RESUME, NOVEMBER 2007 1 A 10b 100MS/s Time-Interleaved SAR ADC Abstract— Analog-to-Digital Converters (ADC) are electronic circuits which translate analog signals to its binary representation (digital). Since the signal processing is easier in digital domain, this type of circuits is extremely important in several areas, for example in the communications sector. This dissertation proposes a new implementation for an Analog-to-Digital Converter (ADC) with 100 MS/s of sampling frequency and 10 bits of resolution. This proposal consists in an 11-channel time-interleaved ADC, each sub-converter with a SAR (Successive Approximation Register) topology. The first part of this work presents the subconverter SAR project. To optimize the sub-converter hardware it is implemented using a segmented topology; it combines different DAC architectures and different bit decoding methods (unary/binary). Moreover, it is provided a performance limitations analysis of the time-interleaved structure. The second part of the thesis describes in more detail the design of the SAR blocks. Firstly the comparator, which consists in a pre-amplifier and a latched comparator. Secondly some digital blocks used to control the converter, namely: (i) a state machine; (ii) a successive approximation register; (iii) a row/column decoder for the capacitor matrix; and (iv) an output register. Finally, some simulations were performed in order to evaluate the proposed architecture. The presented results show that the proposed architecture achieves the initial requirements. Besides, the implementation proposed yield good power performance results as shown by the Factor of Merit (FOM) obtained. Index Terms— Analog-to-Digital Converter (ADC), Successive Approximation Register (SAR), Segmented Topology, TimeInterleaved, Sub-Converter. I. I NTRODUCTION T HE quick development of the wireless communications sector have been demanding for improvements. Wireless communications often require the use of analog signals, since signal processing is easier in the digital domain, circuits capable of convert signals between the two domains are essential. Analog-to-Digital Converters (ADC), produce the binary representation of an analog signal, while Digital-toAnalog Converters (DAC) have the opposite function. The 2 parameters used to characterize any signal converter are: sampling frequency (fS ) and resolution (N ). The fS parameter give the number of results per second and the N parameter corresponds to the number of bits for each sample, thus it determines the minimum amplitude of the signal that can be distinguished. In practice, is hard to improve both parameters at the same time, so a trade-off is needed between them. Nowadays, for wireless applications, it is common to utilize ADCs with a sampling frequency around 80 MHz and 10 bit of resolution. Typically this is obtained with a Pipelined architecture, but it is becoming harder to satisfy the linearity requirements. Therefore, is important to explore other possible implementations. This work proposes a solution for a 10 bit ADC with a fS of 100 MHz. In order to avoid the linearity issues, the new solution is based in a Successive Approximation Register vIN S/H SAR b N-1 vDAC Fig. 1. ... b1 b0 DAC SAR ADC block diagram. (SAR). However this type of converter tends to be slower. Therefore a parallel architecture with an 11-channel timeinterleaved ADC is used to achieve the above requirements. Besides the performance limitations of this parallel architecture will be discussed. The rest of this article is organized as follows: Section II reviews two different SAR ADC topologies. this topologies were merged and optimized to achieve better performance. This section also provides an analysis of a parallel structure. Section III presents the design of the analog blocks. Section IV presents the digital blocks needed to control the ADC. Section V shows some simulation results. Finally, we conclude in Section VI. Luı́s Simões November 2007 II. T OPOLOGY The proposed converter is based on a successive approximation topology comprising four parts (see Fig. 1): a comparator, a S/H mechanism, a DAC and a register. The comparator determines if the sampled voltage (by the S/H) is higher/lower