Partial access mode for dynamic random access memory

Abstract

Some embodiments provide a method to reduce the refresh power consumption by effectively extending the memory cell retention time. Conversion from 1 cell/bit to 2N cells/bit reduces the variation in the retention time among memory cells. The conversion can be realized very simply from the structure of the DRAM array circuit, and it reduces the frequency of disturbance and power consumption by two orders of magnitude. On the basis of this conversion method, some embodiments provide a partial access mode to reduce power consumption dynamically when the full memory capacity is not required. One bit data may be stored into 1 cell for a normal operation mode and stored into 2N cells for a self refresh operation mode for a first partial access mode, while one bit data may be stored into 2N cells for both normal and self refresh operation modes.

Description

BACKGROUND

The present invention relates generally to dynamic random access memory (DRAM). Demands have been placed on DRAM to not only have increased memory capacity and data transfer speed, but also reduced operating and standby currents. When a system uses a DRAM, a refresh operation is necessary because of its data retention time restriction; each bit of the DRAM is stored as an amount of electrical charge in a storage capacitor that is discharged by the leakage current. Power consumption for the refresh operation increases in proportion to the memory capacity.

Japanese Patent application JP 2006-120251 provides a semiconductor memory self-refresh control method which can extend a refreshing cycle and significantly reduce the current consumption of a DRAM when a self-refresh operation is carried out. In that reference, a storage area, which is a memory cell (MC) group on the prescribed number of word lines (SWL) which are data storage objects of the whole memory arrays, is set independently of a copying area which is a memory cell group on the word lines, which are the copying destinations of all the data in the storage area. Prior to the execution of a self-refresh operation, a bit information-copying operation to one memory cell or to each of the memory cells in the copying area in the same bit line (BL) or in the same bit pair lines is executed in terms of each memory cell in the storage area as the copying source. Then, line addresses are sequentially designated in terms of the storage area as the refreshing object, and the corresponding word lines are selected and driven, and, at the same time, one or more word lines of the corresponding copying area are selected and driven to execute the self-refresh operation.

SUMMARY

Certain aspects of embodiments disclosed herein by way of example are summarized below. It should be understood that these aspects are presented merely to provide the reader with a brief summary of certain forms an invention disclosed and/or claimed herein might take and that these aspects are not intended to limit the scope of any invention disclosed and/or claimed herein. Indeed, any invention disclosed and/or claimed herein may encompass a variety of aspects that may not be set forth below.

In some embodiments, a memory device is provided. The memory device includes a plurality of memory cells. The memory device further includes a controller controlling storage of data in the plurality of memory cells. The memory device is provided so that the controller operates a first access mode which stores one bit data into 1 cell of the memory cells, and operates a second access mode which stores one bit data into 2^N cells of the memory cells, where N is a natural number.

In some embodiments, the memory device is provided so that the first access mode is associated with a normal operation mode.

In some embodiments, the memory device is provided so that the second access mode is associated with a self refresh operation mode.

In some embodiments, the memory device is provided so that the controller is configured to operate in the first access mode which activates a word line to access the one bit data and in the second access mode which activates 2^N word lines to access the one bit data.

In some embodiments, the memory device is provided so that the controller is configured in a first partial access mode to operate the first access mode in a normal operation mode and the second access mode in a self refresh operation mode. The controller is further configured in a second partial access mode to operate the second access mode in both the normal operation mode and the self refresh operation mode.

In some embodiments, the memory device is provided so that the controller is configured in a third partial access mode to operate the first partial access mode in both the normal operation mode and the self refresh operation mode.

In some embodiments, the memory device is provided so that the controller storing causes the power consumption of the device to be a first reduced amount for the first access mode, and causes the power consumption of the device to a second reduced amount for the second access mode, the second reduced amount less than the first reduced amount.

In some embodiments, a memory device is provided. The memory device includes a plurality of memory cells. The memory device further includes a controller controlling storage of data in the plurality of memory cells. The controller is configured to store data in the plurality of memory cells as 1 data bit/cell for a normal operation mode and store data in the plurality of memory cells as 1 data bit/2^N cells for a self refresh operation mode in a first partial access mode. The controller is further configured to store data in the plurality of memory cells as 1 data bit/2^N cells for both a normal operation mode and a self refresh operation mode in a second partial access mode.

In some embodiments, the memory device is provided so that an address decoder is coupled to the controller and configured to select wordlines coupled to the plurality of memory cells to store the data in the plurality of memory cells.

In some embodiments, the memory device is provided so that the address decoder is configured to handle bits of a row address as “don't care” according to the first or second partial access mode.

