The Analogue/Digital Converters (ADCs) play a very important role in several wideband applications like wired and wireless high speed telecommunication systems (e.g., 802.11g) or communication over powerlines (IEEE P.1901). High definition TV or high precision real time image processing are also examples of applications that require a conversion rate of several hundreds MSamples/sec or even multi-GSsamples/sec.
While the ADCs may operate in an optimal way when they are initially designed and verified using DC simulation, a transient simulation can designate several problems that appear during the high speed operation. Additional linearity errors are posed by process variations and component mismatches after the chip fabrication. Finally, operating conditions like voltage supply levels and temperature variations can also affect the linearity of an ADC. Several foreground and background calibration techniques have been proposed in the literature. Most of them are developed for specific ADCs and cannot be applied to different ADC architectures.
The most important error sources and the most popular calibration methods for Pipelined, Segmentation/Reassembly and Sigma Delta ADCs as well as a number of generic error compensation methods based on the processing of the ADC output are presented in (Balestrieri et al, 2005). A popular error correction technique used in pipelined ADCs exploits the least significant bit of a “coarse” ADC stage for the error detection and correction. For example, in (Colleran & Abidi, 1993) a 10-bit ADC is constructed by a 4-bit “coarse” and a 7-bit “fine” ADC. The least significant bit of the coarse ADC should match the most significant bit of the fine ADC. Similarly, a 10-bit pipeline ADC consists of a coarse 6-bit and a fine 5-bit ADC in (Sone et al, 1993). Two more recent approaches that are described in (Kurose et al, 2006) and in (Ahmed & Johns, 2005)(Ahmed & Johns, 2008) use 8 stages of 1.5-bit and a 2-bit Flash ADC stage in a 10-bit (or 11-bit in (Ahmed & Johns, 2008)) pipelined ADC architecture. Moreover, in (Ahmed & Johns, 2008), the DAC linearity errors are also taken into consideration. The use of a redundant signed digit also appears at an Analogue-to-Quaternary pipelined converter in (Chan et al, 2006).
The ADC architectures that are based on high precision capacitors suffer from the effects of the mismatch. In (Wit et al, 1993), an additional array of capacitors is used for real time trimming that is performed by an algorithm implemented on-chip in order to handle component ageing. Trimming arrays are also used in (Ohara et al, 1987). A digital calibration of the capacitor mismatch, the comparator offsets and the charge injection offsets in a pipelined ADC is performed in (Karanicolas et al, 1993) for the improvement of DNL errors.
The biasing of the operational amplifiers used in a pipelined ADC according to the power supply, the temperature and the sampling speed is determined by calibration in (Iizuka et al, 2006). The offset of the residue amplifiers is calibrated in the background in (Ploeg et al, 2005) (Van De Vel, 2009). Background calibration is also performed in (McNeill et al, 2005) where two identical algorithmic ADCs operate in parallel, their output is averaged and any difference in their results steers the calibration procedure. In (Wang et al, 2009), a nested digital calibration method is described for a pipeline ADC that does not require an input Sample/Hold Amplifier. A digital background calibration technique is proposed in (Hung & Lee, 2009) to correct gain errors in pipelined ADCs. This calibration technique performs the error estimation and the adaptive error correction based on the concept of split ADCs. In (Sun et al, 2008), a technique called Commutated Feedback Capacitor Switching is used to extract information about the mismatches of the capacitors used and then this information is exploited by a digital background calibration method.
Post processing techniques offer a different approach to the linearity error reduction of the ADCs. While all the aforementioned techniques target to the correction of the error sources, the post processing methods operate on the ADC output. The Differential or Integral Non-Linearity (DNL/INL) errors can be measured in order to estimate correction factors for each output code. These correction factors are stored in large lookup tables and are added to or subtracted from the corresponding output codes at real time. These lookup tables are also subject to real time calibration as described in (De Vito et al, 2007). The estimation of the correction factors can be performed in the simplest case by applying successive DC levels at the ADC input and measuring the DNL of the generated ADC output codes (Provost & Sanchez-Sinencio, 2004). More sophisticated techniques apply a sinusoidal signal to the ADC input and construct a Histogram using the resulting ADC output in order to estimate the DNL errors and consequently the correction factors (Correa-Alegria & Cruz-Sera, 2009).
