GB2352875A - Fabrication of a field effect transistor having a recessed gate electrode - Google Patents

Abstract

The transistor gate oxide is formed by implanting gas atoms (eg using a SIMOX process) in the region between the source and drain regions. The silicon above the implantation region is used to fabricate the gate electrode. An epitaxial device layer is provided overlying the transistor, the device layer is separated from the underlying transistor by an implanted insulating layer.

Description

1 2352875 DYNAMIC RANDOM ACCESS MEMORY

BACKGROUND OF THE INVENTION

The present invention relates to the operation and manufacture of integrated circuits. More specifically, in one embodiment the invention provides improved dynamic random access memories (DRAMs), methods of operating dynamic random access memories and methods of making dynamic random access memories.

I,-, '--he attempt to increase the number of bits -.n present DRAMs, methods are sought for shrinking device dimensions while still maintaining high enough capacitance in the stora-, _ie c;.ipacitors so that data can be reliably c:torei, refreshed and read. For example, new dynamic random access memories are disclosed in Application Serial No. 08/353,788, filed December 12, 1994, and U.S. patent number 5,396,452, each of which is hereby incorporated by reference, wherein a memory cell contains separate read and write transistors.

In another example, substantial space could be saved merely by providing smaller capacitors in present DRAY-s. However, smaller capacitors are generally refreshed more often than larger capacitors, as it can be assumed that the leakage current is the same. The capacitance of a bit line is proportional to the number of bits on the bit line. -The refresh power consumption is proportional to the frequency of refresh cycles multiplied by the number of bits per bit line, where the refresh frequency is. inversely proportional to the cell capacitor size. Therefore, the space advantages of a DRAM with smaller capacitors are mitigated by the higher power consumption of the required additional refresh cycles.

From the above examples it is seen that an improved dynamic random access memory is needed, along with imprcved methods of onerating such memories and improved methods of making such memories.

2 SUKY-22Y CF T-.H-- 7N-VENTTON The present invention is directed, in one embodiment, to a memoz-v structure having shor,_ bit line segments. Each hit line segment is coupled to a separate block of memory cells and a corresponding amplifier. The bit line segments are separated by pass transistors. The amplifiers are activated in all three modes of operation: read mode, write mode and refresh mode, while the -mass transistors are enabled only in connection with data input and io output. Very small cell capacitors can be used in this configuration, making it- possible to use convent4 onal gate capacitors, but Dower consumption is not appreciably increased. Furthermore, the speed.of the memory for read and write operations is faster than present DRAMs with long cont--inuous bit lines coupled to a single amplifier.

In another aspect. of the invention a memory is provided wherein the contents of the memory can be read without any interference from ongoing writing, reading and refreshing. The memory is structured in two tiers, meaning that in addition to a first tier, a DRAM with addressing, reading, writing and refreshing, a second tier with separate addressing from the first tier is used to read the contents of the cells in the first tier.

Improved methods of forming a dynamic random access memory are further provided according to the present invention. For eXample, in one aspect' of the invention, a met'-hod is Drovided for forming a memory cell for use with programmable logic devices that must be con-trolled with relatively large capaci"Cors. in programmable logic devices? pass transistors transfer signals between locations in the device. The controlling capacitor must be significantly larger than 'he gate capacitance of the pass transistor, so that the voltage on the controlling capacitor is relatively constant during the signal transition. A multi-layer approach to forming the memory allows the controlling capacitor to be located underneath the pass transistors and the memory cell transistors.

3 A further understanding of the nature and advantages of the inventions herein may be realized by reference to the remaining portions of the specification and the attached drawings. 5

BRIEF DESCRIPTION OF THE DRAWINGS

Figs. 1A-B show simplified block diagrams of a segmented bit line memory.

Figs. 2A-E show circuit and timing diagrams of latcbing and inverting amplifier circuits.

Figs. 3A-J show circuit diagrams of clamp circuits according to the present invention.

Figs. 4A-C show circuit and timing diag.-ams of a two inverter amplifier.

Figs. 5A-B show simplified block diagrams of control logic circuitry for the memory of the present invention.

Figs. 6A-C show circuit and timing diagrams of a single inverter amplifier.

Figs. 7A-E show circuit and timing diagrams of alternate embodiments of amplifier and control circuits according to the present invention.

Figs. BA-H show circuit and timing diagrams of a two-tier memory.

Figs. 9A-F show process flow and layout diagrams for wafer preparation of the memory according to the present invention.

Figs. 10A-E show alternate process flow and layout diagrams for wafer preparation of the memory according to the present invention.

Figs. 11A-C show layout diagrams of a shared bit line memory.

Figs. 12A-M show circuit and timing diagrams for refresh read and write cycles in the memory of the present invention.

Figs. 13A-G show circuit and timing diagrams of control circuitry for the memory according to the present invention.

4 Figs. 14A-C show the disab-ling of segments above selected segments.

Fig. 15 shows 'L-,R-H cells with separate read and write word lines and a clamDed amulifier.

Figs. 16A-D show memory cells according to the Dresent- invention with a full latch amplifier in each segment.

Figs. 17A-M show the result of simulations using a 0. 5u NMOS Drocess.

Figs. 18A-D show how an ongoing refresh cycle can be interrupted at any time during the cycle without destroying the stored data.

Figs. -atDRAM-having- more- than one stbrage capacitor associated with a single pair of read and write transistors.

DESCRIPTION OF THE PREFERRED EMBODIMENT

Fig. 1A shows a simplified block diagram of a memory including segmented bit lines 102, made up of a read bit line BLR'and a write bit line BLW. Each bit line segment 104 is coupled to a separate block of memory cells 106 and a corresponding amplifier 108. Bit line segments 104 are separated. by pass transistors 110 in li-ne with bit lins 101'2 Amplifiers 108 are activated in all three modes of operation-.

read mode, write mode and refresh -mode. Pass tran--s'istor7i, -110 are enabled only in connection with data input and oiitput.

Address lines 112 select the particular memory calls 2.06 to -'De accessed. A word li-na decode block 114-4 is usad tz-, selac't a desired word line 116, -while segment decode blocN- 1-1-8 to selected a desired bit line segment 104.

Low order address bits select one of the word lines 116 in the segment selected by a segment decoder. The hiaher order address bits select a bit line segment 104. In a refresh mode, the low order bits select one Word in each segment, and the segment decoder is bypassed to enable all segments. It should be noted that in an alternative embodiment, rather than separating the address lines into low and high order bits, a single set of multiplexed address lines may be implemented. The bit lines in the selected segments that are no= ally clamped to a clamp level off Vdd/2 may be released to f1cat during a period when the clamp is off - L - - k.- The selected word line 116 is then broughl: to the first ramp level, and a dummy current is turned on for the selected bit line segment 104. The voltage on the read bit line in the segment changes quickly due to the low bit line capacitance. The voltage displacement due to the cell current can actually be of the same order of magnitude as the final displacement caused by the amplifier, when its power is turned on later in 3.0 the refresh cycle. The amplifier 108 will therefore use mostof its power dissipation to bring the bit lines to the power rails. Both latch type and the inverter type amplifiers, each of which will be described further herein, can be used. As described earlier, the ramp is brought to its high level for a L5 moment, increasing voltage on word line 116, when the high bit line has reached Vdd, completing the refresh operation before power again is turned off. In the very end of the cycle the clamp signal is applied again.

In order to read data from memory 100, all bit line segments 104 are initially disconnected, and the clamp input is brought low on all segments. This permits bit lines 102 in all segments to f loat. only one word line 116 in only one bit line segment 104 is selected in the read mode. When the read cvcle has reached a point where amplifier 108 has been turned on and amplified the signal slightly, the bit line segments 104 on one bit line 102 are connected. The interconnected segment lines initially appear as delay lines consisting of RC elements. The capacitance C corresponds to the capacitance of the individual bit line segments 104, while the resistance R corresponds to the impedance of the coupling transistor. This is however not a regular delay line, as each segment 104 also has its own amplifier 108. This allows a signal to proceed,rom. segment to segment very quickly while also being amplified. in a typical memory configuration, a column decoder would selec:-'L--. the columns to be the subjec-%'-- of the W7-4ting reading or --1 of dat-a. The non selected columns will s-%'--ill be addressed by '"he word line 116 selected by the word line decoder 11-4- When the amDlitude on '"he outmut se=ant 6 has reached the rails, the seg-ment is connected by the column decoder to the output reg4ster of --he memory.

in a write oDeration, data is connected via the column decoder to the. bottom segments of the selected columns.

-he write circuit forces the bottom segment separation to he much larger than the separat-ion of the segmen-t-s in the addressed and -read segment at the time of coupling the segments together. If the read data is different than the data to be written, two signals of different polarity will proceed in opposite directions down the "delay line". The signal with the largest amplitude will determine the final polarity of the addressed segment.

In the refresh mode, the segments remain disconnected during the full read-restore cycle. As mentioned above, one word 14-ne in each segment is selected by the low - -s in the word line decoders and the seg-ment order address b4 4" decoder is bypassed so that all segments are activated the same way. The dummy current circuit is also activated in all segments. The control circuit in Fig. 1B clearly shows how the refresh mode dif fers from the read and write modes in the word line decoding. It should be noted that if the amplifier 108 in each segment 104 consists only of an inverter, the time required for refresh may be longer than the time required for reading and writing. The read bit line in each segment is driven only by the difference of celll.current and dummy current. This current is increased significantly at the end of the cycle, but the signal on the read bit line is not enhanced by amplifier 108. In reading and writing, on the other hand, segments 104 are interconnected, so that the read bit line on the f irst segment is connected to the write bit l1ne of the second segment and so on. This means that the output of one segment amplifier drives the input to the next segment amplifier in the read and write modes. The delay through the "delay line" must of course be taken into account when comparing the timing requirements. It is preferred that all modes have the same timing, as this simplifies the global control circuits.

7 Therefore, very small cell capa C4to_S can be used in this configuration, making it possible to use conventional ga"Ce capacitors. The speed of the memory f or reading and writing is faster than the same type of memory with long continuous hit lines.

The addressing and control of memory 100 is shown in Fig. 1B. Memory sequence generator 140 receives memory control signals such as Memory Request and Read/Write Enable and Drovides amnlifFier control signals 142. Amplifier contrcl signal - s 142, combined with high order segment address bits 112, are input to segment decoder control block 118 (Fig.!A) to control a selected one of memory segments 104. Refresh counter 144 initiates a refresh cycle when necessaij. The output of refresh counter 144, combined with low order word is address bits, are input to word line decoder block 114 (Fig. 1A) to select a particular word line in memory 'i.00. Memory sequence generator 140 also generates additional control signals, such as Copl 1/0, Block Address Enable and 1/0 Latch Clamp, that are input to a column decode block (not shown) that accesses a desired column of memory 100 for reading, writing or refreshing.

The "-,me of amplifier 108 used depends on the applIcation for memory 100. For the highest possible speed, where disturbance on the cell capacitor voltage is permitted to be high, a basic latching amplifier 200 coupled to memory cells 202, as shown in Fig. 2A, may be used. Memory cell 202 includes a storage capacitor 204, a write transistor 206 and a read transistor 208. 2Umplifier 200 includes clamping transistors 210 and cross-coupled inverters 212. The corresponding timing diagram is shown in Fig. 2B. It should he noted that for highest speed with a given transistor size in memory cells 202, a higher initial cell current =-' be obtained if the threshold voltage on write transistor 206 is increased. There is of course a great advantage to using a standard CMOS process for the memory, especially _Lfr -mixed with logic. The seg--mented bit line approach and small cell capacitors 204 are preferred to maintain the highes-L--possible 8 memrory speed. in that case, the segments would be cou--led tZogetzher just after amr)lifier!08 has been nc-wered uD.

Tn oDerat4Lon of -'Che memory shown in Fig. 2A, at the end of a memory cycle, the power to the ampli-fier is turned off and VC1AYP is turned on, clamping the bit lines. to the clamp voltage level (VCLAMPL), which is typically about half of the Vdd level At '"he beginning of a new memory cycle, VCIAXP is brought low, allowing the bit lines BLR and BLW to f 1 o a VWL is then brought to an intermediate level that is at least high enough to draw current from the read bit line BLR if the voltage on cell capacit-or 204 is high (a stored ONE). If, on the other hand, the voltage on cell capacitor 204 is low (a stored ZERO), an excessively high signal level on VWL will charge the cell capacitor through write transistor 206 during the read portion of the cycle. A slight;ncrease of the cell voltage while reading a stored ZERO from cell capacitor 204 is acceptable since the resulting read current is minimal comDared to the cell current while reading a stored ONE. At the same time the first word line voltage is applied, a negative going voltage is also applied to dummy transistor 214, causing it to feed a current to BLR which is deL4igned to be about equal to the average current for reading a stored ONE or for reading a stored ZERO from cell capacitor 204. As shown in Fig. 2B, BIR is pulled in the positive direction when a stored ZERO is read because the dummy current dominates. While reading a stored ONE, however, the cell current dominates and pulls BLR in the negative direction. As soon as the voltage difference between the bit lines is sufficient for reliable operation, the supply -voltages VDDA and VSSA for latching amplifier 212 are applied. Bit lines BLR and BLW wLll then quickly move to their respective rail voltages. Once each bit line is close to its final voltage, the vol-age on the word I-Lne VWL is increased to at leas-'L--- a level of VDD. A previously stored ONE will now be refreshed to approximately one threshold voltage less than VWL, while a previously stored ZERO will be restored to 0 V.

