IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 46, NO. 5, MAY 1999 489 Modeling of CMOS Digital-to-Analog Converters for Telecommunication J. Jacob Wikner and Nianxiong Tan, Senior Member, IEEE Abstract— This paper gives an overview of some of the effects caused by circuit mismatch and parasitics in binary weighted digital-to-analog (DAC) converters, and, as a special case, a current-steering CMOS converter. Matlab is used as a behavior-level simulator. In telecommunications applications, the frequency-domain parameters are of the greatest importance. Therefore, the characterization of the device and its performance is determined by frequency parameters such as the signalto-noise ratio, spurious-free dynamic range, multitone power ratio, etc. In this paper, we show how these frequency-domain parameters are affected when mismatch errors and finite output impedance are applied to a current-steering CMOS DAC. We discuss how static performance is affected when changing the size of the errors and fundamental circuit parameters. The impact of dynamic errors such as glitches, slewing, and bit skew is discussed. Measurement results from 14-bit DAC’s are also shown to illustrate the correlation with the modeling. Index Terms— CMOS, current-steering, digital-to-analog converters, dynamic and static errors, frequency-domain measures, measurement results, modeling, simulation. I. INTRODUCTION I N HIGH-performance telecommunications applications, good linearity and low noise are very important. The digital-to-analog converter (DAC) is a crucial building block, since it puts the information on the line. A wide frequency band and high resolution are needed to meet the requirements on high speed and high accuracy. The time-domain aspects are very important, but most of the characterization is done in the frequency domain [1]. We are using parameters such as the spurios-free dynamic range (SFDR), intermodulation distortion (IMD), signal-tonoise ratio (SNR), multitone power ratio (MTPR), and the signal-to-noise-and-distortion ratio (SNDR) to characterize the device. The characterization can also be divided into static and dynamic properties [1], [2]. The static properties are given by the settled output values, and are often too optimistic to determine the true performance of the converter. Dynamic properties are given by the transition between two states, hence the slewing, glitches, time skew, etc. In telecommunications applications, it is mostly the dynamic performance that determines the quality of the converter. Ad hoc, we can say that the static properties set the best-case performance. With this modeling, we want to Manuscript received January 26, 1998; revised December 31, 1998. This paper was recommended by Associate Editor E. Soenen. J. J. Wikner is with Microelectronics Research Center (MERC), Ericsson Components, Link pings Universitet, S-581 83 Link ping, Sweden. N. Tan is with Globespan Semiconductor, Red Bank, NJ 07701 USA. Publisher Item Identifier S 1057-7130(99)03797-0. show the frequency-domain aspects of errors in static values. Similar discussions have been reported [3]–[5], but none of them have discussed the true impact of different linearity errors on the frequency-domain parameters. We also briefly discuss the impact of dynamic errors. In this paper, we focus on binary-weighted, current-steering DAC’s, as presented in Section II. This type of converter uses a number of weighted current sources, which are switched and summed at a terminating output resistance. Process variations and other parasitics will influence the matching between the current sources. Output impedance is dependent on choice of transistor sizes, etc. In Section III, we show how the linearity of the converter is affected when assuming a finite-output impedance and especially a signal-dependent output impedance of the converter. The mismatch due to process variations of the individual current sources is discussed in Section IV. It is shown that the mismatch errors affect the linearity, and we show how the SFDR is changing with different sizes of mismatch, since it is commonly known that a raw binary structure has a limited performance due to the matching errors. In Section V, we discuss the dynamic errors, such as nonlinear slewing, bit skew, and glitches. Modeling issues similar to those presented in Sections III and IV are presented. In Section VI, we also show measurement results to illustrate the correlation between the calculated, simulated, and measured results. Previous related work has also