* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download A 10nW 12-bit Accurate Analog Storage Cell with 10aA Leakage
Survey
Document related concepts
Electrical substation wikipedia , lookup
Current source wikipedia , lookup
Schmitt trigger wikipedia , lookup
Resistive opto-isolator wikipedia , lookup
Semiconductor device wikipedia , lookup
Earthing system wikipedia , lookup
Voltage regulator wikipedia , lookup
Voltage optimisation wikipedia , lookup
Alternating current wikipedia , lookup
Surge protector wikipedia , lookup
Shockley–Queisser limit wikipedia , lookup
Switched-mode power supply wikipedia , lookup
Power MOSFET wikipedia , lookup
Stray voltage wikipedia , lookup
Mains electricity wikipedia , lookup
Current mirror wikipedia , lookup
Transcript
IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 39, NO. 11, NOVEMBER 2004 1985 A 10-nW 12-bit Accurate Analog Storage Cell With 10-aA Leakage Micah O’Halloran and Rahul Sarpeshkar, Member, IEEE Abstract—Medium-term analog storage offers a compact, accurate, and low-power method of implementing temporary local memory that can be useful in adaptive circuit applications. The performance of these cells is characterized by the sampling accuracy and voltage droop that can be achieved with a given level of die area and power. Hand calculations suggest past implementations have not achieved minimum voltage droop due to uncompensated MOS leakage mechanisms. In this paper, the dominant sources of MOS leakage are experimentally characterized in a standard 1.5- m CMOS process using an on-chip current integration technique, focusing specifically on the 1 fA to 1 aA current range. These measurements reveal an accumulation-mode source-drain coupling mechanism that can easily dominate diode leakage under certain bias conditions and may have limited previous designs. A simple rule-of-thumb is offered for avoiding this leakage effect, leading to a novel ultra-low leakage switch topology. A differential storage cell incorporating this new switch achieves an average leakage of 10 aA at room temperature, an 8 reduction over past designs. The cell loses one bit of voltage accuracy, 700 V on a 12-bit scale and 11.3 mV on an 8-bit scale, in 3.3 and 54 min, respectively. This represents a 15 increase in hold time at these voltage accuracies over the lowest leakage cell to date, in only 92% of the area. Since the leakage is independent of amplifier bias, the cell can operate on as little as 10 nW of power. Initial measurements also indicate the switch’s leakage decreases with the square of process feature size. Index Terms—Accumulation, analog memory, analog storage, attoamp, junction leakage, leakage, low-leakage switch, sampleand-hold, subthreshold, ultra-low current, weak inversion. I. INTRODUCTION M ANY modern analog CMOS circuits employ temporary analog storage to counteract circuit offsets and store signal variables [1], and implement adaptation and learning [2], [3]. Offset correction and signal storage are the most common applications of temporary analog storage due to their widespread use in switched capacitor circuits, and both can be easily implemented using only a capacitor and a MOS switch. This form of storage is a very compact, low-power, and precise storage technique, but is useful only for short time scales since the OFF state leakage of the MOS switch quickly degrades the stored charge [2]. Medium-term capacitive storage cells extend the storage time of short-term cells by reducing MOS switch leakage, and are Manuscript received November 18, 2003; revised May 7, 2004. This work was supported by the Center for Bits and Atoms NSF Research Grant under NSF CCR-0122419, the Office of Naval Research under Award N00014-02-1-0434, the Catalyst Foundation, the Packard Foundation, and the Swartz Foundation. The authors are with the Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA 02139 USA (e-mail: [email protected]; [email protected]). Digital Object Identifier 10.1109/JSSC.2004.835817 often used in adaptive circuit design. Many leakage-reduction techniques have been proposed in the literature [4]–[8]. However, none of these references has presented experimental work focused specifically on characterizing the switch leakage as a function of terminal voltages. As a result, it is not clear if the leakages achieved in these implementations are optimal, and if not, how they may be further reduced. Long-term cells extend the storage times of short-term cells by completely counteracting MOS leakage rather than simply reducing it. Some cells counteract leakage by periodically quantizing the capacitor voltage using some form of analog-to-digital conversion [2], [3], [6], [9], [10]. This scheme has the disadvantages of added circuit complexity and power due to both the analog-to-digital converter and the bussing required for multiplexing the converter between cells. A second long-term storage approach is to eliminate the MOS switch altogether by employing floating-gate technology [2], [3], [11], [12]. The disadvantages of this scheme are that it requires high voltages for writing to the cells, exhibits slow write times, and requires additional supporting circuitry and bussing to accomplish the writing [2]. This paper focuses on extending the hold time of mediumterm storage cells. It is argued in Section II that past mediumterm techniques have most likely not minimized MOS drain leakage. Then the primary sources of drain leakage in MOS transistors are experimentally verified in a 1.5- m process, and an important accumulation-mode conduction mechanism that has not been accounted for in previous designs is characterized. Section III introduces a new ultra-low leakage switch structure based on leakage results from Section II and uses it in the design of two new analog storage cells. Measurements from the fabricated cells are presented in Section IV. Finally, Section V concludes the paper. II. LOW-LEAKAGE MOS SWITCH DESIGN A. Motivation As was mentioned above, medium-term analog storage cells achieve extended hold times by reducing the MOS-switch leakage current that degrades the cell’s capacitively held charge. In the literature, the lowest-leakage storage cell implementations have employed several different techniques to reduce switch leakage. These techniques are briefly reviewed below, and then a unified framework for comparing the performance of past implementations, which were fabricated in various technologies, is introduced. This comparison motivates the remainder of the section. 