR converter, shown in figure 4, conceptually uses a single comparator over many cycles to make its conversion. The SAR converter works like an oldfashioned balance scale. On one side of the scale, we place the sampled unknown quantity. On the other side, we place a weight (generated by the SAR and DAC) that has the value of ½ of fullscale and compare the two values. This first weight represents the most significant bit (MSB). If the unknown quantity is larger, the ½scale weight is retained; if the unknown quantity is smaller, it is removed. This series of steps is repeated n times, using successively smaller weights in binary progression (e.g., 1/4, 1/8, 1/16, 1/32, … 1/2^{n} of full scale) until the desired resolution, n, is attained. Each weight represents a binary bit, with the largest representing the most significant bit, and the smallest representing the least significant bit.
Design Considerations and Implications: A SAR converter can use a single comparator to realize a high resolution ADC. But it requires n comparison cycles to achieve nbit resolution, compared to p cycles for a pipelined converter and 1 cycle for a flash converter. Since a successiveapproximations converter uses a fairly simple architecture employing a single SAR, comparator, and DAC, and the conversion is not complete until all weights have been tested, only one conversion is processed during n comparison cycles. For this reason, SAR converters are more often used at lower speeds in higherresolution applications. SAR converters are also well suited for applications that have nonperiodic inputs, since conversions can be started at will. This feature makes the SAR architecture ideal for converting a series of timeindependent signals. A single SAR converter and an input multiplexer are typically less expensive to implement than several sigmadelta converters. With dither noise present, SAR and pipelined converters can use averaging to increase the effective resolution of the converter: for every doubling of sample rate, the effective resolution improves by 3 dB or ½ bit.
One consideration when using a SAR or pipelined converter is aliasing. The process of sampling a signal leads to aliasing  the frequencydomain reflection of signals about the sampling frequency. In most applications, aliasing is an unwanted effect that requires a lowpass antialias filter ahead of the ADC to remove highfrequency noise components, which would be aliased into the passband. However, undersampling can put aliasing to good use, most often in communications applications, to convert a highfrequency signal to a lower frequency. Undersampling is effective as long as the total bandwidth of a signal meets the Nyquist criterion (less than onehalf the sampling rate), and the converter has sufficient acquisition and signal sampling performance at the higher frequencies where the signal resides. While fast SAR converters are capable of undersampling, the faster pipelined converters tend to be more effective at it.
SigmaDelta
The sigmadelta architecture takes a fundamentally different approach from those outlined above. In its most basic form, a sigmadelta converter consists of an integrator, a comparator, and a singlebit DAC, as shown in Figure 5. The output of the DAC is subtracted from the input signal. The resulting signal is then integrated, and the integrator output voltage is converted to a singlebit digital output (1 or 0) by the comparator. The resulting bit becomes the input to the DAC, and the DAC's output is subtracted from the ADC input signal, etc. This closedloop process is carried out at a very high "oversampled" rate. The digital data coming from the ADC is astream of ones and zeros, and the value of the signal is proportional to the density of digital ones coming from the comparator. This bit stream data is then digitally filtered and decimated to result in a binaryformat output.
Design Considerations and Implications: One of the most advantageous features of the sigmadelta architecture is the capability of noise shaping, a phenomenon by which much of the lowfrequency noise is effectively pushed up to higher frequencies and out of the band of interest. As a result, the sigmadelta architecture has been very popular for designing lowbandwidth high resolution ADCs for precision measurement. Also, since the input is sampled at a high "oversampled" rate, unlike the other architectures described in this paper, the requirement for external antialias filtering is greatly reduced. A limitation of this architecture is its latency, which is substantially greater than that of the other types. Because of oversampling and latency, sigmadelta converters are not often used in multiplexed signal applications. To avoid interference between multiplexed signals, a delay at least equal to the decimator's total delay must occur between conversions. These characteristics can be improved in sophisticated sigmadelta ADC designs by using multiple integrator stages and/or multibit DACs.
Application Examples:
The following three examples illustrate some of the issues described above.