than the DAC generated voltage, which is given according to the SAR value. The successive approximation algorithm has the following steps: (i) the comparator verify if vIN is higher than 12 Vref 1 , to determine MSB; (ii) if the decision of the most significant bit (MSB) is ”1”, the DAC voltage to compare with vIN is 3 1 4 Vref ; (iii) otherwise the DAC will generate 4 Vref ; and so on and so forth. So this algorithm needs N cycles to determine the complete binary word. There are several ways to implement the DAC unit, e.g. with a resistive ladder, or a capacitor array utilizing charge redistribution. The following sections are dedicated to explain the DAC design. 1V ref is the maximum scale voltage. RESUME, NOVEMBER 2007 2 DAC 2 QX = −2N CvIN . N vIN N 2 -1 (1) 2) Charge-Redistribuition: This stage is divided in N steps, where N is the resolution (number of bits to determine). Firstly, to find the value of MSB, it is forced at ”high” in the SAR, this ensure that half of the capacitors switch to VREF P and the other half to VREF N (VREF N = 0). Now the charge of X node is given by, S/H N 2 -2 ... vDAC 3 2 1 QX = 2N −1 C(vX − VREF P ) + 2N −1 C(vX − VREF N ), (2) the charge in X node is constant, so we have SAR −2N CvIN = 2N −1 C(vX −VREF P )+2N −1 C(vX −VREF N ), (3) simplifying its possible to obtain, Fig. 2. SAR ADC block diagram. 1 VREF P − vIN . (4) 2 Clearly the comparator has conditions to verify if vIN is higher than the middle of the scale ( 21 VREF P ). Secondly, if the previous decision was ”1”, vX will be, vX = Array de Condensadores vX 2 N-1 C N-2 2 C ... 2C C C VREFP vIN VREFN 3 VREF P − vIN . 4 and the opposite case (MSB=0), vX is given by, vX = Interruptores SAR Fig. 3. SAR ADC block diagram. A. DAC with Resistive Ladder In this case the DAC is implemented with a resistive ladder and a group of switches. The switches permit to access the voltages generated in the ladder, as shown in Fig. 2. The ladder has 2N resistors and the same number of switches, i.e. 1024 resistors and 1024 switches, becoming impracticable. B. DAC with Capacitor Array Another possible DAC architecture uses a capacitor array as represented in Fig. 3. In this method, power consumption is reduced and the input signal is sampled and holded directly in the capacitor array, so it does not need a dedicated S/H block. As in the previous design (section II-A) the great disadvantage here is the number of components used, because the capacitor array needs 2N capacitors. This architecture also needs switches to control the capacitors. Two different methods can be used, unary and binary. In the first one each capacitor is controlled by one switch, so it requires 2N switches, besides the capacitors. In the binary method it needs just N switches, since the number of capacitors to commute depends on the weight. Bellow are described the sample and charge-redistribuiton functionalities, to simplify we consider binary decoding. 1) Sample: First the input signal is sampled, to do so vX goes to ground (vX = 0), then all capacitors are switched to vIN . This makes the total charge of X: (5) 1 VREF P − vIN . (6) 4 One can easily verify that the comparator is deciding if the input voltage is higher than 43 VREF P or than 14 VREF P , according to the previous decision. This continues untill all N bits are determined. Finally vX is given by "N # X bN −i vX = VREF P − vIN , (7) 2i i=1 vX = each bit-weight contributes to make DAC generate the nearest voltage to vIN . C. Segmented Topology Taking into account the previous hardware limitations, this work adopts a segmented topology. This alternative permits to use a reasonable number of components without modifying the paradigm of the previous methods. The segmented topology uses a resistive ladder to decode 4 bits and a capacitor array to decode the remaining 6 bits. Therefore only 16 resistors and 16 switches are used in the ladder. Concerning the capacitor array, only 64 capacitors are required; and another segmentation is performed (unary/binary decoding) to reduce the number of switches. There is no straightforward decision for the number of bits used for each segment (unary/binary), thus a Matlab model was used to simulate the influence of different segmentations