0018-9200/04$20.00 © 2004 IEEE 1986 Fig. 1. Two low-leakage MOS switches. Two low-leakage switch topologies that have been used in previous analog storage cells are shown in Fig. 1. Part (a) shows and a transmission gate switch and two leakage currents, , that can flow to even when the switch is off. flows from to ground through the NMOS drain-to-bulk flows from to through the PMOS diode and drain-to-bulk diode. Since only the difference between these two , reaches , it is possible to ratio the leakage currents, junction areas to counterbalance these leakages to first order. Note that a given junction area ratio minimizes the net leakage , however typical high-perfor only one particular value of formance storage cell topologies ensure this voltage remains fixed, independent of the stored analog voltage, allowing the . This strucarea ratio to be optimized for one value of ture was implemented in 2 m CMOS in [6] and exhibited a net leakage current of 1 fA (all currents are inferred from measured memory cell voltage leakage rates and storage capacitor sizes). In [5], the same structure implemented in 1.2 m CMOS yielded 4.5 fA of leakage current, while a fully differential storage cell, which only responds to differential leakage, used a pair of these low-leakage transmission gates to achieve a net differential leakage of 0.08 fA. A second low-leakage switch topology is shown in Fig. 1(b). The switch is formed by a single transistor, in this case an NMOS, fabricated inside a well to allow the transistor bulk voltage to differ from the substrate voltage. A disadvantage of this switch is that it lacks the rail-to-rail input-output range of a transmission gate. However, as was mentioned above, typical to remain constant making storage cell topologies force this limitation irrelevant as long as is chosen within the conallowable range of the switch. Additionally, with stant, the bulk can be driven with an identical voltage to minimize the drain-to-bulk diode leakage. In [7], NMOS versions of this low-leakage switch structure were used in a fully differential storage cell implemented in 3- m CMOS. The cell also attempted to minimize subthreshold leakage by IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 39, NO. 11, NOVEMBER 2004 driving the transistor gates below the bulk to turn them off, and achieved a typical net differential leakage of 1.6 fA. It is instructive to compare these measured leakages with back-of-the-envelope estimates of what could theoretically be achieved in each technology. Suppose the switch is implemented using a single minimum-sized MOS transistor and all MOS-related switch leakage, such as subthreshold conduction, is eliminated. Then the only remaining leakage would be due to the drain-to-bulk junction diode. The fully reverse-biased leakage per unit area of this junction in a modern process is roughly 0.02 fA m at room temperature, with this value having decreased by a factor of two to three in the past ten years as technology has improved [13]. Using appropriately scaled versions of this nominal leakage parameter for older processes, the estimated in MOSIS scalable design leakage of a minimum-sized ( rules) reverse-biased drain-to-bulk diode in each technology of interest is calculated in Table I. Before these estimates of nominal leakage current can be compared to the measured results from above, the leakage-reduction techniques employed in each implementation must be taken into account. The following assumptions will be made about the effectiveness of the techniques employed. It will be assumed that both counterbalancing NMOS and PMOS leakages and implementing the storage cell differentially achieve 80% matching accuracy, and therefore applying either technique will reduce the leakage to 20% of its starting value. Drain-to-bulk voltage minimization is typically accomplished using an amplifier in feedback [7], and therefore brings the two terminals no closer than the input-offset voltage of the amplifier employed, which will be assumed to be 5 mV. This level of bias across the diode can be shown to yield an 80% reduction in diode current when compared to full reverse-bias [see (2)]. Applying the relevant reduction factors from above to the nominal reverse-biased diode leakages from Table I, a conservative upper bound on the leakage current expected in each implementation is calculated in Table II. Also listed in the table are the currents that were measured in each implementation, which in all cases are at least 2.5 times larger than the generous estimates of the leakage upper bounds. These results suggest that additional leakage mechanisms, probably due to the MOS structure, exist and have not been compensated for in these implementations. In the remainder of this section, the dominant sources of leakage in MOS devices are characterized in a standard 1.5- m CMOS n-well double-poly process. The goal of these measurements is to establish a list of guidelines that should be followed to achieve minimum switch leakage and thus maximum storage times with a given capacitance. The measurements also reveal an important MOS leakage mechanism that has not been considered in previous designs. B. Leakage Test Structure and Test Procedure Direct measurement of sub-pA level currents is tedious to perform in the laboratory, even with a sufficiently accurate electrometer. At these current levels, the resistance of many common laboratory insulators is nonnegligible, requiring careful design of the entire test setup to ensure insulator leakage does not affect the measurement. Additionally, external O’HALLORAN AND SARPESHKAR: A 10-nW 12-bit ACCURATE ANALOG STORAGE CELL WITH 10-aA LEAKAGE 1987 TABLE I ESTIMATED REVERSE-BIAS LEAKAGE CURRENT OF MINIMUM-SIZED DRAIN-TO-BULK JUNCTIONS FOR SEVERAL TECHNOLOGIES. THE PRE-CONSTANTS IN THE LEAKAGE PER UNIT AREA COLUMN ARE INCLUDED TO ROUGHLY ACCOUNT FOR TECHNOLOGY AGE [13] TABLE II ESTIMATED UPPER BOUND ON LEAKAGE AND ACHIEVED LEAKAGE FOR THREE MEMORY CELLS Fig. 2. MOS leakage test structure. electrical interference must be completely eliminated as even faint signals can easily corrupt the measurement. Within an integrated circuit, however, fA-level currents can be accurately conveyed across the die [13], and the entire die can be shielded from interference by placing it in a grounded metal box. Because of these facts, measuring MOS leakage currents using on-chip circuitry can yield a level of accuracy that may be difficult to achieve with an off-chip electrometer. As was alluded to earlier, capacitive current integration offers a simple on-chip mechanism for indirectly measuring low current levels. It avoids off-chip error sources by first time-integrating the small current within the low-noise on-chip environment to yield voltage movements of a sufficient magnitude that they can be reliably measured in the noisy off-chip environment. The current-integrator leakage tester shown in Fig. 2 conand a PMOS-input fully sists of