Sunteți pe pagina 1din 8

Evolution of accuracy.

Konrad Hejn and Andrzej Pacut

he output spectrum of an analog-to-digital converter (ADC) stimulated with sine wave is useful in ADC testing. An inspection of the spectrum enables detection of various distortions of the ADC, like spurious tones, harmonics, or high noise floor. Both technological improvements and sophisticated DSP cells can correct the imperfections of an ADC. An analog filter can decrease some spurious tones and unwanted harmonics. A two-dimensional lookup table of corrections helps to decrease the effects of the integral nonlinearity in off-the-shelf ADCs. A pseudorandom identification combined with a bank of correlators can improve the differential nonlinearity of two-pass ADCs. We need objective quality measures to assess practical values of these techniques, which aim at reducing the effects of nonlinearities. One such objective measure is provided by the effective resolution (EFR) and, equivalently, by the effective number of bits (ENOB), equal to log 2 (EFR). We will concentrate on the EFR.

Usefulness of the EFR Concept


To illustrate the problem, we analyze the output spectrum S of an off-the-shelf 8-bit ADC clocked at fs = 0.5 MHz (Figure 1). We use a spectrally clean sine wave of a frequency f0 = 49,545.3 Hz as the test signal. The sampling we apply is always coherent,

2001 ARTVILLE LLC. AND COMSTOCK INC

48

IEEE Instrumentation & Measurement Magazine


1094-6969/03/$17.002003IEEE

September 2003

namely, the measurement window T contains an integer number N of sine-wave periods and an integer number M of samples. In other words, T= M N = . fs f0

(1)

If N and M are not relatively prime, then the same phase angle becomes visited more than once, hence enabling averaging of random additive measurement errors. We will always assume that M and N are relatively prime since it leads to the shortest sampling record. Note that, in this case, each sample corresponds to a different sine-wave phase angle, i.e., the angle distance between consecutive samples is equal to 2 / M. We tested the same 8-bit ADC in four different setups. In the first one, we stimulated the ADC with a full-scale sine wave, i.e., of the amplitude equal to the half of ADCs range [Figure 1(a)]. Clearly visible are spurious tones (note the one S[dBc] N = 3,247, M = 32,768, fs = 500 kHz, f0 = 49,545.3 Hz 0 near 161 kHz) and the second, (a) third, and fourth harmonics. The latter may be the result of 51.3 51.7 either the integral nonlinearity 58.9 61.7 of the ADC or the sine-wave 79 dB generator imperfections. In the next setup, we decreased the sine-wave amplitude by half [Figure 1(b)]. A margin between the fundamental component and the second harmonic has increased by 10.8 dB. The margin between the fundamental component and the fourth harmonic has also increased, but only by 4.4 dB. However, the third harmonic behaves differently: the margin has not increased but even decreased by 1.2 dB. This suggests that the even harmonics may be caused by even integral nonlinearity of the ADC. They decrease together with the fundamental component (soft noises [1]). The third harmonic and the noise floor behave in a more stiff way (hard noises). Spurious tones go in between. To examine this hypothesis, in the third setup we applied a band-pass filtering to the sine wave, often advocated as a way to eliminate some harmonics and spurious tones [Figure 1(c)]. The third harmonic has now decreased by 19.7 dB and the spurious tones decreased
100 (b) 0

by 27.1 dB. Accordingly, several aliasing components have also disappeared. However, the second and fourth harmonics have not changed, suggesting that they are contributed by the ADC integral nonlinearity. Small integral nonlinearity is welcome in digital receivers and spectrum analyzers. It is suggested that the integral nonlinearity of an ADC can be decreased with the use of a lookup table [2]. To test the effectiveness of the former method, we applied a code-slope space corrector to the digital output in our fourth setup [Figure 1(d)]. Comparison with the original full-range stimulation spectra [Figure 1(a)] shows that the even harmonics are smaller, the second of order of 18.6 dB and the fourth of order of 18.4 dB. The remaining distortions form a measured noise floor, contributed by the quantization process, the differential nonlinearity, and by thermal noise the last effect is negligible in our experiment.

