GB2310939A - An FPGA DRAM cell sensing circuit with a stepped word line waveform - Google Patents

Description

OPERATING A DYNAMIC MEMORY The present invention relates to the field of semiconductor devices and methods of operating such devices. More specifically, in one embodiment the invention provides a field programmable logic device with dynamic memory control.

The expense of designing integrated circuits for special applications has served as an incentive to develop several ways of using standard circuits that may be modified by a user to perform particular functions. Gate arrays are an example of an integrated circuit that may be modified to satisfy a special user need using, for example, laser programming. This solution is still too expensive for many limited production runs and in prototype design.

To overcome this limitation field programmable gate arrays (FPGA) were developed.

In the earliest field programmable gate arrays, the functions of logic cells and the interconnections between cells were controlled by flip-flops. The flip-flops were loaded with a configuration program entered from the outside of the chip. Once this program was loaded, the standard chip was customized to perform specific functions.

In later designs various means for customizing the array functions have been utilized. Among the methods proposed are the utilization of "anti fuses", which is the reverse of using fuses or "cutpoints." Another common logic circuit today utilizes static ram flip-flops to control the array functions. Typically, long shift registers with shift and hold functions are used in such devices to load the control bits into the arrays. Another type of field programmable gate array utilizes EPROM or EEPROM cells to define the logical function to be performed by the device.

While meeting with substantial success, prior devices have also met with certain limitations. For example, such devices typically occupy more semiconductor area than would be desirable and, therefore, are more costly and difficult to manufacture.

From the above it is seen that improved field programmable gate arrays and methods of operating such devices are needed. The invention provides a method of operating a dynamic memory, which method can, in preferred examples, have particular application in FPGAs.

In accordance with the invention, the method comprises the steps of: precharging a selected bit line in said dynamic memory to a selected value with a precharge circuit, said bit line coupled to a sense amplifier comprising a pull up portion coupled to a selectable high voltage source and a pull down portion coupled to a low voltage source; decoupling said bit line from said precharge circuit, whereby said bit line floats; coupling said bit line to a memory cell by bringing at least one word line to a first word line voltage level while applying a first high voltage to said pull up portion; and thereafter, applying a second, higher voltage to said pull up portion while raising a voltage on said at least one word line to a second, higher word line voltage level.

This method may be employed in improved configurable logic devices in which the charge level of a capacitance is used to control the state of the controlled nodes in a gate array or other logic device. The capacitance's only function is to hold a charge long enough to "remember" if it represents a one or a zero until it is read and refreshed.

The control of gates and switches by the capacitances requires a steady and well defined signal level on the capacitance that is not generally compatible with conventional dynamic ram capacitors.

Examples are described which overcome the effect of the dynamic voltage variations in a storage capacitance so that dynamic memory cells reliably control switches and gates in a programmable logic devices. Several alternative methods are presented, all of which mask the voltage transients imposed on the memory capacitors typical of DRAMs. In preferred embodiments, the refresh of the dynamic memory cells and the "read" of the cells by the logic device may occur simultaneously since voltage transients on the stored capacitance are minimized during refresh.

Accordingly, in most preferred embodiments, the refresh and read need not be synchronized to avoid the simultaneous occurrence of these events.

The described examples can result in programmable logic with substantially smaller memory cell size.

For example, a typical SRAM cell in presently manufactured devices could occupy about 335 square microns. SRAM area on a typical programmable logic chip would be in the range of 31% of chip area. By contrast, the described examples may result in estimated cell areas of between about 45 and 67 square microns. Therefore, device size would be reduced by about 80% of 31%, or about 25%. Obviously, with significant chip size reductions, attendant changes will result in device capability and/or yield.

There follows a description of examples of configurable logic devices and examples of the invention, particularly when applied to such devices.

Figs. 1A and 1B are overall illustrations of a dynamic configurable logic device (DCLD); Figs. 2A and 2B illustrate a look-up table device for use in association with a programmable logic device operating under the control of dynamic memory cells; Fig. 3 illustrates one form of a memory cell; Figs. 4A and 4B illustrate disturbance of a "one and a "zero" on a memory cell; Figs. 5 to 10 illustrate other forms of memory cells; Fig. 11 illustrates aspects of the sense and load portions of the memory in the DCLD; Figs. 12 and 12A illustrate the programming portion of the device in greater detail; Fig. 13 illustrates details of a data loading circuit; Fig. 14 illustrates a half cell design; Fig. 15 illustrates the use of dummy memory cells; Figs. 16A and 16B illustrate an embodiment of the invention; Fig. 17 illustrates a circuit for producing Vdd+; Fig. 18 illustrates a bootstrap circuit; Fig. 19 illustrates a Vdd++ generation circuit; Fig. 20 illustrates a strobe using Vdd++; and Figs. 21 and 22 illustrate another arrangement.

Although the invention claimed herein relates only to a method of operating a dynamic memory, the method has particular, though not exclusive, application to dynamic memory used in programmable logic. Accordingly designs of cells and arrangements of programmable logic devices will be described in the following which are not necessarily examples of the present invention. However it will be apparent to those skilled in the art which of the designs and arrangements can be used with examples of the invention.

Reference should also be made to the following co-pending applications which claim subject matter disclosed in this application: Application No. 9518093.1 (the parent application) which claims programmable logic in an integrated circuit comprising: logic means for outputting signals representative of desired logical functions of inputs to said integrated circuit; means for inputting data to an array of capacitance means for storage of selected voltage levels thereon representing said data, said data being representative of a desired logical function of said logic means; means for refreshing selected voltages in said array of capacitance means, each of said capacitance means comprising a control voltage node coupled to said logic means to supply said data to control said logic means to provide said desired logical function; means for supplying said data from said control nodes while refreshing said selected voltages in said array of capacitance means, whereby logic functions may be performed with said logic means when said capacitance means are being refreshed, and means for reading stored data from said capacitance means whereby data stored in said capacitance means may be verified.