1,f speed is not the primary goal and -t-he disturbance level on the cell capacitors 204 is allowed to be high, the 9 basic inverting ampli-fEier 220 coupled to cells 202, as shown in Fig. 2C, may be used. The operation of the memory shown in Fig. 2C is similar to that of the memory shown in Fig. 2A. Amplifier 220 includes clamping transistor 210 and a simple inverter 222. The clamping to a common reference level, which is chosen to be close to the switching point of the amplifier, is not very accurate, as the switching point of the amplifiers varies with changes in the device parameters. The timing must thlerefore be adeciaate to allow the cell current to disDlace 14 1 0 the read bit -Lne BLR at least as much as the error be-fcre, since the bit line voltages represent the actually stored data. The power to the inverter amplifier 220 should not be turned an before the worst case clamping error has been comDensated. This will prevent the write bit line BLW from temporarily going to a ground level with the word line on, which would cause a stored onE. to be discharged to a zero. A stored zero on the other hand would not be affected if the write bit line inadvertently went to the high level, while the word line voltage is still at its lower level. The preferred method is to wait a sufficient time until a potential clamping error has been compensated for. The amplifier will then consistently move the write bit line in the correct direction. As discussed earlier, power consumption can be minimized if power is turned on even later in the read mode, so that read bit line BLR has r.oved closer to a rail voltage. If the inverting amplifier is used in connection with segmented bit lines, the coupling can be made only when the clamping error has been compensated and the signal has been amplified some-what. Fig. 2D shows the timing for amplifier 220 in Fig. 2C, and Fig. 2E indicates the amplifier's performance as part of a segmented bit line configuration. Note in the latC-er case how the partially amplified signal on both bit, nes BLR and BIM is temporarily reduced when the adjacent segment is connected. Again, the read bit line BLIR on the read seg-ment is connected to the write line BLW on the next segment. The sagments together act as a plurality of latched amnlifiers. it is imnortant t-hat all amplifiers are nowered un before the se-::Tments are coui--led, as this guarant-e-es a gradual amplification of the initially reduced signal before it progresses down the full bit line. Without ampl i f0i cation, the signal on the read bit line BER of the read segment would be reduced to a small fraction of its original value. For a memory structure where a bit line is made up of eight segments, then, the signal would be reduced by a factor of eight. This signal is too weak to overcome the clamping error on some of the segments.

Compensation of Clamping Error In applications where the cell current is low, yet fast memory cycle times are still needed, it is desirable tn minimize the clamping error. Figures 3A through 3G illustrate the different clamping devices and methods. Fig. 3A shows a preferred structu'l7e with clamping transistors 302 and 304, which was also- shown in Figs. 2A and 2C. Both bit lines will reach the reference level VCLAMPL, while VCLAMP is high, but when VCLAMP goes low, the gates of the transistors 302, 304 will couple a small negative charge to each of the bit lines.

The voltage drop on the bit lines can easily be as much as 100 mV, if a short clamp time is desired. To achieve a short clamp time, the clamp transistors 302, 304 must be relatively large. Segmented bit lines, with their lower capacitance, will have smaller clamp transistors than fulllength, unsegmented bit lines, but the ratio of gate capacitance to bit line capacitance is the same and therefore the result is a comparable disturbance. The latched amplifier 200 (Fig. 2A) is relatively immune to the clamping offsets that appears equally on both bit lines. Latched amplifier 200 suffers from a non-correctable offset error due to device parameter variations. The two cross-coupled inverters 212 may have different switching points. One of the bit lines will therefore be moving at a different rate after clamping than the other bit line. The cell current must be larger than the built-in offset current for a correct reading of the selected cell. If used in a segmented bit line configuration, there is a risk that one or more of the non-selected segments may start to switch on their own before the segments are connected, 11 which could result in a read error. The clamn circuit in -or 306 Fig. 3B addresses 'his possibility by adding transist that directly shorts the bit lines during clamping.

The inverting amplifie--r 220 (Fig. 2C) is sensitive both to the difference between its switching point and. the reference VCLAMPL and to the effect of turning off VCIAMP. Instead of waiting for the read bit line to overcome the worst case clamp error, another approach may be taken, as shown in Fig. 3C. VCL-7d!P1 again connec-16--s both hit lines to a common reference voltage VCLAMPL while power is off, but VCIAMPI is turned of f as soon as Dower is turned on. VCLAMP2, on the other hand, turns on at the same time as VCIAMP1, but stays on longez. When power is turned on, the bit lines are at the potential of VCL2-MPL, which may be different than the switching point of the inverter. VCLAMP2 shorts the input and the output of the inverter, so that the bit lines are moved to the switching point, thus putting the inverter in full balance before the cell current is applied to the read bit line. Note that turning off VCLAMP2 causes a negative displacement of the bit lines. The disDlacement on the read bit line (the inverter input) is most critical and must be compensated by coupling a positive charge to the read bi-L- line. The clamp shown in Fig. 3D accomplishes that objective. HP.re, the connection to the reference voltage is done with a P-channel transistor 308. VCLAMPB goes positive at essent-ially the same time that VCLA-MP goes low. The P-channel transistor 308 has a Cively high on resistance compared to the N-channel relat clamping transistor 3o6 and the impedance in the inverter assoc-, ated with searching the switching pcint centering. The --ain the bit purpose of the P-channel transistor 308 is to maint lines close to the switching point at for a longer stand-by perioc. Since the switching point is set to half the Vdd level (Vdd/2), and the bit lines before clamping are sat to he rail volt-ages, the potential of the bit lines a-Eter 3 -5 clamnina will therefore be close to Vdd/2. The width and -or 308 can be chosen such that length of the P-channel transist the requirement- for both high impedance and the recraired coupling of charge to the read b--'-'%-- Ifine BI.R can be met..

12 Ficr. 3E illustrates the addition of compensating capacitors 310 and 312 between VCLAMPB and bit lines BLR and BLW that are chosen to fully compensate at turnoff of the clamp transistor 308. In Fig. 3F, VCLAMP1 holds bit lines BLR and BLW at the reference level in standby mode, While VCLAMP2 remains high after power has been turned on to inverter 220 (Fig. 2C). The charge transfer to the bit lines, when CLA-MP2 is turned off, is compensated by VCIAMPB going positive. The compensation is necessary on the inverter input, but is less critical on its output, so the compensation can be limited to only one capacitor on the read bit line.

As will be discussed herein, the advantageous features of inverting amplifier 220 and segmented bit lines in Fig. 1A can also be applied to conventional DRAM cells. In 1 that case, the DRAM cells are connected to both bit lines. This type of clamp circuit is shown in Fig. 3G. The associated amplifier includes two separately powered inverters and two separately driven compensation capacitors 314 and 316. A similar amplifier arrangement can also be used for the DRAM cells of this invention, except that a common drive signal for the compensating capacitors shown in Fig. 3F can be used. The compensating capacitors used in a conventional DRAM may se--ve two purposes. Firstl they supply the charge to compensate for the negative charge transferred when VCLAMP2 goes negative, but the size of the capacitor is a li:.ttle larger than required for that compensation. The added charge pre-biases the bit line to which the read cell is connected, so that reading a cell will give the same absolute voltage difference between the read bit line and the switching point for both a charged and a discharged cell capacitor. In reading and writing data in a segmented bit line configuration, the segments that are not addressed must have their read bit lines BLR very close to the switching point not to cause errors. The compensating charge transfer for these segments, must therefore be less than for the selected segment. This can be accomplished by using a lower amplitude on the compensating voltage or, of course, a separate compensating capacitor can be used to add the dummy charge on the selected segment.

13 The amplitude of the clamp signal for all the cases discussed above is the full Vdd voltage, resulting in a large capacitive coupling when the clamp signal is turned off. in the initial time period after clamping, the bit lines will still be close to the clamping level (Vdd/2). This means that the clamp signal need not initially go more negative than to ( (Vdd/2) +Vth). As the cell current proceeds to separate the bit lines, the clamp signal must track the negative-going bit line. The arrangement in Fig. 3H shows such an conficraration.

Here, when VCIAMP is turned on, transistor 318 lifts the gate of transistor 320 to (Vdd-Vth), which clamps the bit. lines together. The transistors 322 and 324 act as diodes and are designed as relatively weak transistors in order to limit how much they can pull the clamped bit lines positive while VCLAm-P is is high. Transistors 322 and 324 are actually designed so that the positive excursion during the time of clamping is exactly as much as the negative capacitive coupling when VCLAMP goes low. When transistor 318 is turned off, transistors 322, 324 proceed to Dull the gate of clamp transistor 320 toward the lowest bit line voltage plus the threshold voltage of transistor 322 and transistor 324. With a normal threshold, this tends to hold a slight clamping for a short period, reducing the gain of the inverter for small cell currents. By terminating the clamp signal early or by using lower threshold voli.,-'age on 322 and 324, this problem can be eliminated. When the bit lines eventually move, the low-going bit line will continue to keep the gate of the clamp transistor 320 jusIC below its threshold voltage.

In several of the amplifiers, the assumption has been made that t he amplifier power is permanen-Itly connected. This means that the bit lines are close to Vdd or Vss between cycles. A clamping operation will then have to swing the bit lines as much as Vdd/2 ito the swiiCching point of the amplifie--. The time required for the clamping operation is in this case sligh-tly longer than when the hit lines start cut close to the clamD level. The clamping circuit as shown in F4a-. 31 is used for t-is pur-cose and consists of one N-cnannel clan.z 326 and two caDacitors, here shown, as 14 transistors 328 and 330 with both source and drain connected to the associated bit line, for compensation of the turnoff disturbance.

Fig. 3J shows a single N-channel clamping transistor 332, used in connection with clamping under power. Amplifier power is selectively turned on and off during a memory cycle.

In cycles repeating frequently, the bit lines start out very close to the switching point. At the end of each cycle the two bit lines are clamped together after having been close to the opposite rail voltages. The refresh frequency must satisfy the requirements for the small cell capacitors, so larger bit line capacitances will discharge very little between cycles.

The amplifier's VddA and VssA nodes are both connected to the reference level Vdd/2, which is the nominal switching point level. Any leakage through the transistors would go to this level. only junction leakage could cause the bit lines to drift away from the reference level. Theinverter transistors would however limit the drift to one threshold away from the reference level in either direction.

Starting from this worst case situation, the adjustment toward the switching point afte-,- applying power to the amplifier follows the table below. The inverter has a W/L ratio for the N and P transistors of 1.2/.6 and 2.4/.6, respectively, while the bit line capacitances are 100 fF (100E-15).

Time after power an VBLR - Vref ns mv 0 -600 2 -376 3 -163 4 - 68 - 19 6 5 7 - 1 8 0 Turning off a clamp transistor of with W/L of 2.4/-6 introduces a disturbance of -8OmV. The disturbance must be considered in relation to the cell current. The rate of voltage change on the read bit line is lomv/(pAns) for the bit is line capacitance of 100 fF. A dif 41eren'Lial cell current (!cell - Idummy) of 5AA would change the bit line voltage at a rate of 50mv/ns. Turning off the clamp after Sns, with the cell current on, would result in full compensation 7ns after the start of the cycle, overcoming 80mv plus 19mv. At lower cell currents the clamping will be turned off about ins later, and it would take longer to compensate for the offset caused by the clamp transistor. At a differential cell culrren"C of 1AA, the total time to comnensation wculd be about 14ns. The Polar_44[_--v of the disturbance on the read bit line is such that the write bit line for the lower cell currents may go positive to near the Vdd level. This would not cause any disturbance on the addressed segment. If however the read bit line starts out at Vref + 600mV and the clamp transistor turns off while the read bit line is still more than 80mV above the reference, the write.bit line would go negative.

Only if the clamp is turned of f very early could the write bit line go below the word line voltage, thus turning on the write transistor and write a zero in a cell that was a one.

This case is very remote as the word line voltage would by design not be turned on that early.