62.1

57.7

45.6 66.1

100 (c) 0

161.000

51.3 61.1 79 dB 100 0 78.6 78.8

(d)

69.9 79 dB 100 0

79.9

80.2

80.1

49.5453

99.0906

148.6359

198.1812

250 f [kHz]

Fig. 1. Output power spectrum of the same 8-bit ADC clocked at f s = 0. 5 MHz and stimulated with the spectrally pure sine-wave of frequency f 0 = 49, 545. 3 Hz in four experimental setups. Vertical axes are scaled in dBc units, which guarantees that the power of the fundamental (carrier) component always corresponds to zero decibels. (a) Amplitude of the sine-wave is equal to the half of ADCs full-range. In this way, all the switching points of the ADC characteristic are scanned. (b) Sine-wave amplitude is equal to the quarter of ADCs full range. Here only about half of the switching points are scanned. (c) The amplitude is as in (a) but the sine-wave is first filtered by the bandpass filter. The thin line exhibits its amplitude response. (d) Like in (c) but the ADCs output is corrected with the use of a lookup table.

September 2003

IEEE Instrumentation & Measurement Magazine

49

The noise floor determines the lowest input power level that can be reliably detected at the ADC output. It evidently limits the ultimate ADC sensitivity to weak inputs, since any input whose power is below the noise floor is difficult to recover. Theoretically, for an L-bit perfect ADC, its inherent (reference) noise floor (INF), i.e., the one resulting solely from the perfect quantization process, can be evaluated as [3] INF = ( 6.02L + 176 + 10 log 10 J 10) [dBc] . (2)

e 2 calculated on a base of the output {y} = {y( m), m = 1,K , M}. One may thus define EFR = R
def

record

Ee 2 e2 (4)

where J depends on the type of quantizer. For quantizers without dead zone J = M / 4, and for rounding (mid-tread) quantizers (Figure 2) J = M / 2, where M is the number of samples. Since the exact characteristic of our quantizer was not known, we calculated the INF for both types of quantizers. The record length in our experiment was equal to M = 32,768 samples; hence, in case of the quantizer without the dead zone we obtain INF 79 dBc, which is 9 dB less than the value in Figure 1(d) and for the rounding quantizers INF 82 dBc, which is 12 dB less than the one in Figure 1(d). This proves that a differential nonlinearity exists and should be corrected. The correction effects can be evaluated with the EFR.

where R denotes the nominal resolution and with the assumption that the number of samples is sufficiently high to neglect random fluctuations of e 2 . The nominal resolution can be replaced by 2L where L denotes the nominal number of bits L = log 2 R of the ADC. The above definition proposed by Ochs [4] years ago stimulated extensive research in the area of the ADC testing. The IEEE standard [5] integrates most of the practical results obtained to date and gives a definition of EFR that is a simplified version of (4). The simplification Ee 2 = 1/ 12 leads to EFR ieee =
def

(5)

R 12e 2 (6)

Definition of EFR
Assume that the test signal u(t ) applied to the input of the ADC is stable in time and with temperature. Typical test signals include sine waves, triangle waves, and narrow-band random noises. In this article, we concentrate on sinusoidal test signals as they are easy to generate with high spectral purity. The output signal y( m) is digital, namely both sampled with a period t and quantized with a step Q. These digital data are the only source of measurement information. To access the input signal directly we would need another ADC, of much better specification than the one under test. We use Widrows model to define the discrete-time quantization error e as a difference between the quantized output y and the sampled input signal u at the same moment m, namely (Figure 2) e( m) = y( m) u( m). (3)

and produces biased estimates of EFR. In some cases, the bias is so high that the actual quantizer might even be evaluated as better than the perfect quantizer. On the other hand, the simplification allows simple calculations, as the IEEE definition seemingly does not depend on the test signal parameters. In reality, Ee 2 does depend on parameters of the sine-wave test signal that are only approximately known. The exact actual value of these parameters must be assessed on the basis of the digital output signal. Below, we discuss this problem in detail.