been published in [6] and [7]. In [7], a discussion over circuit noise is also given. In this paper, we show some extensions to these theories as well as measurement results. II. CURRENT-STEERING DAC The use of switched current sources is a straightforward approach in high-speed CMOS DAC’s, since currents are easy to weight, sum, and switch. The structure that is studied in this paper is a binary weighted converter as shown in Fig. 1 and discussed in [6]–[13]. The switches, , in Fig. 1, are controlled by the digital bits, , where is the number of bits. The digital input , , , , with as number is the least significant bit the most significant bit (MSB) and is high, switch is closed and the current (LSB). When is switched to the output. Binary weighting implies that the current source controlled by , i.e., the th LSB current source, is formed by connecting unit current sources in parallel. The 1057–7130/99$10.00 1999 IEEE 490 Fig. 1. An IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 46, NO. 5, MAY 1999 N -bit binary weighted DAC. Fig. 2. Generalized view of a current source with a nonzero output conductance. current associated with this bit position will be denoted (1) is the output current of a single-unit current source. where When considering the static behavior, we only look at time points where the output signal of the DAC has settled to its where is the final value. At a certain time point sample time period, and the code at the input is given by with the corresponding bits . The output current is given by (2) This notation is simplified by writing on the input, it will give rise to a signal-dependent gain, i.e., distortion. Using (5) and assuming a signal-dependent output , the current delivered to conductance of the DAC the load is (6) indicates the DAC’s digital input given by (3) and is the load resistance. The output conductance is determined by studying the DAC structure shown in Fig. 1, and is related to the number of parallel-unit current sources that are currently switched to the output (determined by the digital input). The output conductance associated with oneand unit current source is assumed to be with the th LSB, we have the corresponding conductance . The total output conductance of the converter is given by where (3) (7) Combining (2) and (3), we have the simple description (4) where is given by (3). We will refer to the product between the unit output conductance and the load resistance as the conductance ratio, and denote it as III. OUTPUT-RESISTANCE VARIATIONS (8) The output impedance and the parasitic impedance of interconnections and switches in the converter will strongly determine the performance [4]. Any nonideal current source has a finite output resistance and can be modeled as in Fig. 2. When the different current sources are switched to the output, the total output impedance is changed. When only static values is are considered, the current through the load We will also use the resistance ratio given by (9) Combining (2) and (6)–(8), we have that the load current can be written as (10) (5) is the nominal output current from the DAC given where is the output conductance, is by (2), is the supply the signal-independent load resistance, and voltage. The influence of reactive parts, which influence the dynamic performance, are neglected in this first discussion. From (5), it is seen that if the output conductance of the DAC were constant, there would only be a loss of gain and an additional offset current in the output signal, which would not degrade linearity. If the output conductance is dependent A. SNDR versus Output Resistance is the difference between the The error current discrete-time current (4) and the degraded current (10), i.e., (11) WIKNER AND TAN: MODELING OF CMOS DAC’S FOR TELECOMMUNICATION 491 Fig. 3. View of the wanted continuous-time current and the piecewise linear output current with mismatch error I X , at code X . 