Example 1: Multiple inputs, 16bit resolution
Consider an application that requires 16bit resolution for 4 independent signals with bandwidths of dc to 15 kHz, 15 kHz, 15 kHz, and 45 kHz. The total throughput required to sample these signals under the Nyquist criterion is (2 × 15 + 2 × 15 + 2 × 15 + 2 × 45) ksps = 180 ksps. At first glance, the SARtype AD974, the sigmadelta AD7722, and the pipelined/sigmadelta AD9260 all have the required total throughput capability. But, as has been discussed above, the inherent latency of sigmadelta converters limits their effective throughput when they must continually acquire new signals by multiplexing. Effective multiplexed throughput can be defined as the total throughput of a converter when two or more independent signals are multiplexed. Table 1 compares the total throughput and effective throughput for each converter and indicates the number of converters of its type that would be needed to serve in this application:
Table 1
Converter  Architecture 
Total Throughput (14bit resolution) 
Effective Multiplexed Throughput 
Converters needed for application 
AD974  SAR  200 ksps  200 ksps 
1 
AD7722  SigmaDelta 
195 ksps 
2.3 ksps 
4 
AD9260 
Pipelined/SigmaDelta 
2500 ksps 
75 ksps 
3 
Whether converting a single input or several multiplexed inputs, the AD974 achieves a throughput rate of up to 200 ksps. Since the application requires a total throughput of 180 ksps, the AD974's performance is sufficient. In fact, this is exactly the type of application that the AD974 was designed for: in addition to the SAR converter and reference, it also contains an integrated 4channel multiplexer.
The AD7722 and AD9260 both face the challenges confronted by sigmadelta converters in multiplexing several inputs. The AD7722's throughput is 195 ksps when sampling a single signal, but it drops to just 2.3 ksps when converting multiple signals, due to the settling time which results from oversampling and filtering. To use the AD7723 in this application, four converters (one per channel) would be needed.
The AD9260 combines pipelined and sigmadelta techniques. Its throughput rate of 2.5 Msps makes it ideal for higher throughput single channel systems. But in this application, its settling time of 13.35 µs limits its effective throughput to 75 ksps. To use the AD9260 in this application would require at least 3 converters. Note that if the AD9260 were purely a pipelined flash converter, a single converter would have had the required throughput, assuming that the inputs are periodic.
Example 2: Single input, 16bit resolution
Consider now an application that converts a single 90kHz bandwidth input at 16bit resolution. In this case, all three converters from the first example would work well. Here the choice among converters would be made on other considerations, including ac and dc performance, systemlevel considerations (e.g. is there a great benefit to the antialias performance of sigmadelta converters in this application), latency, and cost.
Example 3: Multiple inputs, 14bit resolution
Consider an application in which 16 inputs, each with a dc to 100 kHz bandwidth, are converted with a resolution of at least 14 bits. Three converters suitable for this application include the SARtype AD7865, the sigmadelta AD7722, and the pipelined AD9240. The total throughput required under the Nyquist criterion is 2 × 100 kHz × 16 = 3.2 MSPS. Table 2 shows the throughput for each converter.
Table 2
Converter  Architecture 
Total Throughput (14bit resolution) 
Effective Mux'ed Throughput 
Converters needed for application 
AD7865  SAR  416 ksps  416 ksps 
8 
AD7722  SigmaDelta 
220 ksps 
2.3 ksps 
16 
AD9240 
Pipelined 
10 Msps 
10 Msps  1 
Of the three converters, only the AD9240 has the throughput needed to convert all 16 channels. The AD7865 has sufficient throughput for 2 inputs per converter. To use the AD7865 in this application, 8 converters would be needed. The AD7722 would need to be used in a converterperchannel implementation; thus 16 converters would be required.
Summary
While not exhaustive, the following table summarizes and ranks (in a generalized sense) the relative advantages of flash, pipelined, SAR, and sigmadelta architectures. A rank of 1 in a performance category indicates that the architecture is inherently better than the others in that category are. An * indicates that the architecture has the capability or characteristic listed.
Table 3
Characteristic  Flash  Pipelined  SAR  SigmaDelta 
Throughput  1  2  3 
4 
Resolution (ENOB)  4  3 
2  1 
Latency  1  3 
2  4 
Suitability for converting multiple signals per ADC  1  2  1  3 
Capability to convert nonperiodic multiplexed signals  1  2  1  3 
Simplified antialiasing  * 

Can undersample  *  *  *  
Can increase resolution through averaging
(with dither noise)

*  *  * 
参考資料
AnalogDigital Conversion Handbook, The Engineering Staff of Analog Devices, Inc. Englewood Cliffs, NJ: Prentice Hall, 1986. ISBN 0130328480. Available from Analog Devices: Phone (781) 4613392.
Practical Analog Design Techniques, edited by Walt Kester. Norwood, MA: Analog Devices, Inc., 1995. ISBN 0916550168. Available free on the Internet as PDF chapters.
High Speed Design Techniques, edited by Walt Kester. Norwood, MA. Analog Devices, Inc., 1996. ISBN 0916550176. Available free on the Internet as PDF chapters.
Linear Design Seminar, edited by Walt Kester. Norwood, MA: Analog Devices, Inc., 1995. ISBN 091655015X.
謝辞
The author is indebted to Alain Guery and Mike Coln for providing valuable direction, and for providing the figures shown.