on the DNL results. The simulations results are shown in Fig. 4. Cases (c) and (d) have missing codes, because the minimum DNL value reaches -1, so, it is only possible to choose between (a) and (b). The first case is clearly better, but its implementation uses more hardware, for this reason the (b) RESUME, NOVEMBER 2007 3 Vreflsb Array de Condensadores Array de Condensadores (dummy) Escada de Resistências SAR Fig. 6. Block diagram of the adopted topology. VCM is a DC voltage and an unwanted charge injection can have a significant impact in the result of expression (9). To avoid this, a dummy capacitor array is used to sample VCM all the time, and both nodes suffer similar charge injections; the blocks diagram is shown in Fig. 6. Fig. 4. DNL histograms to decide the segmentation (unary/binary): (a) - 5+1 (b) - 4+2 (c) - 3+3 (d) - 2+4. Array de Condensadores 60C Descodificação Unária Descodificação Binária Vx b9 ... 4C b6 ... 4C 2C C C Descodificador b5 Escada de Resistências b5 b4 b4 vrefp Vin vrefn 16 15 14 ... Interruptores Descodificador b3 ... b0 3 2 Vreflsb 1 Fig. 5. Connection between resistive ladder and capacitor array and, the different segmentations. segmentation was adopted (i.e. 4 bits unary and 2 binary decoding). Such implementation only requires 18 switches in the capacitor array instead of 64. The final topology is presented in Fig. 5 where the SAR bits b0 to b3 control the resistive ladder switches through a decoder; b4 and b5 directly control the binary switches from capacitor array; and b6 to b9 control the unary switches from the capacitor array through another decoder. The variation of vX is calculated according to the scheme in Fig. 5, the total charge of X node: QX = 2Nc C(VCM − vIN ), (8) where Nc is the number of bits determined by the capacitor array. Following the procedure used in section II-B.2 to determine (7), vX is: # " 10 X b10−i vX = VCM + VREF P − vIN . (9) 2i i=1 Thus vX has a known evolution and the comparator can determine if the tested bit should be ”1” or ”0”. D. Time-Interleaved In order to obtain the sampling frequency of 100 MHz, a parallel structure with 11-channel time-interleaved is used to get one sample per clock cycle, as can be seen in Fig. ??. This architecture permits to use some common blocks, saving power and circuit area. Although, time-interleaving ADCs bring some issues that can limit the overall performance. Thus, some considerations are discussed in the sections below. 1) Channel Offset Mismatches: This type of mismatch can be modeled as a voltage source serially connected with each ideal channel. This means that ADC has a pattern of mismatches along the time with a period of M/fS , this is manifested in the frequency domain as tones at multiples of fS /M . With the Discrete Fourier Transform (DFT), the Fourier coefficients are [1, 2, 3]: Vk = M −1 X Vosi e−j ( M )ki . 2π (10) i=0 This error is independent from the input signal frequency and level. There are two ways to solve this potential problem; an analog method, which consists in a comparator with a lower offset, using more area and power; and a digital method, which generates an offset correction. The latter needs a mechanism to determine the offset in each channel and digitally correct them. 2) Channel Gain Mismatches: Channel gain mismatch can be modeled as the previous mismatch, but with a gain stage instead of the voltage source. Each fixed gain factor is multiplied to the input signal in time domain, obtaining a periodical pattern. The multiplication in the time domain corresponds to the convolution of the gain pattern with the input signal in the frequency domain. The convolution results in side bands centered around multiples of fS /M . Assuming this mismatches are time invariant, the output spectrum Y (jω) is [1, 3]: ∞ 1 X 2πk Gk X j ω − , Y (jω) = T MT k=−∞ (11) RESUME, NOVEMBER 2007 4 where X(jω) is the input spectrum and the Fourier coefficients Gk are M −1 X 2π (12) Gk = gi e−j ( M )ki . VDD M 5a latch M 4a M 3b M 3a M 4b latch M 5b i=0 vON The effect of this error only depends on the input level. Since this issue is caused by mismatches in the reference voltages, all channels need to share the same reference voltages generators. 