1( ) In Fig. 3, we show the situation at code ; the expression in (11) is compared to the ideal wanted continuous-time current, given by (12) The total error power introduced due to the finite output resistance may be described by Fig. 4. Simulated and calculated SNDR versus resistance ratio for 10-, 12-, and 14-bit DAC’s. To illustrate the similarity with the calculated, a piece of the curve has been magnified. Now we can find the SNDR as a function of the output impedance, hence SNDR (19) (13) From (13), we identify the quantization noise power (14) and the time-averaged power of the error current introduced by the degraded converter (15) Let the input signal be given by sinusoid (16) is the dc level of the signal, is the amplitude where is the normalized frequency, and is a of the sinusoid, quantization error which is assumed to be white for converters with a larger number of bits. The ac power of the sinusoid at the output is given by (17) The code-average average value of is found by approximating (11) using the from (16), i.e., (18) In Fig. 4, we show the simulated SNDR of 10-, 12-, and 14. The bit DAC’s when changing the resistance ratio SNDR results are compared with the calculated result from (19), and the formula is well verified. In the simulation, we , V, and A. All use signals fed to the converters are full-scale sinusoids. B. SFDR versus Output Resistance We want to examine how the errors are distributed in the frequency domain. The expression within parenthesis in (10) using (16) as input signal is rewritten as a converging Taylor series [6], and the SFDR is found to be SFDR (20) is the output resistance for a where single-unit current source. For the case with equal amplitude and dc level, i.e., (21) we have SFDR (22) 492 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 46, NO. 5, MAY 1999 Fig. 5. Output spectrum when applying a full-scale sinusoid on a 14-bit 1:0 1 5:106 . DAC. The resistance ratio is Rratio = When is large, in dB this is approximately SFDR dB Fig. 6. Simulated and calculated SFDR versus output over load resistance ratio for 10-, 12-, and 14-bit DAC’s. The curves are saturated since the spurs are hidden below the quantization noise at high resistance ratios. derivations in [6], the distortion with respect to the third harmonic can be found to be (23) SFDR of the signal, we We see that by decreasing the ac level get an improvement of the linearity of the circuit. In the special case with a full-scale sinusoid, i.e., (27) (24) we get (25) SFDR the SFDR given by (27) will be much If larger than the one given in (20). This can be used to relax the design effort on the sources’ output impedance, if the DAC should be designed for a fully differential application. C. Influence of Parasitic Resistance The result found in (25) is based on the knowledge that the second harmonic is dominating the distortion [6]. To illustrate this, we show in Fig. 5 the simulated output spectrum from a 14-bit DAC when applying a full-scale sinusoid and having . If is large, a resistance ratio of (25) can be approximated with SFDR (26) From (26), we realize that with a doubling of the load resistance, the SFDR is decreased by 6 dB. With a maintained resistance ratio, the linearity will also deteriorate with an increased nominal number of bits. In Fig. 6, we show the simulated and calculated SFDR versus resistance ratio for a 10-, 12-, and 14-bit DAC. At high ratios, the simulated values are saturated since the spuriouses are hidden in the noise floor (compare Fig. 4). It is seen that simulated SFDR follows the mathematical result well. When using differential signals, the second harmonic will be cancelled and the third harmonic dominates. Using the The modeling in the previous sections is rather coarse and the resistance of internal wires and switches has not been considered. However, these may be incorporated as well (see the modified current source in Fig. 7). For bit-position , we which is associated also include the parasitic resistance with the interconnection wires from current sources to the output and the on resistance of the switches. Using Norton’s theorem, we may transform the circuit to be similar to the current source as illustrated in Fig. 2. Using that schematic view, we will for the th bit have a current source with the value (28) The output conductance for same bit is given by (29) To find the total output current and conductance for the converter, we have to sum all the contributions from (28) and WIKNER AND TAN: MODELING OF CMOS DAC’S FOR TELECOMMUNICATION 493 Fig. 8. Mismatch modeling with mismatch error current sources, 1I . i IV. CURRENT-SOURCE MISMATCH Fig. 7. Model of the current source at bit position i with parasitic resistance Rpar; i from switches and internal wires. (29) and (30) respectively. Now, we can use (30) with (5) to find the current delivered to the load (31) Due to the sums in (30), this is a rather complex expression. We will, however, assume that the parasitic resistance roughly . By expanding the is equal for all bits, hence terms in (30) to a Taylor series and assuming that , we may further approximate the expressions in (30) with