In some embodiments, the memory device is provided so that the plurality of memory cells are included in memories of a memory module.

In some embodiments, a method for managing storage of data in a device is provided. The method includes storing one bit data into 1 cell of the memory cells when operating in a first access mode. The method further includes storing one bit data into 2^N cells of the memory cells when operating in a second access mode, wherein N is a natural number.

In some embodiments, the method is provided so that storing one bit data into 2^N cells of the memory cells comprises copying a portion of the one bit data into 2^N cells of the device.

In some embodiments, the method is provided to further include performing an all bank refresh operation responsive to a partial access mode exit operation, and storing one bit data into 1 cell of the memory cells.

In some embodiments, the method is provided so that all bank refresh operation responsive to the partial access mode exit operation correspond to exiting a waiting call state.

In some embodiments, the method is provided to further include entering the partial access mode responsive to a partial access mode entry operation corresponding to entering the waiting call state.

In some embodiments, the method is provided so that storing one bit data into 1 cell of the memory cells comprises activating a word line and wherein storing one bit data into 2^N cells of the memory cells comprises activating 2^N word lines.

In some embodiments, the method is provided so that storing the one bit data into 1 cell of the memory cells is associated with a normal refresh operation and wherein storing the one bit data into 2^N cells of the memory cells is associated with a self refresh operation.

In some embodiments, the method is provided so that storing the one bit data into 2^N cells of the memory cells is associated with both a normal refresh operation and a self refresh operation.

BRIEF DESCRIPTION OF THE DRAWINGS

These and other features, aspects, and advantages of the present invention will become better understood when the following detailed description of certain exemplary embodiments is read with reference to the accompanying drawings in which like characters represent like parts throughout the drawings, wherein:

FIG. 1 is a DRAM comparison table in accordance with some embodiments of the present disclosure.

FIG. 2A shows a schematic cross section of a memory cell in accordance with some embodiments of the present disclosure.

FIG. 2B shows a DRAM chip layout in accordance with some embodiments of the present disclosure.

FIG. 2C shows an array circuit in accordance with some embodiments of the present disclosure.

FIG. 2D shows an array circuit waveform in accordance with some embodiments of the present disclosure.

FIG. 3A shows a memory cell structure and leakage current path in accordance with some embodiments of the present disclosure.

FIG. 3B shows a chart of measured error ratio with temperature dependency in accordance with some embodiments of the present disclosure.

FIG. 3C shows a fail bit map in accordance with some embodiments of the present disclosure.

FIG. 3D shows a chart of cumulative fail bit count and the density distribution in accordance with some embodiments of the present disclosure.

FIG. 4A shows a 1-Bank PASR configuration in accordance with some embodiments of the present disclosure.

FIG. 4B shows a chart of measured current value for a 1-Bank PASR configuration in accordance with some embodiments of the present disclosure.

FIG. 4C shows a diagram of a semiconductor device in accordance with some embodiments of the present disclosure.

FIG. 4D shows a schematic view of a DRAM in accordance with some embodiments of the present disclosure.

FIG. 4E shows a diagram of cell phone system and RAM activity in accordance with some embodiments of the present disclosure.

FIG. 5A shows a partial access conversion table in accordance with some embodiments of the present disclosure.

FIG. 5B shows a partial access conversion table in accordance with some embodiments of the present disclosure.

FIG. 6A shows a partial access configuration in accordance with some embodiments of the present disclosure.

FIG. 6B shows a chart of estimated current value of a partial access configuration in accordance with some embodiments of the present disclosure.

FIG. 6C shows a diagram of RAM activity in accordance with some embodiments of the present disclosure.

FIG. 6D shows a time variation of the power consumption in the low power consumption mode in accordance with some embodiments of the present disclosure.

FIG. 6E shows an example of the relationship between the refresh operation interval and the number of memory cells for one bit data in accordance with some embodiments of the present disclosure.

FIG. 6F shows an example of the relationship between the standby current and the storage capacity of the holding area in accordance with some embodiments of the present disclosure.

FIG. 6G shows a relationship between a standby current and storage capacity of the data holding area in accordance with some embodiments of the present disclosure.

FIG. 7 shows a partial access specification in accordance with some embodiments of the present disclosure.

FIG. 8A shows a conventional X-decoder in accordance with some embodiments of the present disclosure.

FIG. 8B shows a partial access X-decoder in accordance with some embodiments of the present disclosure.

FIG. 8C shows a partial access configuration circuit modification in accordance with some embodiments of the present disclosure.