Digital Sine-Wave Fitting


Consider a continuous sine-wave test signal ua ua (t ) = C + V cos( 2f0t + P0 ), (7)

Following Widrow, we assume that the uncertainty expressed by e can be modeled by a random variable that we further denote by e. Its theoretical rms value Ee 2 expresses the inherent noise floor INF of the perfect ADC and remains constant during a test. To assess the quality of an ADC, one may compare this reference rms value with the actual rms value of the ADC under test. The latter is unknown and must be replaced by its empirical value

where f0 is the frequency, P0 is the initial phase, C denotes the dc offset, and V is the signal amplitude. Suppose that the sine wave is sampled at a known rate of fs = 1/ t, and the sampling begins at a moment t 0 , thus giving the sampling times t 0 + ( m 1)t , m = 1,2. The resulting sampled sine wave forms a sequence {u} = {u( m), m = 1,2,K } with u( m) = C + V cos( 2f0 (t 0 + ( m 1)t ) + P0 ) = C + V cos( Fm + P ) (8)

where m is the discrete time corresponding to the analog time t = t 0 + ( m 1)t, P = P0 + 2f0 (t 0 t ) is the discrete initial phase, and F = 2 f0 fs (9)

(10)

Fig. 2. Widrows model of the perfect uniform quantizer of a rounding (mid-tread) type.

is the positive rational discrete frequency that reciprocal can be interpreted as a number of samples taken in one sine-wave period. Note that four parameters, namely C,V ,P and F, fully

50

IEEE Instrumentation & Measurement Magazine

September 2003

characterize the sampled signal. The dc offset C and amplitude V are unchanged by sampling. The sampled sine wave constitutes the input to a quantizer, hence the output of the quantizer can be written in the form

{y} = Q{u}

(11)

where Q denotes a quantization process with a known quantization step Q. The sequence {y} = {y( m), m = 1,2,K } represents the sampled and quantized output, thus forming the digital output. The quantization process does not change either the discrete initial phase P or the frequency F, called later the digital initial phase and the digital frequency, respectively. However, the quantization process influences dc offset C and amplitude V. The sine-wave parameters C and V must be assessed on the basis of the digital data since the analog signal ua is inaccessible. Because of the quantization error (3), the digital output data only approximates the sampled sine-wave. Consequently, calculation of the dc offset C and the amplitude V through a sine-wave fitting of the digital (output) data leads to biased estimates of the analog sine wave [6]. The curve the data must be fitted to is a sine wave distorted by the quantization processQ (see Figure 3). The modified fitting procedure we propose can be understood as a regular sine-wave fitting preceded by a nonlinear mapping of the digital data to the input. This mapping is, in a sense, an operation inverse to the quantizationQ 1 (see Figure 3) and can be derived on the basis of the analytical reference model of the quantization process Q [7]. The next section is devoted to such a model.

calculation of the basic statistics of the quantization process for any input signal with an arbitrary required precision. To give an example of the results, we now focus on the sine-wave test signal (8) and consider the uniform quantizer of rounding (mid-tread) type (Figure 2) for which the alternative Widrows formula exists [11]. The mean value y and the mean square value y2 of the digital output {y} are in this case given by Ey =
1 K0 + K1 1 arcsin k 2 k=K0 +1

(12)

Ey 2 =

2 2 1 K0 + K1 2 k 2 k=K0 +1

1 arcsin k . 2

(13)

The mean Ee and the mean-square value Ee 2 of the quantization error Ee are given by Ee = Ey C Ee 2 =
2 2 V 2 ( C K0 ) ( C K1 ) + + 2 2 2

(14)