A current-source mismatch error can be modeled as an additional current source in parallel with the nominal current source, as shown in Fig. 8. All individual error sources can be summed and modeled as one error-current source connected to the output of the DAC. The ideal output-current signal is given by (2). Suppose there is an error in the current sources. (Now, we assume an infinite output impedance of the current sources). The distorted output current delivered to the \rm load can be written as the sum of the nominal output and the error current current (36) is the digital number given by (3). Suppose that in where . Then the th current source there is an error current the output current is (37) where is given by (38) is the relative error current in the th LSB. In the where for all different inputs . static case, we assume that The total output current can be rewritten as and (39) (32) Using the approximations in (32) we may rewrite (31) as (33) And we may identify a new conductance ratio (34) or the resistance ratio (35) This new ratio (35) can be used in (19), (20), and (27) to find the SNDR and SFDR. This will further be discussed in Section VI, where measurement results are presented. In reality, mismatch of transistors due to process variations - or -mismatch for a are of stochastic nature. The CMOS current source can be characterized by its distribution, transistor sizes, and physical distance [14], [15]. The mismatch may also depend on other aspects such as, for example, the gradient of the oxide thickness over the array of current sources [5], or voltage drops over bias or supply wires [4], [16]. For a unit current source, we associate the relative error . Its mean value is and the standard deviation . A nonzero mean value will only give rise to a dc gain is error in the output and can be neglected in these calculations. First in this discussion, we assume that the current sources are uncorrelated. This is a very coarse assumption, and in reality, transistors close to each other will have highly correlated errors. For the th LSB current source we have a relative 494 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 46, NO. 5, MAY 1999 error current, , with a mean value and a standard , since unit current sources are deviation connected in parallel. With (39), the expectation value of the output current—with respect to the mismatch error—becomes (40) for all . This also implies that the since matching error cannot directly be found by using, for example differential nonlinearity (DNL) or integral nonlinearity (INL), and we have to find other ways to characterize our device. A. SNDR Versus Mismatch Since the mismatch errors are assumed to be uncorrelated, the expectation value of the squared load current is Fig. 9. Calculated and simulated SNDR versus mismatch for a 10-, 12-, and 14-bit DAC. In dB, we have that (41) For the mismatch error, we have (42) Since we get , we have that . Using (42) in (41), SNDR (48) In Fig. 9, we show the simulated and calculated SNDR versus mismatch for a 10-, 12-, and 14-bit DAC. The results are found by taking the average value of 1024 simulations for each mismatch value. The simulated values match the calculated ones well. B. SFDR versus Mismatch (43) We can now find the code-averaged squared value for the expression in (43) (also compare with Fig. 3) and we have that We will now examine how the mismatch error power is distributed in the frequency domain. The Fourier series for bit are coefficients (49) (44) is the average value of the input code, i.e., where the dc value. We have the error power given by is the period in number of samples. The coefficients where , are given by using (39) and (49) for the total error current (45) and we see that by reducing the dc value of the signal, we also reduce the error power. By using (14), (17), (19), and (45), we may form the SNDR as (50) The power of each tone from (50) is given by (51) SNDR Since all mismatch errors are uncorrelated, we have that the expectation value of the power of the th tone is (46) (52) With a full-scale sinusoid, we have from (24) that the SNDR (46) becomes For a full-scale single-tone sinusoid, the ac power is SNDR (47) (53) WIKNER AND TAN: MODELING OF CMOS DAC’S FOR TELECOMMUNICATION 495 The harmonic distortion with respect to the th harmonic is roughly (54) For each bit, we have the error given by a pulse wave, and the Fourier-series coefficients determine the position and power of harmonics. Now if we assume that the matching error dominates in the MSB’s, we