3) Timing Mismatches: Finally, the timing mismatches consists in different sampling times between the channels, this is caused by non-idealities of multi-phases. The fixed timing deviations lead to fixed side bands around fS /M in frequency domain. Considering ideal instants, each clock signal has a ∆Ti delay, the output spectrum is [1, 2, 3]: ∞ 2πk 1 X Φk (ω)X j ω − , Y (jω) = T MT (13) vOP M 2a M 2b vDP vDN vIP M 1a latch Fig. 7. vIN M 1b M6 Dynamic Latched comparator. k=−∞ VDD where X(jω) is the input spectrum and the Fourier coefficients Φk (ω) are given by Φk (ω) = M −1 X M 4a e 2πi 2π j (ω− M T )∆Ti −j ( M )ki e . M 3a M 3b M 4b (14) i=0 q In order to avoid the timing mismatches, all channels need to share a S/H block, this strategy allows to control all the samples with the same phase signal. qz r s M 2a M 2b III. A NALOG B LOCKS In this section is discussed the comparator design. This block is responsible for determining if the input voltage is higher/lower than the voltage generated in the DAC. In order to implement it, a latched comparator is used, as well as a pre-amplifier to reduce the offset contribution. M 1a Fig. 8. M 1b Latched SR. A. Dynamic Latched Comparator This work adopted a latched comparator with a topology of the Fig. 7, based on [4, 5]. When latch is ”0”, the transistors M4a /M4b and M5a /M5b reset the output nodes (q and qz ) and the drains of differential pair (M1a /M1b ) to VDD . M6 is off and no supply current exists. When latch rise, M4a /M4b and M5a /M5b are switched off, and current starts flowing through M6 and through the differential pair, discharging the M1a /M1b drains, this leads to a vDN and vDP reduction. When this voltages become smaller thsn VDD −Vt , M2a /M2b switch on and cause an output nodes voltage reduction. When this voltages reach smaller values than VDD − Vt , transistors M3a /M3b turn on. In this moment, considering a positive input voltage, M2a current is higher than M2b , so vON decrease faster than vOP . This in turn, makes the vGS2a increase faster than vGS2b , which causes a further increase of the difference between the currents in this transistors - there is a positive feedback mechanism. Analyzing the incremental model of the circuit, is possible to get an expression for vod evolution t vod (t) = ∆Vi e τreg , (15) where τreg is τreg = CL + CDBp + CGSp + 2CGDp + CDBn + CGSn + 2CGDn . gmp + gmn (16) Considering the total capacitance of 20 fF and a total gm of 200 µS, it is possible to obtain τreg , 20 f = 0.1 ns, (17) 200 to achieve a voltage level vodf inal = 2 V, with ∆Vi = 1 mV, we can get a regeneration time estimation: vodf inal t = τreg ln = 0.76 ns. (18) ∆Vi τreg = Since the latch impulse is very short, it is needed a circuit to save the decision for the rest of the cycle, so it is used a latch SR (see Fig. 8), this permits SAR to update the new iteration value. When s signal is high and r is low, M1a and M4b turn on, M4b drain current charge qz node that goes to VDD , this in RESUME, NOVEMBER 2007 5 VDD M 2a M 3a M 3b vON vIP M 2b vOP M 1a M 1b vID vIN Fig. 10. I SS Fig. 9. Pre-amplifier simulation with a vid = 1 mV. Considering expression (21), the 95% voltage level is achieved when t = 3τ, (25) Dynamic pre-amplifier. to reach the targeted voltage level in 8 ns, turn, switch on M2a , putting ground at q output node. When the latch goes zero (reset phase of the comparator), s and r signals go high, this way M4a /M4b turn off and M1a /M1b switch on, this leads to M2a /M2b and M3a /M3b connected as a back-to-back inverters, saving the compared result. With a simulation the worst case was when the input changed from vindif f = 2 mV to vindif f = −2 mV and the regeneration time is 646 ps, that is accordingly with the expected. 