1 2V k arcsin k + k = K 0 + 1

1 2 k

(15)

where we denoted 1 K0 = floor C + V 2 1 K1 = floor C + + V 2 1 1 k = k C 2 V

(16)

Model of the Uniform Quantizer


Electronic instrumentation allows for better precision in control of the signal frequency ratio f0 / fs (10) than in measurement of the sampled values y( m) themselves (11). A frequency ratio control is more reliable than digital data processing. Nevertheless, out of the sine-wave parameters, the digital frequency F is the only one that can be precisely estimated on the base of the analog signals. The remaining parameters, namely C,V , and P , must be evaluated on the basis of the digital signal. To be able to derive the estimators without the need for estimating P, we assume that P is modeled by a random variable P uniformly distributed over an arbitrary interval of length2. Mathematical models of various types of static quantizers, including nonuniform and nonmonotonic ones, for deterministic as well as stochastic input signals, are derived in [8]. We used there a direct approach that follows Grays ideas [9] rather than the characteristic functions based on Widrows approach [10]. There are two advantages of the direct approach. First, it can be applied for nonuniform or nonmonotonic quantizers, as opposed to Widrows method that is valid only for uniform quantizers. Second, the direct approach leads to finite-term sums, rather than to infinite sums resulting from Widrows approach. These infinite sums may produce large finite-term approximation errors [7]. Use of the direct approach enables

and floor represents rounding down to the nearest integer. Equations (12)-(15) depend on C and V through K0 , K1 , and k (16). The above formulas do not depend on the digital initial phase P and the digital frequency F. All sums are finite and no approximations are necessary. We distinguish between averaging, which is an actual operation on the measurements, and calculation of the expected value equivalent to the averaging in the probabilistic space. Certainly, if the law of large numbers can be applied, the averages will asymptotically tend to the expected values 1 M 1 M y( m) M y( m) = y = g( C,V ) M M m= 1 m= 1

Ey = lim

(17)

Fig. 3. The quantizers output y is not a sine wave, hence the mean-square $ fitted sine-wave parameters C 0 , V$0 are biased. To obtain a discrete-time sine $ wave with parameters identical to C, V, the signal y must be transformed by a
transformationQ 1 determined with the use of the analytical model of the quantizer.

September 2003

IEEE Instrumentation & Measurement Magazine

51

Ey 2 = lim

1 M 2 1 M 2 2 y ( m) M y ( m) = y = h( C,V ). M m= 1 m= 1 (18)

0.1

0.1

1.5

2.5

3 (a)

3.5

4.5

1.5 1.4 normalized mean square error 1.3 1.2 1.1 1 0.9 0.8 0.7 0.6 1 1.5 2 2.5 3 (b) 3.5 4 4.5 5

If the number of samples is finite but sufficiently large, then the difference between the average value and the expected value will be arbitrarily close to zero. Here we always neglect the error resulting from the finite sampling record, thus identifying Ey and y, as well as Ey 2 and y2 . The theoretical values of the mean and mean square of the quantization error [(14) and (15)] as well as the average and average square of the output (12) and (13) cannot be expressed analytically in the closed form [11]. On the other hand, they can be easily derived numerically. Here we show only some plots to sketch a dependence of mean value Ee = Ey C, normalized mean square value 12Ee 2 , and relative output powerEy 2 / Eu 2 on parameters C and V (Figure 4). All presented curves show a damped periodicity, with the envelopes smoothly shrinking with V. They all are continuous with respect to C and V yet their partial derivatives have singularities. In fact, the limit of Ee 2 for V is equal to 1/12, applied in the IEEE norm, i.e., 1/12 is only an approximation of the exact value. While this fact has been already accepted [12], the evaluation of the dc offset C and the amplitude V has been unclear since these values must be assessed on the basis of the digital signal. The right-hand sides of (12) and (14) differ only by the dc offset C. The difference between the right-hand sides of (13) and (15) is yet more complex since one must take into consideration the cross correlation between u and e.