Application No. SXlCIS 5 (a further Divisional application) which claims a dynamic memory integrated circuit structure comprising: a storage capacitor, a first plate of said storage capacitor coupled to a gate of an access transistor, a second plate of said storage capacitor coupled to a first terminal of said access transistor; a read transistor, said read transistor having a first terminal coupled to a second terminal of said access transistor and second terminal coupled to a bit line; a restore transistor, said restore transistor having a first terminal coupled to said bit line and a second terminal coupled to said first plate of said storage capacitor; and an inverter circuit having an input selectively coupled to said bit line and an output selectively coupled back to said bit line.

CONTENTS I. General II. Memory Cells III. Sense/Load/Refresh IV. Additional Features A. Dummy Cell Embodiment B. Half Cell Dynamic Memory C. Four-Phase/Floating Bit Line Embodiment V. Peripheral Circuits A. Pump Circuit B. Bootstrap Circuit C. Secondary Pump VI. Signal/Noise Level Considerations A. Noise Level Reduction.

I. General Dynamic memory cells are used in the operation of programmable logic such as a prograable gate array. The effect of the dynamic voltage variations in the memory cells of a dynamic memory device can be overcame such that the dynamic memory cells can reliably control switches and gates in programmable logic devices, such as FPGAs.

Several alternative methods are presented. with a common goal of masking out the voltage transients imposed on the memory capacitors typical for dynamic memory. An overriding goal is of course to reduce the circuit area and thereby the cost of controlling the functions of the FPGAs.

Fig. 1A is an overall illustration of one embodiment of a dynamic memory configurable logic device 1. As shown, the device provides an array of dynamic memory cells 3 ttat are periodically refreshed by programming/refresh means 6. all contained on a single integrated circuit structure 10. The integrated circuit strucrure further includes logic device(s) 12 that are controlled by various control nodes CN in the dynamic memory. The logic device 12 provides outputs that are logical functions of inputs wherein the logical functions are based on the state of control nodes CN in the dynamic memory. Of course. in most devices the memory and the logic will be physically intermixed such that the memory will be in close proximity to the controlled device.

Fig. 1B provides a simplified version of a device such as the one shown in Fig. 1A in which logic inputs are selectively passed through switches controlled by dynamic memory cells. As shown therein. various inputs such as inputs INx and. optionally.

orthogonal inputs INy are provided to the device. It will be recognized that many such inputs will normally be provided. although only two are illustrated for simplicity. Of ten. both the true and complement of inputs will be generated and provided to the logic device as indicated in the Figure.

As dynamic memory cells M are used to direct selected inputs to one or both of AND gate(s) 12 and/or OR gate(s) 15. For example. if it is desired to provide the AND gate 13 with INx as a product term. M1, and M3 are programmed high. If it is desired to provide INx as an input to OR gate 15. M7 is also programmed high.

Similarly, M can be used to steer INy' to the inputs of gates 13 and 15, etc. Me may be used to transfer data to/from a horizontal line INx from/to a vertical bus line 14 that intersects many input lines.

M3 and M6 are used for isolation of the gates when a signal is to be transferred .o the vertical bus. but not to the input of a gate. The crevice illustrated in Fig. is can be viewed as utilizing dynamic memory cells =o acdress various inputs @or use in AND or otner logic gates and for connecting segments of data buses. horizontal and vertical.

Conversely, as illustrated in Fig. 2A. the inputs can be used to address selected memory cells. the stored value outputs of which are urilized in a logic function. Like the device illustrated above. this device also includes one or more arrays of dynamic memory cells 3. The memory cells are programmed and periodically refreshed with program/refresh circuit 6. Various inputs I are provided to a cell selection circuit 4. which selects a value stored in one or more of the memory cells for output. For example. if all of the inputs are asserted. this may indicate that the value of the 2n'th cell is to be output. If this cell has a high value. the output will be high. If this cell has a low value stored therein, the output will be low.

The outputs are normally combinatorial functions of the various input values. A simple example is provided below. Assume that two inputs are to be provided. The truth table that a particular user desires is: Incut :1 input 12 Output O 0 0 0 1 1 1 0 0 1 1 1 The combinatorial function described by this table is O-A-S+A-B.

Accordingly, four memory bits would be provided in the DRAM 3 in this simple example. Those with "addresses" 11 (corresponding to A high and B high) and 01 (corresponding to A low and B high) would be set high. Others would be low. When either of these addresses is applied to the inputs. the output will be high: otherwise the output will be low. Accordingly, it is seen that dynamic memory may be used to implement combinatorial logic functions.

Additional functionality will normally be provided by the device. For example. output circuit 9 may be used to select between registered and non-registered output. Memory 3 may, therefore. have additional bits that serve as architecture control bits. For example. one architecture control bit may be used to select between registered and non-registered output in the output block 9.

Since the outputs are generated by providing the various inputs as an address to a particular memory bit in the array, the number of memory bits needed will depend at least partially on the number of inputs that are to be provided. Normally 2n or more memory bits will be provided. where n is the number of inputs to the array.

Of course. the invention may readily be further modified to provide multiple outputs. provide feedback. and the like. in which case much arger numbers of bits may be utilize.

Fig. 2B provides a schematic illustration of hardware that would be used in association with the device shown in Fig. 2A in greater detail. A simple two input logical device is again used for illustration. As shown therein. four memory bits 101 are used in the DRAM 3. The inputs I are connected to define each of the possible combinatorial functions of the two inputs. For example, the output of the line 7 defines the function I1#I2#B4. Accordingly. bit a4 is programmed to provide a high output to its control node CN in tne case where I1#I2 is desired as one of the product terms of the combinatorial function. The output 0 will be the logical OR of all of the product terms designated by the various memory bits B. For the particular truth table shown above. the control bits a2 and B4 would be set high and the remaining bits would be set low to generate the logical function I1#I2+I#I2. Of course. additional elements may be provided. such as buffers on the memory cell control nodes CN or on the output nodes that are not shown for simplicity.

II. Memorv Cells Fig. 3 shows one typical dynamic memory cell 101.

A high or a low voltage is stored in a capacitance 102 depending on whether a ONE or a ZERO.

respectively. is located on tfle memory cell. The memory cell voltage is "read" in the present invention via the control node CN that is located between a memory cell transistor 105 and the capacitance 103.