will principally have pulse-shaped error signals. The Fourier series coefficients for these squared waves are given by if is even (55) if is odd is the period. In the static case, is larger than where We can use the approximation in (55) to rewrite (54) as Fig. 10. Calculated and simulated SFDR versus mismatch for a 10-, 12-, and 14-bit DAC. . (56) In dB, we find that (57) Thereby, we also find the SFDR for the full-scale signal to be ), i.e. given by the minimum value of (57) ( SFDR (58) In Fig. 10, we show the average simulated and calculated SFDR for a 10-, 12-, and 14-bit DAC. The value is found by taking the average from 1024 simulations for each mismatch value. In Fig. 11, we show the typical output spectrum of a 14-bit DAC when applying matching errors with standard %. The SFDR is approximately 83 dBc deviation for a full-scale sinusoid input. C. Impact of Correlated Matching Errors The weighted current sources are constructed by using a number of unit current sources in parallel. If all the matching errors for these unit sources are uncorrelated, we will for the th bit have a relative matching error with a zero mean . In reality, the value and a standard deviation matching errors of the unit sources are correlated if they are densely placed close to each other [5], [14]. The mismatch error for the th unit source associated with the th bit is . The expectation value of the squared output denoted current for bit is Fig. 11. Output spectrum for a 14-bit DAC with mismatch size approximately 1.5%. We denote the covariance or correlation between two unit current sources as and (60) where also can be interpreted as the physical distance. Using (60) in (59), we get (61) where (59) . Equation (61) is rewritten as (62) 496 Fig. 12. IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 46, NO. 5, MAY 1999 Current source with output conductance and capacitance. Fig. 13. Typical behavior of glitches and slewing in the output current at a code transition at the input. Since , we have that at the input will give rise to a glitch and a step function at the output, where the glitch energy and the rise time are dependent on the codes. The first-order model gives the output current through the load [6] to (63) and we may identify the standard deviation for relative matching error in the th bit as (64) , are positive, the standard deviaIf the correlation factors may become larger than the one used in the previous tion modeling. Naturally, there is a correlation of the matching error between different bits as well. The correlation is strongly dependent on the layout style and should be kept as low as possible to achieve a good linearity [8]. V. DYNAMIC PROPERTIES We have discussed how the output current is dependent on variations of static values, hence the value given by the settled signal after a code transition. Basically, this implies that the formulae are only valid at low frequencies (and high oversampling ratios). For higher frequencies, we have to consider the dynamic properties such as, for example, glitching, nonlinear slewing, bit skew, etc. We present some ideas to model these affects in a similar way as was used for static values. We discuss nonlinear slewing and glitches associated with bit skew. (65) . where is the current through the load at the beginning of the codeis the end value, is the transition, is the code-dependent time constant. sample period, and We may slightly inaccurately assume that the start and end values are approximately given by (30), and the time constant is given by [6] (66) If the converter is well designed with respect to the output impedance, (66) can be approximated with (67) The output error in the specific time interval is given by (68) The minimum error is found at , and we have (69) Using (33) in (69), we have that A. Nonlinear Slewing In general, the output impedance of the current-steering DAC is varying with the signal. With each unit current source, a certain output resistance is associated. In the same way, we . Hence, for the th current source, include a capacitance, , and similar to we will have a capacitance we have a total output capacitance of (7), at the code . In Fig. 12, we show this situation, where we have included a parasitic capacitance and resistance and . As illustrated in Fig. 13, a code transition (70) is assumed to be small. From (70), we find that where the size of the error is determined by the difference in code transitions, in some sense the derivative of the signal. We now see that with a higher clock frequency, is