In this chapter, some representative calibration approaches presented in the literature are described emphasising on the more general ones in the sense that they can be applied to different ADC architectures. Moreover, the calibration schemes proposed by the authors in a current mode implementation of a 12-bit ADC with a novel binary tree structure (Petrellis et al, 2010a) as well as in a voltage mode subrange ADC (Petrellis et al, 2010b, 2010c) are also presented since they can also be used in different target applications.
2. Resistor and capacitor trimming
The highest speed ADCs are based on the Flash or Parallel architecture where the input signal is concurrently compared to 2 n reference levels generated by a resistor ladder consisting of identical resistors R. The Flash ADCs cannot offer a high resolution (it is practically lower than 8-bits) since the required area and power is increased in an exponential manner. Linearity is essential in these ADCs in order to prevent the already low dynamic resolution from a further reduction.
A significant linearity error source in these ADCs is the component mismatches in the resistor ladders. If the tolerance in these resistors is expressed as ±kΔR, then it can be reduced to ±ΔR as shown in Fig. 1a where each resistor R has been replaced with a resistor R-kΔR connected in series with 2k trimming ΔR resistors that can be bypassed by the calibration algorithm or at a final stage of the fabrication process.
High precision capacitors are used in several ADC architectures that are based on charge redistribution, integrators, Sigma-Delta ADCs etc (Quiquempoix et al, 2006). Capacitor trimming can be performed in a similar way to the resistors whenever high precision capacitors have to be used (Wit et al, 1993). A simple way to perform such a capacitor trimming is shown in Fig. 1b. If the tolerance of a capacitor C is ±kΔC then, by using a fundamental C-kΔC capacitance and 2k trimming capacitors ΔC that can be potentially connected in parallel, the tolerance can be reduced to ±ΔC.
3. Redundant bits in pipeline ADCs
Pipeline, Subfolder and Subrange ADCs can achieve a descent resolution higher than 8-bit with a conversion speed that is comparable to that of the Flash ADCs. This is achieved by using a number of Flash ADC stages with lower resolution. For example in a two stage Pipeline ADC with m+n bits resolution, the analogue input is connected to a “coarse” m-bit ADC that generates the m most significant bits. These bits are used as input to a DAC in order to reconstruct an analogue signal that is subtracted by the original input and a residue is generated that serves as input to the second “fine” ADC stage that has an n-bit resolution. If the “coarse” ADC generates m+1 instead of m bits but the least significant bit is not input to the DAC as shown in Fig. 2, then this least significant bit should match the most significant bit of the n-bit “fine” ADC, otherwise the subtraction or the DAC operation has not been performed accurately enough (Iizuka et al, 2005)(Van De Vel et al, 2009). In this case, the offset of the subtraction operational amplifier or a resistor/capacitance trimming at the side of the DAC may be necessary.
4. Bias adjustment
The biasing of the operational and differential amplifiers used in several ADC architectures like pipeline or Sigma Delta ADCs often requires an accurate real time calibration around a typical value. An ordinary DAC cannot offer a high resolution adjustment since its dynamic range spans from 0 volts to its maximum range and is not focused around the typical bias
value that is required. For example, if an accurate adjustment has to be performed around 500mV in a range of ±16mV in steps of 1mV, then an ordinary 10-bit DAC would be required with reference voltage of 1024mV. Such an ADC is capable of providing any of the voltages between 0 and 1023mV in steps of 1mV, but most of these output levels would be unused in the specific bias requirement.