an integrating capacitor cascoded wide-output swing operational transconductance amplifer (OTA). In this figure the device under test (DUT) is an NMOS transistor, but the same structure was also used to characterize PMOS leakage. The current-integrator topology, when built using a high-gain amplifier, possesses several key features that are important to the operation of the leakage tester. First, (plus the the negative input terminal of the OTA servos to OTA’s input-offset voltage) as long as the amplifier output is not railed, which allows for indirect control of , the DUT drain that flows into the DUT drain voltage. Second, any current will be immediately integrated on , causing a change in but virtually no change in . Third, the rate of change of is related to and by (1) independent of capacitances from and to ground. Finally, a constant-current load at the integrator output due to causes only a constant an off-chip instrument monitoring voltage offset error, which has no effect on the measured . Using the leakage tester presented above, the DUT drain current can be evaluated as a function of the four MOS terminal voltages by iteratively applying the following procedure. At the start of each test run, initialize the leakage tester so that the four terminals of the DUT are at the required test voltage levels, and the output of the integrator sits halfway between the circuit rails, . This initialization is accomplished by temporarily at 1988 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 39, NO. 11, NOVEMBER 2004 using the DUT as a switch to reset the voltage across . After with an initialization, measure the time rate-of-change of off-chip electrometer. Using this slope measurement and the , the current can be indirectly evaluated using value of (1). In order to obtain accurate results with the leakage tester, must be accurately known. Therefore, a capacitor calibration structure was implemented on each of the test die using two and , with nominally identical package-repackage pins, ’s associated die-level bond pad was lated capacitances. Pin , whose connected to the top plate of a calibration capacitor, layout is identical to that of . The bottom plate of was grounded. Pin ’s associated die-level bond pad was not connected to any other die-level structure. By measuring the and , an estimate of difference in capacitance between (and thus ) can be obtained that should be independent of package and die-level parasitic capacitances to first order. Using this calibration structure and an LCR meter, the fabricated were all measured to lie in the range pF values of pF. The median value of pF was used to infer the leakage data presented in the remainder of this section. C. NMOS Leakage Analysis and Characterization Transistor drain leakage is typically modeled as the combination of two separate phenomena: diode leakage and subthreshold conduction. Drain-to-bulk diode leakage for nonzero drain-to-bulk voltages was discussed earlier and was illustrated in Fig. 1. In an NMOS, this current can be modeled to first order by the ideal diode equation (2) is the diode current and the applied voltage, where is the diode saturation current and is a constant proportional to a linear combination of the junction area and perimeter, and is the thermal voltage (roughly 26 mV at K). The second source of leakage is due to subthreshold conduction. The following equation describes the drain-to-source cur, of a subthreshold ( ) NMOS transistor rent, (3) where is a process-dependent current-scaling constant, is the transistor’s threshold voltage, is its width over length ratio, and is a parameter that is a function of the applied source-to-bulk voltage and has a magnitude less than one [14].1 The threshold voltage is defined by the body-effect relation (4) , which is and is a function of the nominal threshold voltage fixed for a given technology, the applied source-to-bulk voltage , and and , which are process-dependent parameters. Since the drain leakage is the sum of these two leakage mechanisms, subthreshold conduction can be made negligible by reuntil . The at which this happens ducing 1The parameter is equivalent to the reciprocal of the parameter n given in this reference. Fig. 3. Five different minimum-sized NMOS I versus V curves with V = = 150 mV. All I ’s are referenced with respect to the direction shown in Fig. 2. V = 0 V and V has several interesting dependencies. First, it is a weak function – if the diode leakage is reduced must decrease to of ensure that subthreshold conduction remains small enough to be ignored. Second, it depends linearly on the transistor’s nomis a function of the difference inal threshold voltage since and . Therefore, as the threshold voltages of between modern processes decrease to enable low-voltage design, the at which subthreshold conduction can be ignored is steadily decreasing. The net result of this trend is that in most modern processes the subthreshold current and diode leakage equalize V. Due to this effect, most modern NMOS transisat a V) but have nonzero tors that are digitally OFF ( will conduct with a level of several orders of magnitude larger than . The tradeoff between diode leakage and subthreshold conduction outlined above appears to be experimentally verified in Fig. 3. The figure shows the measured versus characteristics of five different minimum-sized NMOS transistors with V and mV. The current (always referenced with respect to the direction shown in Fig. 2) varies exponentially with the gate-to-source voltage for mV indicating subthreshold conduction dominates in this mV region, and remains roughly constant for where diode leakage apparently dominates. Thus, one method of minimizing NMOS subthreshold conduction in this process mV. Note that in this particular example, is to bias the gate would have to be driven to a voltage outside the range of the existing rail voltages to avoid subthreshold conduction. An alternate method of avoiding subthreshold leakage was employed in the transmission gate switch shown in Fig. 1(a), and is also discussed in [13]. Since it is the difference between and that controls , subthreshold current can be reduced by increasing , or equivalently . For large enough , this technique allows subthreshold conduction to be made V. Shifting has the added negligible for a level of V without the need benefit of allowing for levels of O’HALLORAN AND SARPESHKAR: A 10-nW 12-bit ACCURATE ANALOG STORAGE CELL WITH 10-aA LEAKAGE Fig. 4. Minimum-sized NMOS accumulation-mode I versus V curves for 50 mV V 150 mV, with V = 0 V and V = 150 mV. 0 to drive the NMOS gate below the lower rail. Both of these effects alleviate the rail problem, however the disadvantage of inin the case of a switch is that it