reduced and is increased. With a lower oversampling ratio, the derivative is increased. of the input signal is becoming larger, and WIKNER AND TAN: MODELING OF CMOS DAC’S FOR TELECOMMUNICATION 497 Simulation of these errors can be fulfilled by simply adding the minimum error from (70) to the output current. The similarity between (70) and (33) implies that for a sinusoid input signal, there will be the same impact on the distribution as on the error power. Hence, the second harmonic will dominate [9]. Fig. 14. Modeling of the timing error. The mismatch between the ideal switching signal (- -) and the actual signal (-) is given by the t(k )’s. B. Bit Skew and Glitches As indicated in the previous section, the slewing will distort the output signal. However, the occurrence of glitches will further influence the result. The glitches occur due to clock feedthrough and bit skew, i.e., all bits do not switch exactly at the sampling point, and hence it is important to have a proper digital delay for all bits [4], [17], [18]. For example, if the MSB switches faster than all other bits at the code transition 11 100 00, we may for a short period of time 011 11 at the output, giving a glitch (Fig. 13). have the code 111 The common way to guarantee a good design is to guarantee that the glitch energy is kept as low as possible [2], by for example using segmentation and proper switching schemes [18]. The slewing is dependent on the change of the input code, but the glitches are more dependent on the bits changing. A large glitch will also result in a longer settling time. In Fig. 14, we show how this skew can be characterized for the th bit. The errors of switching time are given by the timing errors . The switching activity of bit is determined by (71) Fig. 15. Measured output spectrum for a 14-bit DAC. Sample frequency is 25 MHz and signal frequency 670 kHz. The input signal is 15 dBFS. In the time domain, the error current caused by time skew for the th bit is approximately (the rise and fall times are neglected) for all bits and a full segmentation, we will have the glitch current 0 (75) (72) where otherwise. (73) Hence, for the th bit, we will have a pulse-shaped error current (as shaded in Fig. 14). Using (72), the total skew error current is Hence, the glitch current will be dependent on the code change, similar to the slewing discussion. In simulation models, the result from, for example, (74) and (75) can be used to modify the initial value as used in (65), and the minimum error (70) is thereby changed, further reducing linearity. If the time skew no longer is much smaller than the sample period, the signal will not settle fast enough, and further degrading the result, as discussed in the previous sections. VI. MEASUREMENT RESULTS (74) For a full segmentation of the data, the switching activity would be determined by the signal change or derivative. Now the situation is more complex, since the bits are determining the size of the glitches. The time skew may be in the same order for all bits, but the MSB’s will dominate the result due to their amplitude. If we assume approximately equal time skew The chip used in the measurements is a 5-V 14-bit CMOS DAC. The four most significant bits are segmented. The layout of the converters is a straightforward structure using an array of current sources, as reported in [9]–[11] and [13]. All unit current sources are placed symmetrically to drive their output current in the same physical direction. In Fig. 15, we show the output spectrum of a single-ended output of the converter. The sample frequency is 25 MHz and the signal frequency 670 kHz. The input signal amplitude is 15 dBFS . The distortion with respect to the second harmonic HD is approximately 59 dBc, and for the third harmonic HD we have 57 dBc. In measurements, 498 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: ANALOG AND DIGITAL SIGNAL PROCESSING, VOL. 46, NO. 5, MAY 1999 Fig. 16. Measured distortion with respect to the second and third harmonic versus signal power for a 14-bit DAC. The dc value is held constant and the ac value is varied. we have also found a HD 65 dBc and HD 74 dBc at clock frequencies around 5 MHz [9] and signal amplitudes of 6 dBFS. These results also indicate the impact of the dynamic properties, which degrades the performance with higher frequencies. As a function of the signal amplitude, the measured distortion with respect to the second and third harmonic is also shown in Fig. 16. When varying the ac value of the signal and keeping the dc value