A more serious problem occurs in reading and writing data in a segmented bit line configuration. The d3sturbance introduced by turning off the clamp vol tage would remain on -the non-selected segments. The current to the Write bit line caused by an 80mv differential voltage on the read bit line is approximately 12pA, with the inverter dimension as discussed above. The bit lines would be displaced at a rate of 1203nV/ns and thus completely override the contribution from the selected segment except for rather high cell currents. A cell current of -IpA will change the read bit line voltage at a rate of!Omv/ns,. The displacement on the write bit line after T nS follows the equation Vdispi = T cell (pA) 10 TSquared/ 2mV. A cell current of 10,ua would after 2ns have displaced the read bit to +120mV and the write bit line to -120-mv, while at the same -;C. ime the bilt: lines on the ncn-selected segmants would have -'ive. A simulat4on has shcwn that a cell moved 240m-V positcurrent o-f at least 40LLk would be racffaired to onerate an I - - 16 8-segment memory, ass,.,_m4ng bit!Jne capaci a n e s of loo f=. Bv using different- clamping 14_-nes -Fcselected and non. selected segments, the -problem is Inighly reduced. If the clamping on --ed just before the signal the non selected segment is "Cerminat L - coupling the segments together is turned on, the voltage between the non- selected segments would be approximately 100 mv. Allowing the bit lines in the selected segment to be displaced more than k100 mV before coupling, where k is the number of seaments, would yield more reliable operation. Some of the amnlifiers shown may use a second inverter in each segment which can be selectively engaged. Allowing the second inverter to contribute to the anDl 4 - Lfication in 'he selected segment before the coupling satisfies the above-stated requirement.

It was assumed above that the power to the amp 14 f ie-s L was selectively turned on and of -LE during the memory cycles. The method of clamping to the switching point of -the inverters used in some amplifier configurations, however, makes it natural to onerate the amplifiers under 'Lull power all the time. This means that in standby made, the bit lines will always be at or close to the rail voltages. The clamping will, when used in connection with an inverter, bring the bit lines to the inverter switching point against the transistor that is in the holding mode. The final adjustment to the switching point level will therefore take somewhat longer than if the bit lines start- ou-C. already close to the switching point as described before.

Fig. 4A illustrates an amplifier 400 _-Lntended for segmented bit lines. The basic inverting amplifier consists of transistors 402 and 404 that make up the first inverter, while the second inverter with transistors 406 and 408 may be added for additional amplification later in the memory cycle.

Transisto, r 410, when turned on by the signal VINV2 will connect the outnut of the second inverter to the input of the first.

The clamping is done by turning on the clamping transistor 412, thus shorting '"he bit lines BLR and BLW and adjusting to the switching point level of the first inverter. The selected word lines on the selected 5egments are brought to the reading level 17 at essentiallv the saime -1-4 me as VCLAM_'P goes low. The read bJt line BLR at this time has a predictable offset error as discussed above, due to incomplete adjustment to the switchin'g point during clamping and the capacitive coupling from VCL-AMP, when it goes negative. With only the first inverter connected (VINV2 low), the cell current will both overcome the offset and displace the bit lines sufficiently for a correct refresh level. If limited to only the first inverter and no couDlina between segments, this would be a relatively slow operation of high power consumption. in the refresh mode, as discussed above, the segment--s remain separated and refresh cells on their selected word lines. To speed up the refresh cycle, the second inverter is connected at the time when the bit lines are seuarated on the urder of 100 m-V. The amplifier now acts as a conventional latching amplifier. When the memory is used in read or write mode, the offset, on non-selected segment must still be considered. At the time of connecting the segments, the signals on the non-selec-Ced segments must be an order of magnitude smaller than the signal on the selected segment. As shown in the timing diagrams of Frig. 4B, a different approach is used that guarantees a suitable ratio between the signals. For the selected segment the clamping takes place firs-it.". in the cycle, allowing the signal to grow, driven by the first inverter. Then the second inverter is connected and the higher am-olif- ication proceeds for some time until the segments are interconnected by the signal VCOUPL. A temporary drop in the signal is followed by the mutual amplification from the crosscoupled segment inverters. TI-14tially only the selected segment has its second inverter cont--ributing. The non-selected segments as shown in the lower timing diagram all have a delayed, clamp signal that is turned off just before VCOUPL is turned on. Therefore, the only signal on the bit lines of the non-selec-tad seaments is the offset signal caused by the turnoff of the clamp signal. once the coupling has been establisIned, the second inver-ters on the non-selectEed segments are turned on, introducing additional sneed to the read o-Deration. The lbottozi segments in the selected columns may use sanilpxado ES -5Tj ul umoqs quau5as utoqqoq auq aol oT,6oT aTql r _U'PaB21P atT4 UO UMO'qs 4OU ST-l-'UBTS UOUMOO q PaTTOaqUO0 BUTMTq a-qq. TqqTm 'quau6as uo2a, uT paqPaa,ua,5;--q ao:pojq aq-4 aog uoillmoo aq-,Pm qoiTqm 'paTjdd2 csjP sT dwvdA TeuBTs duiPa V aqPTadozddp a.caqm sasaqquaxed uT umoqs ax2 sTpuBTs asaqq aci suoTq2TAaaqqp a-q,l (am-rjuo,=cfl qMTj uo quBaano Fm=c PUP (aulTluoqm) auTl UO auTa PIOM ' (aaTIAUIT9SUON) auTl UO aa4a9AUi quauLSaS PantlOaT9s-uON (a=IAU:[TaS) am-11 UO aa4-19AUI 4uauLS;S P;9-40aTaS lem7m. dmsTD a-412rI lauTj da-elo AT.XPa 9Xe Sa=519 aTT4 UT STPUbTs CC BUTMT4 P-aTTdd-e a-ql -panoalas aq oq paaapTsuoo iaxe squautSas TTP 9POul aTITO uT ':0010 BuTuuva aaag: P,q paqsanbsa sT qoTqm lucTqPaado qsaagaa aq4 ul -sn_,u9uL6as Pa4391as-ucu PUP P9743aTas aO; AT-4tl!Daa99TP ATqqbTIS P9Sn a:ZP PlUP saaTTOX4UOO 4LIaulblas ITle oq paqrioa eap sltubTs aqj XooTq Aaomam aTqq 04 PaTTddp cz PU12 Pa4l2aaUab 9aP SjPuBTs BuTmT4 aqPTadcadd-2 aqj -aouanbas BuTmT-4 -2 a4PT4TuT PUP (a-4TIM P aO PPal P) uoiqPaadc uaATb P PUP aOMaUI aq4 UT quam6as uaArB P 43aTas 04 aPOO ssaZPP2 Pu'2 uOT40ung: BuTmOOuT aq4 aPcOaP tOS XOOTq aapoosa quauT-SaS PUP Zog XOOlq lap=aa allTU PIOM -T-eubls aTqPua dTqo ao aTqPug OZ Xools.6ui-4dnaaaqUT UP aO 4S9nbaZ qSa_T;a_T PaMT:: A-Eal-41?:rPdas 2 Aq Pa4PT4TUT 9q keM 913.13 AaOM9UI V -aUTT 4Tcl aq4;o 4nO PUP UT P-4PP 5UTaa3;SU2a4 aOg P9Sn ST -42144 'qq -BTZ UT UMOqS lau-cT --Fq e go quauLSas moq4oq zqq qdaoxa 'Vg -LTa uT umoi4s OTBOT aux-2s aqq aApq sluaulb9s TIV -.=G'c;TTdui-t2 pa4PTOOSSe UP ST 5uTTTOa4uO3 -109 4uauL69s qOP9 20; P9=baa OTBOT aq-4 mOT4s qG PUP Vg SqarL6T_z 's4uGu'59s Pa4OIDTas-uOu PUP Pa43@Tas aOg P9.zTr'baa -4 -4u@aa,;-TP 42q4 PauOT.

sT BuTul-L quam ua;aq sp-q ql -paqpuTuiiTa aq uPo aouPqanq_rsTp BuTdmVTO aq4 -c -soul lllm STqq UT UBTsap aqq BUTZTUITqdo,g 'tZtv aO-4STSUPZq 0 1 d puL- oj,7 N qqoq aPnTOuT saO4sTsu2aq BuTTdrLoO P-U'72 aOqSTSuleznt, d PUP ZT, lOqsTsu-Ra4 X qn-rOq an.rITOuT saO4sTsuPz-,.

BuTduipTo eqq IoZt, _-aTgTTdmP ui -zaTTTjdmP aaaasp_uT omq aq- ZrO UOln'2T.ZPA P ST 4vqq OZt aaT_,rTT&a:- S,OTqs Ot' -BTJ -BU74TaM 40aa_-00 0-4 lquau6es Q 2 U10-4-40q atl L:P'qn IialqaO qUaM59S PanloaTaS -2 UO TPUJSTS U_ q4 aaLa-2T FT-+luaTOT-rz7vns ST BUTTdnoo;o alLT-C-4 aT'- q2 2USUL.5aS U-IC-4- 40- aqq qrp -L2,_r.L-rs aqq -,2,q- os TPt'-S'rs aq-c7,1A. au,:; Tdd-2 o- aToio + _r _r Pea-- TTn--r aTqa BuTnaTuiad 'apomm an -am u! T-,ULTS d=_eTO Al-ZPa a-qn 8 T 19 differently than the other segments during the write operation. The enabling of word line and dummy currentt is the same for all segments, but the clamp signal and the connection clt' the second inverter are always early for the bottom segment -1-n the write mode. This ensures that the written data gives a dominating signal to the connected segments. In write mode, input data is coupled to BLR of the bottom segment; in read mode, output data is cou-Dled to BLW of the bottom segment.

Fig. 6A illustrates a memory 600 having two segments 602 and 604 where only one inverter 606 and 608, respectively, in each segment operates under fixed power. Read bit line BLR from segment 602 is coupled to write bit line BLW of the next segment 604 through pass transistors 610 and 612 when VCOUPL is high. Fig. 6B shows the relative timing between selected and non-selected segments. In refresh mode, in a preferred embodinent, all segments are considered selected and the coupling signal remains low during the cycle. Alternatively, the segments may cooperate in pairs, resulting in the selection of every other segment. Fig. 6C shows how the selection of segment for refresh is controlled by the odd/even bit. This decoding may be common for all segments or local as shown. The coupling signal is divided in two. one signal is used only in the refresh mode, while the other is used in read and write modes. The advantage in pairing the segments in refresh mode is that the time and power consumption is reduced due to the higher amplification for the two cooperating inverters. The structure also makes it possible to pre-amplify the signal at reading and writing, before interconnecting all segments. The coupling signal interconnecting the two seg-ments in the pair is made subject to early selection if one of the segments is selected. This increases the amplification of the pair of signals. When a dominating signal amplitude has been reached, all remaining coupling signals are activated. For a preferred margin, clamping is perf'-formed at three different times. The earliest clamD time is for the selected segment in a sea-ment pair, the next clamp time is for the non-selec[--ed segment in '"he segment pair and the third time _J_s for all remaining s ecrment-s.

0 Conventional DRA:,s and. Se.-ment-ed BLt__Lnes Conventional DR_'-_Ms also aain in sneed and marains b%7 using the segmen't"ed bit. line approach. The shorter b4-- 14_eS --ance -11--ha- gives used in conventional DRAMs have a lower canaci 4.

stronger signal for given cell capacitor sizes. This approach actually permits the use of smaller cell capacitors and higher refresh frecmencies, as discussed for the new cell type above. The same methods for minimizing the disturbance caused by the clamping operation apply to conventional DRAY-s. There is, howev-er, a slight difference in the amz)l-ifier design. The cells in a folded bit. line configuration may be alternately connected to one or the other of the two bit lines. All word lines with even addresses may, for instance, be selecting cells on one bit line, while all word lines with odd addresses would select cells on the other bit line. In the simplest amplifier discussed earlier having only one inverter, the coupling between segments must complete the amplifier so that the inverter output of one segment connects to the input of the inverter in the next segment. Each segment must therefore have two inverters, each selectively used according to which bit line is addressed.

Fig. 7A shows memory cell 700 including an amplifier where while reading cell 701 power is first turned on to inverter 702, whose input is coupled to read bit line BLR holding the cell to be addressed by the ' word line. The bit lines are clamped while power is on, adjusting to the switching point of inverter 702. Due to canacitive couplLng when VCIAAY1P goes low, tHe displacement on read bit line BLR is compensated by VCOMPODD going positive, as shown in Fig. 7B. The second inverter 704 in a selected segment may be powered up before the segments are coupled together by the VCOUPL signal. For reading and writing operations, this gives a large enough Si gnal to overcome the combined error signals from the non-selected segments as discussed above. The non-selected segments will typically have the bit lines clamped late on the primary inverter 702 so that at time of coupling a very smal! signal has developed. The secondary inverter in non-selected segments is powe-red up only after the segments have been 21 connected. Only in the refresh mode are both inverters 702 and 704 used the same way in all segments. The same principles regarding the inconvenience of turning power on and of-f applies to conventional DRAMs, so the same approach discussed for Figs.

4A and 4B above can also be used here. The only difference is that, when used in connection with a fixed Vdd supply, the order of connection of the inverter amplifiers is determined on which bit line holds the addressed cell.

