Flash eeprom memory systems and methods of using them

The invention relates to a method of altering a memory state of an addressed cell of an array according to the preamble of claim 1 and to an array of electrically alterable memory cells according to the preamble of claim 15.

A method of altering according to the preamble of claim 1 and a four-state EEPROM using floating gate memory cells according to the preamble of claim 15 are already known from "IEEE Journal of Solid-State Circuits", Vol. SC-22, No. 3, June 1987, pages 460-464. One memory state is assigned to a positive floating gate voltage and three further states are assigned to negative floating gate voltages. There is no disclosure of a net positive charge on the floating gate being responsive for programmable threshold states of the cell.

In connection with EEPROMs (US-A-4 357 685) it is already known to program the cells by applying charging pulses increasing in height to the floating gate of the cell transistor, and to erase analogously by applying discharging pulses increasing in height. During periods between the single pulses, the cell currents are read to check whether the desired state has been reached. No further condition is disclosed in the prior art.

US-A-4 087 759 discloses a two-state EEPROM having floating gate transistors of the split channel transistor type. The memory states are realized either by positive or by negative charges.

The threshold voltage characteristics of a transistor is dependent from the amount of charge that is retained on the floating gate. That is, the minimum amount of voltage (threshold) that must be applied to the control gate before the transistor is turned "on" to permit conduction between its source and drain regions is determined by the level of charge on the floating gate. A transistor is programmed to one of two or more states by accelerating electrons from the substrate channel region, through a thin gate dielectric and onto the floating gate.

The memory cell transistor's state is read by placing an operating voltage across its source and drain and a selected voltage on its control gate, and then detecting the level of current flowing between the source and drain which determines the programmed state of the transistor.

One example of such a memory cell is a triple polysilicon, split channel electrically erasable and programmable read only memory (EEprom). It is termed a "split channel" device since the floating and control gates extend over adjacent portions of the channel. This results in a transistor structure that operates as two transistors in series, one having a varying threshold in response to the charge level on the floating gate, and another that is unaffected by the floating gate charge but rather which operates in response to the voltage on the control gate as in any normal field effect transistor.

Such a memory cell is termed a "triple polysilicon" cell because it contains three conductive layers of polysilicon materials. In addition to the floating and control gates, an erase gate is included. The erase gate passes through each memory cell transistor closely adjacent to a surface of the floating gate but insulated therefrom by a thin tunnel dielectric. Charge is then removed from the floating gate of a cell to the erase gate, when appropriate voltage are applied to all the transistor elements. An array of EEprom cells are generally referred to as a Flash EEprom array if an entire array of cells, or a significant group of cells, is erased simultanaeously (i.e. in a flash).

EEprom's have been found to have a limited effective life. The number of cycles of programming and erasing that such a device can endure before becoming degraded is finite. After a number of such cycles in excess of 10,000, depending upon its specific structure, its programmability can be reduced. Often, by the time the device has been put through such a cycle for over 100,000 times, it can no longer be programmed or erased properly.

This is believed to be the result of electrons being trapped in the dielectric each time charge is transferred to or away from the floating gate by programming or erasing, respectively.

It is a primary object of the present invention to provide an EEprom array with increased storage capacity and life.

It is a further object of the present invention to provide techniques for increasing the number of program/erase cycles that an EEprom can endure.

It is an advantage of the present invention to provide EEprom semiconductor chips that are useful for solid state memory to replace magnetic disk storage devices.

The features of the invention are specified in claims 1 and 15.

The present invention relates, briefly and generally, to an EEprom memory cell which is caused to store more than one bit of data by partitioning its programmed charge into three or more ranges. Each cell is then programmed into one of these ranges. If four ranges are used, two bits of data can be stored in a single cell. If eight ranges are designated, three bits can be stored, and so on.

An intelligent programming and sensing technique is provided which permits the practical implementation of such multiple state storage. Further, an intelligent erase algorithm is provided which results in a significant reduction in the electrical stress experienced by the erase tunnel dielectric and results in much higher endurance to program/erase cycling and a resulting increased life of the memory.