A much lower area/power 5-bit DAC would be sufficient if it could provide an offset of 484mV and a dynamic range of 32mV above the 484mV level. This can be achieved e.g., with a weighted current source DAC with offset like the one presented in (Petrellis et al, 2010a). An example of such a 3-bit DAC is shown in Fig. 3. The output current range of such a DAC is [I c ..I c +7I p ] according to the switch configuration. Of course, this current range can be mapped to a voltage one through the current to voltage converter I2V that can be simply a resistor R. In this case the voltage range is [RI c ..R(I c +7I p )]
5. Averaging the output of a pair of identical ADCs
Another method for the detection and the correction of errors at an ADC output is based on the generation of a duplicated output by a pair of identical ADCs (McNeil et al, 2005). For example, if two ADCs accept the same input they should generate the same digital output. Nevertheless, their outputs may differ slightly due to component mismatches and process variations. The averaging of these outputs can lead to an error reduction. Assuming that the ADC digital outputs are D+ε1 and D+ε2 respectively, where D is the ideal output and ε1, ε2 are the signed errors of each ADC output, the averaged output is
If ε1<0 and ε2>0 the averaged output has lower error than any of the two separate ADC outputs. Moreover, there is a great possibility that ε1=-ε2 and in this case the error is totally eliminated. In the worst case where ε1 and ε2 have the same sign, the averaged output has smaller error than the higher ε1 or ε2 error. More specifically if |ε1|<|ε2| then
The main drawback of this approach is the required die area and power duplication. A difference in the ADC outputs may trigger a more sophisticated calibration algorithm that corrects the error at its source instead of simply using the average of these outputs.
6. Lookup tables with DNL error correction factors
Another error correction technique that is based on the processing of the ADC output, estimates the DNL error of each output code and a corresponding correction factor. All of these correction factors are stored in a lookup table and are accessed at real time in order to determine how the current output code should be altered to improve linearity.
The DNL error is defined using the ADC transfer function shown in Fig. 4. In the ideal case, any output code should have the same width as the Least Significant Bit (LSB):
The parameter Vref is the maximum input voltage of the ADC and n is its resolution. The DNL error represents the relative difference of the actual code width from the LSB:
If the D i code is corrected by a factor f i that is defined as:
the corresponding DNL i error would be eliminated. Nevertheless, it is easier to estimate the initial DNL i error of a code instead of its width D i using for example the Histogram method (Correa-Alegria & Cruz-Sera, 2009). The correction factor can be estimated in this case as:
The lookup tables with the correction factors may also require real time calibration as described in (De Vito et al, 2007).
7. Current mode circuit calibration
In current mode implementations of ADCs, the current mirrors play a very important role. For example, in current mode Flash ADCs, the input current is compared to a number of current levels that are generated from a single reference level using appropriately scaled mirrors. The input (I in ) and the output current (I out ) of a simple current mirror consisting of transistors with the same length and width W in and W out respectively, that operate in saturation mode can be approximately expressed as:
Nevertheless, this scaling is not as accurate as indicated by equation (7) due to component mismatches, while it is also affected by temperature. Cascode current mirrors offer a higher accuracy and temperature stability than simple current mirrors due to higher output resistance but their usage leads to slower implementations.
In (Petrellis et al, 2010a) two versions of an ADC that is based on a current mode integer division are presented. Higher speed can be achieved by using simple current mirrors instead of cascode ones for the generation of reference currents and the implementation of operations like subtraction and multiplication/division by a constant. Since simple current mirrors are faster but more sensitive to component mismatches and temperature variations than cascode ones, replacing some critical simple current mirrors with gain-boosted ones in such an ADC can reassure its correct operation without sacrificing speed. The biasing of these gain-boosted mirrors is controlled by an appropriate calibration algorithm.
The novel ADC architecture presented in (Petrellis et al, 2010a) is based on the integer division of an input current I in by the reference current I ref and is defined using the following relation (q represents the integer quotient):
The current mode integer division can be implemented by the circuit shown in Fig. 5. If the relation (8) holds, then q current comparators are active connecting the same number of I ref current sources at the output. Thus, the quotient q is expressed as a multiple of the reference
current I ref . All the I ref current sources are implemented using current mirrors with equally sized transistors, while the 2I ref , 3I ref , etc, levels are generated by current mirrors with output transistors that have twice, or three times respectively the width of the input transistor as shown at the bottom of Fig. 5.