FIG. 9A shows partial access waveforms in accordance with some embodiments of the present disclosure.

FIG. 9B shows functional waveforms of a partial access configuration in accordance with some embodiments of the present disclosure.

FIG. 10A shows an example of the relationship for a partial access mode between normal operation current and X-addressing interval in accordance with some embodiments of the present disclosure.

FIG. 10B shows a chart of estimated current values for different partial access modes in accordance with some embodiments of the present disclosure.

FIG. 11 shows a partial access specification in accordance with some embodiments of the present disclosure.

FIG. 12A shows functional waveforms of a partial access configuration in accordance with some embodiments of the present disclosure.

FIG. 12B shows functional waveforms of a partial access configuration in accordance with some embodiments of the present disclosure.

FIG. 13 shows an example of the relationship for different partial access modes and the relationship between normal operation current and X-addressing interval in accordance with some embodiments of the present disclosure.

FIG. 14A shows a time variation of the power consumption for a partial access mode in accordance with some embodiments of the present disclosure.

FIG. 14B shows a time variation of the power consumption for a partial access mode in accordance with some embodiments of the present disclosure.

FIG. 15 shows a partial access configuration circuit modification in accordance with some embodiments of the present disclosure.

FIG. 16A shows a memory module in accordance with some embodiments of the disclosure.

FIG. 16B shows a memory capacity allocation for the a memory module in accordance with some embodiments of the disclosure.

FIG. 17 shows a diagram of cell phone system and RAM activity in accordance with some embodiments of the present disclosure.

FIG. 18A shows a cell circuit in accordance with some embodiments of the present disclosure.

FIG. 18B shows a graph of the measured waveforms in accordance with some embodiments of the present disclosure.

FIG. 18C shows a graph of the measured signal amount in accordance with some embodiments of the present disclosure.

FIG. 19A shows a fail bit map in accordance with some embodiments of the present disclosure.

FIG. 19B shows a graph of error ratio in accordance with some embodiments of the present disclosure.

FIG. 20A shows a comparison of current between PASR and PAM in SELF operation in accordance with some embodiments of the present disclosure.

FIG. 20B shows a comparison of current between PASR and PAM in normal operation in accordance with some embodiments of the present disclosure.

FIG. 21A shows a delay circuit with copy flag implementing the copy operation of PAC in accordance with some embodiments of the present disclosure.

FIG. 21B shows waveforms indicating an operation of the delay circuit in FIG. 21A in accordance with some embodiments of the present disclosure.

DETAILED DESCRIPTION

One or more specific embodiments of the present invention will be described below. These described embodiments are only exemplary of the present invention. Additionally, in an effort to provide a concise description of these exemplary embodiments, all features of an actual implementation may not be described in the specification. It should be appreciated that in the development of any such actual implementation, as in any engineering or design project, numerous implementation-specific decisions must be made to achieve the developers' specific goals, such as compliance with system-related and business-related constraints, which may vary from one implementation to another. Moreover, it should be appreciated that such a development effort might be complex and time consuming, but would nevertheless be a routine undertaking of design, fabrication, and manufacture for those of ordinary skill having the benefit of this disclosure.

Some embodiments provide a method to reduce the refresh power consumption by effectively extending the memory cell retention time. Conversion from 1 cell/bit to 2^N cells/bit reduces the variation in the retention time among memory cells. Although active power increases by a factor of 2^N, the refresh time increases by more than 2^N as a consequence of the fact that the majority decision does better than averaging for the tail distribution of retention time. The conversion can be realized very simply from the structure of the DRAM array circuit, and it reduces the frequency of disturbance and power consumption by two orders of magnitude. On the basis of this conversion method, some embodiments provide a partial access mode to reduce power consumption dynamically when the full memory capacity is not required.

DRAM has served as the main role of storage in computer systems including high performance systems, personal computers, and mobile phones for more than 30 years. The memory capacity of a DRAM has increased to meet system demands, supported by the development of a semiconductor process technology that follows Moore's law, whereby the number of elements on a fixed silicon die doubles every 18 months. In fact, needs from the system side have included not only memory capacity but also data transfer speed, operation current reduction, and standby current reduction. In a DRAM, each bit is stored as an amount of electrical charge in a storage capacitor, and the increase in memory capacity has directly caused two problems: disturbance and power consumption. Both of these problems are attributed to the rewrite operation to a memory cell associated with a finite data retention time.

In particular, standby power consumption has become one of the most serious problems for using a DRAM in mobile applications. Following the constant-electric-field scaling theory, the voltage should be lowered at the same rate by which the dimensions are reduced. However, the rewrite voltage is saturated by the difficulty associated with the real operation. This means that the standby power increases with the memory capacity.

