A 700uW 1GS/s 4-bit Folding-Flash ADC in 65nm CMOS for Wideband Wireless Communications
Bayan Nasri, Sunit P. Sebastian, Kae-Dyi You, RamKumar RanjithKumar, Davood Shahrjerdi
AA 700 µ W 1GS/s 4-bit Folding-Flash ADC in 65nmCMOS for Wideband Wireless Communications
Bayan Nasri, Sunit P. Sebastian, Kae-Dyi You, RamKumar RanjithKumar, Davood ShahrjerdiElectrical and Computer Engineering, New York University, Brooklyn, NY 11201Email: [email protected]
Abstract —We present the design of a low-power 4-bit 1GS/sfolding-flash ADC with a folding factor of two. The design ofa new unbalanced double-tail dynamic comparator affords anultra-low power operation and a high dynamic range. Unlikethe conventional approaches, this design uses a fully matchedinput stage, an unbalanced latch stage, and a two-clock operationscheme. A combination of these features yields significant reduc-tion of the kick-back noise, while allowing the design flexibility foradjusting the trip points of the comparators. As a result, the ADCachieves SNDR of 22.3 dB at 100MHz and 21.8 dB at 500MHz(i.e. the Nyquist frequency). The maximum INL and DNL areabout 0.2 LSB. The converter consumes about 700 µ W from a1-V supply yielding a figure of merit of 65fJ/conversion step.These attributes make the proposed folding-flash ADC attractivefor the next-generation wireless applications.
I. I
NTRODUCTION
Next-generation 5G wireless communications would usemillimeter wave bands between 30 and 300GHz [1]. Theuse of such wide spectrum allows the implementation oftransceivers with high-dimensional antenna arrays for analogor digital beamforming. For many applications, however, theoverall power consumption of the transceiver would be akey design parameter. In analog beamforming, the incomingsignals from the antenna array are combined in the analogdomain and processed by a single pair of ADC. In contrast,the digital beamforming uses a pair of ADC for each antennaelement. The architecture of the digital beamforming is moreflexible than the analog beamforming, thus making the digitalbeamforming more popular for the implementation of cellulartransceivers [1]. However, for a transceiver with large antennaarray, the high number of ADCs might lead to significantincrease in the power consumption. A recent report by Orhan et. al shows the possibility of reducing the overall power con-sumption of a fully digital transceiver without compromisingits performance by reducing the bit resolution of the high-speed ADCs [2].Previous works have shown high-speed flash ADCs withlow-bit resolution for applications in wideband transceivers[3], [4]. Although conventional flash ADCs have high speed,their power consumption is high because they require N − comparators for an N-bit conversion. The application of signalfolding technique in flash ADCs can reduce the numberof comparators while maintaining the high conversion rates,thereby giving rise to significant improvement of key designparameters including the power consumption, the kickbacknoise, and the chip area [5], [6]. In this work, we introduce a low-power 4-bit 1GS/s folding-flash ADC with folding factor of two. We propose an un-balanced comparator, which significantly improves the powerconsumption and the kick-back noise of the ADC. The ADCconsumes as low as 700 µ W from a 1V supply, giving a figureof merit (FoM) of 65fJ/conversion step. The paper structure isas follows: section II describes the system architecture, sectionIII describes the key design considerations of the comparator,and section IV presents the simulation results.II. S
YSTEM A RCHITECTURE
Fig.1 shows the architecture of the proposed folding-flashADC. The ADC comprises a track-and-hold (T/H) circuit, a1-bit folding stage, a 3-bit flash ADC, and a digital encoder.Two important design parameters in high-speed folding-flashADCs are the linearity of the folding stage and the kick-back noise, where the kick-back noise arises from the highfrequency switching in comparators. To increase the linearity,the sampling capacitor ( C s ) should be larger than the totalparasitic capacitance at the input nodes of the comparator [5].A sufficiently large C s can also reduce the kick-back noise,generated primarily by the comparators in the 3-bit ADC [6].The optimal value of C s in our design is 500fF.Fig. 1: Architecture of the proposed folding-flash ADC.Fig.2a illustrates the timing diagram of the clock signals. Allclock waveforms are generated by feeding an external clock toan inverter chain to produce the desired timing of these signals,shown in Fig.2b. To improve the accuracy of the ADC, CK and CK signals are generally out of phase with respect tothe CK T R signal. In our design, CK occurs 100ps after thehold phase to accommodate for the settling time of C s , while CK takes place 100ps before the next track cycle to accountfor the 3-bit ADC decision time. a r X i v : . [ c s . A R ] D ec a)(b)Fig. 2: (a) Timing diagram of the folding-flash ADC (b)Inverter delay chain to produce different clocks for ADC.The chopper circuit shares the charge on the samplingcapacitor C s with the parasitic input capacitance of the 3-bitflash ADC during CK . The parasitic input capacitance of the3-bit ADC is reset to zero during the tracking phase to removethe residual charge from the previous sample. In our design,the comparators in the 3-bit ADC are unbalanced with built-in references to quantize the input signal. The output signalsof the 3-bit flash ADC are in the form of a thermometriccode. The code is then passed to a digital encoder, whichincorporates a first-order bubble correction for producing amore accurate gray code.III. C IRCUIT DESIGN
