Circuit and method for reading a memory cell that can store multiple bits of data

CROSS-REFERENCE TO RELATED APPLICATION

The following pending U.S. Patent Application by Cristiano Calligaro, Vincenzo Daniele, Roberto Gastaldi, Alessandro Manstretta, Nicola Telecco and Guido Torelli entitled: “Serial Dichotomic Method For Sensing Multiple-Level Non-Volatile Memory Cells, And Sensing Circuit Implementing Such Method,” Ser. No. 08/593,650 (Attorney's Docket No. 853063.420), which has the same effective filing date and ownership as the present application, and to that extent is related to the present application, and which is incorporated herein by reference.

TECHNICAL FIELD

The present invention relates to a sensing circuit for serial dichotomic sensing of multiple-levels non-volatile memory cells.

BACKGROUND OF THE INVENTION

The market demand for non-volatile memories with higher and higher storage capacity is forcing the semiconductor manufacturers to a continuous effort in scaling the devices and in increasing the chip size.

As an additional possibility to increase the memories' capacities, it has been proposed to store more than one bit per memory cell: a memory device with memory cells capable of storing two or even four bits has a storage capacity two or, respectively, four times higher than that of a memory device with the same chip size but with memory cells capable of storing only one bit each.

Non-volatile memory cells (i.e., memory cells which retain their programming state even in the absence of power) are generally represented by MOS field-effect transistors; data can be programmed in non-volatile memory cells by changing the threshold voltage of the MOS field-effect transistors: in the case of ROMs this is done during their fabrication, while in the case of EPROMs, EEPROMs and Flash EEPROMs, the change in the threshold voltage is achieved by properly biasing the MOS field effect transistors to cause an injection of charges in a floating gate.

To determine the programming state of a non-volatile memory cell, i.e., to “read” or to “sense” the contents of the memory cell, a fixed voltage VG is applied to the control gate of the MOS transistor: the programming state of the memory cell can thus be determined by detecting the position of the threshold voltage of the MOS transistor with respect to said fixed gate voltage.

In the most common case of non-volatile memory cells that are capable of storing only one bit of information, a memory cell can show two different programming states (logic levels), corresponding to two different threshold voltage values; hereinafter, such a cell will be called a “two-level memory cell.” The reading of the memory cells is performed by a so-called “sensing circuit,” which delivers a voltage signal having two distinct possible values, corresponding to the two logic levels.

In the case of non-volatile memory cells that are capable of storing more than one bit of information, a memory cell must be able to show m=2ⁿdistinct programming states or levels, where n represents the number of bits which can be stored in the memory cell; in the following, such a cell will be called a “multiple-level memory cell.” As in the case of two-level cells, each level corresponds to a different value for the threshold voltage of the MOS transistor.

The discrimination of the m different programming levels can be performed by means of a either a voltage-mode sensing technique or a current-mode sensing technique. In the latter case, for example, the allowed threshold voltage range of the memory cell is divided, on the basis of the electric and physical characteristics of the memory cells, into m sub-intervals, each corresponding to one of the different m levels to be discriminated. The memory cell is then programmed in a desired one of the m different levels by properly adjusting its threshold voltage, so that when the memory cell is biased in the prescribed sensing conditions, it sinks a current corresponding to the desired programming level.

Two sensing techniques have been proposed for multiple-level memory cells: parallel-mode sensing and serial-mode sensing.

Parallel-mode sensing is for example described in A. Bleiker, H. Melchior, “A Four-State EEPROM Using Floating-Gate Memory Cells,” IEEE Journal of Solid State Circuits, vol. SC-22, No. 3, July 1987, pp. 460-463. This technique is the natural extension of the conventional technique used for two-level memory cells, and provides for generating m−1 distinct predetermined references (current references for the current-mode approach, or voltage references for the voltage-mode approach), and for performing m−1 simultaneous comparisons of such m−1 distinct voltage or current references with a current (or a voltage) derived from the memory cell to be read.