The voltage stored on the memory cell capacitance will determine the logic function to be performed by a logic device. The capacitance 103 may take the form of for example. a conventional capacitor or parasitic capacitance.

Due to the leakage cf charge from the capacitance its charge must be refreshed periodically. Refreshing is accomplished by addressing the memory cell via a word line. 'reading" the charge level stored thereon. and applying the correct refresh voltage to the bit line. For example. in the embodiment shown in Fig. 3 the word line is brought high, rendering transistor 105 conductive and coupling the capacitor to the bit line. The appropriate high voltage is developed on the bit line via a sense amplifier 107 for restoration of the full charge on the capacitor when a 1 is stored.

or rhe bit line is brought to ground when a O is stored.

At the moment of reading before refreshing a small amount of charge is drawn from the capacitor. This charge is replaced quickly. but the voltage on the capacitor is momentarily disturbed as indicated in Fig. 4A. wherein tne voltage level at the storage node is shown to be disturbed downwards when a "1" is stored on the capactor. Conversely, as indicated in Fig. 43, the voltage at the contro noae o= a capacitor wtn a "@" stored thereon 5 no=naily disturbed upwards since the bit lines are typically precnaraed to an intermediate voltage between the 0 and 1 levels. The cell as shown in Fig. 3 can be used directly if the circuit logic is synchronized with the refresh cycle so that the logic system reads the capacitor voltage during times when the capacitor voltage is not disturbed for refresh. In practice this limits operating speed. increases circuit ccmplexity, and otherwise negatively impacts the device. Therefore.

it is desirable to utilize other methods of refreshing the circuit while still maintaining the ability to utilize the capacitor voltage in a logic circuit, even when the cell voltage is impacted during a refresh operation.

The noise amplitude is reduced to acceptable levels such that the logic circuit will reliably read the capacitor voltage during either the disturbed or undisturbed portion of the cycles shown in Fig. 4. In general larger capacitors and shorter bitlines are used, but will still result in smaller overall chip areas than other logic devices.

Fig. 5 shows another memory cell that may be utilized.

In this embodiment. an extra transistor 501 is introduced with its gate connected to a "strobe" signal. The transistor 501 separates a memory capacitor C2 from a refresh capacitor C1 except when the strobe signal is activated. Therefore. the circuit separates capacitor C2 from the refieShcspacitor C1 during the time of the refresh disturbance.

C2 may be a very small capacitor (such as about 10.20 fF using current device technology). and in some embodiments may only be the load capacitance of the controlled logic gate, while c1 is large enough (e.g., 30-50 fF using current device technology) to provide a reliable reading in connection with the refresh cycle. During refresh. transistor 501 is off, and isolates the memory capacitor from the refresh disturbance. Therefore. also during refresh the voltage on capacitor C2 may be used in the device logic as a control signal. The strobe signal goes active after the capacitor C1 has been refreshed. such that the capacitor C2 is also refreshed. When the strobe is active. the voltage on C2 is supported by the charge from Ci. The strobe can be a global signal or it can be acrive only for a row that has j just been refreshed.

In either case the programming logic will preferably enable the strobe function so that both C2 and C1 would be fully charged when the original control patter: or data are written into the array. The advantage of this arrangement is that even if C2 is unusually large. which could occur for long connecting lines. C2 would be fully charge. Individual strove lines for each row will reduce the dynamic power consumption.

Fig. 6 illustrates another memory cell that may be used In this case. the word line is used as a strobe by replacing the N-channel transistor 501 with a P-channel depletion mode transistor 601. Accordingly, during refresh. the word line is high. providing charge to the capacitor C1.

After refresh, the word line is brought low. permitting charge to flow from C1 to C2. Again. the voltage at C1 may be used at any time (during refresh or not) in the device logic.

The refresh disturbance on the memory capacitor can also be filtered out using the traditional RC filter as indicated in Figs. 7, 8, and 9. In the embodiment in Fig. 7, a resistor R is formed between the memory cell capacitor C and the call transistor 105. The relation between R and C will be set such that R is small enough to give sufficient signal to the sense amplifier. but large enough to filter out the voltage swing on the bit line during the read refresh cycle such that the swing does not exceed a desired level. For example, using conventional. present technology R will be about 100 kn. The associated amplifier's speed of response will be increased when the bitline capacitance is low and the transistors in the amplifier are large. This is accomplished if the number of bits per bit line is minimized. The memory cell capacitor and the resistor must still be relatively large.

Zn the embodiment shown in Fig. 8, a resistor R is provided between memory capacitor C2 and refresh capacitor C1. The arrangement according to Fig. 8, with C1 and C2 of approximately the same size, has the advantage that the instant signal to the amplifier is much larger when the wordline is accessed. Therefore the amplifier response time is shorter. shortening the duration of the disturbance on C1. The filtering R-C2 time constant can therefore be shorter. The size of resistor R depends to a high degree on the available processes, but with present device technology will be about 100 kO. The resistor can of course be replaced with a long and narrow switching transistor 105 as indicated in Fig. 9.

Fig. 10 illustrates another memory cell according to one aspect of the invention. In this embodiment. depletion mode.

N-channel transistors 1001 and 1002 are utilized. In essence. the devices 1001 and 1002 make up a resistor/capacitor circuit similar to that shown in Fig. 8. The transistor 1001 has its gate tied to the source/drain of transistor 1003. The resistor can be made up of two such devices connected as indicated in Fig. 8.

III. Sense/Load/Refresh A sense. load. and refresh circuit is illustrated in rig. 11. This particular embodiment is adapted for the memory cell arrangement shown in Fig. 5. As shown there are cwo rows of sense amplifiers 1101 and 1103, one row for the upper half of the array and one row for the lower half of the array respectively. The device utilizes @folded bit lines" in which one bitline reads and restores cells addressed by even numbered word lines and the other bitline reads and restores cells addressed by odd numbered word lines. The bit lines extend in the same direction from the sense amplifier, wnich provides alpha particle error protection and other benefits.