In some embodiments, a method is provided to reduce the refresh current in a DRAM by extending the retention time effectively when the amount of the data to be stored is small. This low-power mode may be referred to as the partial access mode (PAM). The retention time has been shown to exhibit both tail and main distributions. Most of the cells belong to the main distribution and have retention times significantly higher than the product specification. Only a minor portion suffers from increased leakage. The short retention time with an extremely low probability (approximately 5σ from the median for the retention tail) determines the refresh interval of the memory cell. Conversion from 1 cell/bit to 2^N cells/bit reduces the variation in the retention time among memory cells. Although the active power increases by a factor of 2^N, the refresh time increases by more than 2^N as a consequence of the fact that the majority decision does better than averaging for the tail distribution of retention time. The conversion can be realized very simply from the structure of the DRAM array circuit. This method can reduce the frequency of the disturbance and its power consumption by two orders of magnitude.

A DRAM article of manufacture according to some embodiments is fully compatible with a conventional DRAM. In its usual operating mode, the full memory capacity is used. In the PAM, the capacity is limited to 2^N of the total capacity. However, memory cells are fully used to share the storage charge to extend the retention time.

DRAM Operation and Power-Reduction Mode: DRAM Refresh Operation

DRAM memory capacity has been increasing, even though its die size has almost remained constant, as summarized in FIG. 1. In 2011, a 2-Gb DRAM was fabricated by a 30-nm [minimum feature size (F) value] process and was placed on the market. A DRAM stores a single bit in a memory cell as an amount of electrical charge on a storage capacitor. Charge is lost by the current of the p-n junction, sub-threshold current, and gate-induced drain (GIDL). FIG. 2A shows a schematic of a DRAM cell with a leakage current. The DRAM cell has a capacitor electrode 200, an accumulated charge 210, a bit line 220, a word line 230, and a leakage current 240. The loss of charge means that a DRAM requires a rewrite operation before the memory cell loses its storage charge. This rewrite operation is called refresh. Refresh is performed by issuing an auto-refresh command (AREF). Because refresh is a type of disturbance in the system where sense amplifier (SA) activation, pre-charging, and read or write operations are forbidden, the frequency of the AREF command should be minimized.

In this disclosure, reference will generally be made to a refresh operation using a 256-Mb DRAM as an example. However, the present disclosure is relevant to any other DRAM capacity as would be understood by those of ordinary skill in the art. An example of a 256-Mb DRAM with 4 banks is shown in FIG. 2B. The refreshing of all memory cells is completed by issuing 2¹³=8,192 AREF commands, called an 8-K AREF operation. The data retention time of each memory cell is expressed by t_ret. The minimum retention time of the memory cells, during which all memory cells maintain their own charges, should be longer than 64 ms, standardized by the cell refresh time t_ref. Thus, the maximum time interval of an AREF command, t_REF, is 64 ms/2¹³=7.8 μs in an 8-K AREF operation.

The 256-Mb DRAM in FIG. 1 consists of four banks, each of which is composed of 16×16 mats. One mat is 256 Kb, including 512 wordlines (WLs) and 512 bitlines (BLs), as shown in FIG. 2A. In an 8-K AREF operation, a single AREF command is accomplished by one refresh operation for all four banks. In each bank, one refresh operation is applied to 16 mats located in the WL direction. The corresponding 16 WLs are selected at the same time, and memory node voltages are read to the BLs through transfer NMOSFETs. These signals are amplified by sense amplifiers that perform rewrite operations of 8-K memory cells. Once an SA is activated, 16 WLs allocated on the same X address of the 16 mats in one bank are selected at once, and data are read to 8-K BLs. This number, 8-K, is called the “page size” and shows the maximum data size written or read during one SA activation. If the data size for writing is greater than this page size, the system must issue a pre-charge command and another activate command.