also reverse-biases creasing . the drain-to-bulk diode, increasing The options presented above for avoiding subthreshold conduction in NMOS transistors in this process are unattractive. However, if it were possible to fabricate the NMOS inside a well, the switch topology presented in Fig. 1(b) could be biased near to simultaneously minimize used with subthreshold leakage, avoid the rail problem, and minimize drain-to-bulk leakage. The obvious alternative in an n-well process is to implement the switch using a PMOS transistor. Before examining the leakage of a PMOS version of this switch, we present observations of an accumulation-mode leakage mechanism that can easily dominate diode leakage in both PMOS and NMOS transistors. D. Accumulation-Mode Source-Drain Coupling Equation (3) is valid only when the NMOS structure is biased in weak inversion, which was true in the above experimental mV. However, for mV data for the transistor is biased in a different region of MOS operation known as accumulation-mode. In this regime, the subthreshold mechanisms that couple the source and drain in the MOS channel are negligible. As a result, it is typically concluded the source no longer significantly interacts with the drain in accumulation mode. The measurements presented in Fig. 4 not only show that this assumption is false, but that the source-drain interaction is still quite strong. In this experiment, the accumuversus characteristic of a minimum-sized lation-mode with NMOS transistor was measured at multiple values of V and mV. The plot shows that as approaches , the curves asymptote to a -independent aA – much lower than the apparent lower drain current of bound seen in Fig. 3. reFig. 5 characterizes the accumulation-mode versus sponse of the same minimum-sized NMOS transistor at a fixed 1989 Fig. 5. Minimum-sized NMOS accumulation-mode I versus V 300 mV, with V = 0 V and V = 1 V. 150 mV V 0 curves for Fig. 6. Empirical model of accumulation-mode NMOS leakage sources. V, for multiple values of . The plots show that mV mV the drain current still defor pends exponentially on , changing an order of magnitude for , but is independent of the source-to-bulk a 60 mV change in mV. Also, the residual leakage values voltage for , most likely due to the drain-to-bulk weakly increase with diode’s reverse current increasing with reverse bias. Modeling this response as an ideal diode plus a drain-dependent offset current gives data fits with correlation coefficients above 0.99, suggesting that a significant component of the accumulation-mode leakage current is due to a coupling of the source-to-bulk diode current to the drain terminal. Finally, since the source and drain are symmetric terminals in the MOS transistor, the observed interaction will also be symmetric. An empirical model of this accumulation-mode conduction mechanism in the NMOS is shown in Fig. 6. The distributed drain-to-bulk and source-to-bulk diodes are modeled as two and pairs of lumped diodes. The first diode in each pair, , are lumped models of the portions of the distributed diodes that only interact with the bulk. The second diode in and , represent the portions of these each pair, distributed diodes that only interact with the opposite MOS terminal. It is postulated that the mechanism linking these two 1990 diodes is carrier diffusion under the accumulation layer. This theory is bolstered by the fact that in Fig. 4 the drain current in becomes more negative. The the upper curves decreases as empirical model explains this behavior as a reduction in the and due to the increasing cross-sectional areas of . depth of the accumulation layer with decreasing A second interpretation of this empirical model is to view the and as a single symmetric lateral coupled diodes bipolar transistor acting between the source, bulk, and drain terminals. A detailed device-physics-based characterization of this leakage mechanism could be developed based on either of these equivalent views. However, since the primary goal of this section is only to understand how to minimize MOS leakage, a simple rule-of-thumb for avoiding this leakage mechanism is offered and the effect is not characterized further.2 From the symmetry of the empirical leakage model, it is expected that the accumulation-mode source/drain interaction . Under these can be minimized by ensuring that and will be conditions, the currents flowing in equal in magnitude and opposite in direction, summing to zero like in traditional subthreshold conduction or symmetric lateral bipolar conduction. This was previously observed in Fig. 4 for mV, and resulted in a residual leakage due to only . Additionally, if either (or both) or mV, the model predicts the net leakage from that node will depend on the node voltage only weakly through the diode’s reverse-bias current dependence on reverse-bias voltage, as seen in Fig. 5. The implications of the accumulation-mode leakage effect on is the switch topology of Fig. 1(b) are clear. In this switch, intentionally biased as close to 0 V as possible to minimize the drain-to-bulk diode leakage. Thus, to ensure that the accumulation-mode conduction mechanism characterized above does not dominate diode leakage when the switch is turned off, we reV. quire that E. PMOS Low-Leakage Switch The final measurements in this section characterize the leakage of a PMOS version of the switch topology shown versus curves of five in Fig. 1(b). The measured minimum-sized PMOS transistors are shown in Fig. 7 for mV mV, with V, a bulk-to-substrate V, and V. voltage of These bias voltages eliminate subthreshold conduction, minimize drain-to-bulk diode leakage, and minimize accumulation-mode conduction. The leakage is characterized as a because this voltage will eventually depend on function of a random amplifier offset, much like it did in the leakage tester of Fig. 2. Thus, the sweeps shown in Fig. 7 represent the range of leakages that most likely will be encountered in a storage cell using a minimum-sized PMOS transistor as a switch. 