Additional objects, features and advantages of the present invention will be understood from the following description of its preferred embodiments, which description should be taken in conjunction with the accompanying drawings.

Brief Description of the Drawings

Figure 1 is a cross section of an example split channel Eprom or EEprom.

Figure 2a is a schematic representation of the composite transistor forming a split channel Eprom device.

Figure 2b shows the programming and erase characteristics of a split channel Flash EEprom device.

Figure 2c shows the four conduction states of a split channel Flash EEprom device in accordance with this invention.

Figure 2d shows the program/erase cycling endurance characteristics of prior art Flash EEprom devices.

Figure 2e shows a circuit schematic and programming/read voltage pulses required to implement multistate storage.

Figure 3 outlines the key steps in the new algorithm used to erase with a minimum stress.

Figure 4 shows the program/erase cycling endurance characteristics of the split channel Flash EEprom device using intelligent algorithms for multistate programming and for reduced stress during erasing.

Description of the Preferred Embodiments

Referring initially to Figure 1, the structure of a split-channel Eprom or EEprom cell is described that is suitable for use in the improved memory array and operation of the present invention. A semiconductor substrate 11 includes source region 13 and drain region 15, usually formed by ion implantation. Between the source and drain is a channel region 17. Over a portion of the channel region designated as L1 is a floating gate 19, separated from the substrate by a thin layer of gate oxide 21. Over a portion of the channel region designated as L2 is formed a control gate 23, separated from the substrate 11 by a thin gate oxide layer 25. The control gate 23 is also electrically isolated from the floating gate 19 by an oxide layer 27.

It is the amount of electrical charge on the floating gate 19 that is programmed in accordance with the state desired to be stored in the cell. If the charge level is above some set threshold, the cell is considered to be in one state. If below that threshold, it is designated to be in its other state. The desired charge level is programmed by applying an appropriate combination of voltages to the source, drain, substrate and control gate, for a designated period of time, in order to cause electrons to move from the substrate 11 to the floating gate 19.

The floating gate is confined to its one memory cell and is electrically isolated from all other parts of the structure. The control gate 23, on the other hand, extends across a large number of cells sharing a common word line. As described hereinafter, the split-channel has the effect of providing two field-effect-transistors in series, one with the floating gate 19 as its control gate and the other with the control gate 23 as its control gate.

The generic split-channel Eprom or EEprom structure of Figure 1 becomes a Flash EEprom device when an erase gate (not shown) is added. The erase gate is a separate electrode positioned near a portion of the floating gate 27 and separated from it by a tunnel dielectric. When the proper voltages are applied to the source, drain, substrate, control gate and erase gate, the amount of charge on the floating gate is reduced. A single erase gate extends to a large number of memory cells, if not the entire array, so that they may be erased all at once. In some prior art Flash EEprom cells the source or drain diffusions underneath the floating gate are used also as the erase electrode, while in other cells the erase electrode is implemented either in the same conductive layer as the control gate or in a separate conductive layer.

Multistate storage

The split channel Flash EEprom device can be viewed as a composite transistor consisting of two transistors T1 and T2 in series - figure 2a. Transistor T1 is a floating gate transistor of effective channel length L1 and having a variable threshold voltage V_T1. Transistor T2 has a fixed (enhancement) threshold voltage V_T2 and an effective channel length L2. The Eprom programming characteristics of the composite transistor are shown in curve (a) of figure 2b. The programmed threshold voltage V_tx is plotted as a function of the time t during which the programming conditions are applied. These programming conditions typically are V_CG = 12V, V_D = 9V, V_S = V_BB = OV. No programming can occur if either one of V_CG or V_D is at OV. A Virgin (unprogrammed, unerased) device has V_T1 = +1.5V and V_T2 = +1.0V. After programming for approximately 100 microseconds the device reaches a threshold voltage V_tx ≧ +6.0 volts. This represents the off ("0") state because the composite device does not conduct at V_CG = +5.0V. Prior art devices employ a so called "intelligent programming" algorithm whereby programming pulses are applied, each of typically 100 microseconds to 1 millisecond duration, followed by a sensing (read) operation. Pulses are applied until the device is sensed to be fully in the off state, and then one to three more programming pulses are applied to ensure solid programmability.