A. Conventional double-tail comparator
Fig.3a shows the transistor-level architecture of a double-tailcomparator. The double-tail comparator is commonly used indata converters due to its high speed, low offset, and low staticpower consumption [7].To eliminate the conventional resistor ladder in the flashADC architecture, it is desirable to implement built-in ref-erences. This is typically done by introducing an intentionaloffset at the input of the comparator. There are differentmethods for implementing this offset. One approach involvesusing different size input transistors [3], [4] to make theirtransconductance different from one another. This, however,makes the input capacitance highly imbalanced, which can re-sult in unpredictable kick-back noise and degrade the linearity.Another approach to program the offset is by varying thecapacitive load difference at the mid p and mid n nodes of thecircuit ( C diff = C midp − C midn ) and keeping the input M and M transistors balanced. The shift in the trip point (i.e. V offset ) is given by the following expression [9]: V offset = I D g m , . C diff C sum = V ov , . C diff C sum (1) (a)(b)Fig. 3: Schematic of (a) conventional double-tail comparatorand (b) proposed double-tail comparator.where C sum is the total load capacitance in the balancedcase, I D , g m , and V ov , are the drive current, transcon-ductance, and overdrive voltage of the input pair in satu-ration region. To implement the capacitive load difference,Verbruggen et al. added an MOS capacitor at the mid p nodeof the circuit [5]. This modification changes the slew rateof the input transistors, and the resulting regenerative actionof the cross-coupled inverters for implementing the built-inreference. The potential drawbacks of this approach are thereduced linearity and larger comparator size. Alternatively,D’Amico et al. has implemented the offset by mismatching thesize of the M and M transistors [6]. One shortcoming of thisimplementation is the difference in the amount of the chargenjection at the mid p and mid n nodes during the dischargephase, which can potentially give rise to an erroneous decisionby the comparator. B. Proposed double-tail comparator
To overcome the aforementioned shortcomings of the previ-ous implementations, we introduce a new double-tail compara-tor, shown in Fig.3b. The new features of our design include:(1) implementation of the built-in reference by choosingdifferent size reset transistors M R and M R , (2) reduction ofthe kick-back noise by adding the intermediate M k transistorsand implementing a two-clock operation. It should be notedthat the use of a two-clock operation without adding the M k transistors results in the increase of the power consumption.We describe these features in the subsequent sections. Implementation of built-in reference : The 3-bit ADCconsists of 7 comparators with different trip points. For eachcomparator, we implement the offset by mismatching the sizeof the reset transistors. To explain the mechanism for creatingthe built-in offset in our circuit, we refer to Fig.3a. Transistors M , M , M , M , M R and M R contribute to the totalparasitic capacitance ( C sum ) at the mid p and mid n nodes.Assuming that M , M , M , and M transistors are fullymatched, only the reset transistors contribute to the differencein the capacitive loads at those nodes: C diff = C gsR + C gdR (1 + A cc ) − [ C gsR + C gdR (1 + A cc )] (2)where A cc is the gain of the cross-coupled inverter. Further,given the large gain of the cross-coupled inverter, we ignorethe contribution of the M , M , M , M transistors to C sum .We can therefore express C sum as: C sum = C gsR + C gdR (1 + A cc ) + [ C gsR + C gdR (1 + A cc )] (3)Since the reset transistors are operating in the triode region,the parasitic capacitance of the transistors is given by: C gs = C gd = C ox × W × L (4)Combining the equations 2, 3, and 4, and also assuming thatthe reset transistors have similar gate length L , we re-write theequation 1 as: V offset = V ov W R − W R W R + W R (5)We use this simplified model to estimate the size of thereset transistors. Although the offset can be easily imple-mented by mismatching the size of the reset transistors, itis critical to match the capacitive load at the out p and out n nodes. Otherwise, this might result in a significantly large,undesirable offset [9]. Therefore, to match the load at thesenodes, we connect the output nodes of each