2Further characterization of the effect is also not trivial. The drain terminal leakage is easy to measure using the current-integrator structure because the transistor can also be used as a switch when resetting the integrator between test runs. However, this ability to easily reset the integrator is lost if the source terminal is simultaneously connected to a second integrator, or if the bulk terminal leakage is being characterized. IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 39, NO. 11, NOVEMBER 2004 Fig. 7. Five minimum-sized PMOS I versus V = 1:65 V, and V = 3:3 V. V curves with V = 0 V, An interesting result seen in Fig. 7 is the nonzero current V. This highlights an additional leakage when mechanism unique to transistors fabricated inside wells. In these transistors, the well-to-substrate junction can interact with the drain through minority carrier diffusion, just as the source-to-bulk junction does in accumulation-mode. Equivalently, this leakage can be viewed as the result of a vertical bipolar transistor acting between the drain, well, and substrate terminals. Since the well-to-substrate junction is heavily reverse-biased and the drain-to-well junction is not, a net differential leakage current flows from the drain to the substrate, accounting for the observed offset. This theory is supported by measurements showing this offset leakage approaches zero approaches 0 V. The first storage cell in the next as section accepts this offset as unavoidable, while the second cell attempts to eliminate it. III. ANALOG STORAGE CELL IMPLEMENTATION A. Single-Sided Storage Cell Overview A top-level view of the single-sided switched-capacitor storage cell topology is shown in Fig. 8. This cell is based on a standard high-accuracy sample-and-hold circuit presented in [15]. The analysis of cell operation that follows ignores switch charge injection and device noise, but these error sources will be considered in a moment. Additionally, for the moment and assume that V. ignore The sampling phase of circuit operation occurs while the digital signals and (a delayed version of that occurs later is low. Under these conditions in time) are both high and switches and are closed and switch is open. This ties to the the amplifier in unity-negative feedback and connects . The net voltage apleft side of the sampling capacitor during this phase is , plied across where is defined to be the offset between the negative and positive terminals of the amplifier in this unity-negative feedback configuration. At the instant in time when switches from O’HALLORAN AND SARPESHKAR: A 10-nW 12-bit ACCURATE ANALOG STORAGE CELL WITH 10-aA LEAKAGE 1991 Fig. 8. High-level view of switched-capacitor storage cell. high to low, switch opens and the sampling phase ends. Asinto the virtual ground suming zero charge injection from will be node of the amplifier, the voltage held across (5) Notice that if there is a nonzero parasitic capacitance at and althe node, the capacitor divider formed by lows movements in to alter even after opens. is also forced to However, because of the capacitive divider, . When switch eventually closes move in proportion to and the hold phase begins the resulting negative feedback loop to return to a voltage nearly identical to its samwill force pling phase value. This can happen only if the left terminal of returns to the voltage that was present at at the instant when opened. This independence from movements at once is opened is important when switch topologies are chosen. opens and the input voltage has been samAfter switch pled, the delayed control signals and cause to open to close. This connects the left terminal of to the and , configuring the circuit in the same output of the amplifier, topology as the current integrator of Fig. 2 and highlighting that it is switch that must be low-leakage. For the moment, ignore the transient behavior of the circuit after closes and focus on the final dc voltages that will exist after the system settles in this new configuration. Intuitively, negative feedback will force the virtual ground node to a hold-phase voltage of (6) where is due to the amplifier’s finite open-loop dc gain The hold-phase output voltage is . (7) Using (7), can be expressed in terms of the change in from the sample to the hold phase (8) which simplifies to give (9) Fig. 9. PMOS-input fully cascoded wide-output swing OTA. All devices are sized 9 m/4.5 m. causes to differ from the deSince the error term , the amplifier’s dc gain must sired held voltage tolerable at the extremes of . be large enough to keep To this end, the amplifier was implemented as a subthreshold fully cascoded wide-output swing OTA, shown in Fig. 9, and . All dedesigned with an open-loop dc gain vices in the amplifier were sized with m m. V and V, (9) predicts that the With absolute value of the sampling error due to finite gain will be for V . B. Charge Injection, Leakage, and Switch Implementations When MOS switches are turned off they inject channel charge and parasitic capacitive charge into the circuit nodes attached to their drain and source. Since the stored information of the above cell is represented by the sampled charge at the virtual ground node of the amplifier, any charge injection into the viropens will alter the stored analog tual ground node when value. To minimize this charge injection, the standard dummy transistor compensation technique is employed in constructing [16]. The low-leakage, low-charge-injection switch structure used for is shown in Fig. 10(a). Note that a single-delay version of the control signal is used in the switch control, for reasons that will be clear when the transient response is discussed. connects to while terminal Referring to Fig. 8, terminal connects to . Transistor forms the core of the switch and is the PMOS equivalent of the low-leakage NMOS switch implements charge-injection cancelof Fig. 1(b). Transistor lation by nominally extracting the same charge from the virtual injects into the node, and is delayed approground node as , priately. The bulks of both of these devices are biased at as this is the best available estimate of the voltage that will be during the hold phase of the cell. This minimizes present at the deterministic component of the source/drain-to-bulk diode and , but does not counteract the random voltages of or the input-dependent offset amplifier input-offset voltage . Choosing significantly below also ensures that is biased deep in accumulation-mode when it is off, minidisconnects mizing its subthreshold conduction. Transistor from during the hold phase so that the source of can connect this node to , minimizing accumulation-mode 1992 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 39, NO. 11, NOVEMBER 2004 M Fig. 10. (a) Low-leakage low-charge-injection switch for single-sided cell use. All transistors are minimum-sized (2.25 m/1.5-m) except , which consists of two minimum-sized transistors in parallel. (b) Low-leakage switch for differential cell use. Compared to the switch in (a), has been removed and now consists of only a single minimum-sized transistor. M source-drain conduction in . Transistor ’s bulk is biased at to ensure that this transistor’s drain-to-bulk diode cannot during the hold phase. be forward biased by a high level at can be biased at either or , whichever The bulk of is more convenient during layout. It should be emphasized that past switch designs have also bideep in accumulation-mode