mean error

1.4 1.3 relative output power 1.2 1.1 1 0.9 0.8 0.7 0.6 1 1.5 2 2.5 3 3.5 4 4.5 5

$ $ To obtain the correct estimates C and V one must solve numerically the system of (12) and (13) taking (17) and (18) into account, with respect to C and V, namely G( C,V ) y g( C,V ) = 0 H( C,V ) y2 h( C,V ) = 0 (19) (20)

The Algorithm

sine-wave amplitude V (c)

Fig. 4. Quantization statistics (all normalized to Q = 1) for the perfect uniform quantizer. (a) Mean value Ee = E y C versus the sine-wave amplitude V the dc offset C = 0: 0. 02: 1 serves as a parameter. (b) Normalized mean square value : 12Ee 2 versus the sine wave amplitude V with the dc offset C = 0: 0. 011 as a
parameter. (c) Relative output power y 2 / u 2 versus the sine-wave amplitude V for the dc offset C = 0: 0. 011 as a parameter. :

Both G( C,V ) and H( C,V ) are continuous in C and V, yet some of their partial derivatives may be singular [6], namely discontinuous or infinite on certain lines. The singularities are caused by changes in the number of summation terms in (12) and (13). The set D of the singularities forms a rectangular grid of unit grid length, rotated by 45 with respect to the coordinate axes in the C V plane. The existence of derivative singularities may create difficulties in numerical maximization procedures. Robustness of the numerical procedure seems to be of prime importance. To solve the problem one may apply MATLABs function fsolve, which finds the root of a system of nonlinear equations. The function cv = fsolve(yfun, cv0, avary, avary2) starts at cv0 and attempts to solve the equations for avary=y and avary2=y2 . While the choice of an initial value is not important, we may suggest some possibilities. The initial estimate C0 of the dc offset C may be equal to the average of the maximum and the minimum values of the

52

IEEE Instrumentation & Measurement Magazine

September 2003

digital record, i.e., C0 = ( y max + y min ) / 2. Another possibility is to take the average over an integer number of cycles (1), namely C0 = 1 M y( m). M m= 1

Note that while the phase P belongs together with the digital frequency F to the time axis parameters, it cannot be easily measured practically on the analog side of ADC. Phenomenon of the fractional delay makes the necessary hardware too expensive.

(21)

A reasonably accurate initial estimateV0 of V is equal to the half of the digital data span, namely V0 = ( y max y min ) / 2. Another possibility is based on the histogram of the probability density function of the digital codes. Yet other choices based on DFT or LS algorithms both produce exactly the same result V0 = where s= 1 M $ y( m) sin Fm M 1 s 2 + c2 (22)

Measurement Results
Our measurement experiment was designed to capture a data record at the output of a L-bit ADC. We repeated the experiment K times (say, K = 1,000) with slightly different values of the dc offset C and the amplitude V, with variations comparable to the analog quantum Q. These variations can be in part intentional, e.g., modified by generator settings, and in part natural, e.g., modified by slow thermal processes appearing in the ADC under test. Ideally, the results of the EFR measurement would be identical, regardless of the particular dc offset C and amplitude V of the test signal (7). Variations of C and V may cause some variations in the EFR measurements. These variations can be visualized in the histogram, which, in the ideal case, would have the Kronecker delta look, with the delta placed at the point corresponding to the nominal resolution R = 2L . A high-quality ADC with programmable differential nonlinearity DNL would be very welcome for testing; only then we can assess the moment when the quantization error correction begins to be dominated by hard noises resulting from the differential nonlinearity. A typical off-the-shelf ADC usually guaranties DNL < 0.5, which is too large for our purposes. An ad hoc solution that we applied here is to analyze only certain most significant bits (MSB) of the digital words captured, namely the first eight, first six, first four, and so on. The EFR measurement results should then be ideally equal to, or slightly less than, 256, 64, 16, respectively. Ignoring several least significant bits is equivalent to decreasing the differential nonlinearity at the cost of lower resolution. Figure 6 shows eight selected histograms of the EFR that have been obtained for the folding flash ADC (TLC5510: 8 bit/20 MHz) (top plots) and for the multipass flash ADC