Fig. 7C shows a memory 710 including this type of amplifier 712, with capacitive compensation cf the disturbance a'" the termination of the clamping. The first inverter 714, with its input connected to the f irst bit line BL1 has its output connected to the second b.it line BL2 when VINVEVEN goes high, while the second inverter 716 connects its output to the first bit line BLI at the time when VINVODD goes high. If, for example, the input of inverter 714 is coupled to the bit line of the selected cell, then inverter 714 is activated at the start of the memory cycle and is used in the clamping operation, which adjusts the bit line voltage to become close to the switching point of the inverter. The second inverter 716 is then activated late in the cycle, when the output from the first inverter 714 is much larger than any offset error due to device variation between the first and the second inverter. For selected segments, the second inverter 716 is activated some time before the.signal VCOUPL is applied. This signal connects the segment to the neighboring segments for read and write operations only. In the refresh mode, all segments are selected and the read restore oneration is completed with help of both inverters. The VCOUPL signal remains low during the refresh operation. In the read and write made, as described above, the second inverter in the non-selected segments is turned on after VCOUPL has connected the segments. Some time is allowed to permit "he first inverters in all the non-selected segments, in cooperation with the two inverters in a selected segmen't, to amplify the temporarily reduced signal enoug.n. to domina"Ce over the coupling transient when the seccnd inverters are activated. As indicated in Fig. 7C the succeeding segment mizrrors the previ6us segment, so that an 22 even addressed cell -z's located on -'I---he right bit 14ne -L i f the even cell on the -,Drevious segment was located on the left. Likewise the inverter direction is reversed Ln the succeeding segment (not shown in the f igure). The timing rela tions between the signals for selected and non selected secrments are shown in Fig. 7D.

Fig. 7E shows the segment control logic for memories 700 and 710 discussed above. The segment control logic of Fig. 7E is similar to that shown in Figs. 5A, 5B and 6C above, e--cep-IL-- for the separate control of odd and even inverter and clamp compensation.

Two Tier Memory The DRAM of the present invention by its nature makes it nossible to design a memory where the content of the memory can be read without any interference from ongoing writing, reading and refreshing. The memory is structured in "two tiers," meaning that in addition to 'he first tier, a regular DRAM of the present invention with normal addressing, reading, writing and refreshing, a second tier with separate addressing is used to read the contents of the cells in the f irst tier.

Fig. 8A shows a tvo tier memor,%, 800 with first tier memory cell 801. The DRAM cell, which includes transistors 802, 804 and 806 with memory capacitor C1, is addressed by its word line WLi and is treated as part of ' an independent memory in the reading, writing and refreshing operations. As discussed earlier, one of the features of '"his DRAM is that the stored data in the memory capacitor C1 is disturbed very lit-t-le during the operation of the memory. As a matter of fact, the high state need not be disturbed at all, while the disturbance in the low state (the voltage on C1 close to zero) can be limited to around Vth. For higher operating speeds a slightly higher disturbance is acceptable- The first tier memory 80! can have any of the features discussed above, including being 35 of the segmented bi-L--- line t-y-De.

In the second tier memory cell 810, transistor 812 is used to sense the state of C1 and transistor 814 acts as the read transistor. The second tier cell 810 is addressed by 23 WLTi, by applying a high signal thereon. This connects second tier cell 810 to the bit line BLT, which in turn connects to a sense amplifier 816. Amplifier 816 may, merely by way of example, be a single-ended amplifier, as the signal amplitude is such that folded bit lines are not necessary for common mode noise suppression. However, folded bit line configurations 820 and 830, such as those shown in Figs. 8B and 8C, respectively, are also possible. Fig. 8D shows a pair of memory cells 842 and 844, each having first and the second tier cells, 846, 848 and 850, 352, respectively. The first and second tier cells share a common ground bus GND. The connections to the bit lines are common for neighboring primary cells. The second tier cells have separate connections to BLT, but using three level metal and different layouts, common connection points can be used.

Ic should be noted that the sense transistor in all the previous figures has the source connected to ground as in most cells discussed above. The reversed cell, where the select transistor has the source connected to ground and the sense transistor connected between the select transistor and the BLT, can also be used. This configuration is shown in Fig. BE. As a rule, this cell requires a larger storage capacito- C1 than the cell with both sense transistors connected to ground. The gate capacitance of the sense transistor in the second tier cell in read mode will couple a disturbance to C1, which may cause an error in -the memory functions of the first tier cell. This is contrarv %to the normal (non-reversed) cell, where the major part of Ci actually is the sum of the gate capacitances of transistors 802 and 812 (Fig. 8A).

Again, the amplifier for the second tier may be of the folded bit line type, as indicated in figures 8B and 8C. This am-c-l'ifier would use the same techniques discussed above in conjunction with memories having only one tier. Seg-men"C'ed bit lines -may be used in this configuration as well.

A single-ended ampliffier 850 with a data out latch 852 is indicated in Fig. SE, and a corresponding timing diagra--, is shown. in Fig. 8F. The signal VBIAS is held at an 24 intermediate level concurrently with VWL'-". The current from p channel transistor 854 is aDnroximately half of the sun of t-h.e cell currents - for a zero cell and a cne cell. Li-miting the negative swing of VBIAS makes P transistor 854 act more like a constant current source than if the swing is larger. The circuit can be made to operate with somewhat lower margins if designed so that VBIAS goes to ground when activated. The D innut on the flip flop is designed to have its switching point. close to Vdd/2, and as long as the bit line voltage in the two states deviates a minimum amount from Vdd/2, the operational margins are satisfied. The actual moment of reading may in t-Ah-e wors-[.: case coincide with the reading of the first tier memory. At this moment C1 may have a disturbed Zero level. The amplifier design must take this into account and also plan 'Eor changes in the charge levels on Cl between refresh due to leakage. in the amplifier indicated in Figs. 8E and 8F, no clamping of BLT to a neutral midpoint is made. Adding this feature reduces the access time of the second tier memory 810, but would not reduce the power consumption. The folded bit line approach has both speed and power advantages, but the gain is highly dependent on the bit pattern in the read data.

Denser layouts using a two-tiered memory may be achieved by using nondedicated bit lines, which means that a given bit line can be used as a read bit line or a write bit line depending on if the address is odd. or even. In this type of arrangement, the bit line capacitances remain equal. Fig. 8G shows part of such a memory using the DRAM cell of the Present invention e>manded to include a sense transistor and a select transistor for the second tier. In the configuration shown, each column of basic DRAM cells is combined with two folded bit lines for read out in the second tier. Read onerations from this memory configuration are very fast. The bit lines' are clamped to a mid level, the word lines swing to Vdd when reading, the amplifier is disconnected from the b4t lines at an optimum time when the bit lines have just separated from the clamp level, and the reclamping is completed while data is read out from the amplifier. A preliminary layout fcr this circuit is show,-! in Fig. 8H. Other conf igura-'.- ions may of course yield a denser layout.

Process for Reduced Area and High Speed One of 1--he features of the memory of this type is that t ' he conventional CMOS Drocess with three metal layers can be used for production of the memory. For dedicated memory chips, the cell area can be reduced substantially if the fused and thinned wafer approach is taken, as also described in U.S.

patent number 5,396,452. in logic circuits where memory and logic is mixed, even larger benefits can be achieved. The approach would be 'to prepare a wafer that, later in the process, would strictly follow a standard fabrication process.

The added cost in the wafer preparation would be well compensated by the savings in area and the gain in speed.

Fig. 9A illustrateE stcps in the wafer preparation as follows:

1. A support wafer 902 of low resistivity is used. The areas 904 where capacitors are to be located are etched down. A thin dielectric layer 906 is deposited or grown- 2. Polysilicon layer 908 is deposited over thin oxide layer 906.

3. The wafer is polished, so that the support wafer 902 and the polysilicon 908 have a common smooth surface.

4. Another thin oxide layer 910 is grown on the polished surface.

5. The epi layer 912 c.-,Alf a second wafer faces the -he support wafer and is fused thereto. The asser-bly is top of t then processed as described above so that only the epi layer 30 remains on top of the support wafer.

6. Contact holes 914 are etched and plugs 916 are deposited to establish contact points to Ithe embedded capacitor 918 and to the substrate.

7. If ci--,her means have not beer. used in the sten35 above to isolate the plugs from the surrounding silicon, L- - --hed. in trencl---s 920 surrounding the plugs are defined and et the same st-eD, the individual transistors in the memory cell and in the rest of the lcg- ic are als,:'separat'--d from each other 26 by trenches 920. The trenches are preferably back-:ffilled W;th a form of glass.

S. P-ch-anne! transistors are exposed to N-type imDlant and are heat treated to a sufficient depth. The sur-fface is next Drepared for the gate oxide. Tf tHe capacitor is to be connected to the gate of a transist-or, the gate oxide 922 is removed on top o-f the contact plug for the capacitor. To more closely follow a standard fabrication process, the ca.nac itor connection can wait for the metal 1 step to connecILthe gate and the capacitor.

The wafer is now ready to be run in a known three layer metal process. The LOCOS isolation process step need not be completed, as the area under the poly contact can be isolated by a trench in step 7 above. This results in a planar metal 1 layer.

Multi-laver Memory Cell Controlling Logic Programmable logic devices such as Field Programmable Gate Arrays (FPGA) have devices that must be controlled with relatively large capacitors. The pass transistors transferring signals from one data bus to another must be large enough for minimum voltage drop in passing the signals. The controlling capacitor must be significantly larger than the gate capacitance of the pass transistor, so that the voltage on the controlling capacitor is relatively constant during the signal transition. Using the multi-layer approach described above lends itself ideally to this purpose. The total area under the pass transistor and under the memory cell transistors can be used for the controlling capacitor.

Fig. 9B shows one DRAM cell 930 controlling a pass transistor N4. The cell 930 includes transistors Nl, N2, N3 and a large capacitor C1 located under both the cell and the pass transistor. The source of N1 is connected to the gates of N2 and N4 through first metal layer MI, with a silicon island used to suport a feed through to Ithe underlying capacitor. (The vertical connection is indicated on the left side of the gure.) The combined structure has three vertical busses, which may be in the second metal layer M2, with two busses 27 reserved -L"cr the Inemory bit lines and one reserved for a vertical interconnect. bus for data. The word line and local connections use layer Mi, while ground and a horizontal data bus use the top metal layer M3. '. "here are several options to 5 make the connections between the different layers, including feed-"---.hroughs in the back-filled areas between transistors. The epi layer remaining after thinning the top wafer is usually thicker than a normal implanted source-d-rain diffusion. The transistors therefore have a rema-Lning substrate portion.

Provisions have therefore been made so that this substrate layer may be connected to the underlying support wafer which is at ground level for the N-channel transistors. The substrate layer -may be connected to the underlying support wafer which may be at a ground level (Fig. 9C). The support wafer does have N wells to support the connection of the substrate of the P transistors for this reason.

The same procedure discussed in conjunction with Figs. 9A and 9B can alsobe used for the memory cell shown in Fig. 9C. Adding a recessed capacitor 936 increases the storage time o.-'L' the cell, which reduces the refresh frequency. The gate capacitance of the sense transistor is generally sufficient for reliable operation. With all three cell transistors in the top layer, no additional area is gained by adding a recessed capacitor.

Fig. 9D shows in greater detail a compact cell utilizing a recessed transistor and a double-sided capacitor.

The thin transistors in this figure are not fully deple'Ced, which means that the transistor substrate must be tied Zo ground. It should also be noted that the transistor substrate is one side of a double sided capacitor and for that reason mutt have a relatively high conductivity. The surface of the wafer to be fused should therefore have a P+ top layer. With the transistor substrate grounded, the source and drain junction capacitances are as high as in the standard processes, while the area eff-4ciency is increased as the devices are separated bv trenches cut through the top silicon laver. The top transistor laver in this case is assumed to be fo--,',Ied by de-fining an e-t-ch st-on bv an epi lave-i: with a much different 28 imourity concentration and a different tvpe. The etch ra,"e n the bulk is 4C-herefcre much hiaher With the righ+ choi-- of e"L:chant. T.he remaining layer is therefore too thick to permit, -full depletion by the junction implants while still maintaining 'he small device dimensions.

An alternative approach has been taken in Fig. 9E. Here the top wafer used for fusion is of a SIMOX type. The toD of this wafer has a thin silicon layer separated from the bulk by an imvlant-ed laver of oxygen. Such a wafer after adding a relatively thick oxide layer, can be fused and thinned down to the buried oxide layer with high accuracy. The transistors formed in this thin layer will have fully depleted junc:1---ions and small junction capacitances. They will therefore be much faster than -the transist--ors in Fig. 9D. The recessed capacitor iill however have a thicker oxide layer on top, resulting in a value close to half of the fully double sided capacitor in Fig. 9D.