8 ns . (26) 3 To obtain a first iteration, some parameters were assumed, τ< as CLtot = 100 fF (27) gm3 = 0.6gm2 . (28) and a gm relation, The differential gain was assumed to be Avd > 5. B. Pre-Amplifier To reduce the offset error, the comparator is enhanced with a pre-amplifier as the one presented in the Fig. 9. Analyzing the small signal model of the circuit it is possible to obtain the gain expression, which is given by: vod Avd = = gm1 Req , (19) vid where Req is Req = 1 . gm2 − gm3 (20) The goal of the pre-amplifier is to achieve 95% of VF IN AL in 8 ns, considering the parasitic capacitances, it is possible to write the vod (t) expression: t gm1 vod (t) = Vid 1 − e− τ , (21) gm2 − gm3 where τ is τ= Ca + 2Cb , gm2 − gm3 (22) with, Ca = CL + CDB1 + CGD1 + CGS2 + CDB2 + CGS3 , (23) and Cb = 2CGD3 . (24) (29) Based on [6] for the strong inversion, it is considered Vod = 200 mV. Assuming this values allowed to achieve a first shot and after some iterations to obtain the results presented in Fig.10. The final W/L relations are shown in Table I. TABLE I F INAL W/L RELATIONS . Transistor M1 M2 M3 W/L 10 10.56 5.56 Finally the simulation results show a pre-amplifier gain of Avd = 5.186 and an amplification time tamp95% = 6.35 ns. IV. D IGITAL C ONTROL A. Global State Machine In order to control the sub-converter states, it was created one global state machine, this measure allowed to reduce circuit area and power consumption. Since we have 11 channels and each one has 11 states, the state machine has 11 flip-flops organized as a shift register. This way, each sub-converter only need to see a different state machine output, i.e. each subconverter is working in a different state. RESUME, NOVEMBER 2007 Fig. 11. Output spectrum without random deviations. 6 Fig. 12. Output spectrum with Monte Carlo parameters. Fig. 13. sation. Output spectrum with Monte Carlo parameters and offset compen- B. Successive Approximation Register This block goal is to save the successive approximations and consequently the final result of each channel. It is implemented by a register with 10 flip-flops, each flip-flop is single controlled by a state signal to know when it needs to save a result coming from the comparator block. C. Row/Collumn Decoding The unary decoding mechanism needs a complex control system for the capacitors, to implement it, the capacitors are placed with a matrix format, and the SAR’s bits connected to row/column decoder. This permits to switch only the required capacitors, avoiding the binary procedure. D. Output Register Intuitively, to obtain a converter with a single output, the ADC requires an output register. The 11 channels are connected to a multiplexer to correctly handle the several results. V. S IMULATION R ESULTS In order to evaluate the proposed architecture some simulations were performed. The simulations allowed to verify the performance of the complete converter. Since this siluations take lots of time it was used HSIM, which permits some calculus simplifications. A. Output Spectrum Without Monte Carlo Distributions The first simulation has a sinusoid with fi = 10449219 MHz as input signal. The simulation results have been analyzed by a Matlab script in order to obtain the output spectrum. As can be seen in Fig. 11 the tones coincide with the input frequency multiples, as expected and the ENOB is 8.9 bit. Only an ideal converter could reach 10 bit. B. Output Spectrum With Monte Carlo Distributions In this simulation, the input wave used is the same of the previous simulation (section V). The output spectrum was obtained with the same Matlab script and can be seen in Fig. 12. Analyzing the results, they clearly are lower than expected, since the ENOB is only 6.55 bit. This can result from random deviations, but also from the simplifications taken by HSIM. As indicated in Fig. 12, the higher tones coincide with fS /11 multiples, which correspond to channel offset mismatches. However, other tones result from timing mismatches. In order to solve the offset problem, a calibration can be implemented. To comfirm this an offset calibrations was performed with Matlab, resulting in the output spectrum of Fig. 13, as can be seen the ENOB increased from 6.55 to 8.07. Fig. 14 presents both spectrums (Fig. 12 and Fig. 13) becoming easier to verify that the offset’s tones had a significant reduction and the timing mismatches tones did not suffered any changes. C. Factor of Merit In order to compare the several performances of many ADC types, it is possible to calculate the Factor of Merit (FOM), RESUME, NOVEMBER 2007 7 Fig. 14. Output spectrum with Monte Carlo parameters, with and without offset calibration. to obtain a better converter without significant disadvantages. An