FIG. 2C shows an array circuit in accordance with some embodiments of the present disclosure. The array circuit includes a sense amplifier 250, common power supply 252, mat 254, sub word driver 256, transfer NMOS and memory cell capacitor 25, and other elements there illustrated. As shown in the figure, each sense amplifier SA 250 located next to the mat 254 has common power supplies CSP 252 for a PMOSFET and CSN for an NMOSFET; these perform pre-charging and SA amplification, as illustrated in FIG. 2D. In the pre-charge mode, CSP, CSN, and the BL (BLT and BLB) voltages are set to VARY/2, which is half of the voltage of VARY. The high (H) and low (L) levels of the BL voltage are VARY and VSS (0 V), respectively. First, the SA amplification voltage VDD is applied to BLEQB, which reduces BLT/B from VARY/2. Then, the selection of the WL connects the memory node to the BL. The charge in the storage capacitor changes the BL voltage to a higher or lower value than VARY/2. The difference from VARY/2 is called the “signal amount.” After the voltage of the BL has stabilized, CSP is changed to VARY, and CSN is changed to VSS at the same time. The SA increases the voltage difference between BLT and BLB to VARY. In the pre-charge mode, the WL voltage is changed to VSS to reduce the memory node from the BL voltage, and then, CSP and CSN are reduced from VARY and VSS, respectively. Finally, BLEQB is lowered to set CSP, CSN, and the BL voltage to VARY/2. The WLs are driven by a sub word driver, which receives RXT signals and FXT signals. The RXT signals and FXT signals can be produced from X address signals. The RXT signals and the FXT signals can be called main word lines and word line driving signals respectively.

DRAM Oration and Power-Reduction Mode: Charge Retention Time

The AREF interval time, t_REF, should be as long as possible to meet the system requirements. For example, a user may set this time to 7.8 μs, which is the maximum value defined from the specification, t_ref of 64 ms. The retention time t_ret depends on the characteristics of a memory cell that has a leakage current that reduces the charge on the storage capacitor. This leakage current is caused by the diffusion and generation of electrons and holes at the p-n junction in the silicon substrate, the sub-threshold current, and GIDL. FIG. 3A shows the structure of the storage capacitor and transfer NMOSFET in a memory cell. The leakage current is influenced by the voltages of the WL, BL, and body (p-well). This causes a variety of t_ret values among several conditions. There are two states that hold the storage charge in the memory cell, described as follows.

State 1: the observed memory cell is not selected, and all memory cells in the same mat are not selected. WL voltage=VSS, and BL voltage=VARY/2. If the memory node voltage is H in the observed memory cell, leakage current flows from the memory node to the body. The left panel in FIG. 3A shows this state. The memory node voltage H is changed to L in t_ret, which is the specific finite time of the cell. This destruction mode can be referred to as the all cell high (ACH) pause. “All” in ACH means that the memory node voltages in all DRAM memory cells are H at the last restore time. If the memory node voltage is L, there is no destruction to H, as the voltage of the body is the same voltage, L.

State 2: the observed memory cell is not selected, and one of the other WLs in the same mat is selected. WL voltage=VSS, and BL voltage=H or L. If the memory node voltage is H and BL voltage is L, leakage current flows not only into the body but also into the BL. The electric field induces a current into the BL. This destruction mode can be referred to as ACH disturb. If the memory node voltage is L and BL voltage is H, only leakage current into the BL appears, and there is no current into the body. This can be referred to as all cell low (ACL) disturb. The case of ACH disturb is the worst, because there are two current paths. The right panel in FIG. 3A shows this state.

The measured t_ret data in state 1 is obtained from the time when the fail-bit judgment of the read command appears after a pause time in the pre-charge mode from the write operation of memory node voltage H. FIG. 3B shows a chart of measured error ratio with temperature dependency in accordance with some embodiments of the present disclosure. FIG. 3B expresses the pause time. For a 256-Mb DRAM, the fail-bit count of ACH pause denotes the number of cells of signal L in 256 Mb after the write operation of H. In FIG. 3B, 1 cell/bit measurement data show the error ratio of the 256-Mb DRAM from 100 ms to 100 s, where the error ratio is defined by the number of fail bits divided by 256 Mb. The error ratio is converted with an inverse cumulative distribution function to check whether the distribution coincides with a standard normal distribution. If the distribution appears as a standard normal distribution, the line becomes straight. FIG. 3B shows that all lines have a kink, indicating the existence of two kinds of standard normal distributions, tail and main distributions. The variation of t_ret is greater than three orders of magnitude at a temperature of 85° C., and t_ret has one order of variation in the area whose error ratio is less than 0.1%. This means that, although t_ret of 99.9% for 256 Mb is longer than 3 s, t_ret,min of the 256-Mb DRAM is 100 ms. This difference is caused by the tail distribution, whose error ratio is less than 0.1%. Furthermore, the variation among temperatures is very large. According to FIG. 3B, t_ret,min of 100 ms at a temperature of 85° C., which is the worst value, is improved to 1 s at 45° C. In the DRAM specification, t_ref ensures a t_ret,min of 64 ms at temperatures between 0° C. and 85° C. A DRAM supplier can replace the worst 20 or 30 memory cells with other good ones prepared in advance. FIG. 3B shows the effect of 20-bits replacement.