The advantages of the parallel-mode sensing technique are its high speed and the independence of the sensing time from the programming state of the memory cell; a disadvantage is the large area required by the sensing circuit, since m−1 distinct comparison circuits are necessary to perform the m−1 simultaneous comparisons.

Differently from parallel-mode sensing serial-mode sensing requires just one reference (current or voltage), which can be varied according to a prescribed law. This single reference is used to perform a series of successive comparisons, and is varied to approximate the analog current or voltage derived from the memory cell to be read. A serial-mode sensing circuit is simple to implement, and requires only a small area.

Two different kinds of serial-mode sensing methodologies are known, which differ in the law according to which the reference is made to vary.

The first methodology, also called “sequential,” described for example in M. Horiguchi et at., “An Experimental Large-Capacity Semiconductor File Memory Using 16-Levels/Cell Storage,” IEEE Journal of Solid State Circuits, vol. SC-23, No. 1, February 1988, pp. 27-32, consists of a succession of comparisons (at most m−1 ) between a fixed quantity (voltage or current) and a variable quantity (voltage or current) which is sequentially varied starting from an initial value.

For example, the fixed quantity can be the current sunk by the memory cell to be read (subject to a prescribed biasing condition), while the variable quantity can be a current supplied by a digitally-driven generator. The (constant) current sunk by the memory cell to be read is compared with a reference current which takes successively increasing (or decreasing) discrete values starting from a minimum (or maximum) value; said discrete values are ideally chosen in such a way as to fall between the different current values corresponding to the m programming levels of the memory cell, so that the result of a comparison is negative (or positive) as long as the reference current is lower (or higher) than the cell's current. The series of successive comparisons stops after the first positive (or negative) result; the last value of the reference current represents the current of the memory cell, except for a constant term associated with the position of the reference current value with respect to the programming levels of the memory cell.

It appears that the time required to read a memory cell with the serial sequential method is not uniform, but depends on the particular programming level of the memory cell and on the starting value of the reference voltage or current (the sensing time depends on the distance between the programming level of the cell to be read and the starting value of the reference voltage or current): from a minimum of one to a maximum of m−1 comparison steps are necessary to determine the programming state of an m-level memory cell. The sensing time soon becomes excessive with an increase in the number of bits stored in a single memory cell.

The second serial-mode sensing methodology, also called “dichotomic,” is described in the co-pending European Patent Application No. 95830023.8 filed on January 27, 1995 in the name of the same applicant. This methodology consists of a successive approximations search that, starting from an initial value for the reference current, finds the value of the memory cell current after a succession of iterations. At each step of the iterative search, the (constant) memory cell current is compared with the variable reference current, whose value is chosen according to a dichotomic or “binary search” algorithm. The initial interval of possible memory cell current values is divided in two parts: depending on the result of the comparison, the successive dichotomy will be applied to only that part of the initial interval wherein the memory cell current falls; the iterative search is recursively repeated until the value of the memory cell current is determined.

Using the serial dichotomic method, the programming state of a memory cell with m=2ⁿdifferent programming levels is determined in n comparison steps, independently from the particular programing state of the memory cell.

In EPA 95830023.8, a sensing circuit suitable for actuating the serial dichotomic method is also described. The sensing circuit comprises a variable reference current generator controlled by a successive approximation register supplied with an output signal of a current comparator; the successive approximation register comprises a sequential network that, starting from a predetermined initial state, evolves through a succession of states, each one corresponding to one step of the serial dichotomic search.

The circuit implementation of the sensing circuit strongly depends on the structure of the variable reference current generator: this in fact affects the structure of the current comparator and of the successive approximation register.

According to the implementation described in EPA 905830023.8, the variable reference current generator comprises m−1 distinct current generators which are activated in a mutually exclusive way; each one of the m−1 current generator corresponds to one of m−1 values which can be taken by the reference current (absolute current generators technique).