overview of the steps performed in this project is presented below: • In order to design an ADC sub-converter, this work studied some possible topologies and adopted the one with less circuit area and most power efficient. • This work analyzed some issues of a parallel structure, namely, the influences of offset, gain and timing mismatches. • To implement this project some analog parts needed a more careful design, namely the latch comparator and the pre-amplifier. In the end the expected results were achieved. • To control the entire converter some blocks were designed, e.g. a global state machine and a SAR (register to contain the intermediate iterations and the final result). This work always intended to save circuit area and power consumption, so when possible, the channels share blocks, as well as reference voltages generators. • Finally, combining all converter parts, some simulations were effectuated to verify its performance. The results were a little below expectations, mainly because of the used simulator, due to some calculus simplifications. Some improvements can be considered, as shown in section V-B an offset calibration is important to improve the overall performance. R EFERENCES Fig. 15. FOM comparison between this project and some projects presented in ISSCC’07. its value is given by the following expression [7]: Ptotal [pJ/conv], FOM = ENOB 2 fS (30) where ENOB is the effective number of bits. It is noticeable that lower FOM values correspond to better performances. Applying to the obtained simulation results, with 2.6 mA in the analog voltage source, 490 µA in the digital voltage source and considering 9 effective bits, we have: FOM = (2.6 mA)2.5 V + (490 µA)1.2 V = 0.138 pJ/conv (31) 29 100 MHz In the Fig.15 is represented some FOM values in order to the frequency from ADC projects presented in ISSCC’07. As can be seen the obtained FOM is better than the published results. VI. C ONCLUSION This work consists in an ADC project, with a parallel structure, it was verified that with this topology it is possible [1] Lauri Sumanen, Pipeline Analog-to-Digital Converters for Wide-Band Wireless Communications, PhD dissertation, Helsinki University of Technology, Espoo, Dec. 2002. [2] Y.-C. Jenq, Digital Spectra of Nonuniformly Sampled Signals: Fundamentals and High-Speed Waveform Digitizers, IEEE Trans. Instrumentation and Measurement, vol. 37, no. 2, pp. 119-122, June 1988. [3] A. Petraglia, S. K. Mitra, Analysis of Mismatch Effects Among A/D Converters in Time-Interleaved Waveform Digitizers, IEEE Trans. Instrumentation and Measurement, vol. 40, no. 5, pp. 119-122, Oct. 1991. [4] P. Figueiredo, High-Speed CMOS Analog-to-Digital Converters, PhD dissertation, Instituto Superior Tcnico, Lisboa, Jun. 2006. [5] T. Kobayashi, K. Nogami, T. Shirotori and Y. Fugimoto, A currentcontrolled latch sense amplifier and a static power-saving input buffer for low-power architecture, IEEE J. Solid-State Circuits, vol. 28, pp. 523-527, Apr. 1993. [6] Y. Tsividis, Mixed Analog-Digital VLSI Devices and Technology, World Scientific, 3rd ed., 2005. [7] R. Walden, Analog-to-Digital converter survey and analysis, J Selected Areas in Comunications, pp. 539-550, Apr. 1999. [8] M. Yoshioka, M. Kudo, T. Mori, A 0.8V 10b 80MS/s 6.5mW Pipelined ADC with Regulated Overdrive Voltage Biasing, ISSCC, Session 25, no. 1, 2007. [9] D. Huber, R. Chandler, A. Abidi, A 10b 160MS/s 84mW 1V Subranging ADC in 90nm CMOS, ISSCC, Session 25, no. 2, 2007. [10] Y.-D. Jeon, S.-C. Lee, K.-D. Kim, J.-K.Kwon, J. Kim, A 4.7mW 0.32mm2 10b 30MS/s Pipelined ADC Without a Front-End S/H in 90nm CMOS, ISSCC, Session 25, no. 3, 2007. [11] S.-C. Lee, Y.-D. Jeon, K.-D. Kim, J.-K.Kwon, J. Kim, J.-W. Moon, W. Lee, A 10b 205MS/s 1mm2 90nm CMOS Pipeline ADC for Flat-Panel Display Applications, ISSCC, Session 25, no. 4, 2007. [12] B. Hernes, J. Bjrsen, T. Andersen, A. Vinje, H. Korsvoll, F. Tels, A. Briskemyr, C. Hold, . Moldsvor, A 92.5mW 205MS/s 10b Pipeline IF ADC Implemented in 1.2V/3.3V 0.13m CMOS, ISSCC, Session 25, no. 6, 2007. [13] C.-C. Hsu, F.-C. Huang, C.-Y. Shih, C.-C. Huang, Y.-H. Lin, C.-C. Lee, B. Razavi, An 11b 800MS/s Time-Interleaved ADC with Digital Background Calibration, ISSCC, Session 25, no. 7, 2007.