when it is off to minimize ased ’s drain-tosubthreshold conduction, and have minimized bulk diode voltage to minimize diode leakage. It is the novel adand combined with the switching and biasing dition of scheme employed that minimizes accumulation-mode conducand allows this switch to achieve even lower leakage tion in compared to past designs.3 Since movements at don’t affect the cell’s sampled inis opened, the charge-injection of switches formation once and don’t affect the cell’s sampled value. However, the leakage of these switches does introduce an additional amplifier voltage offset defined by the ratio of the leakage current to the amplifier transconductance, m . Fortunately, this offset is less than 1 V for standard MOS switch leakages ( fA) and amplifier bias currents greater than 500 pA. Thus, the only restriction placed on these two switches is that they both must pass rail-to-rail voltage levels to maximize the inputoutput voltage range, and therefore they were implemented as minimum-sized transmission gates. C. Random Noise An additional sampling error can be attributed to random circuit noise, which introduces an uncertainty in the voltage apduring the sampling phase. When transiplied across will tions from high-to-low, the voltage sampled across be the sum of the value given in (5) plus the instantaneous value of this noise voltage. In this circuit the noise was determined to be dominated by the amplifier thermal noise. Since the amplifier is effectively tied in unity negative feedback during all phases of circuit operation, the total thermal noise is similar in form 3To be fair, an equivalent switch topology was mentioned in [3] during a discussion of the circuit implemented in [8]. However, the topology was not actually employed in [8], and was only mentioned in [3] as a simple method of avoiding accidental forward-biasing of the switch’s lateral bipolar transistor. The data presented in Section II clearly illustrates that even cutoff-region operation of the lateral-bipolar transistor can significantly affect switch leakage. M to that of a standard single-pole RC circuit, and is described by [17] m (10) m is the effective number of devices conIn this equation, in this amplifier), tributing noise from the amplifier ( is the electron charge, m is the subthreshold transconductance of the amplifier, and is the total effective capacitance present at the output of the amplifier. The RMS voltage noise of the circuit can be evaluated by taking the square root of (10) and simplifying to get (11) where represents the expected standard deviation of the noise voltage, and . During the sampling phase while during the hold phase alone – does not affect the hold-phase noise. Capacitor serves the added role of increasing the loop’s phase margin to 90 . The capacitor values were chosen based on noise, chargeinjection, and frequency response constraints to be pF and pF. Assuming , (11) preV, dicts a sampling-phase standard deviation of V. Since and a hold-phase standard deviation of thermal noise is Gaussian, the sampling error will lie within a window over 99% of the time [18]. The random sampling voltage error window expected in this cell is V. Coupled with the finite-gain and charge-injection errors, this still theoretically allows the cell to achieve 12-bit accurate sampling. D. Transient Operation and Reset Phase The transient response of the storage cell during its transition from the sampling phase to the hold phase requires special consideration due to the topology chosen for switch . In the , the voltage stored across after case where opens will be negative. Ignoring for the moment, the voltage stored on at the end of the sampling phase will . The problem occurs when switch connects the equal to the output of the amplifier, and thus left terminal of . Since , the to O’HALLORAN AND SARPESHKAR: A 10-nW 12-bit ACCURATE ANALOG STORAGE CELL WITH 10-aA LEAKAGE resulting capacitive divider causes ’s voltage to only deat the instant when closes. As a crease slightly below transiently rises from to result, the left terminal of , and the capacitive divider formed by and forces above . The voltage at remains above while the amplifier’s output slews to the equilibrium dc output voltage determined previously in (7) and (9). Referring to the diagram in Fig. 10(a), it is clear that if is pulled above , the of and will forward bias alsource/drain-to-bulk diodes of lowing charge to flow out of , corrupting the stored analog information. to 0 V One solution to this problem is to always reset opens, but before closes, which will ensure that after during the initial transient and amplifier is always less than slewing phases. This is accomplished using an NMOS transistor , as shown in Fig. 8. The digital connected in parallel with high after has gone low but becontrol circuitry pulses goes low, with the pulse lasting . The charge fore during slewing injected by ’s reverse-biased diodes into still introduces an additional sampling offset error, but calculations show that it is negligible. 1993 Fig. 11. Modifications of original cell to implement differential leakage and charge cancellation. E. Differential-Leakage Cancellation The leakage of the storage cell in Fig. 8 is expected to be significantly larger than the values measured in Fig. 7 as a result of the dummy-device charge cancellation employed in . Referring again to Fig. 10(a), this technique required to be con’s structed from two parallel minimum-sized transistors and source and drain to both be connected to the virtual ground node of the amplifier. This makes the source/drain-to-bulk area connected to the virtual ground node a factor of four larger than that of a minimum-sized device. Since diode saturation current is proportional to junction area, both the offset and slope of the leakage current of this switch will be approximately four times larger than the values measured in Fig. 7, increasing to approximately 70 aA and 4 aA/mV, respectively. The actual increase in these parameters is not so simple to predict because the mechanisms involved exhibit more intricate dependencies on junction sizing and layout geometry, but it is certain that they will increase. The estimated leakage floor of the single-sided cell can be reduced by modifying the original cell of Fig. 8 to employ differential charge-injection and leakage-cancellation techniques. The necessary modifications to the original cell are highlighted directly to the posiin Fig. 11. First, rather than connecting , tive terminal of the amplifier, this voltage is sampled using a minimum-sized PMOS, onto which is identical to . The left terminal of could be tied to any dc is preferred as it is a quiet reference voltage voltage, but and thus offers a