A novel ADC architecture based on integer division was presented in (Petrellis et al, 2010a) and is shown in Fig. 6. A binary tree structure is used and each node of the tree implements an integer division by a number of the form: , where L is the level of the tree (leaves are assigned to level L=0). The quotient and the residue of such a division are the outputs of each node and are connected to subtrees that correspond to ADCs with lower resolution. The residue can be estimated by subtracting the quotient from the original input of the integer divider. For example, if a copy of the input signal is available at the output of a PMOS current mirror and a quotient copy is available at the output of a NMOS current mirror, the subtraction can be carried out by simply connecting these two outputs and driving the residue to the input of a second NMOS current mirror.
Simple current mirrors with small transistors can be used to implement time critical operations of the ADC architecture that is shown in Fig. 6. At such critical nodes like the ones at the root of the binary tree of Fig. 6, gain-boosted current mirrors can be used like (M5, M6, IC0) and (M7, M8, IC1) that are shown in Fig. 7.
The α g and α off gain factors of the corresponding A0 and A1 operational amplifiers are determined by the voltages V g and V off respectively, through an appropriate calibration method that detects whether the residue signal has been shifted up e.g., due to high temperature or if its amplitude is different than the one expected. This conclusion can be extracted by the digital codes at the output of the ADC (e.g., missing codes). If the calibration method detects through missing output codes that the residue signal has been shifted up, it can remove higher offset α off I off current by increasing α off through V off . Similarly, if the residue amplitude is detected to be smaller than the one expected, it can be increased through the appropriate adjustment of α g .
The gain-boosted current mirrors can also be used in different ADC architectures like current mode pipeline ADCs. The adjustment of the subtraction outcome between the DAC output of a pipeline stage from its input in order to generate the residue can be carried out by a gain boosted current mirror arrangement like the one presented in Fig. 7.
8. Non-monotonic error elimination
Non-monotonic errors are a significant issue at several architectures but fortunately in most cases they do not appear at random transitions. For example, in two-stage pipeline ADCs such monotonic errors may appear during the transition of the residue signal between two peak values. For example, the sawtooth signal shown in Fig. 8 may represent two periods of a pipeline ADC residue. As can be seen in this figure the linearity of this residue is not very good since this signal does not start to rise immediately. Moreover, non-monotonic errors appear during the falling edge of each tooth since it does not fall immediately from its peak value to its minimum.
In ADCs like the ones presented in (Petrellis et al, 2010a) the severe non-monotonic errors appear whenever the bit No. 4 changes. Generally we assume that non-monotonic errors appear when the bit B i of an ADC changes. These non-monotonic errors are caused by slow falling edges like the one shown in Fig. 8 at the residue of the root ADC divider of Fig. 6.
A simple way to handle this kind of problem is to detect the changes in the bit No. 4 and keep the previous ADC output stable for an interval equal to the falling edge of the residue. Although, this technique does not lead to a linear solution, it eliminates most of the non-monotonic errors that are more important than the ones of the linearity. A simple analogue circuit capable of performing this non-monotonic error elimination is shown in Fig. 9.
When the input of each XOR gate in Fig. 9 rises from 0 to 1 the connected capacitor is charged almost immediately but when the input changes from 1 to 0, the capacitor is discharged through the resistor connected in parallel with the capacitor. During the time that it takes the capacitor to discharge the inputs of the corresponding XOR gate are different and its output is 1 generating a pulse with a duration that is determined by the RC constant. A pair of RC components are driven by the Q and the Q~ output of the D-flip flop that is used at the input of this circuit in order to generate a pulse at the BUSY signal both at the rising and the falling edge of the observed bit B i of the ADC output. The BUSY signal indicates to the system that reads the ADC output, should not take into consideration this output as long as the BUSY signal is active.