Prior art split channel Flash EEprom devices erase with a single pulse of sufficient voltage V_ERASE and sufficient duration to ensure that V_T1 is erased to a voltage below V_T2 (curve (b) in figure 2b). Although the floating gate transistor may continue to erase into depletion mode operation (line (c) in figure 2b), the presence of the series T2 transistor obscures this depletion threshold voltage. Therefore the erased on ("1") state is represented by the threshold voltage V_tx = V_T2 = +1.0V. The memory storage "window" is given by _ΔV = V_tx("0") - V_tx("1") = 6.0-1.0=5.0V. However, the true memory storage window should be represented by the full swing of V_tx for transistor T1. For example, if T1 is erased into depletion threshold voltage V_T1 = -3.0V, then the true window should be given by _ΔV = 6.0 - (-3.0) = 9.0V. None of the prior art Flash EEprom devices take advantage of the true memory window. In fact they ignore altogether the region of device operation (hatched region D in figure 2b) where V_T1 is more negative than V_T2.

This invention proposes for the first time a scheme to take advantage of the full memory window. This is done by using the wider memory window to store more than two binary states and therefore more than a single bit per cell. For example, it is possible to store 4, rather than 2 states per cell, with these states having the following threshold voltage: <ul id="ul0001" list-style="none"><li>State "3": - V_T1 = -3.0V, V_T2 = +1.0V    (highest conduction) = 1, 1.</li><li>State "2": - V_T1 = -0.5V, V_T2 = +1.0V    (intermediate conduction) = 1, 0.</li><li>State "1": - V_T1 = +2.0V, V_T2= +1.0V    (lower conduction) = 0, 1.</li><li>State "0": - V_T1 = +4.5V, V_T2 = +1.0V    (no conduction) = 0, 0.</li></ul> To sense any one of these four states, the control gate is raised to V_CG = +5.0V and the source-drain current I_DS is sensed through the composite device. Since V_T2 = +1.0V for all four threshold states transistor T2 behaves simply as a series resistor. The conduction current I_DS of the composite transistor for all 4 states is shown as a function of V_CG in figure 2c. A current sensing amplifier is capable of easily distinguishing between these four conduction states. The maximum number of states which is realistically feasible is influenced by the noise sensitivity of the sense amplifier as well as by any charge loss which can be expected over time at elevated temperatures. Eight distinct conduction states are necessary for 3 bit storage per cell, and 16 distinct conduction states are required for 4 bit storage per cell.

Multistate memory cells have previously been proposed in conjunction with ROM (Read Only Memory) devices and DRAM (Dynamic Random Access Memory). In ROM, each storage transistor can have one of several fixed conduction states by having different channel ion implant doses to establish more than two permanent threshold voltage states. Prior art multistate DRAM cells have also been proposed where each cell in the array is physically identical to all other cells. However, the charge stored at the capacitor of each cell may be quantized, resulting in several distinct read signal levels. An example of such prior art multistate DRAM storage is described in <u>IEEE</u><u>Journal</u><u>of</u><u>Solid-State</u><u>Circuits</u>, Feb. 1988, p. 27 in an article by M. Horiguchi et al. entitled "An Experimental Large-Capacity Semiconductor File Memory Using 16-Levels/Cell Storage". A second example of prior art multistate DRAM is provided in <u>IEEE</u><u>Custom</u><u>Integrated</u><u>Circuits</u><u>Conference</u>, May 1988, p. 4.4.1 in an article entitled "An Experimental 2-Bit/Cell Storage DRAM for Macrocell or Memory-on-Logic Applications" by T. Furuyama et al.