Fig. 9E is a supplement to Fig. 9D, where marker line 942 denotes the location of the capacitor and marker line 944 the location of the junctions in the support wafer. In the cells of figures 9D and 9E, the bit line connections are not side by side. This permits the layout of a narrower ce-11 than Lf the connection are side by side, which must be the case f or single layer cells.

Fig. 9F. shows another construction for a capacitor in combination with a pass transistor in a programmable logic device. In Fig. 9B, the capacitor was formed underneath the pass transistor. Here, the canacitor is instead formed above the pass transistor. In Fig. 9F, a flat capacitor covers most of -the pass transistor area and typically comprises two polysilicon layers separated by a thin dielectric layer. one layer is coupled to a fixed lootential, for example ground or Vdd, while the other layer is coupled to the control node from the DRAM ciruu.4;_4[_-. The gate of the pass transistor is likewise coupled to the control node of the DRAM circuit.

29 Multi-lave-1---Desians using Epita,-,-7 and Simox Figs. IQA-10D show altL..ernat-e methods of submerging capacitors and transistors below a thin ePi layer added directly on top of a supporting wafer.

Fig. 10A shows the steps for making a capacitor with both electrodes accessible. The sequence of processing is as f ollows:

a) A recess 1002 is etched in the support wafer 1004, which is assu:med to be P-type.

b) N-;- material 1006 is implanted but not, necessarily heat treated at this time. The purpose is to form an N-channel junction, which will isolate the bottom electrode of the capacitor from the support wafer when operating at levels positive in relation to the support wafer.

C) A thin oxide 1008 is grown which will become the capacitor dielectric layer. Polysilicon layer 1010 is then deposited until the recess is f illed. This can be a blanket deposit (as shown) or a local deposit.

d) The waf er is polished so that the remaining surface 1012 is a continuous flat surface, but with no possibility to contact the buried N+ junction.

e) An area covering the capacitor and the intended junction contact regions is defined and another N+ implant 1006 is made.

f) An epitax-Jal layer 1014 is grown, which may require two steps to obtain a usable laver on top of the polysilicon. in the first step the layer is grown vert-ically and should give single crystal areas, at least where the' starting surface is single crystal. In a second step some form of Zone crystallization nay be used to include the areas above the poly into the single crystal.

g) In this st-en 0 2 is im-olanted to a denth ecrua' to the deposited erl laver. After heat treatment, an insulating is formed. The remaining laver of the eni layer layer of Si02 35 is '"hen used for MOS devices of the tv-oe used in Silicon on eV4Ce insulator technology. Connections from the d layer to the canacit-or elect-rodes uses known methods.

F-;'.g. 10B illustrat-es how a t-ransistor such as N1 in the memory cell can be submerged into the sunnort wae- as follows:

a) A recess 1016 is etched in the P- wafer, poly is 5 denosi4Z'ed and the wafer is polished.

b) The transistor area is defined and N+ 1018 is implanted.

C) The epi layer 1020 is grown as discussed above.

d) 02 is inDianted. When the wafer is heat treated, Si02 is formied, leaving a device laver insulated from the submerged transistor. The transistor channel goes from the edge of the source im-plant to the edge of the drain implant. The channel will go in a vertical direction close to source and drain, if the depth of the diffusions are less than the depth of the original recess Since there is some uncertainty about the silicon quality in the epi layer above the submerged poly layer, it is desirable to consider the approach illustrated in Figs. 10C and 10D, where the capacitor electrode is made from silicon. The process for Fig. 10C is as follows:

a) A deen N+ diffusion 1022 defining the area of the caDacitor and its connection regions is made in a P type support wafer.

b) To form a capacitor, dielectric layer 1024, or 0 is implanted in a defined which may be, for example, N4 2 area.

C) To isolate the capacitor electrode from the rest of the wafer, a trench 1.026 i1s. cut around its edges. The trench is refilled W4 th SO using known techniques.

L L_ - 2 d) Fields of epi layers 1028 are next grown on all silicon areas.

e) oxygen or nitrogen is now implanted to the depth of the epi layer and the completed wafer is heat treated. The ton layer will be used for devices and conventional stevs are taken to contact the devices to the capacitor electrodes.

The method in Fig. 10D shows the process steps used to form a transistor with a large gate capacitor. As in Fig. 10C, the caDacitor electrode (gate) is 4 n crystalline 31 silicon, isolated from the rest of the su-pport wafer by implanting N4 or 02 and by trench isolation. it should be notted that a capacitor can be made using the same configuration, hut by excluding the drain diffusion and connection. 5 Fig. 10E shows a conservative layout of cells using submerged transistors. Shared Bit Lines and Amplifiers in the single layer cells the bit line connections are as mentioned above, side by side. Figs. 11A-C show an alternate method by which bit lines are shared. This is possible if different word lines access successive cells and three metal interconnect layers are utilized. In Fig. 11A, each bit line can be a read or a write bit line depending on whether an odd or an even word line is addressed. For an odd word line (WLI), the left-most bit line serves as a write bit line and is therefore connected to the write bit line input of the amplifier (control line "odd wlns"). The middle bit line serves as a read bit line and is connected to the read bit line inDut of the same amplifier. For an even word line (WL2), on the other hand, the middle bit line serves as a write bit line and is connected to - the write input of the amplifier and the third bit line serves as a read bit line and is connected to the read line inDut of the a=lifier. At the same time, the left most bit line also acts as a read bit line and is connected to the read bit line input of the amplifier on the left (only the read line input is shown). in the figure, the word lines are routed across the array in MetCal 2. The Metal I is then used to reach the Poly gates, with the gate connection located in -the area of local oxide between the write and read transistors. In the serially segmented bit line configuration, the ou4%-., Out from each segment amplifier connects to the bit 14 _Lnes in the neXt-, segment via a pair of pass transistors as shown for the dedicated bit line configurations discussed earlier. In Fia. 11B the lines are dedicated for read line olw_-ite line, but still with the f eature of sharing. Here, the szacond bit line is used as a write bit line both for the cell or, is le-F- and 'he cell to its i-ght. This bit line 32 _;lways connects to -write bit 14ne i-Put on the same ampli r, while the read bit line innut'_- to the anpiJL-fier is connected to the first or the third bi-tZ line deDending on if an even or an odd word line is addressed.

The cells in Figs. 11A and 11B were assumed to be in a single level process. Further area savings will be achieved if the multi- layer process is used also for th " e shared bit line configuration, as shown in Fig. 11C.

Refresh Read and Write in a Seamented Conficrurat-ion Fig. 12A indicates a por"Cion of representative column 1202 in a segmented memory 1200 that may he selected for reading and writing. In the figure JLt is assumed that the f irst data bit in a data word of N bits can selectively be connected to one of the first eight columns in the memory. Three column address bits, BO, B1 and B2, and their comnlements are used for addressing. Data can be read from or written to either bit line in the last segment 1204 (SEGMENT 0) in column 1202, but there is a small advantage in writing to the read bit line and reading from the write bit line. The coupling is therefore done accordingly, controlled by the signals COPLIOREAD ILND COPLTOWRITE.

The timing of the clamping and coupling signals for the segments are optimized for speed and performance. In read mode, it is preferred that an addressed ' segment be allowed to amplify its signal before any additional segments are connected. In one embodiment, a first increment of time is set aside for the addressed seamenl_-, such as, for example, segment 1204, to establish a signal. At the end of this first time increment, the second segment in the pair, which would be segment 1206 in this example, is coupled to first segment 1204 by a signal COPL. At the end of a second time increment, the remaining segments are interconnected.

However, if an inverting amplifier is used in the memory segments, the segment pairs may be allowed to be connected at the outset of a read operation. Furthermore, i-A -the dimensions of coupling transistors 1208 are chosen correctly, all segments may also be connected at the outset of 33 a read oneration. This results in simplified logic and in a higher operating speed. In this simplified scheme, all segments are clamped at the same time at the beginning of the cycle followed immediately by an active coupling signal COPL that connects the segments. With inverting amplifiers, memory refresh occurs in segment pairs; all segment pairs are refreshed at the same time. The signal COPL connects the two segments in the pair. This signal may also be applied after a delay that giving the selected segment signal time to establisn itself, compensating for any disturbance introduced by the clamp signal. on the other hand, coupling the two sec rments n a pair with COPL jiist after the CLMP signal is active is equally reliable, as both segments are disturbed equally. The fact that read bit line BLR in one segment connects to write bit line BLT-7 in the next segment results in nearly full compensation, without using a compensating signal as discussed earlier. The COPLR signal, which is used to connect one segment pair to an adjacent segment pair, is not applied in the refresh operation. For reading and writing, both COPL and COPLR are applied after CLMP at the beginning of a cycle. A write operation must. apply a signal to secment 1204 which is larger than tthe signal in the addressed cell if the cell happens to hold data of opposite value. If the topmost segment in column 1202 is being written to, the written signal would have to travel through all the segments in column 1202 to reach the segment at the top of the column. At the same time, a signal of the opposite value may be travelling from the top segment in column 1202 toward the bottom segment 1204. At some point, probably near the middle of column 1202, the two opposite-going signals would meet, and the difference between the two signals would then proceed in both directions. If the written signal dominates, the addressed segment at the top of column 12'02 will have the correct data written. The larger the Written signal, the faster the write bit line in the addressed 3 5 segment will reach a safe write level.

On1v selected columns are addressed for the purposes of reading and writing data to the memory segnnent-s in the selected colummn. In both cpera-!Cions;- the st-c-red da-'Ca in the 34 non-sel ected colurms is not modified. 1-11 cc!ulmns are ex'Dosed "horizontall! sianals, such as COPL, COPLER, C.TMLP, WL to the same and DUM. (defining the dummy current level).

Figs. 12B through 127- are simulation outputs. A memory array was simulated that included 256 columns with 16 segments per column arranged in 8 segment pairs, where each segment had 64 memory cells. A 0.5 micron industrial process at normal operating conditions was used. A 15ns cycle.time was simulated, whiLch provides sufficient margins for reliable operation.

In Fig. 12B, the simulation shows the results when a memory cell storing zero volts on the cell capacitor is addressed by a word JLine in the top segment. The write signal is applied to the bottom segment, with a polaritv such that the stored zero will become a one. After the clamn and activation of the coupling signal, the dummy current dominates over the cell current and the read bit line for the addressed segment (BLRIS) starts going positive and write bit line for the addressed segment (BLW15) goes negative. At the same time BLWO is pulled low, causing BLRO to go high, thereby initiating writing of a one in the addressed cell. After approximately 9ns the written signal reaches the top segment, causing BLR15 and BLW15 to change direction of transition. BLW15 reaches the full 5 V level at the end of the cycle and the cell is charged to about 3.3 V, which is well above t-helevel of 2.5 V that is generally required to store a one in the cell capacitor.

Tin Fig. 12C, the simulation result is shown for the oDeration of writing a zero over a stcred one in the cell capacitor in the top segment. The cell has a stored voltage level of 3.5 V, which is the highest stored one voltage level that can be written without boosting the word line level. This gives the highest possible gain contribution from the stored cell, which is acting against the written data input applied to the bottom segment. The direction of BLW15 starts turning around after 8ns and reaches 0. 3 V at the end of the cycle. The cell voltage tracks B.LW15 after 11 ns and reaches 1. 5 V after 13ns. A cell voltage of 1.5 V is marginally acceptable for a stored zero level, so the 15 ns cycle time provides sufficient margin at normal operating conditions.

The stored data in non-selected columns is not disturbed, as illustrated by the simulation results shown in Figs. 12D and 12E. In Fig. 12D, a stored zero at a level of 1 V is originally present in the addressed cell in the top segment. B-itlines BLRIS and BLW15 are the first to move due to the dominance of the dummy current to '"he selected segment over the current. from the selected cell. BLW15 falls below I V af ter 9 ns and reaches close to 0 V at the end o-E tha cycle. There is less than a 2 ns delay between BLWIS and BLRO and an actual crossover between BLR15 and BLWO, caused by the fact tha'L the coupling signal amplitude is limited to 5 V -in the simulation. The coupling transistors are not conducting higher than one threshold voltage below 5 V. Due to a high body effect of the transistor, the threshold voltage Vth at 3.5 V is actually 1.5 V. As the same control signals are used for both reading and writing on nonselected columns, Fig. 12D also illustrates the read operation. The differential signal between BLWO and BLRO revresents the stored data. Fig. 12E shows the simulated result when a stored one at 2.5 V is read. The differential signal between BLWO and BLRO also here represents the stored data.

It is apparent that the read data can be read out long before the end of the 'Lull memory cycle. This is a very important feature of this type of memory. A memory read or refresh cvcle can be interruo-Ced at any time in the cycle without destroying the stored data. A read cycle can therefore be terminated as soon as a detectable differential signal between the bit lines in the bottom segment is transferred to the outnu-'L,--. The only disadvantage is that the read cell level will not be refreshed in an interrupted cycle, but the level will remain the same as at the start of the cycle The voltage level in the cell will therefore remain the same until t-h-e next refresh cycle.