Each amplifier is a CMOS sense amplifier of the type well known to those of skill in the art of dynamic memory design and contains two cross-coupled pairs of transistors including a cross-coupled pair of NMOS transistors 1107 and a cross-coupled pai; of PMOS transistors 1109.

Each row of amplifiers is powered by two buffer drivers 1111 and 1112 with complement outputs. When the amplifiers are active. the sources of the N-channel transistors N1 and N2 are connected to ground and the sources of the P-channel transistors P1 and P2 are connected to a positive supply voltage. As will be discussed later. this positive supply voltage is preferably higher than the Vdd normally used n the chip. Accordingly, when the amplifiers are active (i.e., when 1112 produces a high potential and 1111 produces a low potential), the higher of the two bit lines is brought to Vdd or above. while the lower of the two bit lines is brought to ground.

When the polarity of the buffer drivers is reversed. the bit lines will be precharged to an intermediate voltage between 0 and V:c near the switching points of the cross-coupled pairs. typically a potential close to half of the supply voltage. Transistor N3 is also turned on during a precharge mode, bringing the bitline pair to the same potential.

During a refresh of a memory cell. its word line is brought high and the sense amplifier is activated. At essentially the same time the selected word line is brought high. the strobe signal is brought lou. The selected memory capacitor. if at low potential will pull its bitline low enough to cause the flip-flop to switch to the low potential on the selected bit line. On the other hand if the selected memory capacitor is at high potential the selected bitline will swing to the positive potential. thus refreshing tne high charge on the memory capacitor.

Programming data are fed into the chip via a shift register 1112. When the register has been filled. the shifting stops and the register holds its data until te refresh starts. At the same same as the amplifiers 1111 and 1i2 are activated, a "Write Upper" or a @write Lower" signal is applied. This will connecr the appropriate nodes on the shift register to the sense amplifier and force the sense amplifier to store a 1 or 0 state in the memory capacitors along tne activated word line. For example. if the data shifted into the shift register produce a low voltage at node 1115 and WRITE UPPER is activated while WL1 is active. a low voltage level will be stored in memory cell 101. In alternative embodiments, reading of data by the logic array from the capacitors during the loading of data will be disabled.

The polarity requirement is different for the bit lines associated with the even and the odd word lines since. in preferred emoodiments. a single bit will comprise change of opposite polarity stored on both of the adjacent bit lines. Data are therefore complemented before feeding the shift register for either the even or odd word lines. The lower half of the array is a mirror image of the upper half.

Fig. 12 illustrates a dynamic shift register 1201 used for sequential addressing of the word lines as part of a refresh and data load function. In this particular embodiment. each word line WLi has its own stro@e line Si. The shift register 1201 is set up so that it can only circulate one ZERO bit. with all other bits I ONE (i.e.. high). The position wh

Fig. i2A also shows the preferred timing for a global strobe signal. which replaces the individual stro@es S#...Sn. The strobe returns positive while the word line is still high, securing full refresh on both C1 and C2. The same timing on individual strobes can be arranged at a slightly higher cost.

Fig. 12 also shows a two stage shift register 1205 that is used to keep track of odd and even addresses. The shift register is clocked with #1 and 2. If. for example, EVEN begins high. and after #1 and f2 have been high. the register will clock the signal ODD high. This will repeat until LAST WL bar has gone high, resetting the register to its initialized state via NANDs 1207 and 1209. Alternative circuitry such as a toggle flip-flop may also be used for this purpose. It is assumed that the address shift register is duplicated for both halves of the array, for the case when the arrangement n Fig. 11 is used. An even number of word lines is Illustrated. so that when the last word line (WLn), which is odd.

shifts the circulating low co the first word line (WLO), the odd-even register shifts to generate the signal EVEN. Both shift registers have provisions to block forbidden bit combinations by setting a "1" to all but the S rs: position at the last word line time. The registers can also be set to WLO and EVEN when a load request is sent to the chip.

Fig. 13 shows an example of the type of control that can be used to communicate with outside for loading of data onto the chip. A load request signal LOAD REQ is applied by the user and is synchronized to the refresh clock which is internally generated on the chip. A signal LOAD ENABLE is returned to the programmer and will scay on one clock cycle beyond when LOAD REQ scays active. LOAD REQ is clocked into flip-flop 1301 with clock #2 and. thereafter, to register 1303 with the next 2 clock. The user also provides a signal to indicate whether data are to be written to the upper half of the chip. WRITE UPPER HALF. Using NAND 1315, when WRITE UPPER HALF and LOAD REQ are high. the WRITE UPPER signal will be high during #2' allowing data to be written into the upper half of the array. Alternatively, when WRITE UPPER is low. data may be written to the lower half of the array since WRITE LOWER will be high using NOR gates 1305 and 1307, respectively.

This circuit also generates SET WL0 bar. When LAST WL bar is received. the WRITE UPPER and WRITE LOWER signals will be deactivated by te controlling system bumping load request low.

The data shift register is advanced by clock signals from the programmer. #1D and 2D This clock will be run in bursts. each burst transferring as many bits as te register holds (as many as the number of amplifiers in each row). As shown. the data are provided at DATA IN, and are inverted with XOR gate 1309 with a signal EVEN, which indicates that an even row is being addressed. For test purposes it is desirable to read stored data. READ TO SR is used to check the stored data values. READ DATA OUT may optionally be used to check the data after passing through the registers.

There is a window in the refresh cycle during which data can be loaded. This is defined by the LOAD WINDOW signal. that is produced by NAND 1311 and inverter 1315 based on 02 bar. As as long as LOAD REQ and the accepted LOAD REQ output from flip-flop 1301 are active, the LOAD WINDOW is open ance each refresh cycle (during 02 bar). Naturally, the clock frequency for loading data must be high enough to complete the loading before the window is closed.

When the last word line has been written a STOP LOAD signal is sent to the programmer via NOR gate 1313 based on LAST WL bar and the data outputs of the flip-flops 1301/1203. If at this time both the upper and the lower halves have been programmed the programmer brings the LOAD REQUEST signal down.