FIG. 3C shows a fail bit map of 256-Mb DRAM in each pause time after H write operation to all memory cells at 85° C., which may represent the worst condition for the leakage current in DRAM specifications. Cumulative distribution in FIG. 3D shows the fail bit counts of 256-Mb DRAM in each pause time after H write operation to all memory cells at 85° C. Fail bit appears at pause time=100 ms, but error ratio (=fail bit count/total capacity) begins to increase at pause time=4 s, which is 40 times larger than t_ret,min. Density distribution in FIG. 3D is defined from the difference of fail bit counts and of pause time (ΔCumulative distribution/Δpause time), which clarifies the distribution feature. The variation of t_ret is larger than 3 orders and the average value of t_ret is around 10 s. This indicates that the t_ret,min of 256-Mb DRAM is 100 ms although t_rets of almost all memory cells in 256-Mb is longer than 4 s. This is caused by the t_ret variation among memory cells.

DRAM Operation and Power-Reduction Mode: DRAM Standby Current

A DRAM consumes standby power, mainly by refresh, even when there are no read and write accesses to memory cells. The storage capacitor needs around 20 fF, independent of F, as t_ref is always 64 ms through all generations. If the minimum feature size is cut in half by process improvement, the memory capacity of the DRAM for the same die size increases by a factor of four. The voltage to rewrite data should be lowered to maintain the same power consumption. However, the BL rewrite voltage VARY saturates at around 1 V caused by real operation difficulties. FIG. 1 shows the external voltage saturation which is caused by SA amplification where VARY is the source voltage. Therefore the amount of refresh current increases by a factor of four as well. In fact, the total refresh current is less than this, because the other currents used to drive signals for the refresh operation become less than half owing to the miniaturization. Taking this factor into account, the standby power increases in every generation.

If there is no access to the DRAM, t_ret is determined only by ACH pause (1) without disturb (2) discussed previously. The DRAM has a SELF mode in the specification, which means that the DRAM performs a refresh operation by itself in the longer interval t_ref,etl. The oscillator circuit in the DRAM can adjust the period for t_REF in the SELF mode. This adjustment based on the measurement of the manufactured DRAM enables a greater reduction in power consumption in the SELF mode. This t_REF adjustment is set on the basis of the t_ret-measured DRAM data that were compiled during the manufacturing process of the silicon wafer. If it is known in advance that the system has no access to the DRAM in a certain interval, a user issues a SELF ENTRY command, which is the longest time interval during which storage charge is not lost. Once the DRAM receives a SELF ENTRY command, it remains in the SELF mode until receiving a SELF EXIT command, and it accepts no commands except SELF EXIT.

Conventional PASR Method

System side requests contradictory performance, e.g., power consumption should be reduced in spite of the increase in the memory capacity. Recent smartphones require a long battery life as well as a significant amount of memory to enable better video on a more vivid and detailed screen. Even in the standby mode, smartphones periodically check for incoming calls and utilize a certain amount of memory. The system demands much greater power reduction than when it is in the SELF mode. The conventional power reduction method is partial array self refresh: PASR, as shown here for a 256-Mb DRAM with four banks. The refresh operation is stopped for two or three of the four banks. Both the memory capacity and the refresh current decrease by a factor of two or four in the SELF mode. PASR has no regulation in the normal mode and two problems from the system-side viewpoint may exist.

Problem 1: there is mismatch in the number of banks between the AREF operation in the normal mode and PASR. FIG. 4A is a block diagram that illustrates a one-bank PASR. At the first refresh after SELF EXIT, an AREF command is issued to select WLs of all four banks, although three of those do not store data. This means that power consumption is four times larger than needed, as a refresh operation for three banks is not necessary. FIG. 4B shows power consumption using PASR as given by the measured standby current of 256-Mb DRAM. The memory capacity in SELF is limited to 256-M/n where n=2^N. The constant power consumption in AREF is estimated from the maximum AREF interval t_REF.

Problem 2: the system usually activates the SA for all banks and accesses four pages (8-K×4=32-K size) of the four banks, because the frequency of the disturbance factor, i.e., pre-charge, should be less. However, the PASR mode cannot preserve all 32-K data in four banks, so four pages must be activated separately with four pre-charge operations before entering the SELF mode to preserve the 32-K accessed data.