( )

(23)

c=

1 M $ y( m) cos Fm . M 1

( )

(24)

We recommend using the balance of the power principle to obtain the appropriate formulas from Parsevals theorem, namely V0 = 2( y2 y 2 ). (25)

The initial estimate V0 obtained from (25) is always slightly larger than the one resulting from (22), due to the power balance. All the above choices produce the initial points ( C0 ,V0 ) not far from the desired solution ( C,V ). Listings of MATLABs m-files that evaluate functions (19) and (20) and, if required, their four-element Jacobian, and draw the trace of iterations are presented in [13]. Discontinuities of the derivatives do not interfere with fsolve, which is very welcome. This may be related to the fact that the derivative of G( C,V ) with respect to C and the derivative of H( C,V ) with respect to V (although singular) are both of constant sign, independent of the value of the remaining arguments. However, the number of iterations increases if the desired solution becomes closer to the grid of derivative singularities on the C V plane (see Figure 5) and the number of iterations may become quite large. Fortunately, such a situation is unlikely in practice.

The Initial Phase


Finally, we propose a method to estimate the phase P. We minimize the error index e2 = $ $ $ y( m) C V cos(Fm + P )
M 2

m= 1

(26)

$ $ $ with respect to P, where C,V, and F are known or already calculated. If (26) has a minimum, then e 2 s $ = 0 P = arctan . P c

(27)
Fig. 5. Number of iteration depends on the solution location on the C V plane; darker area corresponds to higher number of iterations.

This formula is a well-known result, typically being obtained from DFT.

September 2003

IEEE Instrumentation & Measurement Magazine

53

(TLC876C: 10 bit/20 MHz) (bottom plots). Both the classical IEEE method as well as the analytical method have been applied. Note that the same data are needed for both methods. The proposed analytical method behaves very well when only the six MSB of the captured words are analyzed (for 10-bit ADC), see bottom left plots in Figure 6. In this case, the maximum differential nonlinearity is very small (DNL < 0.03125), hence the 6-bit converter under test can be treated as a perfect one. Consequently, the corresponding histogram shows very small empirical variation and the average effective resolution EFR = 63.12 is almost equal to the nominal resolution R = 64. On the contrary, the histogram obtained with the IEEE method is smeared over a wide range, thus giving nonrepeatable results; some of them are exceedingly incorrect, such as for EFR > 64. For the 8-bit ADC, the situation is slightly different; see top left plots in Figure 6. As compared to the previous case, the maximum differential nonlinearity is approximately four times higher (DNL < 0.125). Consequently, average EFR = 62.01 is smaller than the nominal resolution R = 64 and, while the results look sensible, the histogram corresponding to the new analytical method is smeared over some interval. This is caused mainly by the influence of a higher dif-

ferential nonlinearity. The results obtained by the IEEE method are again unusable and hardly repeatable, and again it happens that EFR > 64. For the 10-bit ADC, the differential nonlinearity (DNL < 0.125) is similar as in the former case, yet the new method exhibits only slightly better results than the classical IEEE method; see the bottom right plots in Figure 6. Both histograms are spread over similar intervals. This means that the 10-bit converters specification seems to be rather optimistic. A relative decrease in EFR is six times higher as compared with the 8-bit converter of the same nominal differential nonlinearity (top left plots in Figure 6). In the case of the differential nonlinearity of order of 0.5 LSB (that is, 50%), both methods produce comparable yet unsatisfactory results; see the top right plot in Figure 6. Low repeatability of EFR measurement is not caused by the hardware instability, which could have a remarkable influence only for more precise converters. The total measurement error introduced by the measurement setup was significantly smaller than the error of the classical IEEE method introduced by the estimation bias. In the case presented in the top right plots, the differential nonlinearity totally dominates the effect of better estimates and masks the effects created by the bias of the quantization error. Note finally that both methods lead to almost identical average EFR, but the standard deviation S for the new analytical method is always smaller, and sometimes considerably smaller. In the experiments reported in the bottom left plots in Figure 6, the standard deviation was 21 times smaller for the analytical method, as compared to the IEEE method.