It may be desirable to verify that the correct data has been stored in the control memory. Read and modify operations may also be used to modify only selected cells. To accommodate this.

provisions are added to the on-chip control logic and to the DATA IN shift register. For the external programming system where the chip is used to "know" the row address. the last WL signal and one of the refresh clocks are connected to I/O pins. This is illustrated in Fig. 12. The system keeps a separate row count. thus permitting the system to read or write data from or to any row synchronously with the refresh cycle.

To read data out, the procedure is as follows. The DATA IN shift register is shifting slowly during normal refresn cycles under control of the refresh clocks (this prevents floating gate situations to cause excessive current drain). At the end of the time associated with the row to be read. a READ To SR pulse is applied to an input pin. As exemplified in Fig. 14, this copies the data read by the sense amplifier to the shift register. After the end of #2 and after the READ To SR pulse, the internal 1 and 2 clocks are replaced with the fast shift clocks #1d and 02dw causing the read data to be shifted out.

A read-modify-write operation can be arranged as follows.

The 1 and 2 clocks can be timed far enough apart so that two complete shifts of data can take place without interfering with either clock. Data are shifted out to the system bit by bit and returned to the input of the data register on the chip. As the bits pass through the system, old data bits can be replaced with modified ones, if desired. After one complete cycle a data load pulse.

replacing the #2 2.DATALD pulse indicated in Fg. 14, transfers the new data to the sense amplifier. The new data is now on tne sense amplifiers output before the arrival cf tne 1 clock used to restore charge to the memory capacitor. Alternatively the full content of the shif register can be shifted out. modified and shifted in during one refresn cycle. Only if more time consuming data modifications are made. will it be necessary to wait for the row to return after a full sequence of refresh cycles.

rv. Additional Features A. Dummy Cell Embodiment Fig. 15 illustrates an arrangement using the basic DRAM cell combined with "dummy cells." A wordline will in a typical array be very long and with high capacitive load. The line. preferably driven from a pumped voltage source. will have a slow rise cime. The time connecting a capacitor at low voltage level to its bit line is much earlier than the time at which a capacitor at high voltage is connected to its bit line.

This time difference can be critical for the following reason. The clamping of the bitline pairs brings both to the same potential. Just before the word lines are brought high, the clamping is reversed. The flip-flop goes ideally to a meta stable state with both nodes at the same potential. at the flip-flops switching potential. The switching potential is where the P-channel and the N-channel transistors draw the same current. Due to imprecise design and processing the ratio between device sizes in one leg of the flipflop may be different than that of the opposite leg. After clamping, this unbalance may be equivalent to a signal already applied to a flip-flop node.

Left alone the flip-flop nodes will proceed to switch in the direction of the unbalance. To minimize the effect of the unbalance the signal from the capacitor should be applied as early as possible after switching the amplifier from the clamping state, making it desirable to read both high and low voltage on applied storage capacitors at the same time. The fact is that if there is no signal from a low capacitor. then the associated capacitor must be at a high level. Adding a dummy capacitor on the not addressed bit line minimizes the delay between the two readings. The disadvantage of a dummy memory cell is that the effective signal from the storage capacitor is "reduced' by the signal from the dummy capacitor. The charge transferred from che dummy capacitor should be half of the charge transferred from the storage capacitor.

As shown in Fig. is. single capacitor memory cells are provided with single access transistors in each memory cell 1400. As snown in Flg. is. tne voltages on the bitlines are normally clamped t= an n ermediate voltage. such as Vd/2. with a signal CLAMP.

CLAMPbar S generated with a large buffering inverter i401.

Accordingly, both of the bit lines are connected together via transistor .402 and the sources of pull-up transistors 1403 in sense amplifier 1404 are brought low when clamp is high.

When a word line is addressed for refresh, the ODD DUMMY WL signal will be generated when the word line is odd. or the EVEN DUMMY WL signal will be generated when the word line is even.

act'vating the access transistor in the corresponding dummy memory ceil 1407 or 1409. respectively. These dummy capacitors will have low or ground voltage stored thereon. When the word line is addressed. the sources of PMOS transistors 1403 will be brought high and the bit lines will be disconnected. The.word line (WL) and dummy word line (DWL) voltages will be brought high. Preferably the word line is raised above vcc. In the case where the addressed storage capacitor has a low voltage thereon. the addressed storage capacitor voltage (C) will increase slightly. while drawing charge from the bit line. However, the charge drawn from the opposite bit line by the dummy memory cell will be less. Accordingly, the dummy bit line will be higher than the bit line with the addressed cell. The switching will then proceed to completion.

As pointed out earlier the noise on a low capacitor can be held low with the right combination of capacitor size, number of bits on each bit line and the size of the transistors in the amplifier flip-flop. There will be no noise on the high capacitor when the dummy capacitor is used.

In addition to the amplifier flip-flop and the dummy cells s Fig. 15 also shows the circuitry 1411 for writing in new data and to read previous stored data. Such circuitry is needed only if the user does not complement the input data.

B. Half Cell Dvnamic Memory In the embodiment below, a very small storage capacitor can be used since the cell has its own amplifier. The bitline capacitance can be large. as there are no speed requirements for this application. Accordingly, a half cell system may be used.

Returning to Fig. 14. che memory cell 1501 includes transistors N1. N2. N3 and capacitor C. which is preferably the extended gate of transistor N2. Two word lines are used for each row of cells, one for reading (READ WL) and one for restoring (RESTORE WL) te charge on capacitors C. The amplifier is basically an inverter, so tat the read signal on the bit line can be reinvested and returned to the bit line at the restore crime.

The operation o the read restore and data write operations are as follows. The address shift register works the same as described in the above embodiments. The first inverter in the selected shift register stage is set @y the #1 clock the second inverter is reset by the following #2 clock. The selected REAAD WL is brought high by the same #2, clock that resets the second inverter.

The READ WL turns on tne cell transistors N1. If the capacitor has a hign charge. transistor N2 is in the on state, which causes the bit line BL to be pulled low. During 2 transistors N4 and N5 in the amplifier 1503 are also turned on. N4 is a weak transistor. serving as a pullup load in case the cell capacitor is in a low charge state.