FIG. 4C illustrates an exemplary semiconductor device 10 implementing some of the features just described. The semiconductor device 10 can be called a semiconductor chip. A memory module includes a plurality of semiconductor chips like SIMM (Single Inline Memory Module), DIMM (Dual Inline Memory Module) and RIMM. The memory module can be mounted on a mother board so that the memory chip is electrically connected to a CPU including memory control function or a memory controller chip, each of which is mounted on the mother board. Control circuit 13 controls the operation of the semiconductor device 10. The control circuit 13 is provided in the device 10 and responds to read or write command input from outside of the device 10 for conducting write or read operation from or to the memory cells. The control circuit 13 controls a partial access based on the information (data set) stored in the mode register 14. The control circuit 13 generates a burst signal, selection setting signal N0-N4, a copy pulse CP and other control signals. The control circuit 13 receives a burst signal stop signal BURST_STP and so on. Mode register 14 holds the data set, which is information that defines the behavior of 10 semiconductor device. Row address buffer refresh counter 15 temporarily holds the address signal inputted from the outside, based on an output from the control circuit 13, to generate an address signal for refresh operation. The row address buffer refresh counter 15 functions as an address signal generation circuit. Row decoder 17 decodes the address signal from the row address buffer refresh counter 15, and activates the word line corresponding to the decoded address signal. In other words, the row decoder 17 functions as a word line activation signal generation circuit. The row decoder 17 activates plural word lines at the same time as the access unit based on control of the control circuit 13.

Memory cell array 18 has plural banks BANK_A-D. The Banks BANK A-D are formed on the same substrate. Each bank is divided into a plurality of blocks. Each block includes a plurality of memory cells connected to plural bit lines and plural word lines. Each of memory cells is assigned a unique address and selected by the address signal.

In order to reduce the power consumption in low power consumption mode, semiconductor devices mounted on the mobile phone performs a self refresh operation which is called a partial array self refresh (PASR). The PASR performs self refresh only for a portion of the memory cell array. For example, the PASR performs the self refresh for only one bank against a memory cell array having four banks BANK_A-D. The method allows its power consumption needed for self refresh operation to reduce ¼. The semiconductor device 10 has a partial access mode which sets a portion of each of banks A-D as accessible area (data holding area) and operates self refresh only for the set portion of each of banks A-D.

FIG. 4D shows an example of a DRAM, which has 256M bit capacity and 32-bit input and output data configurations. The DRAM is capable of selecting each memory cell using Y address Y0-Y7 and X address X0-13 (mat address X0˜X8, X9˜X13 mat outside address). The DRAM has a page size of 8 kbit and a refresh frequency of 8 Kref.

FIG. 4E shows a cellular phone periodically communicating with the base station in standby state, that is, low power consumption mode. For example, after the entry into the low power consumption mode, the portable phone repeats transmission (communication access period 280 μs) with the base station every 2 seconds. As shown in the right column of the figure, the semiconductor device mounted in the mobile phone repeats self refresh (280 μs) every 2 seconds.

Partial Access Mode

Embodiment 1

Mismatches between the AREF operation in the normal mode and the preserved bank data in PASR have been explained. PAM eliminates these mismatches and uses all memory cells efficiently. One characteristic of the PAM is that it holds data using 2^N cells/bit to extend t_ret. Another characteristic is the control method between 1 cell/bit and 2^N cells/bit that is located higher in the hierarchy than the control of the normal and SELF modes. The PAM reduces the refresh operation frequency for both the normal and the SELF modes. As a result, the AREF command frequency in the normal mode and power consumption in the SELF mode are reduced at once. In contrast, PASR reduces the power consumption in the SELF mode only.

A PAM ENTRY operation indicates a conversion from 1 cell/bit to 2^N cells/bit. This may simply be a copy operation from the memory cell connected with one WL to 2^N−1 memory cells connected with 2^N−1 WLs in the same mat. FIGS. 5A and 5B show how this operation is simply achieved by the DRAM array architecture in FIG. 2C. This operation is performed by only a delayed WL selection. A PAM ENTRY operation is completed through 8-K/2^N copy operations applied to all memory cells in the 256-Mb DRAM.

In the PAM according to some embodiments, 2^N memory cells compensate for each other's storage charge by selecting 2^N WLs at the same time in both the normal and the SELF modes, as shown in FIG. 5. In the PAM according to other embodiments, 2^N memory cells compensate for each other's storage charge by selecting 1 WL in normal mode and 2^N WLs in the SELF mode for a first PAM mode, and by selecting 2^N WLs at the same time in both the normal and the SELF modes for a second PAM mode, as will be discussed later in the present disclosure. The PAM according to the disclosed embodiments enable an extension of t_ret, which is also discussed later in the present disclosure.