constant, it is seen that HD is approximately constant and HD is decreasing by 6 dB for each doubling of the signal’s amplitude. This follows the result found in (20). We see from the equation that if the dc value value, the SFDR should be is small compared to the decreasing by 6 dB with a 6 dB increase of the signal. The same measurements can be made for a varying dc value and constant ac value [9]. It can be seen that the distortion is increased with a higher dc value as predicted in (46). (Note that the whole impact of the matching error is not represented by the third harmonic only). From circuit-level simulators (spectreS), we have a parasitic from switches and wires. resistance of approximately 400 Using (35) in (23) and the 5-MHz measurement results, we identify the output resistance of a unit current source to be approximately Fig. 17. Simulated output spectrum for a 14-bit DAC. introduce a higher than the actual one given by the processing. In Fig. 17 we show the simulated output spectrum of a single-ended 14-bit converter, where the output impedance of each unit current source is 2.9 G , the load resistance is 25 , parasitic resistance is approximately 375 , the capacitance associated with each current source is approximately 4 fF, and the parasitic capacitance 40 pF. The sampling frequency is 25 MHz and the signal frequency is 670 kHz, and the ac signal is 15 dBFS. Compare with the measured spectrum in Fig. 15. The performance of the converter may be predicted using Matlab simulations. VII. CONCLUSION The influence of circuit imperfections on the important frequency-domain performance parameters has been modeled. Behavior-level models have been derived and used in Matlab simulations. The modeling is used to predict the performance of an implementation and to help the designer in his construction. We have seen that the calculated models can be verified by simulations and measurements. Matlab simulations are much more time effective than circuit-level simulations. The freedom of varying circuit parameters is large as well. ACKNOWLEDGMENT G (76) Using the 25-MHz measurement results, the output impedance would be approximately 250 M . Circuit-level simulations show a dc output impedance of over 20 G . This implies that the dynamic properties still affect the result at frequencies as low as 5 MHz. To find the size of the matching error (58) is used, and for the 5 MHz measurements we have approximately . This is a rather high value, and we realize that the dynamic properties are influencing, as well as the correlation of the unit current sources. Hence, the correlation factors The authors would like to thank M. Gustavsson, Globespan Semiconductor, L. Wanhammar, Department of Electrical Engineering, Link pings Universitet, for valuable discussion and hints. They would also like to thank P. Pettersson, Ericsson Radio Systems, and J.-E. Eklund, Ericsson Components, Stockholm, Sweden, for help with measurements. REFERENCES [1] P. Hendriks, “Specifying communication DAC’s,” IEEE Spectrum, vol. 34, pp. 58–69, July 1997. [2] R. J. van de Plassche, Integrated Analog-to-Digital and Digital-toAnalog Converters. Norwell, MA: Kluwer, 1994. WIKNER AND TAN: MODELING OF CMOS DAC’S FOR TELECOMMUNICATION [3] E. Liu and A. Sangiovanni-Vincentelli, “Verification of Nyquist data converters using behavioral simulation,” IEEE Trans. Computer-Aided Design, vol. 14, Apr. 1995. [4] T. Miki, Y. Nakamura, M. Nakaya, S. Asai, Y. Akasaka, and Y. Horiba, “An 80 MHz 8-bit CMOS D/A converter,” IEEE J. Solid-State Circuits, vol. 21, pp. 983–988, Dec. 1986. [5] H. J. Schouwenaars, D. W. J. Groeneveld, and H. A. H. Termeer, “A low-power stereo 16-bit CMOS D/A converter for digital audio,” IEEE J. Solid-State Circuits, vol. 23, pp. 1290–1297, Dec. 1988. [6] J. J. Wikner and N. Tan, “Influence of circuit imperfections on the performance of current-steering DAC’s,” Analog Integrated Circuits and Signal Processing, pp. 7–20, Jan. 1999. [7] J. J. Wikner and N. Tan, “Modeling of CMOS digital-to-analog converters for telecommunication,” presented at IEEE Int. Symp. Circuits Syst., Monterey, CA, May/June 1998. [8] C. A. A. Bastiaansen, D. W. J. Groeneveld, H. J. Schouwenaars, and H. A. H. Termeer, “A 10-b 40-MHz 0.8-m CMOS current-output D/A converter,” IEEE J. Solid-State Circuits, vol. 26, pp. 917–921, July 1991. [9] J. J. Wikner, “Measurement and simulations of a CMOS DAC chipset,” Linköping Univ., Internal Rep., LiTH-ISY-R-2704, Dec. 1998. [10] N. Tan, E. Cijvat, and H. Tenhunen, “Design and implementation of