However, with a high charge on the cell capacitor the gates of transistors P1 and N6 are pulled to close to ground level and remain at this level when the gates of N4 and N5 are returned to ground ar the end of a signal #2xDATALD*. At the next.#1 the output of the inverter 1505. formed by transistors P1 and N6. is connected to the bit line via transistor N7.

with a short delay after #1 (to permit the bit line to change state. . RESTORE WL is brought high via NOR 1507 and connects tne bit line to the storage capacitor C via transistor N3 restoring a high or low. as appropriate.

When data are written into a selected row, the content of the DATA IN shift register is similarly applied to the input of the amplifier inverter by the signal 2xDATALD.

The supply voltage for most signals is the regular Vdd' but in order to get the lowest impedance and highest speed on the circuits controlled by the cell capacitors some of the circuits are connected to a pumped supply Vdd+. This supply, which will be described later. delivers approximately 8 volts when Vdd = 5 volts.

The inverter in te amplifier and the NOR driving the restore word line are supplied by Vdd+. The X signal that is applied to the gate of N7 is also at Vdd+ level. The hign voltage on the storage capacitor will therefore be one N- channel thresnold below vdd+ or approximately 7 volt. Optionally. bootstrapping can be added to the signal on the gate of N7 and on the RESTORE WL.

C. Four phase/Floating Bit Line Embodiment Figs. 16A to 16B illustrate an em@odiment cf the invent on. In particular. Fig. 16A illustrates the relevant portions of a four-phase circuit. Fig. 16B illustrates the relevant signals in the circuit on a tirilng diagram. One important feature illustrated by way of the circuit shown in Fig. 16 is that the bit lines are clamped and then left to float during initial sensing wit both the high (VddA) and low (VssA) supplies to the sense emplifier 602 left at a clamp voltage (e.g.. i.S v) during initial sensing.

UPPER HALF FIELD is the output from flip-flop, which is set when the address shift register reaches the first position in the upper field. This field is true when this field is addressed and remains set until the address shifts out of the UPPER FIELD. The signal CLAMP LEVEL is at about 1.5 v. Clocks CLl-CL4 are high (S v) during the times indicated at the bottom of the timing diagram (Fig. 16B). CL2 also serves as the CLAMP signal.

CLAMP sets the bit lines BL1 and BL2 at this same voltage via clamping transistors 1603. During this time. the high and low voltage supplies to the sense amplifier are also set to the CLAMP LEVEL. The low voltage supply is set to CLAMP LEVEL since transistor 1605 is on. CL2 resets thrce flip-flops in the circuit. This brings nodes 1623. 1631 and 1643 low. and 1625 and 1635 high. With 1625 high, nodes 1629 and 1627 are connected to the CLAMP LEVEL via transistors 1609 and 1605. Node 1627 is V5sA. which supplies the sources of the amplifiers N-channel transistors, which are now at about 1.5 v.

With node 1635 high. node 1637 (VddA) is also at 1.5 v.

The CLAMP signal (CL2) turns on transistors 1603 also bringing both bit lines BL and BLd to 1.S v. The CLAMP also turns on transistor 1615. discharging the DUMMY capacitor to ground.

The DUMMY WL and WL are both at ground level during Ct2, because 1641 was brought low and 1637 was brought high by the previous CLI. rt? pulls node 1643 (WL) low.

During CL3. both the selected word line WL and the dummy word line for the UPPER HALF FIELD DUMMX WL are brought to 1.5 v via circuits 1621 (WL connected to 1641) and 1613. respectively. A low value stored in the memory cell is illustrated in Fig. 16S and.

accordingly, the bit line BL is brought down significantly (due to the relatively large capacitance of the memory cell and the low value stored therein). Since a low value is stored in all dummy memory cells. the voltage of the dummy bit line BLD is also brought down.

although not as low due to the relatively smaller capacitance of the dummy memory cell. An important feature is that. with the voltage on WL during CL3 and CL4 at the CLAMP LEVEL. the voltage on CN should never be higher than CLAMP LEVEL minus the threshold voltage on the selected transistor. This means that even if CN is small relative to the bit line capacitance. the disturbance on CN will barely reach the threshold level on the circuits controlled by CN. On the other hand, if CN is larger than the bit line capacitance. then the disturbAnce will be less than the threshold level on the controlled circuits.

CL4 changes the state of nodes 1623 and 1625 bringing V5sA to ground and VddA high to about 4 v via transistor 1607. The WL and DUMMY WL both remain at 1.5 v during CL4. The application of power on the amplifier causes the bit lines to switch in the relative direction preconditioned by the earlier displacement. BL will be pulled to ground. bringing CN along and BLd goes to 4 volt. while the dummy cell voltage remains at .7 v. CL4 according to one embodiment is relatively longer than the other clocks to permit complete switching of the high going bit line. while VddA at 4 v.

Thereafter. during CLI. the supply vddA is switched to Vddt (a v), pulling the high bit line up to Vdd+. CLI brings node 1631 to Vdd. which in turn brings node 1637 (VddA) to Vdd+. The word line is also brought to Vdd during CL1. but is preferably delayed enough relative to the positive going bitline so that a high storage node remains undisturbed. The dummy word line has already served its purpose during this time and may optionally be returned to ground at CL4. CL2 rerurns the word line to ground just before is returned to the clamp level. As indicated in dashed lines. VddA may alternarively be brought up continuously over CL4 and CL1.

In more detail. the switching of the voltages described above is accomplished as follows. Circuit 1619 controls VddA and VssA levels for the amplifier in the selected field. Circuit 1613 controls the dummy line voltage. The address shift register 1617 cogether with the circuits 1621 and 1611 selects and drives the word line.

Circuit 1619 is acrive only when the field is selected.