high level of power supply rejection. Switch , shown in Fig. 10(b), is a modified version of the origtopology with transistor eliminated and transistor inal formed from a single minimum-sized PMOS rather than two in parallel. The difference between the charge injections and onto and , respectively, define of the net charge injection, which should nominally be zero due to matching. Additionally, the net hold-phase leakage will be and , two set by the difference between the leakage of Fig. 12. Difference between input and stored voltage in single-sided cell. minimum-sized PMOS transistors that are both biased in the ultra-low-leakage configuration. Assuming 80% matching, the offset component of the leakage observed in Fig. 7 should be aA, while the random component due reduced to less than to the amplifier input-offset voltage should remain at aA -mV range. for offsets in the IV. MEASURED PERFORMANCE A. Single-Sided Cell The measured dc input-output response of a representative V, V, single-sided storage cell with and an amplifier bias current of 1.5-nA is shown in Fig. 12. The plot shows the difference between the input and stored voltage as a function of the input voltage, as well as a first-order fit below 0.2 V and to the relevant portion of the data. For above 3.1 V the error rapidly increases, indicating these sampled voltage levels are outside the saturation region of the amplifier’s output leg. However, inside these limits the absolute value of 1994 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 39, NO. 11, NOVEMBER 2004 TABLE III MEASURED PERFORMANCE OF FABRICATED SINGLE-SIDED STORAGE CELLS Fig. 13. Single-sided cell leakage. the sampling error is less than 600 V, including sampling V—roughly noise. The sampling error when 200 V based on the data fit—represents the constant offset error due to charge injection, since at this input level the finite-gain induced error is zero. The fit also shows that at the the sampling error lies within V of the extremes of nominal offset of 200 V, indicating a maximum finite-gain induced error of V. The measured dc sweep parameters of the four storage cells that were fabricated are listed in Table III along with the statistical results. Since the absolute V V, the sampling error of all cells is less than cells achieve 12-bit accurate sampling. The leakage responses of the four cells are shown in Fig. 13. The leakage was verified to be virtually independent of the sampled voltage level, as expected, so the sampled voltage level was arbitrarily chosen to be 1 V. The measured amplifier offsets and inferred leakage currents of each cell are listed in Table III. The measurements show that the average leakage of the cell is roughly a factor of three larger than the value seen in Fig. 7, and the change in leakage for a given change in offset voltage is between four and five times larger than that seen in Fig. 7, in agreement with the arguments presented in Section III-E. Fig. 14. Difference between input and stored voltage in differential cell. B. Differential Cell The measured dc input-output response of a representative V, V, and differential storage cell with an amplifier bias current of 1.5 nA is shown in Fig. 14, along with a first-order fit. This differential cell exhibits an inputoutput range spanning 0.2 to 3.1 V, a charge-injection offset of V. 75 V, and a maximum finite-gain error of The measured dc sweep parameters of all four differential cells are summarized in Table IV. Notice that the input-output range and finite-gain error are similar for the differential and singlesided cells, as expected, but the charge injection offset is consistently smaller in the differential cells. Since the absolute samsampling pling error is now less than 400 V, including noise, and the input-output range remains greater than 2.9 V, these cells achieve 12-bit accurate sampling. The leakage responses of the four differential cells are shown in Fig. 15, and a summary of the inferred leakage currents and measured amplifier offsets of each cell are listed in Table IV. Clearly the differential cell leakages still vary with amplifier offset, but the standard deviation of this dependence has been significantly reduced by the decreased junction area. The average leakage current of the differential cells is 9 aA per cell, O’HALLORAN AND SARPESHKAR: A 10-nW 12-bit ACCURATE ANALOG STORAGE CELL WITH 10-aA LEAKAGE 1995 TABLE IV MEASURED PERFORMANCE OF FABRICATED DIFFERENTIAL STORAGE CELLS Fig. 15. Differential cell leakage. with only one cell exceeding this level of leakage due to its large amplifier offset voltage. This was verified by externally remV, which reduced this cell’s ducing the amplifier offset to leakage to under 10 aA. C. Discussion To compare these results with past cells, the framework introduced in Section II is used. The nominal leakage of the singlesided cells should be less than the theoretical leakage given in Table I multiplied by the appropriate scaling factor for drain-tofA aA. The bulk voltage minimization, or measured leakage is less than 80 aA in the worst case, 45 aA in the average case, and is clearly under the estimated bound. These cells’ worst-case leakage is already comparable to the leakage of the leading implementation in the literature [5]. The differential cells implement drain-to-bulk voltage minimization as well as differential leakage cancellation, and their leakage should be bounded by fA aA. The achieved leakage is less than 15 aA in the worst case, 9 aA in the average case, again under the estimated upper bound. The average leakage is more than 8 lower than the leading implementation in the literature [5]. A die photo of the differential cell is shown in Fig. 16. Additionally, other cell parameters of interest are included in Table V. Fig. 16. white. Die photograph (2.2 mm 2 2.2 mm). The differential cell is circled in TABLE V ADDITIONAL CELL PARAMETERS In particular, the differential cells on average lose one bit of voltage accuracy, 700 V on a 12-bit scale and 11.3 mV on an 8-bit scale, in 3.3 and 54 min, respectively. The hold time of this cell at these voltage accuracies is 15 longer than the cell in [5], using only 92% as much area. Since junction leakage nominally 1996 IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 39, NO. 11, NOVEMBER 2004 decreases with the square of the feature size [13], the design is expected to achieve even lower leakage and smaller cell sizes in more modern processes. This claim is supported by preliminary measurements taken from several minimum-sized 0.5- m versions of the low-leakage switch, which exhibit over an order of magnitude less leakage than the 1.5- m versions presented here. These results agree with the expected leakage reduction due to