Final Remarks and Conclusions


Before we derived (12) and (13), we designated a heuristic two-point method [14] based on (14) and (15), which employs a symmetry of (14). The EFR measurement results agreed quite well with simulations, but their repeatability for precise ADCs were still not satisfactory. We then tried to employ the model (12) and (13) through a lookup table, yet the time and memory needed were too restrictive. Further theoretical investigation [6], [7] resulted in derivation of (12) and (13) and the new analytical method detailed in this article.

Fig. 6. EFR histograms. Each histogram has been calculated on the base of K = 1000 experiments to show variability , of the EFR of (a) the folding flash ADC (TLC5510 8 bit/20 MHz), and of (b) the multipass flash ADC (TLC876C 10 bit/20 MHz) calculated with the proposed analytical method (thick lines) and the classical IEEE method (filled areas). For both converters, the 6 MSB (left plots) and 8 MSB (right plots) ADCs were analyzed, corresponding to DNL < 0. 125 and DNL < 0. 5 for (a) 8-bit converter, and DNL < 0. 03125 and DNL < 0. 125 for (b) 10-bit converter. Here EFR is the average EFR while S denotes its empirical standard deviation.

54

IEEE Instrumentation & Measurement Magazine

September 2003

As we showed here, the basic source of the erroneous behavior of the classical IEEE method for the EFR evaluation is the assumption of the uniform distribution of the quantization error and the lack of cross correlation between the input and the quantization error. There are two consequences of this fact: first, fitting of the input signal is based on oversimplified estimates of C and V, and second, these estimates lead to wrong value of the theoretical rms value Ee 2 in the numerator of the definition (4). The method, recommended currently by the IEEE for sine-fit testing, evidently leads to biased EFR measurement, especially for ADCs of high precision (in the sense of small differential nonlinearity). In other words, EFR measurements are marred by systematic errors. This happens even if the sine-wave source has a good short-term stability and is filtered to remove harmonics and spurs. We have shown examples of the erroneous behavior of the classical IEEE method and compared it to the results obtained for the new analytical method, proposed in this article. The efficiency of the new analytical method results from the exact fitting of the input test signal (7) and leads to repeatable results, independent of measurement conditions (including a temperature drift of the dc offset C and amplitude V). We must stress that the EFR, as a characteristic of the ADC, is useful if the differential nonlinearity does not exceed 20%. If the nonlinearity reaches 50% or more, the EFR shows strong random variations and should rather be specified through its distribution. We verified the quality of the proposed analytical method and the IEEE method in hundreds of thousands experiments. The analytical method always produces smaller bias and smaller variance of the EFR measurements, thus serving as a single quality index that enables to compare objectively the ADCs. Moreover, the analytical method produces results that agree with the theory of the perfect ADC. The proposed method is addressed to designers and users of computer-aided mixed-signal electronic systems. We also advocate here for including the analytical method into the text of the current norm [5], hence substituting the sine-fitting procedure based on DFT and/or LS algorithms. The analytical method requires one to solve a system of two nonlinear equations (19), (20) for parameters y and y2 obtained from measurements. The existing numerical procedures easily enable solving such equations. Moreover, (21) and (25) can be used as the initial conditions for the numerical procedure, which are easy to evaluate and lie close to the exact solution. ADCs have always been bottlenecks in electronic systems. Recently, the ADCs are often integrated on a semiconductor substrate with other mixed-signal circuits. This creates a need for clever strategies for error measurement and correction. The analytical method allows implementation of the necessary test on-cell. The main challenge is to produce a high-quality sine wave on the board. The parameters (17), (18), (23), (24) can be easily (in real time) accumulated by a typical digital circuitry.