comparator to abuffer inverter. Reduction of kick-back noise : The kick-back noiseoccurs due to high-frequency voltage swings across the inputtransistors of a comparator. The cumulative kick-back noiseof all comparators in the 3-bit ADC can be large enough tocorrupt the sampled signal. Therefore, it is essential to reducethe kick-back noise of each comparator. We now proceed toexplain the origin of the kick-back noise and strategies formitigating it.In a double-tail comparator, the decision phase starts whenthe CK signal transitions from the low state to the high state.At this time, the mid p and mid n nodes begin to discharge intothe ground. This subsequently lowers the drain-source voltageof the input transistors and pushes them from the saturationregion into the triode region. The change in the operatingregion of the input transistors creates a kick-back charge thatresults in a noise (i.e. kick-back noise) at the input nodesof the comparator. Therefore, the kick-back noise corrupts thesampled signal in a single-clock operation scheme. To mitigatethis problem, we used a two-clock scheme for operating the3-bit ADC comparators, shown in Fig.3b. In our scheme, thereis enough time to refresh the input of the 3-bit ADC during t rfsh before the decision phase, thereby mitigating the effectof the kick-back noise. However, using the two-clock operationfor a conventional double-tail comparator will significantlyincrease the static power consumption due to the direct pathbetween the supply voltage V DD and the ground during t rfsh .To alleviate this problem, we have added the intermediate M K transistors. The size of these transistors influence the kick-backnoise. Therefore, we optimized the size of these transistors todiminish the kick-back noise during the decision phase.Fig. 4: Effect of the kick-back noise on the sampled signal atthe input node of the proposed and the conventionaldouble-tail comparators.We also used the proposed architecture shown in Fig.3bfor implementing the front-end folding comparator. Unlikethe comparators in the 3-bit ADC, this comparator is fullybalanced (no mismatch between M R and M R ) and uses asingle-clock operation scheme. Finally, we have verified thepossibility to calibrate the ADC against any process variationsusing the bulk voltage trimming [10].IV. S IMULATION RESULTS
The proposed ADC was designed in a standard 65nmCMOS process with a supply voltage of 1V to operate atig. 5: DNL and INL for different output code words.1GS/s. A differential signal of mV p − p is given to the inputof the ADC. Fig.4 shows the effect of the kick-back noise at1GS/s. According to our simulation results, the conventionaldouble-tail comparator exhibits a kick-back noise of around20mV (0.64LSB) while this value is about 3mV (0.1LSB)for the unbalanced comparator with a double-clock operationscheme. Fig.5 shows the simulation results for the systemlinearity, where the maximum DNL and INL are less than0.2LSB.The FFT plots for the input frequencies of 100MHz and500MHz are shown in Fig.6. The SNDR and ENOB are22.3dB and 3.42 bits at 100MHz while they are 21.8dB and3.34 bits at 500MHz. TableI summarizes the performance ofthe ADC. The ADC consumes about 700 µ W of which theT/H circuit, the comparators, the clock generator, and the en-coder consume about 10%, 25%, 45% and 20%, respectively.Figure of Merit (FoM) of the system was evaluated using thefollowing expression [11]:
F oM = power ENOB × f sample (6)We deduced the FoM of the ADC to be 65fJ/conversion step.TableII summarizes the comparison of our folding flash ADCwith other high-speed ADCs.Fig. 6: Power spectral density of 100MHz and 500MHzinput signals sampled at 1GS/s.C ONCLUSION
We introduced an unbalanced comparator architecture forrealizing ultra-low power high-speed ADCs with low-bit res-olution. We designed a 4-bit 1GS/s ADC in 65nm CMOS,which consumes 700 µ W. This translates into an FoM of TABLE I: Performance summary
Technology 65nm CMOSSupply voltage 1 VSampling rate 1GS/sNumber of Bits 4Input Swing mV p − p ENOB at 100MHz 3.42 bitsENOB at 500MHz 3.34 bitsINL/DNL 0.2 LSBSNDR at 100MHz 22.3 dBSNDR at 500MHz 21.8 dBPower 700 µ W TABLE II: Comparison of the proposed folding-flash ADCwith other high-speed ADCs
Ref. Architecture Power(mW) Fs(GHz) Res.(Bits) SNDR(dB) FoM(fJ/conv.)[3] Flash 2.5 1.25 4 23.8 160[5] FoldingFlash 2.2 1.75 5 28.5 50[6] Folding &Int. Flash 7.65 1 5 27.4 390[8] Delay line 1 1 4 21.3 126ThisWork FoldingFlash 0.7 1 4 22.3 65
EFERENCES[1] C. N. Barati et al., “Directional Cell Discovery in Millimeter WaveCellular Networks,” IEEE Transactions on Wireless Communications,December 2015.[2] O. Orhan et al., “Low Power Analog-to-Digital Conversion in MillimeterWave Systems Impact of Resolution and Bandwidth on Performance,”Information Theory and Applications Workshop (ITA), 2015.[3] G. Van der Plas et al., “A 0.16pJ/conversion-step 2.5 mW 1.25 GS/s 4bADC in a 90 nm digital CMOS process,” IEEE ISSCC, Feb. 2006, pp.566567.[4] P. Nuzzo et al., “A 10.6 mW/0.8 pJ power-scalable 1 GS/s 4 b ADC in0.18 µµ