In practice, each current generator is formed by a reference memory cell identical to the memory cell to be read. The reference memory cells are programmed in m−1 distinct states which, however, do not coincide with any of the m programming levels of the memory cells to be read, since the reference current values shall fall between the cell current values; in this case, the current comparator can be balanced (i.e., the currents to be compared are supplied to the inputs of the comparator in a 1:1 ratio).

Due to the fact that the reference memory cells are not programmed at the same levels as the memory cells to be read, there is not a perfect equivalence between the circuit branch containing the reference memory cells and the circuit branch containing the memory cell to be read. This means that there is a poor tracking between these two circuit branches with respect to process of biasing condition variations.

As an alternative, if the current comparator is properly unbalanced (i.e., if the currents to be compared are supplied to the inverting and non-inverting inputs of the comparator in a ratio equal to f different from 1, e.g., f=0.7) the reference memory cells can be programmed in m−1 distinct programming levels which coincide with the programming levels of the memory cell to be read.

Thanks to this, there is a substantially perfect equivalence between the circuit branch containing the memory cell to be read and the circuit branch containing the reference memory cell; also, every possible variation in process parameters or biasing conditions between the memory cell to be read and the reference memory cell is treated by the comparator as a common-mode contribution.

The drawbacks of having an unbalanced current comparator are that the values of the reference current do not fall exactly in the middle of each interval between successive memory cell current values; in fact, the ratio f causes a linear reduction of the current values. Moreover, in an unbalanced comparator, it is more critical to control the switching characteristics than it is in a balanced comparator.

SUMMARY OF THE INVENTION

In view of the state of the prior art described, it is an object of the present invention to provide a sensing circuit for the serial dichotomic sensing of multiple-level non-volatile memory cells that overcome the above-mentioned drawbacks.

According to the present invention, said object is achieved by means of a sensing circuit for the serial dichotomic sensing of multiple-level memory cells that can take one programming level among a plurality of m=2ⁿ(n>=2) different programming levels, comprising biasing means for biasing a memory cell to be sensed in a predetermined condition, so that the memory cell sinks a cell current with a value belonging to a plurality of m distinct cell current values, each cell current value corresponding to one of said programming levels, a current comparator for comparing the cell current with a reference current generated by a variable reference current generator, and a successive approximation register supplied with an output signal of the current comparator and controlling the variable reference current generator, characterized in that the variable reference current generator comprises an offset current generator permanently coupled to the current comparator, and m−2 distinct current generators, independently activatable by the successive approximation register, each one generating a current equal to a respective one of said plurality of cell current values.

In a sensing circuit according to the present invention, the reference current is given by the sum of an offset current plus a current equal to one of the possible values of the memory cell current. By properly adjusting the value of the offset current, it is possible to have reference current values which are exactly central between adjacent memory cell current values. This allows the use of a balanced current comparator, which is advantageous with respect to an unbalanced one.

BRIEF DESCRIPTION OF THE DRAWINGS

These and other features and advantages of the present invention will be made more evident by the following detailed description of some particular embodiments, described as non-limiting examples with reference to the annexed drawings, wherein:

FIG. 1 schematically shows a multiple-level non-volatile memory cell under sensing conditions, and a reference current generator used to sense the memory cell according to a serial dichotomic sensing method;

FIG. 2 diagrammatically shows the distribution of currents sunk by a four-level memory cell in its four different programming conditions, and the distribution of reference currents used to sense the memory cell according to the serial dichotomic sensing method;

FIGS. 3 and 4 diagrammatically show the steps of the serial dichotomic sensing method for two different programming conditions of the memory cell;

FIG. 5 schematically shows a sensing circuit for sensing multiple-level non-volatile memory cells according to the present invention;

FIG. 6 schematically shows a variable reference current generator for the sensing circuit of FIG. 5;

FIG. 7 is a circuit diagram of a Successive Approximation Register (SAR) of the circuit of FIG. 5, suitable for sensing four-level memory cells;

FIG. 8 is a circuit diagram of a current comparator of the sensing circuit of FIG. 5;

FIG. 9 is a truth table of the SAR of FIG. 7;

FIG. 10 is a state-transition diagram of the SAR of FIG. 7; and

FIG. 11 is a time diagram of some signals of the SAR of FIG. <b>5</b>.