To take full advantage of multistate storage in Eproms it is necessary that the programming algorithm allow programming of the device into any one of several conduction states. First it is required that the device be erased to a voltage V_T1 more negative than the "3" state (-3.0V in this example). Then the device is programmed in a short programming pulse, typically one to ten microseconds in duration. Programming conditions are selected such that no single pulse can shift the device threshold by more than one half of the threshold voltage difference between two successive states. The device is then sensed by comparing its conduction current I_DS with that of a reference current source I_REF, i (i=0,1,2,3) corresponding to the desired conduction state (four distinct reference levels must be provided corresponding to the four states). Programming pulses are continued until the sensed current (solid lines in figure 2c) drops slightly below the reference current corresponding to the desired one of four states (dashed lines in figure 2c). To better illustrate this point, assume that each programming pulse raises V_tx linearly by 200 millivolts, and assume further that the device was first erased to V_T1 = -3.2V. Then the number of programming/sensing pulses required is: <ul id="ul0002" list-style="none"><li>For state "3"(V_T1 = -3.0V)    No. of pulses = (3.2-3.0)/.2 = 1</li><li>For state "2" (V_T1 = -0.5V)    No. of pulses = (3.2-0.5)/.2 = 14</li><li>For state "1" (V_T1 = +2.0V)    No. of pulses = (3.2-(-2.0))/.2 = 26</li><li>and for state "0" (V_T1 = +4.5V)    No. of pulses = (3.2-(-4.5))/.2 = 39.</li></ul> In actual fact shifts in V_tx are not linear in time, as shown in figure 2b (curve (a)), therefore requiring more pulses than indicated for states "1" and "0". If 2 microseconds is the programming pulse width and 0.1 microseconds is the time required for sensing, then the maximum time required to program the device into any of the 4 states is approximately 39x2 + 39x.1 = 81.9 microseconds. This is less than the time required by "intelligent programming algorithms" of prior art devices. In fact, with the new programming algorithm only carefully metered packets of electrons are injected during programming. A further benefit of this approach is that the sensing during reading is the same sensing as that during programming/sensing, and the same reference current sources are used in both programming and reading operations. That means that each and every memory cell in the array is read relative to the same reference level as used during program/sense. This provides excellent tracking even in very large memory arrays.

Large memory systems typically incorporate error detection and correction schemes which can tolerate a small number of hard failures i.e. bad Flash EEprom cells. For this reason the programming/sensing cycling algorithm can be automatically halted after a certain maximum number of programming cycles has been applied even if the cell being programmed has not reached the desired threshold voltage state, indicating a faulty memory cell.

There are several ways to implement the multistate storage concept in conjunction with an array of Flash EEprom transistors. An example of one such circuit is shown in figure 2e. In this circuit an array of memory cells has decoded word lines and decoded bit lines connected to the control gates and drains respectively of rows and columns of cells. Each bit line is normally precharged to a voltage of between 1.0 V and 2.0 V during the time between read, program or erase. For a four state storage, four sense amplifiers, each with its own distinct current reference levels IREF,0, IREF,1, IREF,2, and IREF,3 are attached to each decoded output of the bit line. During read, the current through the Flash EEprom transistor is compared simultaneously (i.e., in parallel) with these four reference levels (this operation can also be performed in four consecutive read cycles using a single sense amplifier with a different reference applied at each cycle, if the attendant additional time required for reading is not a concern). The data output is provided from the four sense amplifiers through four Di buffers (D0, D1, D2 and D3).

During programming, the four data inputs Ii (I0, I1, I2 and I3) are presented to a comparator circuit which also has presented to it the four sense amp outputs for the accessed cell. If Di match Ii, then the cell is in the correct state and no programming is required. If however all four Di do not match all four Ii, then the comparator output activates a programming control circuit. This circuit in turn controls the bit line (VPBL) and word line (VPWL) programming pulse generators. A single short programming pulse is applied to both the selected word line and the selected bit line. This is followed by a second read cycle to determine if a match between Di and Ii has been established. This sequence is repeated through multiple programming/reading pulses and is stopped only when a match is established (or earlier if no match has been established but after a preset maximum number of pulses has been reached).