The BUSY signal duration cannot be determined very accurately in the way described above because it depends on the RC constant. A higher precision digital BUSY signal generator is shown in Fig. 10. The XOR gate of Fig. 10 compares two successive values of the observed ADC output bit B i . If they are different, a timer is enabled activating the BUSY signal for a specific time period that is determined from simulation results. The effect of the BUSY signal use is shown in Fig. 11 where the ADC output is reconstructed by an ideal DAC at the top of this figure and the non-monotonic error reduction that occurs when the BUSY signal is taken into consideration is shown at the bottom of this figure.
9. Correction of differential signals in voltage mode ADCs
High speed conversion is achieved by ADCs that operate on differential signals. The differential amplifiers are faster because the stage that converts the differential signal to a single-ended one is omitted. Moreover, differential signals are more immune to noise interference.
In voltage mode ADCs like the one presented in (Petrellis et al, 2010b, 2010c), the differential amplifiers that are used to perform addition or subtraction are also sensitive to component mismatches that can lead to the drifting of the output differential signals, away from their predefined levels as well as the modification of their amplitude. The authors propose a calibration method that continuously observes such differential signals and shifts them appropriately to their correct positions. Auxiliary components like draft frequency detectors and digital to analogue converters that generate fine voltage levels around an offset, are also required in such ADC architectures and are described in this paragraph.
A monitoring circuit can be used to decide whether two differential signals overlap or not. For example, if Fig. 12 shows the optimal differential residue signals form, then a voltage comparator that accepts as input these differential signals can decide whether they overlap or not.
The circuit shown in Fig. 13a monitors two differential signals (V + , V - ) and can control the level shifter logic of Fig. 13b. The circuit shown in Fig. 13b can insert delay and shift appropriately the inputs (V i + , V i - ) of a differential amplifier stage. Delaying and shifting just one of the differential amplifier inputs (V i + in the case of Fig. 13b) is adequate in order to set the amplifier outputs at the correct level. More specifically, the offset of the signals V i + and V i + can be determined by the bias voltages V b1 and V b2 respectively. The bias voltage V b2 can be connected to a constant source while V b1 can be controlled by the Digital to Analogue Converter (DAC) of the circuit presented in Fig. 13a.
When the calibration starts, the counter of Fig. 13a is cleared, shifting away the signals V i + and V i - . Consequently, the outputs V + and V - of the following differential amplifier stage are also shifted away. The comparator output at the input stage of the differential signal monitor shown in Fig. 13a is low, enabling the counting operation. The increasing counter output values force the differential amplifier input and output signals to get closer. When these signals start overlapping, the D-flip flop output will be set, disabling a further level shifting and after this time point, the differential amplifier output will have the desirable form of Fig. 12.
A higher resolution can be achieved at the output of the DAC in the range that we are interested in, if a DAC with offset is used like the one presented in paragraph 4. An undesirable phase difference in the signals V p + and V p - can be corrected by connecting a combination of capacitors (0..7C) in parallel as shown in Fig. 13b.
The undesirable phase shift of the signals V p + and V p - is usually dependent on the frequency of these signals. Consequently, a module that detects the draft frequency range of such signals could be employed to decide the appropriate capacitance combination through the switches S0-S2 of Fig. 13b and such a circuit is shown in Fig. 14.
The comparator of Fig. 14 and the 2nd Counter are used to enumerate how many times a monitored differential signal like V p + crosses a reference voltage V cmp in a time interval specified by the 1st Counter of Fig. 14. The Combinatorial Logic circuit disables both counters after the specified time interval. The output of the 2nd Counter can directly give an indication of the draft frequency range that was detected if this time interval has an appropriate duration. The switches S0-S2 of Fig. 13b can be directly connected to the output of the 2nd Counter.