DETAILED DESCRIPTION OF THE INVENTION

First of all, the serial dichotomic sensing method will be described in the particular case of a four-level memory cell (a cell capable of storing two bits of information). A memory cell MC to be read is biased with a fixed, prescribed control gate voltage VG (FIG. <b>1</b>). The memory cell MC shown in FIG. 1 is a floating-gate MOS field effect transistor, such as an EPROM, EEPROM or Flash EEPROM memory cell. Nevertheless, the memory cell could be a simple MOSFET with threshold voltage adjusted during fabrication, as in the case of a ROM memory cell. When the memory cell MC is biased with said fixed control gate voltage VG, it sinks a current IC whose value depends on the particular programming state of the memory cell itself, i.e., on the memory cell's threshold voltage. In FIG. 1, a current generator G is also shown supplying a reference current IR; IR is not constant, but can take values belonging to a discrete set, as will now be explained.

In FIG. 2, four distinct values IC<b>0</b>-IC<b>3</b> for the current IC are shown: each value corresponds to a respective one of the four different programming states of the memory cell MC (practical values can be, for example, IC<b>0</b>=0, IC<b>1</b>=30 uA IC<b>2</b>=60 uA and IC<b>3</b>-90 uA). FIG. 2 also shows, on the branches of a decision tree, the different values I<b>0</b>-I<b>2</b> that can be taken by the reference current IR: by choosing the reference current values in such a way that they are central between successive values of IC, only three different reference current values I<b>0</b>-I<b>2</b> are necessary (for example, I<b>0</b>=15 uA, I<b>1</b>=45 uA and I<b>2</b>=75 uA).

Let's assume that the programming state of the memory cell MC corresponds to a current IC=IC<b>2</b> (FIG. <b>3</b>). The first step S<b>1</b> of the sensing method provides for a comparison of the memory cell current IC with a reference current IR=I<b>1</b>, which is the central value in the set of values IC<b>0</b>-IC<b>3</b>; this is the best choice from the point of view of the efficiency of the method. The comparison tells that the cell current IC is higher than I<b>1</b>: a priori, it could be equal to IC<b>2</b> or IC<b>3</b>. In the second step S<b>2</b> the current IC is compared with a reference current IR=I<b>2</b>, which is the central value between IC<b>2</b> and IC<b>3</b>, and it is found that IC is lower than I<b>2</b>: necessarily, IC must be equal to IC<b>2</b>. The programming condition of the memory cell MC has thus been determined in only two steps.

Let's now assume, as a second example, that the programming state of the memory cell MC corresponds to a current IC=IC<b>0</b> (FIG. <b>4</b>). In the first step S<b>1</b> the cell current IC is again compared with the reference current IR=I<b>1</b>, to find that IC is lower than I<b>1</b>: this means that IC could be either IC<b>0</b> or IC<b>1</b>. In the second step S<b>2</b> the current IC is compared with a reference current IR=I<b>0</b>, which is the central value between IC<b>0</b> and IC<b>1</b>: since the comparison tells that IC is lower than I<b>0</b>, IC must necessarily be equal to IC<b>0</b>. Again, the programming condition of the memory cell MC has been determined in two steps.

The number of steps required to determine the programming condition of the memory cell MC is uniform, i.e., it does not depend on the programming condition itself, and it is always equal to two in the example illustrated in FIGS. 2-4. It is straightforward to show that the programming condition of a sixteen-level memory cell is determined in four steps. In general, the serial dichotomic sensing method allows one to determine the programming condition of an m-level memory cell (with m=2ⁿ) in n steps, independently from the programming condition of the memory cell.