The result of such multistate programming algorithim is that each cell is programmed into any one of the four conduction states in direct correlation with the reference conduction states I_REF, i. In fact, the same sense amplifiers used during programming/reading pulsing are also used during sensing (i.e., during normal reading). This allows excellent tracking between the reference levels (dashed lines in figure 2c) and the programmed conduction levels (solid lines in figure 2c), across large memory arrays and also for a very wide range of operating temperatures. Furthermore, because only a carefully metered number of electrons is introduced onto the floating gate during programming or removed during erasing, the device experiences the minimum amount of endurance-related stress possible.

In actual fact, although four reference levels and four sense amplifiers are used to program the cell into one of four distinct conduction states, only three sense amplifiers and three reference levels are required to sense the correct one of four stored states. For example, in figure 2c, I_REF("2") can differentiate correctly between conduction states "3" and "2", I_REF("1") can differentiate correctly between conduction states "2" and "1", and I_REF("0") can differentiate correctly between conduction states "1" and "0". In a practical implementation of the circuit of figure 2e the reference levels I_REF, i (i=0,1,2) may be somewhat shifted by a fixed amount during sensing to place them closer to the midpoint between the corresponding lower and higher conduction states of the cell being sensed.

Note that the same principle employed in the circuit of figure 2e can be used also with binary storage, or with storage of more than four states per cell. Of course, circuits other than the one shown in figure 2e are also possible. For example, voltage level sensing rather than conduction level sensing can be employed.

Improved Charge Retention

In the example above,states "3" and "2" are the result of net positive charge (holes) on the floating gate while states "1" and "0" are the result of net negative charge (electrons) on the floating gate. To properly sense the correct conduction state during the lifetime of the device (which may be specified as 10 years at 125′C) it is necessary for this charge not to leak off the floating gate by more than the equivalent of approximately 200 millivolts shift in V_T1. This condition is readily met for stored electrons in this as well as all prior art Eprom and Flash EEprom devices. From device physics considerations, retention of holes trapped on the floating gate should be significantly superior to the retention of trapped electrons. This is so because trapped holes can only be neutralized by the injection of electrons onto the floating gate. So long as the conditions for such injection do not exist it is almost impossible for the holes to overcome the potential barrier of approximately 5.0 electronvolts at the silicon-silicon dioxide interface (compared to a 3.1 electron volts potential barrier for trapped electrons).

Therefore it is possible to improve the retention of this device by assigning more of the conduction states to states which involve trapped holes. For example, in the example above state "1" had V_T1 = +2.0V, which involved trapped electrons since V_T1 for the virgin device was made to be V_T1 = +1.5V. If however V_T1 of the virgin device is raised to a higher threshold voltage, say to V_T1 = +3.0V (e.g. by increasing the p-type doping concentration in the channel region 560a in figure 5a), then the same state "1" with V_T1 = +2.0V will involve trapped holes, and will therefore better retain this value of V_T1. Of course it is also possible to set the reference levels so that most or all states will have values of V_T1 which are lower than the V_T1 of the virgin device.