Tf an 1/0 latch 1210 is added to memory 1200 from Fig. 12A, as indicated in Fig. 12F, the delta V in Figs. 12D and 112E car be allowed to be as low a-t!00 mV. This would mean 36 that data could be transferred to t1n.e latcli after only 5 -,s.

Additional time is naturally -required to forward the data to the out"put, but a read access time of less than 8 ns is practical for %'-his array configuration.

The number of cells per segment has an 1'MnaC4C-. on speed, which -makes it possible to reduce the cycle time for reading and refreshing if the number of cells per secment are reduced. The delay through the coupling transistors combined with -a limited gain in the amplifiers makes difficult to write over vreviously stored data in the most remote segment without making special arrangements. simulations run on a memory structure with 256 columns, 8 segment pairs (16 segments) and 32 cells per segment are shown in Figs. 12G-12L. If a cell in the most remote segment is addressed for writing, a read cycle to this segment will progress quite far before a write signal of opposite polarity reaches the addressed segment. on the other hand, if the bitlines in the segments of the addressed pair are held clamped for an extended time, while awaiting the arrival of the write signal, then the change of polarity between the bit lines will be faster. In Fig. 12G, the clamp signals CLMP and CIMPR for the selected pair are allowed to s-t--ay on 2ns longer than the clamp signals for all other segments. All segments are coupled together at the same time indicated by COPL becoming active. The write inputs are applied to the bit lines in the bottom segment (BLRRO and ELWRO). The write signal proceeds from segment to segment toward the top of the array. Before it reaches the addressed segment, its clamping is terminated and "&---he write signal overcomes the small signal from the cell. At a time of 7.5 ns, the previously stored zero has been changed to an acceptable one. After lOns, a maximum level for a stored one is reached. Fig. 12H shows the signals for writing a zero over a stored one.

It is important that the columns not selected by the column decoder for writing still preserve the stored data in their addressed cells. Fig. 121 shows how a stored one is actually restored to its maximum value in the non-selected column. BLR and BLW in the ton segment still have enough time 37 to reach their -full value. There is however not enough time for '"he signal generated in the top segment to proceed down to the bottom segment and overcome a noise-generated signal therein. The polarity of the bit line signals in this case 5 does not matter as the output is not used in this mode. Fig. 12J shows how a stored zero is restored in a non-selected column during the write operation. in this case, the noise signal in the bottom segment happened to go in the same direction as the signal coming down the line from the add-ressed segment. The noise source for the bottom segment will be discussed in connection with the read mode below.

Figs. 12K and 12L show simulations of the read mode. To make sure that a signal from the top segment reaches the bottom segment and is correctly presented to the 1/0 circuits some improvements have been made compared to the sequence used for the 64 cells/segment seqaence described in connection with Figs. 12B-12E. The negative-going clamp signal will, as discussed earlier, introduce a small negative displacement of both bit lines. In a disconnected segment this would cause the write bit line to move positive with the read bit line remaining stationary. In a long chain of segments, write and read bit lines are alternately connected, so the inside of a long chain is more or less self compensated. The problem arises in the segments at the ends of the chain. For all segments the coupling signals adds a positive compensation. The top segment is exposed to one half compensation from the coupling signal, while the internal segments each are exposed to full compensation. The bottom segment in read mode is coupled to the 1/0 circuit through some form of decoder or to 1_; 0 an output latch. In either case the coupling transistors controlled by the signal COPLIN connect to nodes with unknown potential and capacitance. it is therefore desirable to connect these nodes to the bit lines during clamping to ecrualize the Doten'C-ials- The COPLIN signal is turned on momenta--4Lly at 1:he beginning of CLMP and turned on again at the same time as all other coupling signals. Using this technique, 'Figs. 12K and 12L show how the bottom segment bit lines BLILRO and BL:rqR-O corz-ectly reflect the s4C_crdd da4%,--a.

In the _,_eV 4 ous d4scussion, it was assumed that segments operated in pairs, so it- is reasonable to use a segment pair decoder 12220 shown in Fig. 121H. The low order bit BO in the segment address is used to select one or the other segment after the higher order bits select the pair (B1.... Bn). The dummy current for instance is anDlied only to the selected segment, so each secriment has its own dummy voltage arenerator. The dummy transistor in the selected seg-ment actually mirrors the current flowing through the dummy voltage generator. The current through the dummy voltage generator 1222 is primarily a unction o' the 'RAMP voltage and increases the dummy current -ional to how the cell current varies with the RAMP proDor4. voltage. The two transistors 1224 and 1226 between the RAMPcontrolled t-ransistor 1228 and ground are larger and serve mainly as switches. Transistor 1224 is used to define how long the dummy current is to stay on during the cycle and transistor 1226 is turned on only during memory activity and when the specific segment is selected. A three input NOR ga"C'e. 1230 combines the selected pair signal with the ODD/EVEN selection by B0 and its complement to select the segment and associated dummy voltage generator. The fourth N-channel transistor 1232 in the dummy voltage generator is very small and in standby barely keeps the P-channel mirror transistor 1234 conducting. A number of global timing signals are connected to all segment pairs. SpecifLically, DUY-TM.defines the active dummy current time, ECLTM defines early clamp time, LCLTM defines late clamp time, ECP. TT-M defines early coupling time, and LCPLT-M defines late coupling time. The complement of the refresh command (R:-:FRB) is also a global signal. REFRB (low level) bypasses the segment pair decoder and applies an early clamp signal to both segments through signals CLMP and CIMPR. An early coupling signal is also applied -t-o the top segment in each pair, connecting the two segments in the middle. The LCPLTM signal is not delivered in the refresh cycle, thus 35 keeping all pairs separated.

39 Control ol. Lcaic Cells with DRAMs U.S. Patent Numbers 5,375,086 and 5,317,212, both hereby incorporated by reference, describe methods for controlling the function of logic by using the capacitors in DRAMs as the controlling element. The DRAM of the present invention lends itself to control of logic with some advantages over the previously disclosed method.

The control capacitors that- are also storage caDaci-t-ors in the memory cells are typically larger than what is recruired for memory functions. With the leakage currents in the dielectric of the capacitors much smaller than the leakage in the write transistor, successive refresh cycles can occur less frecTuen-t-ly. As most of the power at refresh is consumed moving the bit lines, the refresh power is inherently low.

However the power consumption can be further reduced by using segmented bit lines, as the power consumption is inversely proportional to the number of segments.

The control capacitor cc, which principally replaces an SRAM cell. must in its high state be at least at a level of Vdd. The cell capacitor is typically connected to the gate of a large pass transistor, which has a gate capacitance of Cgp. Any signal Vsig passed by the pass transistor will be coupled 4 -through a capacitive voltage divider. The variation on to Cc the voltage on a programmed cell is described by Dvc = Vsig x Cgp/ (Cgp + Cc). The signal Vsig typically has an amDlitude of Vdd. If Vdd is 5 V and Cc = 4 x Cgp, then Dvc will be 1 V. The programming is asynchronous to any signals in the controlled system. If programming occurs when the passed sianal is at Vdd and the programmed cell voltage is Vcellp, then the cell voltage will be reduced by Dvc, when the passed signal has returned to ground. The gate to source/drain voltage will go from (Vcellp - Vdd) to S4 (Vcellp Drvc) during the negative tran _Ltion. conversely at a a L ollowing transition 'the gate to source/drain voltage will go f-rom (Vcellp - Dvc) to (Vcellp - Vdd).

if -t-he programming occurs when the passed signal is at'- ground level, then the cell voltage will increase by Dvc when the Dassed signal goes from 0 to Vdd. The gate to source/drain voltage will go from Vcellp to (Vcellp + transition. 71 -vc - Vdd) during the positive -- the design goal is to ecuai 'he perfo=mance of SRIM cont--ol, then Vcellp is a a level of Vdd. The SRAM has a relatively high output resistance, so in a negative transition the gate voltage will drop and recover wilt.:h a tine constant of (Rou4L--pu4L--. x Cg-p) If it is assumed that this slows the negative transi1--ion approximately as much as the transition would be slowed by reducing the gate to source voltage to Vdd - Dvc at the end of the transition, then a relatively high value such as 2 V could be chosen for Dvc - When the positive transition occurs (assuming that the programming was done when the signal was at, Vdd level) the gate to source voltage would go from 3 V to 0 V, while the SRAM-controlled gate to source voltage would go from !E; Vdd to a positive voltage caused by the same time constant of (Routput x Cgp) - but limited by the forward-bi.ased diode from the drain of 'he P-channel transistor to its N-well (approximately 0.6 V). if the programming had occurred at a low signal, both transitions would have been faster than for the SRAM case, with the gate to source voltages (assuming Vdd - 5 volt) 5 V and 7 V in the two extremes of the transitions. To program Vcellp to 5 V, the supply voltage to the word line of the amplifier must be approximately 6 - 6 V due to the high body ef f ect on the short channel trans is[--ors. If Dvc = 2, the cell voltage would peak at 7 V. The pass transistor gate would however be exposed to a maximum of 5 V, while the full 7 V would stress the gate of the sensetransisto-'r and the eXtended storage capacitor. If 7 V is the specified maximum voltage, the word line voltage would also be permitted to go that high. The amplifier supply could then be 5.4 V and a Ervc of 1. 6 would be chosen. The cell capacitance is determined by the relationship Cc = (5 1.6 - 1) x Cgp = 2.125 x Cgp. A pass transistor may have W 20 p and L =.5 p, or a total gate area of 10 gsq.

The cell capacitor area would be 21.25 jusa or in the form of a full t-ransistor with W = 4.6 y and L = 4.6 g. The pass transis-%'--or would occupy about 70 psa, while the capacitor would occupy 35 Asq. A cazacitor with a source 4 Ln the middle would A I only occupy 25 4sq. It is apparent that Dvc could be reduced even furt- Inler at relativelv small area cost. The actual voltages and capacitor sizes depend on process limits and design criteria. 5 The DRAM cell of the present invention has been described as using three N-channel transistors, but it should -ood that three P-channel transistors can naturally be unders4k also be used, with associated circuits and signals modified to fit 'he reversed requirements.

The addressing of a memory used for control of logic need not be as fast as f or a conventional memory. Data and addresses can be fed serially through shift registers in both vertical and horizontal directions. Recent requirements for higher speed of loading uses byte wide data shifting and for more advanced applications random addressing for the writing of control data is used.

Fig. 13A shows a decoder 1300 that may be used to address the word lines. The decoder is operated by N pairs of signals, both true and complement, selecting a desired word line. The SEGMSEL signal may be global for non-segmented memories or nay be derived from a combination of a global enabling signal and a segment decoder. The low order bit An+! and its comDlement selects one segment in a pair to be fully decoded for read and write operations, while the higher order bits are bypassed for the refresh operation. The input to the first inverter 1302 in the seginent;.: select circuit is assumed to have a weak pullup device normally. With the decoder activated, the input node is pulled low and the output of the first inverter 1302 enables all word line decoders in the selected segment. However only the word line selected by AO to An'and their complements will track the global VRAMP signal. The outnut of second inverter 1304 goes low on the selected sec Tment, thus allowing 'he selected word line to go high.

The active time for the cycle is superimposed on An-! L. -- both are low between cycles. Then and its comple-ment such that L at the end of the cycle, the in-put to the first. inverter 130-4 will ao high, as will 'he outnut of the second inverter 1304.

Transistor!306 will then ioull down the selected word line al_ld 4 4 2 hold it and all other word lines to ground. To reach a wcrd line voltage of 7 V, VR._MP need nct go higher lt_-han 7 V, bu_ the address inDu-Cs and the out'-put of the first inverter must reach. one threshold higher, which means close to 9 volt. This type of decoder 1300 would therefore be used in connection with a large ratio between the cell capacitance and the load capacitance (gate capacitance of the pass transistor), so that the operating voltage could be reduced.

Fig. 13B shows a decoder 1310 with a boosting feature WL-ch- allows most of the circuits to onera"Ce at normal Vdd levels. A WLENABLE and a complement WLRESET is applied globally or from a segment select circuit. in the enable mod4.. LIN on the decoded circuit is Dulled low, setting the previous reset latch. When the LIN signal is low and the LOUT signal is high, nodes DN2 and DN3 are elevated to a level that is one threshold voltage below Vdd (e.g., 3.5 volt). The timing diagrams in Fig. 13B show the levels of the different nodes. The RAMP signal starts out at a low voltage level of approximately 1.2 V, and WL is brought to that level by fully turning on MNWL. Late in the cycle, RAMP moves from 1.2 V to 8 V, in this particular case. The capacitor MNC is much larger than the capacitance of node DN2, so DN2 will track the rise of the RAMP and the word line. DN2 will end up higher than the RAMP by one threshold voltage, so the word line will go as high as the RAMP signal (e.g., 8 volt). DN2 'at 9.6 V might expose the gate oxide of MN3 -to the break down limit. Thus, MN2 acts as a barr-Ler in that its gate is at 5 V, as is the gate of MNI.