FIG. 5 schematically shows a sensing circuit according to the present invention. The circuit substantially comprises a digitally-driven variable reference current generator G for generating a variable reference current IR, a current comparator <b>1</b> for comparing the reference current IR with a current IC sunk by a memory cell MC to be read, and a Successive Approximation Register (“SAR”) <b>2</b>.

The current comparator <b>1</b> has an inverting input connected to the drain electrode of a memory cell MC to be sensed, and a non-inverting input connected to the variable reference current generator G; the comparator <b>1</b> has an output signal CMP which is supplied to the SAR <b>2</b>. The SAR <b>2</b> is further supplied with a preset signal PR and with a clock signal CK (timing signal), and supplies a group CNT of control signals (in digital format) to the variable reference current generator G; the SAR <b>2</b> also generates a group OUT of output signals carrying in digital format the programming state of the sensed memory cell MC.

FIG. 6 schematically shows the variable reference current generator G in the particular case of a sensing circuit for four-level memory cells MC (m=2ⁿwith n=2 and m=4). The reference current generator G comprises three distinct current generators IR<b>0</b>, IR<b>1</b>, and Ioff. Ioff is an offset current generator, generating a constant current, and is permanently connected to the non-inverting input of the current comparator <b>1</b>. IR<b>0</b> and IR<b>1</b> are instead connectable to the non-inverting input of the current comparator <b>1</b> by means of respective switches SW<b>0</b> and SW<b>1</b>, activated by respective control signals Q<b>0</b> and Q<b>1</b> of the group CNT.

The value of Ioff is equal to the reference current I<b>0</b> shown in FIG. 2, i.e., it is equal to (IC<b>1</b>+IC<b>0</b>)/2, here IC<b>1</b>/2, because in the case IC<b>0</b>, the minimum memory cell current, is zero. The values of IR<b>0</b> and IR<b>1</b> are respectively equal to IC<b>1</b> and IC<b>2</b>, i.e., to the currents sunk by a memory cell MC in two particular programming states.

In practice, the current generators IR<b>0</b> and IR<b>1</b> are implemented by means of two reference memory cells, identical to the memory cell MC to be read, programmed in two of the four possible programming states of the memory cells MC, more precisely in the states corresponding to the memory cell current values IC<b>1</b> and IC<b>2</b>, respectively.

With these choices for the values of loft, IR<b>0</b> and IR<b>1</b>, the variable reference current IR can take the following values:

SW<b>0</b> and SW<b>1</b> open: IR=Ioff=(½)*IC<b>1</b>=I<b>0</b>;

SW<b>0</b> closed, SW<b>1</b> open: IR=Ioff+IR<b>0</b>=({fraction (3/2)})*IC<b>1</b>=I<b>1</b>;

SW<b>0</b> open, SW<b>1</b> dosed: IR=Ioff+IR<b>1</b>=IC<b>2</b>+(½)*IC<b>1</b>=I<b>2</b>.

These values are central with respect to the memory cell current values IC<b>0</b>-IC<b>3</b>, and are obtained by an additive process, adding a constant offset to the values IC<b>1</b> and IC<b>2</b>. The reference current values can be adjusted by simply varying the value of the offset current generator Ioff.

Thanks to this, the current comparator <b>1</b> can be balanced (i.e., the currents to be compared are supplied to the inputs of the comparator in a 1:1 ratio). A balanced comparator is better than an unbalanced one from the point of view of the circuit and layout symmetry. Furthermore, a balanced comparator has a better common-mode rejection ratio, a better controlled switching characteristic, and lower mismatch errors.

The characteristics of the current comparator <b>1</b> depend on the number of programming levels to be discriminated. In particular, the input sensitivity (i.e., the minimum current difference which can be detected by the comparator) must be lower than the difference between the currents of two adjacent programming states, taking into account the spreading &amp;values due to process tolerances.