Intelligent

Erase

for

Improved

Endurance

The endurance of Flash EEprom devices is their ability to withstand a given number of program/erase cycles. The physical phenomenon limiting the endurance of prior art Flash EEprom devices is trapping of electrons in the active dielectric films of the device. During programming the dielectric used during hot electron channel injection traps part of the injected electrons. During erasing the tunnel erase dielectric likewise traps some of the tunneled electrons. The trapped electrons oppose the applied electric field in subsequent write/erase cycles thereby causing a reduction in the threshold voltage shift of V_tx. This can be seen in a gradual closure (figure 2d) in the voltage "window" between the "0" and "1" states. Beyond approximately 1x10⁴ program/erase cycles the window closure can become sufficiently severe to cause the sensing circuitry to malfunction. If cycling is continued the device eventually experiences catastrophic failure due to a ruptured dielectric. This typically occurs at between 1x10⁶ and 1x10⁷ cycles, and is known as the intrinsic breakdown of the device. In memory arrays of prior art devices the window closure is what limits the practical endurance to approximately 1x10⁴ cycles. At a given erase voltage, V_ERASE, the time required to adequately erase the device can stretch out from 100 milliseconds initially (i.e. in a virgin device) to 10 seconds in a device which has been cycled through 1x10⁴ cycles. In anticipation of such degradation prior art Flash EEprom devices specified to withstand 1x10⁴ cycles must specify a sufficiently long erase pulse duration to allow proper erase after 1x10⁴ cycles. However this also results in virgin devices being overerased and therefore being unnecessarily overstressed.

A second problem with prior art devices is that during the erase pulse the tunnel dielectric may be exposed to an excessively high peak stress. This occurs in a device which has previously been programmed to state "0" (V_T1 = +4.5V or higher). This device has a large negative charge Q. When V_ERASE is applied the tunnel dielectric is momentarily exposed to a peak electric field with contributions from V_ERASE as well as from Q. This peak field is eventually reduced when Q is reduced to zero as a consequence of the tunnel erase. Nevertheless, permanent and cumulative damage is inflicted through this erase procedure, which brings about premature device failure.

To overcome the two problems of overstress and window closure a new erase algorithm is disclosed, which can also be applied equally well to any prior art Flash EEprom device. Without such new erase algorithm it would be difficult to have a multistate device since, from curve (b) in figure 2d, conduction states having V_T1 more negative than V_T2 may be eliminated after 1x10⁴ to 1x10⁵ write/erase cycles.

Figure 3 outlines the main steps in the sequence of the new erase algorithm. Assume that a block array of mxn memory cells is to be fully erased (Flash erase) to state "3" (highest conductivity and lowest V_T1 state). Certain parameters are established in conjunction with the erase algorithm. They are listed in figure 3: V₁ is the erase voltage of the first erase pulse. V₁ is lower by perhaps 5 volts from the erase voltage required to erase a virgin device to state "3" in a one second erase pulse. t is chosen to be approximately 1/10th of the time required to fully erase a virgin device to state "3". Typically, V₁ may be between 10 and 20 volts while t may be between 10 and 100 milliseconds. The algorithm assumes that a certain small number, X, of bad bits can be tolerated by the system (through for example error detection and correction schemes implemented at the system level. If no error detection and correction is implemented then X = 0). These would be bits which may have a shorted or leaky tunnel dielectric which prevents them from being erased even after a very long erase pulse. To avoid excessive erasing the total number of erase pulses in a complete block erase cycling can be limited to a preset number, n_max. _Δ V is the voltage by which each successive erase pulse is incremented. Typically, _ΔV is in the range between 0.25V and 1.0V. For example, if V₁=15.0V and _ΔV=1.0V, then the seventh erase pulse will be of magnitude V_ERASE=21.0V and duration t. A cell is considered to be fully erased when its read conductance is greater than I_"3". The number S of complete erase cyclings experienced by each block is an important information at the system level. If S is known for each block then a block can be replaced automatically with a new redundant block once S reaches 1x10⁶ (or any other set number) of program/erase cycles. S is set at zero initially, and is incremented by one for each complete block erase multiple pulse cycle. The value of S at any one time can be stored by using for example twenty bits (2⁰ equals approximately 1x10⁶) in each block. That way each block carries its own endurance history. Alternatively the S value can be stored off chip as part of the system.