References
[1] H. Hilton, A 10-megasample-per-second analog-to-digital converter with filter and memory, H-P J., vol. 44, pp. 100-104, Oct. 1993. [2] D. Hummels, Linearization of ADCs and DACs for all-digital wide-bandwidth receivers, in Proc. IWADC99, 4th Int. Workshop on ADC Modelling and Testing, Bordeaux, France, Sept. 9-10, 1999, pp. 145-152. [3] M. Mahoney, DSP-Based Testing of Analog and Mixed-Signal Circuits. Los Alamitos, CA: IEEE Computer Society, 1987, chaps. 2, 4. [4] L. Ochs, Measurement and enhancement of waveform digitizer performance, in Proc. IEEE ELECTRO 76 Professional Program, May 11-14, 1976. [5] IEEE Standard for Terminology and Test Methods for Analog-to-Digital Converters, IEEE Std 1241-2000, 2000. [6] K. Hejn and A. Pacut, Improved definition of ADC effective resolution, Comput. Stand. Interfaces, vol. 23, no. 2, pp. 137-147, May 2001. [7] A. Pacut and K. Hejn, Reference properties of uniform quantizersComparison of Widrows and direct approaches, Comput. Stand. Inerfaces, vol. 25, no. 25, pp. 3-13, Mar. 2003. [8] A. Pacut and K. Hejn, Analog-to-digital converters: Towards a generalization of Widrows theorem, in Where Instrumentation Is Going, Proc. IEEE IMTC98, St. Paul, MN, May 18-21, 1998, vol. 2, pp. 1190-1197. [9] R.M. Gray, Quantization noise spectra, IEEE Trans. Inform. Theory, vol. 36, pp. 1220-1244, Nov. 1990. [10] A.B. Sripad and D.L. Snyder, A necessary and sufficient condition for quantization errors to be uniform and white, IEEE Trans. Acoustics, Speech, Signal Processing, vol. ASSP-25, pp. 442-448, Oct. 1977. [11] A. Pacut and K. Hejn, Equivalence of Widrows and Grays approaches to uniform quantizers, Comput. Stand. Inerfaces, vol. 19, nos. 3 and 4, pp. 205-212, Sept. 1998. [12] J. J. Blair and T. E. Linnenbrink, Corrected rms error and effective number of bits for sinewave tests, in Proc. ADDA and EWADC 2002, Prague, Czech Republic, June 26-28, 2002, pp. 177-183. [13] K. Hejn and A. Pacut, Sine-wave parameters estimationThe second source of inaccuracy, in Proc. IEEE IMTC 2003, Vail, CO, May 20-22, 2003, vol. 2, pp. 111-999. [14] K. Hejn, A. Pacut, and L. Kramarski, The effective resolution measurements in scope of sine-fit test, IEEE Trans. Instrum. Measur., vol. 47, pp. 45-50, Feb. 1998.

Konrad Hejn received his M.Sc. in control and computer engineering in 1966, his Ph.D. in electronics in 1975, and his D.Sc. in measurement and instrumentation in 1999. Since 1966 he has been with the Warsaw University of Technology. He is currently a professor and the leader of Automatic Measurement System Group. He is interested in analog-to-digital conversion, computer-aided design of mixed-signal circuits, and related areas.

Andrzej Pacut received his M.Sc. in control and computer engineering in 1969, his Ph.D. in electronics in 1975, and his D.Sc. in control and robotics in 2000. Since 1969 he has been with the Warsaw University of Technology. He is currently a professor and vice-director of the Institute of Control and Computation Engineering. He is a Senior Member of the IEEE and serves as the vice-president of the Poland Section of the IEEE. He is interested in learning systems, neural networks, biometrics, identification, stochastic modeling, and related areas.

September 2003

IEEE Instrumentation & Measurement Magazine

55

S-ar putea să vă placă și