All non selected fields are in the clamp mode. CL2 applied to the circuit 1619 causes node 1623 to go to ground and node 1625 to go to Vdd. This connects VssA and node 1629 to the CLAMP LEVEL. CL2 also brings node 1631 of circuit 1633 to ground. toggling the high voltage latch. so that node 1635 goes to Vdd+. The high node 1635 toggles the flip flop with node 1629 connected to V55 so that the node 1637, which is supplying vdd to the amplifier. is connected to the CLAM LEVEL When the field is selected CL4 pulls node 1625 in the circuit 1619 low. joggling node 1623 high. This connects node 1627 to ground and lifts node 1629 to one threshold below 5 v (about 4 v). Node 1635 is still at vdd+, so node 1637 goes to 4 v. If the field is selected. CL1 will finally pull node 1635 low, making node 1639 high.

This in turn toggles the latch for VddA. so that the node 1637 goes to Vdd.

Circuit 1613 in the clamped mode, initiated by CL1. is holding the DUMMY WL at ground. With the UPPER HALF FIELD flip flop set (not shown) node 1637 is pulled low at CL3. This makes node 1639 high. connecting the DUMMY WL to the CLAMP LEVEL. Circuit 1621 functions similarly to circuit 1613 except that node 1641 is connected to the CLAMP LEVEL only when the associated address shift register position is set. Circuit 1611 is the latch. whose output.

node 1643. is the word line. CLI AND a set address shift register position brings node 1643 to the Vdd+ level. while following CL2 returns node 1643 to ground. During 2 node 1643 remains at ground and will stay at this level until the address shift register returns to this position. The latches used to switch the V5sA between ground and the CLAMP LEVEL cf the different latches are used to maKe sure that one gate is turned off before the other is turned on. This prevents glit='es on the CLAMP LEVEL voltage. Of course. a simpler arrangement may be used if more current crain in the circuit is acceptable.

V. Peripheral Circuits A. Pumo Circuit Fig. 17 is a preferred pump circuit 1701 to provide the Vdd+ voltage. The goal of the design is to avoid any risks of charge injection into the substrate. The oscillator includes of three inverters 1703. 1705. and 1707 with the last (1707) being a Schmidt Trigger. The oscillating frequency generated by oscillator 1709 is relatively high, at least 100 MHz. Two powerful inverters 1703 and 1705 drive the pump circuit. which may or may not be symmetric.

The symmetry configuration was chosen here to minimize the noise generation. The physical location of this circuit should be as close co the power pins as possible.

The current requirement has two components. First a DC component to supply gates and inverters. that have inputs at Vdd level and the P - channel devices connected to Vdd+. The transient current associated with charging word and bit lines must also be considered. Fortunately refresh cycles are very long, permitting long rise and fall times. so the AC component is very small. The size of capacitors C1 and C3 can be approximated by the formula I X 2xCxDVxf. where DV is the voltage drop below the voltage at no load. and f is the frequency. Additional voltage drops are due to the transistors N1 througn N4. These transistors should be large to minimize this voltage drop component. For a load current of 2mA eacn capacitor 1 and C3 should be approximately 10pF at 100 MHz.

Capacitor C2 can be smaller then the other t'wo.

B. Bootstrap Circuit It was mentioned earlier that some signals (e.g., wordline and strobe) could benefit from bootstrapping. Fig. 18 shows a bootstrap circuit that nay be used.

The input IN is applied directly to the gate of a source follower 1801. which after a small delay in the first inverter 1803 lifts the output to approximately on threshold below the supply voltage. After a further delay tnrough the second inverter 1805 the lower end of C swings positive to the supply level.

A voltage of VddxCl/(C1 - C2) is therefore added to the first level. In most cases the load capacitance C2 is quite large.

requiring a large C . The source f=llcwer ril also ave : drive vert @arge instantaneous current to precharge the cutput node. If a bootstrap circuit is used. when the circuit is supplied from a pump.

very high peak currents must be avoided. A small source driver transistor must be used in combination with a long delay in the second inverter. The second inverter would in that case be replaced with a delay circuit. This delay circuit may be a gate utilizing some of the clocks in the dynamic control circuitry.

C. Secondary oumc In Fig. 19 the pump 1707 has been duplicated to generate a voltage. Vdd " which is a few volts higher than the Vd+ voltage, i.e.. Vdd++. Using Vdd++ will permit the word line and transfer signals to go high enough to achieve gate control signais at Vdd+ levels without resorting to Bootstrap circuits.

Fig. 20 shows an alternative way to utilize the Vdd++ voltage. In Fig. 20 STROBE. which is supposed to go low when the word line is selected. is pulled down by transistor N1. ?1 is still turned on with a gate voltage I Vdd - Vdd.+ since the high shift register output is only at the vdd level. To minimize the associated current P1 is long and narrow. Fortunately no current 5 drawn by non-selected wordlines since NI is then turned off and the strobe is pulled to the Vdd++ level.

The NAND gate connecting to the word line is supplied from Vdd++ and ground. so full CMOS perform pulse of Vnoise peak amplitude = 2+Vth (1.4 v) . If the threshold voltage for the selected transistor at 4 volt t i volt and VT for the non-selected transistor = 0.7 volt. the output would drop l volt to 3 volt. There are of course worse cases. where more than one non selected transistor with their inputs at ground all are refreshed at the same time With 2 transistors pulling down under those conditions the voltage would drop approximately 1.5 volt to 2.5 volt and with 4 transistors pulling down, the drop would be 2 volt.

brnging the output to 2 volt. The stated numbers are generally based on the equations applicable to MOS transistors.

If pumped voltages are used to bring the voltage on the control capacitors above Vdd. then the situation changes. With the control voltage at 7 volts. an incoming signal at 5 volt and one disturbing input would drop the output to 4.6 volt. Two disturbed signals as discussed above would drop the output to 4.3 volt and four disturped inputs would drop the output to 3.8 volt.