feature-size scaling. Before concluding, a few final issues should be briefly addressed. The first is the storage cells’ slow 10-ms sampling time. In order to demonstrate that the proposed leakage reduction technique can be used in low-power design, the cells were biased at low current levels, making them slow. However, the usual power versus speed tradeoff can be applied to these cells without affecting their leakage performance. A second issue worth addressing is the temperature dependence of cell leakage. Measurements revealed that the residual leakage is thermally activated, increasing by roughly a factor of 3 per 10 C. Since this behavior is similar to that of a standard diode, the data is not shown for brevity. The final issue that should be mentioned is that capacitor soakage—the voltage-dependent polarization of a capacitor’s dielectric—fundamentally limits the performance of any sample-and-hold circuit [19]. This effect is typically ignored in analog storage cells because, due to the applications in which they are used, the stored voltages usually remain within a small neighborhood of the full input-output voltage range of the cell. However, for true rail-to-rail sample-and-hold operation, capacitor soakage will limit the sampling accuracy and initial leakage of the cell, and must be considered. V. CONCLUSION Using an on-chip current integration technique to characterize the dominant sources of MOS drain leakage, an accumulation-mode leakage mechanism that can dominate junction diode leakage has been observed. A simple rule-of-thumb was offered for avoiding this leakage mechanism, leading to a novel ultra-low leakage switch topology. A differential storage cell incorporating this new switch achieves an average leakage of 10 aA at room temperature, an 8 reduction over past designs. The cell loses one bit of voltage accuracy, 700 V on a 12-bit scale and 11.3 mV on an 8-bit scale, in 3.3 and 54 min, respectively. This represents a 15 increase in hold time at these voltage accuracies over past designs. Since the leakage is independent of amplifier speed, the cell can operate on as little as 10 nW of power. REFERENCES [1] C. C. Enz and G. C. Temes, “Circuit techniques for reducing the effects of op-amp imperfections: Autozeroing, correlated double sampling, and chopper stabilization,” Proc. IEEE, vol. 84, pp. 1584–1614, Nov. 1996. [2] Y. Horio and S. Nakamura, “Analog memories for VLSI neurocomputing,” in Artificial Neural Networks: Paradigms, Applications, and Hardware Implementations, E. Sánchez-Sinencio and C. Lau, Eds. New York: IEEE Press, 1992, pp. 344–363. [3] E. Vittoz, H. Oguey, M. A. Maher, O. Nys, E. Dijkstra, and M. Chevroulet, “Analog storage of adjustable synaptic weights,” in VLSI Design of Neural Networks, U. Ramacher and U. Rückert, Eds. Norwell, MA: Kluwer, 1991, pp. 47–63. [4] K. Stafford, P. Gray, and R. Blanchard, “A complete monolithic sample/hold amplifier,” IEEE J. Solid-State Circuits, vol. SC-9, pp. 381–387, Dec. 1974. [5] M. Ehlert and H. Klar, “A 12-bit medium-time analog storage device in a CMOS standard process,” IEEE J. Solid-State Circuits, vol. 33, pp. 1139–1143, July 1998. [6] G. Cauwenberghs, “Analog VLSI long-term dynamic storage,” in Proc. IEEE Int. Symp. Circuits and Systems, vol. 3, 1996, pp. 334–337. [7] P. Heim and E. Vittoz, “Precise analog synapse for Kohonen feature maps,” IEEE J. Solid State Circuits, vol. 29, pp. 982–985, Aug. 1994. [8] M. Nahebi and B. Wooley, “A 10-bit video BiCMOS track-and-hold amplifier,” IEEE J. Solid-State Circuits, vol. 24, pp. 1507–1516, Dec. 1989. [9] B. Hochet, “Multivalued MOS memory for variable-synapse neural networks,” Electron. Lett., vol. 25, no. 10, pp. 669–670, May 1989. [10] R. Nelson and C. Harbourt, “Long term storage circuit using a sampling technique,” IEEE J. Solid-State Circuits, vol. SC-2, pp. 49–53, June 1967. [11] A. Kramer, V. Hu, C. K. Sin, B. Gupta, R. Chu, and P. K. Ko, “EEPROM device as a reconfigurable analog element for neural networks,” in Tech. Dig. IEEE IEDM, 1989, pp. 259–262. [12] C. Diorio, S. Mahajan, P. Hasler, B. Minch, and C. Mead, “A high-resolution nonvolatile analog memory cell,” in IEEE Int. Symp. Circuits and Systems, vol. 3, May 1995, pp. 2233–2236. [13] V. Linares-Barranco and T. Serrano-Gotarredona, “On the design and characterization of femtoampere current-mode circuits,” IEEE J. SolidState Circuits, vol. 38, pp. 1353–1363, Aug. 2003. [14] Y. P. Tsividis, Operation and Modeling of the MOS Transistor, 2nd ed. New York: McGraw-Hill, 1999. [15] D. A. Johns and K. Martin, Analog Integrated Circuit Design. New York: Wiley, 1997, pp. 346–347. [16] J. McCreary and P. R. Gray, “All-MOS charge redistribution analog-todigital conversion techniques–Part I,” IEEE J. Solid-State Circuits, vol. SC-10, pp. 371–379, Dec. 1975. [17] R. Sarpeshkar, R. F. Lyon, and C. A. Mead, “A low-power wide-linear-range transconductance amplifier,” Analog Integrated Circuits and Signal Processing, vol. 13, no. 1/2, May/June 1997. [18] J. J. Sit and R. Sarpeshkar, “A micropower logarithmic A/D with offset and temperature compensation,” IEEE J. Solid-State Circuits, vol. 39, pp. 308–319, Feb. 2004. [19] R. A. Pease, “Understand capacitor soakage to optimize analog systems,” in Proc. EDN, Oct. 13, 1990, pp. 125–129. Micah O’Halloran received B.S. degrees in electrical and computer engineering from the University of Florida, Gainesville, in 2000, and the M.S. degree in electrical engineering from the Massachusetts Institute of Technology (MIT), Cambridge, in 2002. He is now a Ph.D. candidate working with the Analog VLSI and Biological Systems Group at MIT. His research interests include low-power analog and mixed-signal design, adaptive systems, and control theory. Rahul Sarpeshkar (M’97) received Bachelor’s degrees in electrical engineering and physics from the Massachusetts Institute of Technology (MIT). After completing his Ph.D. at Caltech, he joined Bell Labs as a member of the technical staff. Since 1999, he has been on the faculty of MIT’s Electrical Engineering and Computer Science Department where he heads a research group on Analog VLSI and Biological Systems, and is currently the Robert J. Shillman Associate Professor. He holds over a dozen patents and has authored several publications, including one that was featured on the cover of Nature. His research interests include analog and mixed-signal VLSI, ultra-low-power circuits and systems, biologically inspired circuits and systems, and control theory. Dr. Sarpeshkar has received several awards, including the Packard Fellow Award given to outstanding young faculty, the ONR Young Investigator Award, and the NSF Career Award. He was recently awarded the Junior Bose Award for Excellence in Teaching at MIT.