high-performance CMOS D/A converter,” in Proc. 1997 IEEE Int. Symp. Circuits Syst., Hong Kong, vol I. pp. 421–424. [11] N. Tan, “A 1.5-V 3-mV 10-bit 50 MS/s CMOS DAC with low distortion and low intermodulation in standard digital CMOS process,” in Proc. 1997 IEEE Custom Integrated Circuits Conf., Santa Clara, CA, May 1997, pp. 599–602. [12] M. Otsuka, S. Ichiki, T. Tskuada, T. Matsuura, and K. Maio, “Lowpower, small-area 10bit D/A converter for cell-based IC,” in Proc. 1995 IEEE Symp. Low Power Electron., 1995, pp. 66–67. [13] J. J. Wikner and N. Tan, “A CMOS digital-to-analog converter chipset for telecommunication,” IEEE Circuits Devices Mag., vol. 13, pp. 11–16, Sept. 1997. [14] M. J. M. Pelgrom, A. C. J. Duinmaijer, and A. P. G. Welbers, “Matching properties of MOS transistors,” IEEE J. Solid-State Circuits, vol. 24, pp. 1433–1439, Oct. 1989. [15] J. Bastos, M. Steyaert, A. Pergoot, and W. Sansen, “Mismatch characterization of submicron MOS transistors,” Analog Integrated Circuits and Signal Processing, vol. 12, pp. 95–106, 1997. [16] B. E. Jonsson, “Design of power supply lines in high-performance SI and current-mode circuits,” in Proc. 15th Norchip Sem., Tallinn, Estonia, Nov. 10–11, 1997, pp. 245–250. [17] J. M. Fournier and P. Senn, “A 130-MHz 8-b CMOS video DAC for HDTV applications,” IEEE J. Solid-State Circuits, vol. 26, pp. 1073–1076, July 1991. [18] D. Mercer, “A 16-b D/A converter with increased spurious free dynamic range,” IEEE J. Solid-State Circuits, vol. 29, pp. 1180–1185, Dec. 1994. J. Jacob Wikner was born in Borgholm, Sweden, on July 13, 1973. He received the M.Sc. degree in computer science and engineering in 1996, and the Lic.Eng. degree in September 1998, both from Link pings Universitet, Sweden, where he is currently working toward the Ph.D. degree in electronic engineering. During the academic year 1994–1995, he studied telecommunication theory and high-speed electronics at the Technical University in Darmstadt, Germany. He is currently with Microelectronics Research Center (MERC), Ericsson Components, Link pings Universitet. His main research interests include analog and mixed-signal integrated circuits, especially current-mode and switched-capacitor techniques, as well as converter circuits for digital subscriber line applications. Modeling and behavior-level simulation techniques for converters and analog circuits are also within his field of research. 499 Nianxiong Tan (S’91–M’95–SM’98) was born in Sichuan, China, on March 31, 1966. He received the B.Eng. and M.Eng. degrees in electronic engineering from the Department of Electronic Engineering, Tsinghua University, Beijing, China, in 1988 and 1991, respectively, and the Lic.Eng. degree and Ph.D. degree in applied electronics from the Department of Electrical Engineering, Link pings Universitet, Sweden, in 1993 and 1994, respectively. From 1987 to 1991, he was with the Department of Electronic Engineering, Tsinghua University, Beijing, China. His research involved analysis and design of bipolar circuits, CMOS operational amplifier design, switched-capacitor network analysis and design, and digital ASIC design. From 1991 to 1994, he was with the Department of Electrical Engineering, Links pings Universitet, Sweden, working on oversampling data converters, current-mode techniques, and low-power design of DSP circuits . From 1995 to 1998, he was with Microelectronics Research Center of Ericsson Components AB, Sweden, responsible for developing analog and mixed-signal circuits for telecommunication. He managed, coordinated, and participated in various projects, while also holding the position of Adjunct Professor parttime at the Department of Electrical Engineering, Link pings Universitet, leading the Analog Design Group and supervising Ph.D. students. Since April 1998, he has a Manager with GlobeSpan Semiconductor, Inc., Red Bank, NJ, leading a group developing analog front ends for high-speed internet modems. His research interests include mixed-analog and digital design methodology, high-speed analog circuits, data converters, and low-power design. His recent focus is analog front ends for high-speed internet modems. He has published more than 60 journal and conference papers, and holds 15 patents granted or pending. He has authored one book and has given many presentations and lectures for industries and research institutes on analog design. Dr. Tan was Session Chair for the 1995 and 1997 IEEE International Symposium on Circuits and Systems. He is currently coeditor for IEEE Circuits and Devices Magazine.