In the discussion above. the assumption was made that the distur:ance was much slower than the speed of che circuits that are controlled. with the attempts to make a fast amplifier. a noise pulse may actually be filtered in a multiplexer. It is also assumed that the disturbance on the controlling capacitors peaked at 0.7 volts above the threshold level of the controlled transistors. If the threshold voltage on the transistor is 0.7 volts. the peak would be at 1.4 volts. Disregarding the discharge contribution from the amplifier itself on the bitline and assuming that the cell capacitance equals the bieline capacitance. this would mean that the clamped voltage would equal 2.8 volt. By lowering the clamped voltage and making sure that the clamped voltage equals the potential at the switching point of the amplifier flip-flop. the disturbance level n the storage capacitor can be reduced. In the example above.

reducing the clamp level to 2 volt would drop the noise peak to 1 volt. (VG-VT) is then 0.3 volt compared to 0.7 volt before. This educes the noise current to 18% of that in the example above.

Fig. 21 shows partly in block form an alternative to the devices discussed above. A pump 2101 is continuously running and supplies power to N-wells associated with the high voltage latches 2103 and the Vdd supply for latch 2102. Latch 2102 has an output 2109, =at is low during clamp time and high during read time.

Output 2109 supplies vdd for sense amplifiers 2104 and for word line select latcnes 2103. A recirculating shift register 2121 selects the word line to oc used. The selected late is set by the C OCX signal and reset by is complement CLOCK. Output 2109 of latch 2102 goes high on CI CLOCK for every selected word line and goes low for every LOCK. DUMMY WL is criven directly from tne normal VA (not shown ne figure). An O@@/EVEN counter 2122 is used to selectively trigger the ODD or the EVEN dummy word line latc@. A data input shift register 2108 is also shown.

The operation of the refresn cycle of the device in Fig. 21 is best further described in connection with Fig. 22.

Clamping action takes place during CLOCK. Node 2109 is low during CLOCK. so the power is turned off to all high voltage latches and co the sense amplifiers. CLOCK also pulls all bit lines to ground.

Fig. 22 shows two cases. In the first case. word line 2114 is selected and the associated bit 2113 is low. After the clamping. which brought both bit lines 2112 and 2111 to ground level.

the CLOCK signal starts the read-refresh phase. Node 2109 goes high, but its rise time is controlled by careful selection of the P-channel transistor pulling up node 2109 in relation to the load imposed by the latches connected thereto. The word line 2114. which is very long and connects to a large number of cells. has a slower slope than node 2109. The dummy word lines DUMMY WL have about the same rise time as the word lines. In this case. node 2109 moves positive until the P-channel transistors are turned on. Nodes 2111 and 2112 are pulled up in parallel except for a slight difference caused by size differences between the P-channel transistors. When nodes 2111 and 2112 reach the threshold level of the N-channel transistors. the pullup of nodes 2111 and 2112 is slowed down. If the word line had not been made high, nodes 2111 and 2112 would have reached the level of the flip-flops switching point and stayed there until the slight unbalance had caused the flip-flop to flip to one side. The word line 2114 potential is. however. increasing concurrently with node 2109. As soon as 2114 reaches the threshold level. the transistor 'n the cell connects 211 and the storage capacitor.

This slows node 2111 in relation to node 2112. As 2114 goes higher. the unbalancing current increases. When this current exceeds the build in unbalance of the flip-flop, the desired switching proceeds at an accelerated rate. The advantage of this approach is that at this relatively low voltage on node 2109. the current in the P-channel transistors is small and thereby the differential current is also small. The current from the storage capacitor need only overcome this small differential current.

The dummy word line (EVEN) goes positive at the same or slower rate than te word line. The dummy capacitor is only half as large as che storage capacitor. so the capacitor, so the capacitor at 2113 dominates.

In the second case. storage node 2116 s high. word line 2115 is selected and the ODD dummy word line is selected. The dummy capacitor in 2106 now holds back node 2:'1 and causes node 2112 to fli thigh. Node 2115 is disturbed only If 2115 at any time is more t@an one vth alcove 2112. This could happen only if the slope of 2115 is very close to that of 2109. But even at that eventuality, the negative excursion of node 2116 would be minimal compared to the case without dummy capacitors. Any charge lost by capacitor 2116 due to leakage or the potential cause discussed above will be replaced at the time wordline 2115 reaches the V +.

The voltage will be (Vdd+ - Vth).

Claims (9)

1. A method of operating a dynamic memory comprising the steps of: precharging a selected bit line in said dynamic memory to a selected value with a precharge circuit.

said bit line coupled to a sense amplifier comprising a pull up portion coupled to a selectable high voltage source and a pull down portion coupled to a low voltage source-; decoupling said bit line from said precharge circuit, whereby said bit line floats; coupling said bit line to a memory cell by bringing at least one word line to a first word line voltage level while applying a first high voltage to said pull up portion; and thereafter, applying a second, higher voltage to said pull up portion while raising a voltage on said at least one word line to a second, higher word line voltage level.

2. A method of operating a dynamic memory as recited in Claim 1 wherein said word line is brought to said first word line voltage level and maintained at said first word line voltage level for a selected period of time.

3. A method of operating a dynamic memory as recited in Claim 1 or Claim 2 wherein said step of coupling said bit lien to a memory cell further comprises the step of coupling another bit line to a dummy memory cell by applying said first word line voltage level to a dummy word line.

4. A method of operating a dynamic memory as recited in Claim 3 further comprising the step of lowering a voltage applied to said dummy word line when said voltage on said at least one word line is raised.

5. A method of operating a dynamic memory as recited in any preceding claim wherein said memory cell is coupled to a logic circuit and further comprising the step of inputting a voltage level in said memory cell to said logic circuit, whereby a selected logic function is performed.

6. A method of operating a dynamic memory as recited in any preceding claim wherein said step or precharging applies a voltage of about 1.5 v to said selected bit line.

7. A method of operating a dynamic memory as recited in any preceding claim wherein said first word line voltage level is about 1.5 v and said second word line voltage level is about 7 v.

8. A method of operating a dynamic memory as recited in any preceding claim further comprising the step of boosting a supply voltage level in said dynamic memory, and wherein: said step of applying a second word line voltage level is a step of applying said boosted voltage level; and said step of applying a second, higher voltage level to said pull up portion is a step of applying said boosted voltage level.

9. A method of operating a dynamic memory as claimed in Claim 1 and as further described hereinbefore with reference to Figures 16A and 16B.