The sequence for a complete erase cycle of the new algorithm is as follows (see figure 3): <ul id="ul0003" list-style="none" compact="compact"><li>1. Read S. This value can be stored in a register file. (This step can be omitted if S is not expected to approach the endurance limit during the operating lifetime of the device).</li><li>1a. Apply a first erase pule with V_ERASE=V₁ + n _ΔV, n=0, pulse duration = t. This pulse (and the next few successive pulses) is insufficient to fully erase all memory cells, but it serves to reduce the charge Q on programmed cells at a relatively low erase field stress, i.e., it is equivalent to a "conditioning" pulse.</li><li>1b. Read a sparse pattern of cells in the array. A diagonal read pattern for example will read m+n cells (rather than mxn cells for a complete read) and will have at least one cell from each row and one cell from each column in the array. The number N of cells not fully erased to state "3" is counted and compared with X.</li><li>1c. If N is greater than x (array not adequately erased) a second erase pulse is applied of magnitude greater by _ΔV than the magnitude of the first pulse, with the same pulse duration, t. Read diagonal cells, count N.</li></ul>

This cycling of erase pulse/read/increment erase pulse is continued until either N ≦ X or the number n of erase pulses exceed n_max. The first one of these two conditions to occur leads to a final erase pulse. <ul id="ul0004" list-style="none" compact="compact"><li>2a. The final erase pulse is applied to assure that the array is solidly and fully erased. The magnitude of V_ERASE can be the same as in the previous pulse or higher by another increment _ΔV. The duration can be between 1t and 5t.</li><li>2b. 100% of the array is read. The number N of cells not fully erased is counted. If N is less than or equal to X, then the erase pulsing is completed at this point.</li><li>2c. If N is greater than X, then address locations of the N unerased bits are generated, possibly for substitution with redundant good bits at the system level. If N is significantly larger than X (for example, if N represents perhaps 5% of the total number of cells), then a flag may be raised, to indicate to the user that the array may have reached its endurance end of life.</li><li>2d. Erase pulsing is ended.</li><li>3a. S is incremented by one and the new S is stored for future reference. This step is optional. The new S can be stored either by writing it into the newly erased block or off chip in a separate register file.</li><li>3b. The erase cycle is ended. The complete cycle is expected to be completed with between 10 to 20 erase pulses and to last a total of approximately one second.</li></ul>

The new algorithm has the following advantages: <ul id="ul0005" list-style="none" compact="compact"><li>(a) No cell in the array experiences the peak electric field stress. By the time V_ERASE is incremented to a relatively high voltage any charge Q on the floating gates has already been removed in previous lower voltage erase pulses.</li><li>(b) The total erase time is significantly shorter than the fixed V_ERASE pulse of the prior art. Virgin devices see the minimum pulse duration necessary to erase. Devices which have undergone more than 1x10⁴ cycles require only several more _Δ V voltage increments to overcome dielectric trapped charge, which only adds several hundred milliseconds to their total erase time.</li><li>(c) The window closure on the erase side (curve (b) in figure 2d) is avoided indefinitely (until the device experiences failure by a catastrophic breakdown) because V_ERASE is simply incremented until the device is erased properly to state "3". Thus, the new erase algorithm preserves the full memory window.</li></ul>

Figure 4 shows the four conduction states of the Flash EEprom devices of this invention as a function of the number of program/erase cycles. Since all four states are always accomplished by programming or erasing to fixed reference conduction states, there is no window closure for any of these states at least until 1x10⁶ cycles.

In a Flash EEprom memory chip it is possible to implement efficiently the new erase algorithm by providing on chip (or alternatively on a separate controller chip) a voltage multiplier to provide the necessary voltage V1 and voltage increments _ΔV to n_ΔV, timing circuitry to time the erase and sense pulse duration, counting circuitry to count N and compare it with the stored value for X, registers to store address locations of bad bits, and control and sequencing circuitry, including the instruction set to execute the erase sequence outlined above.

While the embodiments of this invention that have been described are the preferred implementations, those skilled in the art will understand that variations thereof may also be possible. Therefore, the invention is entitled to protection within the full scope of the appended claims.