A suitable current comparator structure is shown in FIG. <b>8</b>. The circuit comprises two load MOSFETs DL and DR (P-channel type) performing a current/voltage conversion of the memory cell current IC and of the reference current IR, respectively. The drain voltages <b>4</b> and <b>5</b> of MOSFETs DL and DR control the gates of two cross-connected MOSFETs MS<b>1</b> and MR<b>1</b> (P-channel type) forming a latch. The source electrodes of MS<b>1</b> and MR<b>1</b> can be connected to a power supply line VDD through two respective P-channel MOSFETs T<b>3</b> and T<b>4</b> which are commonly driven by a signal CKS derived from the clock signal CK (CKS can for example be the logic complement of CK). The drain electrodes of MS<b>1</b> and MR<b>1</b> can be short-circuited to each other by the activation of an N-channel MOSFET TE driven by the signal CKS. The drain electrode of MR<b>1</b> forms the comparator output CMP. Experimental tests have shown that this circuit is quite fast even if a power supply VDD of 3 V is used, and has an input sensitivity of about 10 uA. These characteristics make the shown comparator structure particularly suitable for the sensing of four-levels memory cells.

The SAR <b>2</b> comprises a sequential network (or state-machine) implementing the successive approximation search algorithm. The state of the sequential network at a given step of the successive approximation search depends on its state at the present step, and on the result of the comparison between the cell current IC and the reference current IR at the preceding step. The sequential network activates in the correct sequence the switches SW<b>0</b>, SW<b>1</b>, depending on the results of the comparisons.

FIG. 7 is a circuit diagram of the sequential network of the SAR <b>2</b> in the case of a sensing circuit for four-level memory cells. The sequential network comprises two Delay-type (“D-type”) flip-flops FF<b>0</b>, FF<b>1</b>. Each flip-flop FF<b>0</b>, FF<b>1</b> has a clock input CK and a preset input PR; the clock inputs CK and the preset inputs PR of the flip-flops FF<b>0</b>, FF<b>1</b> are commonly connected to the clock signal CK and to the preset signal PR, respectively; more precisely, FFO receives the logical complement of PR (as indicated by the inverting dot at the input PR of FFO). Each flip-flop has a data input D<b>0</b>, D<b>1</b>, a “true” data output Q<b>0</b>, Q<b>1</b>, and a “complemented” data output Q<b>0</b>N, Q<b>1</b>N which is the logic complement of Q<b>0</b>, Q<b>1</b>; as known to anyone skilled in the art, in a D-type flip-flop the true data output after a clock pulse takes the logic value of the data input during said clock pulse. The data input DO of the first flip-flop FF<b>0</b> is supplied with the complemented data output Q<b>0</b>N of the first flip-flop FF<b>0</b>. The data input D<b>1</b> of the second flip-flop FF<b>1</b> is supplied with an output of a NOR gate <b>5</b> whose inputs are represented by the signal CMP and by the complemented data output Q<b>0</b>N of the first flip-flop FF<b>0</b>.

The true data outputs Q<b>0</b>, Q<b>1</b> of the flip-flops FF<b>0</b>, FF<b>1</b> form the group of digital control signals CNT for the variable reference current generator G in FIG. <b>6</b>. The switches SW<b>0</b>, SW<b>1</b> close when the respective control signal Q<b>0</b>, Q<b>1</b> is a logic “1”, otherwise the switches SW<b>0</b>, SW<b>1</b> are open.

The signal CMP, complemented by an inverter <b>3</b>, forms a least significant bit OUT<b>0</b> of a two-bits output code OUT<b>0</b>, OUT<b>1</b>; the true data output Q<b>1</b> of flip-flop FF<b>1</b> forms a most significant bit OUT<b>1</b> of the two-bits output code OUT<b>0</b>, OUT<b>1</b>. OUT<b>0</b> and OUT<b>1</b> represent the group of signals OUT shown in FIG. <b>5</b>.