By returning the IRA-HP voltage to 1.2 volt before resett-ing the word line, the protection of the devices is maintained.

The voltage of a stored one with this conditions would be approximately 6. 4 V, provided the amplifier supply voltage is also at 6.4 V. The maximum stress on gate oxides in the cell is 6.4 volt plus Dvc.

When controlling logic, the high voltage on the cell is of most concern, even when an stored zero at a level of about I V may be acceptable. Another alternative is, therefore, to implement a decoder using P-channel devices only. Fig. 13C shows decoder 1326 that is the equivalent of Fig. 13B, 43 but with the polarities reversed. The common node for the sense transistors and cell cat)acitors in the cells are connected to the 6.5 V supply indicated in the figure. ThF-. low.RAMP level is -1.5 V in relation to 6.5 V or the 5: V Vdd supply. To approach the Vss level in the controlled circuits, it is desirable f or the RAMP to go to -1. 5 V, thus overcoming the high threshold volt'-age due to the body effect of the P-channel write transistor in the cell. The negative swing on the RAMP signal could be limited to 0.5 volt below Vss. The refreshed high voltage on -'%zhe controlling capacitoz will be at 6.5 V, but in the read phase of the refresh operation the voltage will drop slightly. For cells with large capacitors that are used for controlling pass transistors, this voltage drop is very limited due to the long time constant in the combination:of the write transistor and the cell capacitor. In other words, the write bit line will reach the 6.5 V level before the cell caDacitor has been disturbed more than a small fraction of 1 V. The ref resh phase with the RAMP and word line at its negative extreme is made long enough to fully refresh both a high and a low cell level The original storing of control data may require extra long write times or multiple write cycles for very large storage capacitors. All cells can also be precharged to an intermediate level by turning on all word lines, with the bit lines clamped to the intermediate level. The writing will then only need to displace the cell voltage slightly to indicate the stored data. The refresh operations will later gradually charge the capacitors to their full values.

Switched Power Ant)lifier In order to further reduce -%'--he power consumption in the control memory, it is desirable to let the cell current or the dummy current displace the read bit line as much as possible be-fcre the amplifier is powered up. Fig. 13D shows a 33 5 memory 1330 where each pai-r of segments 1332 contains only one common cla=) circu-it. The associated timing diagrams are shown in Fig. 13EE. Clamming to an in termed i ate clamp level is done without ioower - to the amplifier and 'ftilrther timed so that at 44 least -'%---he two seg-ments in each pair are cou.-Iled together during $the clamping. All four bil lines will there'ore be floating close to the clamp level at- the beginning of the memory C:ycle. Applying the low ramp level to the word line starts the displacement- of the read bit line in the selected segment. When the selected bit line has been displaced sufficiently for good operatt_-ing margins, the power is applied to the selected segment- This is controlled by an early timing signal, EPERB, which is negative true. When the inverting amplifier in the I - - selected secmen4L- has disnlaced its write bit line a certain amount, power is applied to the non selected segment in the pair (controlled by LPWR) and the two segments 1332 in the pair are couDled together to form a full amplifier. After increasing the ramp voltage and holding it at its high level 1 S for a time, the cell is refreshed and the cycle is terminated by turning off power, clamping, and coupling all segments together. In the read and write modes, indicated as REFRB, all segments are coupled together at the early coupling time (ECPLTM). As described in connection with a regular memory, a very large signal is applied to the bottom segment of a selected column which dominates over the signal detected at the selected segment.

Mixed Cell TyDes on Common Bit Lines Typically large cell capacitors are required to control the gate level on pass transistors. At the same time, the bit line canacitances can be kept low by utilizing the segmented bit line approach. The ratio of the cell canacitance and the bit line capaci-Cance will in such cases be high enough to impose a large bit line displacement at reading a conventional DRAM cell. In other circuits controlled by the control memory, such as look up tables, the cell capacitor can be very small and the Drefer.-ed cell is of the new DRAM type. It is T)OSS4 L_ - -1-ble to have both types of cells on common bit line pairs, as indicated in memory 1340 shown in Fig. 13F. Here, an approach with non-dedicated bit lines has been used, which 14 means that a given hit _Lne can be used as a read bit line o_r a write bit line denending on if the address is odd or even. in this type of arrangement, the bit line capacitances remain ecrual. The dummy current is applied to the bit line that is active for reading a new type of cell and is applied to the bit line to which the selected conventional DRAM cell is connected.

The dummy charge a'" reading a conventional DP-A-M cell is the product of the dummy current and the dummy current on time, which operates with the same margin as for the new DRAM cell of the present invention. The bit lines are clamped to an intermediate level (e-g., 3.5 V) at the start of the cycle, which is higher than the first RAMP level of 1.2 V. A fully charged conventional cell capacitor (e.g., 5.6 V) will also be well above the word line voltage of 1.2 volt. The cell transistor is off and the dummy current increases the bit line voltage at the same rate as at reading a new type of cell.

If, on the other hand, the cell capacitor is discharged to a low level (e. g., 0 V), charge is drawn from the bit line, reducing the bit line voltage more than the dummy current. is able to compensate before the amplifier dominates the displacement of the bit lines. With a normal RAMP signal, the word line voltage is increased later in the cycle bringing the cell capacitor to its refreshed level, also for a stored one. in t-he word line, addressing is sequential without regard to what type of cell is used. In Fig. 13F, WLAO and WIAI indicate even and odd word lines addressing conventional cells, and WLBO -e even and odd word lines addressing the new and WLB1 indicat type of cells. The dummy voltages DUMO and DUMI are act- ive negative and "Curn on the dummy current on BL2 and BLI res-oectivelv. Note that to write a one (high voltage 6n the capacitor) on an even cell of the new type, BLI must- be forced high by the write circuit, while on the other hand BL2 must be forced high to write a one on an even cell of in a conventional DRAM. The program.mer must therefore take this into account when preparing the control data. only if cells of the same type are used for a given range of addresses an alternative approach can be taken. in that case, using the figure as a basis for the d4scussJon, even addresses would apply to conventional cells z1aced on ELI and DUMI would he ac-tivated,or all even addresses call4ng this type oil cells. This would 46 however -lace unnecessary restrictions on '"he cell distribution.

The timing for refreshing a stored zero and a s4L--.cred one in a conventional cell is shown in Fig. 13G. Note for the stored zero that the exchange of charge continues only until the cell capacitor has been charged to one threshold voltage below WLAD. This also Doint-s out the need for frequent refreshing as a cell charged by leakage current to close to one th--eshold below the initial word line volt'-age will cause a read error. in reality, the large cell capacitors used in this configuration at the refresh frequency dictated by the new type of DRAM cells wii.1- be charged very little between cycles. The COUPL signal and the disturbance on the early bit line signals have not been shown in this figure as this is of interest only in the i nitial wriiLe mode and in the read mode used for checking that the written data was correctly stored (mainly in production tests).

Disabling Segments Above Selected Secrments in both read and write operations, the segments above the selected segment. need not be activated. This is done by forwarding the segment select signal to the segment above, as depicted in Fig. 14A. In each segment an OR is formed by the segment select signal and the selected signal from the segment below this or signal is then forwarded to the segment above, where it is treated as the select signal from the segment below. one disadvantage with this app-M-oach is the serial delay time to the last segments in the chain. It is desirable to be able to interrupt the clamping signals in all segments before they occur. A parallel approach as indicated in Fig.

14B addresses this concern. In this case each select signal is forwarded to all segments above and each segment has an OR gate as wide as the numbe-- of segments tha-L- are located below.

If power consumption is not a concern, only the segment directly above the selected one will need to be disconnected fron the segment below as indicated in Fig. 14C.

He--re the coupling signal is disabled for SET-un-1.

47 DRA241 cells with serarate read and write word lines and clainned amolifier The DRAIM cell shown in Fig. 15 requires a larger chip area and more complex word line drivers, but offers advantages in two areas. The cell has the same advantages as the simpler cell described above in that the content of the cell is not disIC-urbed during the refresh cycle. The fact that the word line voltage at reading has the full amplitude means that the read tim-e is shcTrrter than when the word line is left at an intermediate level. There is no feedback from the write transistor as the write word line is low during reading. if the retresh frequency is high enough, a stored ZERO will stay very close to Vss and a stored ONE will stay one threshold under WLW. Boosting the voltage of WLW can maintain the cell voltage (if a stored ONE close to Vdd. This property makes the cell useful for control of logic, where the control node indicated in the figure connects to the logic to be controlled.

The amplifier can have all the forms as discussed before and can also be used in the segmented configuration.

Bit lines need not. be dedicated as indicated in the figure, but will in the non-dedicated case recTuire dummy trans is-Itz ors on both- bit. lines under control of the address logic. The assumption is that the layout is rare efficient with every other cell reversed around a vertical axle. 25 The am-Dlifier in a different. mode of operation can act as a latch, which is first reset, indicating a ZERO, and at reading a ONE is set. This requires lChat the cell with full cell voltage is strong enough to set the amplifier flip- flo-p. The segmented memory approach can still be taken. 30 Full-Amplifier DR.AM Conficrarations in situations where fast access is of prime importance, the single inverting amplifier in each segment may be replaced with a full latch amplifier in each segment. This allows for simpler control with fewer control signals. The relative timing is however slight-ly more critical. Tl.ae DRM with a full- am- ilifier can operate in basically two modes. In 2-1-ode 1, shown in Fig. 16A, Dower is 48 always az::)lied 1--o ' e anDli-Pi er, and the bi ines are cl -Irmed th 4- Che same Dotential, which is the swit'-ching point of the latch. In the Mode 2, shown. in Figs. 16B and 16C, power to the amplifier is turned off during the clamping and the first mortion of the read cycle. In Fig. 16B, the amplifier is OFF in standby mode with the bit lines clamped to a reference cla-mr) level. In Fig. 16C, on the other hand, power is ON in standby leaving one hit line at Vdd and the other at Vss level. The single clamp transistor is turned on at about the same time as nower is turned off. If power is turned of f before applying the clamp signal the bit lines will reach a voltage of approximately Vdd/2 due to charge sharing. 11L power is still on, the bit lines will eventually reach the switching poin-'C level of the amplifier latch. By designing the amplifier for a _5 switching point level of Vdd/2, the relative timing at this time is not critical. Power is turned on again, when the input. signal has been integrated to a safe level, to guarantee that the amplification will go in the right direction.

The bit lines as shown in Figs. 16A-C are dedicated to reading and writing, which means that all cells in a segment are oriented the same way, with the write transistors connected to the write bit line (BLW) and all read transistor connected to the read bit line (BLR). To allow more options for cell layout, it may be advantageous to have non-dedicated bit lines as illustrated in Fig. 16D. There is a minor added cost in each amplifier in having two dummy transistor and controls to select the one corresponding to the address of the selected cell.

Figs. 17A-M show the result of simulations using a 0..5u NMOS process. The memory module consists of 8 segments per column, with 64 bits/segment, and 128 columns (65,536 bits). I Fig. 17A illustrates reading a ONE in a cell located in the top segment, with the read data propagating from BLR and BLW in the tolD segment, to BLRO and BLWO in the bottom segment.

The data on BLRO and BLWO is forwarded to the 1/0 decoder at about 8 ns (not. shown). Power in this case is initially turned off at the same tine as clamping takes place (compare to Fig.

49 16C) and is turned on again when reading starts by h-ringing the word line (WL) to the f irs4L-- -ramp level (1. 8 V). The signal in the selected segment (the top segment in this case for worst case test) is amplified slightly before the COPL signal connects the selected segment to the rest of the segments that were interconnected earlier by the signal COPLN. In this simulation the non- selected segments were clamped a little longer by the signal COPLN, which is 0.55 ns longer than COPL, the s--:Lgnal clampinq the selected seament. The stored enerav 4n 10 the bi-lines of the selected sec L_ _rment is now transferred to the lower seg-inents at the same time as each of the segments contributing to the amplification of the transmitted signal. At about 6 ns, the delayed signal has reached the bit lines BLRO and BLWO of the bottom segment and is reaching half of full amplitude at a delay of less than 7.5 ns. All bit lines are very close to the supply rails at the nominal end of the cycle (10 ns). The voltage difference between BLWO and BLRO at 8 ns is more than 2. 5 V, which is suf f icient to supply an output signal via decoder and output amplifier within 2 ns.