The 2nd Counter output of the draft range detector of Fig. 14 can also control the bias of an ADC pipeline stage that accepts as input the signals V + and V - . More specifically, if this stage is implemented as a Flash ADC, the resistor ladders that generate the Flash comparator reference voltage levels may require different bias for different signal frequencies. This is due to the fact that the amplitude of this Flash ADC input signal is usually reduced at high frequencies because the previous pipeline stages may not be fast enough to reach their maximum amplitude as the frequency increases. Instead of attempting to adjust the same signal amplitude in all frequency ranges, the resistor ladder bias can be adapted to the input signal amplitude at a specific frequency. The output S0-S2 of the frequency detector shown in Fig. 14 can adjust this bias by connecting different voltage dividers or active DC voltage level generators.
The use of the Differential Signal Monitor and the Delay Insertion/Level Shifting circuit of Fig. 13 as well as the draft Frequency Range Detector of Fig. 14 are demonstrated in Fig. 15. The differential sawtooth curves at the bottom of that figure represent the monitored residue signals (V + and V - ) that are initially shifted away when the V b1 bias voltage of Fig. 13b that is driven by the DAC output of Fig. 13a, has its lowest value. The DAC output is the curve that appears at the top of Fig. 15 and is increased forcing the V + and V - distance to be reduced. When these differential signals start to overlap, the DAC output level is stabilised since the counter operation of Fig. 13a is disabled. The frequency range detector of Fig. 14 estimates then the frequency of the residue signals for the period indicated by the signal in the middle of Fig. 15.
10. Post processing techniques for linearity improvement
The use of the correcting factors stored in lookup tables that were presented in paragraph 6 is a type of post processing technique. The authors are currently developing different post processing techniques that are general enough to be used for the linearity improvement of several ADC architectures. These techniques are based on the fact that often the high DNL errors have a periodic form and appear at output codes with a specific format. For example, in a two stage pipeline ADC the residue that serves as input to the “fine” ADC stage may not consist of identical teeth in the sense that some teeth may have different amplitude or offset as shown in Fig. 16.
If the differential signals of Fig. 16 are input to a 4-bit “fine” Flash ADC, then the two resistor ladders that generate the voltage levels of each differential comparator at the input stage of this ADC have to be biased appropriately. Consider for example the differential signal at the top of Fig. 16. If the input range of this signal is assumed to be 680mV..880mV in order to cover the minimum/maximum peaks of all teeth, then some codes will be missing at the ADC output since some teeth do not span at the whole range (they have a lower than 200mV amplitude). In order to avoid missing codes, a smaller range can be assumed for this specific differential signal e.g., 700mV..840mV. A similar biasing approach may be chosen for the differential signal at the bottom of Fig. 16 to avoid missing codes. Nevertheless, in this case the codes of the binary form x0000 and x1111 will have a significantly higher DNL error than the others due to the teeth clipping. In fact, the DNL of these output codes will probably be higher than 1 LSB.
A post processing technique is under development by the authors that corrects such a high DNL error by detecting the erroneous codes at the ADC output and replacing them with successive codes of 1-bit higher resolution. The rest of the codes are simply shifted appropriately and their resolution is also extended by 1-bit. In order to decide the duration of the inserted codes an averaging of the ADC output code duration is continuously performed by a digital circuit. Simulation results show that the SNDR of the 8-bit ADC described in (Petrellis et al, 2010c) can be increased in this way by up to 6dB.
In a more general approach, the average duration of the ADC output codes can be continuously estimated, and a correction of the successive codes’ duration can be carried out. For example, if a code appears in average 5 consequent times while the codes X and X+1 appear 7 and 3 times respectively, then the last 2 appearances of X can be replaced with X+1.
The appropriate calibration techniques allow the ADCs to operate at the extremely high conversion rates required by the nowadays applications. Although many ADC architectures require customised solutions it was attempted to select and present the most popular and general ones. A number of calibration and post processing techniques that have been developed by the authors have also been presented. These techniques include current mode calibration based on the use of gain boosted mirrors as well as voltage mode calibration methods that perform differential signal monitoring, level shifting, delay insertion, frequency range detection and bias adjustment.