The preset signal PR is used at the circuit power-up to assure that the starting condition of the flip-flops FF<b>0</b>, FF<b>1</b> is that corresponding to Q<b>0</b>=1 and Q<b>1</b>=0. This condition corresponds to the switch SW<b>0</b> being closed, i.e., to a reference current value:

<maths><formula-text>IR=Ioff+IR<b>0</b>=Ioff+IC<b>1</b>=({fraction (3/2)})*IC<b>1</b>=I<b>1</b></formula-text></maths>

which is the central value between IC<b>0</b> and IC<b>3</b>.

The operation of the sensing circuit will be now described with reference to the truth-table of FIG. 9, to the state-transition diagram of FIG. <b>10</b> and to the time diagram of FIG. <b>11</b>. As previously described in connection with FIGS. 2 to <b>4</b>, the sensing of a two-levels memory cell MC is carded out in two steps. At the beginning of the first step (t₀in FIG. 11) Q<b>0</b>=1 and Q<b>1</b>=0, SW<b>0</b> is closed and SW<b>1</b> is open, so that IR=I<b>1</b>; on the rising edge of the clock signal CK the comparator <b>1</b> compares the cell current IC with the reference current IR=I<b>1</b>: if IC is higher than IR, CMP=0, while if IC is lower than IR, CMP=1. On the falling edge of the clock signal CK, the logic state of Q<b>0</b>, Q<b>1</b> changes to Q<b>0</b>=0 and Q<b>1</b>=1 if CMP=0, or to Q<b>0</b>=Q<b>1</b>=0 if CMP=1 (see FIGS. <b>9</b> and <b>10</b>): in the first case, SW<b>0</b> opens and SW<b>1</b> closes, so that:

<maths><formula-text>IR=Ioff+IR<b>1</b>=(½)*IC<b>1</b>+IC<b>2</b>=I<b>2</b></formula-text></maths>

which is the central value between IC<b>2</b> and IC<b>3</b>, while in the second case both SW<b>0</b> and SW<b>1</b> opens, so that:

<maths><formula-text>IR=Ioff=(½)*IC<b>1</b>=I<b>0</b></formula-text></maths>

which is the central value between IC<b>1</b> and IC<b>0</b>.

On the next rising edge of the clock signal CK, IC is compared to the new value of IR: if IC is higher than IR, CMP=0, while if IC is lower than IR CMP=1. On the basis of the logic state of CMP and Q<b>1</b>, it is possible to determine the programming state of the memory cell MC; the valid output data OUT<b>0</b>, OUT<b>1</b> are available at t₀+({fraction (3/2)})T (where T is the period of the clock signal CK), i.e., before the end of the second clock pulse. On the next falling edge of the clock signal CK, the flip-flops FF<b>0</b>, FF<b>1</b> are automatically preset to the state Q<b>0</b>=1, Q<b>1</b>=0 (self-preset), and the circuit is ready to perform a new sensing.

The SAR <b>2</b> of the sensing circuit of the present invention is much simpler than that described in the already-cited European Patent Application No. 95830023.8: since in the variable reference current generator G one of the three current generators (Ioff) is permanently connected to the current comparator, only two control signals Q<b>0</b> and Q<b>1</b> (in the case of a four-level memory cell) are necessary to control the reference current generator G (in general, for an m-level memory cell, m−2 control signals are necessary). Furthermore, in the case of four-level memory cells, the output digital code OUT is obtained directly from the sequential network that implements the dichotomic search algorithm, no combinatorial network being necessary.

The previous description has been made taking as an example the case of a sensing circuit for four-level memory cells. Extending the case of a four-level memory cell to that of an m-level memory cell, the variable reference current generator G should comprise an offset current generator Ioff with value equal to (IC<b>0</b>+IC<b>1</b>)/2 (IC<b>0</b> and IC<b>1</b> being the two lowest currents of an m-level memory cell), and m−2 distinct current generators with values equal to IC<b>1</b>, IC<b>2</b>, . . . , ICm−2.

From the foregoing it will be appreciated that, although specific embodiments of the invention have been described herein for purposes of illustration, various modifications may be made without deviating from the spirit and scope of the invention. Accordingly, the invention is not limited except as by the appended claims.