The initial cell voltage was assumed to have degenerated due to leakage after the last refresh cycle to 2.5 V, but is refreshed in the shown read cycle to 3.5 V. The cell capacitance in the simulation is 2e-15 Farad plus the gate capacitance of the sense transistor (MS in Figures 1-3). There is a small coupling from the word line, which can be noticed in the figure, so the final cell voltage when WL reaches Vdd is slightly below 3.5 V. The maximum voltage is determined by the max value of WL less the threshold voltage. The short channel process has a strong body effect on the threshold. The cell current in the read phase increases very slowly with the cell voltage, when above 2.5 V, so boosting the WL voltage above Vdd would have a small effect on speed, but would of dourse allow for more leakage in the cell capacitor.

Fig. 17B shows the reading of a stored ZERO degenez-ated due to leakage from 0 V to 1 V. CnlV a 14-nit-ed number of control signals are shown in this figure. I L: i s a clear indication in this figure that the switching point of the I ier is below "Che clamp level def J ned by the charge amp.,.L - shar 4 ng between the hit lines. This causes the bit lines to dri-;"t in the nega"1_'-_ive direction, before the DOS 4 1 i Ve L L__ current,c-om the durvmy transistor less the low cell current causes tna read bit line (BLR) to go positive. In the full aAmplifier this drift is acceptable as long as the device parameters in the amplifier are consistent.

In the write operation (Fig. 17C) the data to be written is applied via the column decoder to BL-40 and/or BLRo on the bottom segment. This signal is large enough to be amplified and propagate to the selected segment, which has started a normal read sequence, and override this Signal. Fig. 17C shows how a stored ZERO is written over by a ONE. BLRO and BLv7O are separated by the input signal already during the clamping operation. With all non-selected segments interconnected, tfl-ie signal is quickly propagated to t1he selected segment, causing the direction of BLR and BLW to reverse, so that BLW will drive the cell voltage to 3.5 V when WL goes high. Fig. 17D illustrates how a stored ONE is written over by a ZERO.

Turning off power during clamping saves power, as shown in Figs. 17E-N, but adds to circuit complexity.

Figs. 17E-G all have a first ramp level of 1. 8 V, while the following simulation used a first ramp level of 1.5 V. The circuit for the dummy current control adjusts the dummy current to an optimal value for a given f irst r ' amp level. The current difference between the cell current and the dummy current varies very little depending on the ramp level, but there is less dependence on device parameter variations at 'he lower fi_-s4[- ramn level value.

Comparing the initial clamp level of Fig. 17E with the previous simulations shows that the clamp level has dronDed. This is caused by a change of the cell parameters equalizing the sizes of the read and write transistors and their junction areas. The higher junction capacitance at 0 VIF compared to 5 V, brings the clamping level below 2.5 V. The switching point should also be set equal to the clamp level. The clamD Dulses CLNP and CIMPN in Fig. 17E terminated before the word line had reached the first ramn level. This could 51 cause the dummy current, if it had built up earlier to start the displacement of the read bit line prematurely. The margins otherwise are sufficient for correct reading of a ONE: as the figure shows. To eliminate the possibility of the problem discussed above, the length of the clamp pulses in the following simulations were increased. Fig. 17F also displays the dummy voltage, that is applied to the gate of the P-channel dummy transistor. The dummy voltage is also close tc its nvyimum negative value at the end cf tts clang siSnals. A CYZ is read in Fig. 17F and a ZERO is rsad in Pig.!7G. Figs. 17'.' and 171 show normal read restore cycles for ZERO and ONE respectively.

It should to pointed out that in all oases a refres-,.. cycle is the same as a read cycle except. that all segments operate simultaneously fully disconnecto? from sAch other. signals COPL and COPLN are therofore inactive in tne refresb. cycles. All segments are enabled and the low order address bits addresses "the same" word line in each segment. With the timing for the clamping operation used in Figures 17E-Q, CLMP and CLMPN have the same duration and is in reality the same pulse. Figs. 17J and 17K illustrate refresh cycles with the segments disconnected restoring ZEROs and ONEs. Comparing Figs. 17H and 17J shows how the bit lines move faster without the added load of the additional segments. This indicates that the refresh cycles can be made shorter than a read or write cycle. Figs. 17L and 17h show the write cycles for storing ZERO and ONE.

Memory Refresh interruDtion without Delav One reasons why SRAMs are typically preferred over conventional DRAMs is that synchronous systems require that a memory 7eq-uest is served without the delays that a DRAM refresh cycle would cause. A great advantage to the DRAM of the present invention is that an ongoing refresh cycle can be interrupted at any time during the cycle without destroying the stored data. Data stored in the call capacitor remains intact most of the time or is rafrashal during the end of the cycle. in ongoing refresh cycle can be stapptd as long as the word 52 line is brought low as the first step. Frigs. _18.711-B show -,he sirmulations of interrup'Ling refresh cycles aftt_'-er 6 ns and starting read cycles on a d-ji-flerent address, w'nile Fias. 18C-D show an interrupting write cycle.

In these s imulat- ions the CLMP signal starts 1 ns later than in the v.-evious simulations, but ends at the same time - The shorter pulse is sufficient for the clamping of the bit lines and the delay makes sure that the bit line voltages have not changed before the word line active during the refresh has been brought down. it is assumed that the refresh and the reading or writing is done in '"he top segment, with the bit lines BLR and BLW. The word line addressed during the refresh is WLA and the word line addressed during read or write is WLB. Data inDu-;"-- and output connects to the bit lines of the bottom segment BLRO and BLWO. The interrupting memory cycle starts with signal INT, which in the simulation was used to force the WLA low, while the decoder was selecting WLB. In Fig. 18A the cell addressed by WIA (CELLA) is in the ZERO state, starting at 0.5 V and during the read phase of the refresh cycle is lifted to 0 - 6 V and returned to 0. 5 V when WLA is turned of f (capacitive coupling between WLA and the cell capacitor). The cell to be read is at 2. 5 V and is brought to 3.5 V at the end of 'he interrupting memory cycle. In Fig. 18B the cell to be refreshed (CELLA) is at 2.5 V and the cell to be read in the interrupting cycle (CELLB) is at 0.5 V.. The reading returns CELLB to 0 V.

In Figs. 1SC-D the refresh cycles are interrupted by write cycles. in Fig. 18C the cell to be refreshed (CELLA) is at 2.5 V and remains at the same level after interrupting the refresh before WIA has been brought to 5 V. CELLB starts at its previous value of 0. 5 V and at the end of the write cycle is at 3.5 V. In Fig. 18D CELLA remains at the 0. 5 V it was at the start'of the refresh cycle, while the voltage of CELLB is changed from its previously stored value of 3.5 V to 0 V. In the write operation a signal COPLIN occurring at the same time as the signal COPLN is applying the input data to the bottom segments bit '- lines BLRO and BLWO. As can be seen these bit lines proceed in time the bit lines BLR and BLW in the 41---op 53 segment, contrary to the situation in Figs. ISA-B, where B!R and BL-W come first.

Figs. 18A-D illustrate how a memory cycle can interrupt a refresh cycle without destroying the stored data for an arbitrarily chosen time in the refresh cycle. In the example a previously degenerated cell voltage was chosen (some time after last refresh). An incomplete refresh due to an interrupting memory access will cause further degeneration until new uninterrupted re-f--ash cycle is performied. The address o-L2 the cells of the interrupted refresh would be retained during the memory access (accesses) and would be the first address used when the refresh cycle is resumed. With the very low refresh frequency compared to the high memory access rate very large blocks of data with very short interruptions would have to be transferred before the refresh integrity would be in danger. As a backup for all eventualities a long delay circuit could be used to interrupt the flow occasionally.

* Related to the idea of interrupting an ongoing refresh cycle at any time without destroying the stored- data is the concept of performing a fast burst read. A fast burst read operation involves reading data from the memory cell without complet-ing a refresh cycle following the read operation. In the DRAM of the present invention, this is possible sincethe data stored in a memory cell is not destroyed when the contents of the cell are read, eliminating the need to refresh the contents of the cell. By reading data from a memory cell without a subsequent refresh cycle, the amount of time needed to perform a read operation is typically halved. Thus, the possible frequency of read cycles is doubled, improving the efficiency of memory operat-ion. of course, -i'L',. should be understood that a standard DRAM refresh cycle must still be performed. The fast burst read operation essentially turns the customary read/refresh operation into a stric-tC read operation, allowing a read operation to be performed without an associated refresh oPeration.

DRAM. Cluster Cell In another embodiment, -'---he DR1,-M o-.L:' t-he Dresent invention may include more than cne storage capacitor associated with a single pair of read and write transistors. Fig. 19A shows a simpli.-,Eied circuit diagram 1900 of a "cluster cell" using four storage nodes Cl-C4. instead of having one storage node at the common (COM) node, any of the four node caDacitors CI-C4 can be connected to COM during a read or write o'Deraticn. The COM node can be looked upon as a very short bit line in a conventional DRAIM and all the techniques for layout and processing can be applied to the cells on the COM "bit line."

The nurber of cells that can be placed in the clu..3ter -ion between the memory cell capacitance and der)ends on the rela4l.'he capacitance of the COM node. The diagram in Fig. 19A shows four memory cells in the cluster, merely by way of example. The COM node capacitance includes the junction capacitance of the COM bit line, the source junction of write transistor 1902, the drain junction on CLMPCOM transistor 1904 and the gate of sense transistor 1906. The COM bit line may have relatively high resistance, so minimum active junctions can be used.

in operation, '"he primary DRAM cell, which normally reads and restores the charge stored in a capacitor connected to the COM node, has been modif ied so that any of the cell capacitors CI-C4 on the COM bit line can be connected to COM bef ore the reading and restoring of the charge on the selected cell. Due to charge sharing between a cell capacitor and the COM node, there is a slight modification of the cell voltage after the connection. The actual vol tage on the COM node depends on whether a ONE or a ZERO was read in the previous cycle. In order to make the effect of the charge sharing more predictable, the COM node is clamned to a fixed level between cycles. Once the cell is connected to COM, a normal cycle is then execu'ted. Fig. 19A also shows a portion of a word line decoder 1910 coupled to master word line WL and four slave word 3 5 lines WLC1. WLC2, WLC3 and WLC4. A number of global signals are also shown that define active time for the different horizontal lines. An active subcell is selected by a decoder (not shown) and applied at a specified time on ClTIMEB, C2TIMEB, C3TIY.---B or C4TIMEB. An amplifier 1920 is also shown in Fig. 19A, and its performance is similar to that described above.

Fig. 19B shows the timing relationship between the different signals. For example, the reading and restoring of a ZERO and oif a ONE is shown. The charge sharing effect between the cell capacitor and the COM node, when the COM node caDacitance when the capacitances are equal is also i Llustra-11--ed.

With the COM node cla=ed at 2 V between cycles, the displacement on the stored value while reading is the same, but in opposite directions. The cell current to BLR is normally limited by the low WL voltage on read transistor i908 and therefore increases less than linearly with the cell voltage.

The cost of adding a separate write word line 1920 is minimal to address this limitation, as shown in an alternate cluster cell configuration 1950 in Fig. 19C. The high WLR voltage (5 V) renders the cell current quadratically dependent of the active voltage on sense transistor 1906 (the COM node). The much higher read current on BLR makes for an even faster read operation and higher margins. The timing diagrams for the circuit of Fig. 19C is shown in Fig. 19D.

7f a cluster cell is read before a refresh cycle, the cell voltage will have been modified by charge sharing with the COM node, as described above. A second reading may still be correct if the COM node capacitance is very small compared to the cell capacitor. The mod ifica,"Zion is however progressive from cycle to cycle, so a burst read could only be permitted once or twice 'before the cell is refreshed. The same restriction applies to interrupted refresh cycles for reading the same or other addresses and writing to other addresses. Note that the two tier configuration described earlier is still valid in't-he cluster cell as long as the second tier design 4. ffect. takes into account the charge shar-ng e- 56 The above description is illusnraOve and rsstrictiva. Kany variatians of the apparent to those of skill in the art upon review of disclosure. The scope of the invention should, therefore, be determined not with reference to the above descrip-tEicn, hut instead should be determined with referz.-,ce to the appended claims along with their full scope of equivalents.

S7 -5-8-

Claims (5)

CLAIMS:

1. A method of forming a submerged transistor and a capacitor on a semiconductor support wafer comprising the steps of:

forming a first and second diffusion area in the support wafer; implanting a first gas in the support wafer to form a dielectric layer coupling said first and second diffusion areas.

etching an isolation trench in the support wafer; implanting a dopanu layer to form the capacitor; growing an epitaxial layer; and implanting a second gas to separate said epitaxial layer from the suppo2t wafer.

2. The method of claim 1 further comprising the step of heat-treating the support wafer.

3. The method of claim 1 wherein said step of implanting a first gas in the support wafer further comprises the step of implanting oxygen to form said dielectric layer.

4. The method of claim 1 wherein said step of implanting a first gas in the support wafer further comprises the step of implanting nitrogen to form said dielectric layer.

5. The method of claim 1 wherein said step of implanting a dopant layer further comprises the step of implanting an N+ layer.