A PAM EXIT operation indicates a reconversion from 2^N cells/bit to 1 cell/bit, which is necessary because the storage charge in one memory cell is not enough to read one bit. 2^N times the charge in the PAM is used for one bit. A PAM EXIT operation is simply continuous 8-K/2^N refresh operations to all memory cells. The array circuit in FIG. 2C illustrates this operation very simply.

One refresh operation takes 100 ns, meaning that continuous 8-K refresh operations to all 256-M memory cells take 8-K×100 ns=819.2 μs. The turn-around time for PAM ENTRY or PAM EXIT is approximately 819.2 μs/2^N, because the memory capacity is changed from 256-M to 256-M/2^N as well. The PAM ENTRY operation can be referred to as a partial access conversion, being a continuous copy in the entire capacity, and the PAM EXIT operation can be referred to as a partial access reconversion, being a continuous refresh in the entire capacity.

FIG. 6B shows power consumption using PAM as given by the measured standby current of 256-Mb DRAM. The memory capacity in SELF is limited to 256-M/n where n=2^N. The constant power consumption in AREF is estimated from the maximum AREF interval t_REF.

Referring now to FIG. 5A, the memory cells are indicated by ∘ or •. The word lines are provided along the vertical direction in the figure and are connected to an associated one sub word driver (SWD) located above and below the figure. The sub word driver constitutes a part of the row decoder 17. The bit line is provided along the lateral direction of the figure and is connected to an associated one sense amplifier (SA) 19 located at the left of figure.

When entry to the partial access mode, as shown in the left panel of FIG. 5A, semiconductor device 10 copies the data of each memory cell (source) connected to the word line of the holding region to one or plural memory cell(s) (destination) corresponding to and connected to the word line in the copy area.

This is achieved by reading out data held by the memory cells connected to the source word line and to a corresponding bit line and amplifying the data on the bit line, and activating the word line connected to the destination memory cells (connected to the same bit line) in the same mat. The copy operation is carried out continuously until all the source data is copied to the destination memory cells. In the partial access mode, as shown in the center panel of FIG. 5A, the memory device simultaneous access to the plurality of memory cells that hold the same data. In other words, plural word lines connected to the plurality of memory cells are activated at the same time. For example, to activate at the same time eight word lines, row address X0, X1, X2 row address may be neutralized (Don't Care).

When exiting from the partial access mode, as shown in the right panel of FIG. 5C, all bank refresh is performed against all the memory cells in the chip as plural word lines as an access unit.

Referring now to FIG. 6A, a case of setting 32M bit memory cells as data holding area in the 256M bit DRAM is illustrated. Each bank of memory cells is set 8M bit memory cells as the holding area (address space being accessed), the remaining 48 M bit memory cells as the copy area. The refresh number, needed for refreshing all memory cells one time as a refresh target, can be reduced by limiting to the accessible area by this method. As a result, not only the refresh number of self refresh but also the interrupt counts can be reduced. In the case where small data transmission only takes place as the idle mode of the mobile phone, no problem occurs even if the accessible region is limited to a portion of the memory cell array. The data holding area is set in the semiconductor device 10. Also, the copy region other than the data holding area is set in the semiconductor device 10. The memory cells in the copy region are used as memory cells independently in the normal operation mode, and used for retaining data stored in the memory cells in the holding region in the partial access mode.

That is, in the partial access mode, the same data is held by plural memory cells (for example, eight memory cells), which are connected to different word lines one another and connected to the same bit line. During the partial access mode, the plural word lines connected to the plurality of memory cells holding the same data are activated/deactivated simultaneously. In other words, during the partial access mode, the word lines or memory cells constitutes an access unit for the memory. The number of memory cells (word lines) as the access unit is 2^N (N is a natural number). The number 2^N is achieved by “don't care” handling of some bits of one or more lower side of the row address signal.

As described above, the data retention time and the refresh cycle can be longer by holding the same data in plural memory cells. By this, the refresh interval of the self refresh and auto refresh can be increased and the current consumption required for these refresh can be reduced.

The retention time may be changed by varying the number of memory cells for one bit data. Therefore, even though the capacity of the memory cell array can be increased, by increasing the number of word lines (memory cells) as the access unit, power consumption required for the auto refresh and self refresh can be suppressed.

FIG. 6C shows that when the mobile phone enters to the low power consumption mode, the semiconductor device 10 enters to the partial access mode. By this, the semiconductor device 10 limits memories accessed when the mobile phone communicates with the base station to the memory cells for the data holding area. One bit data is held by the plural memory cells.

After the entry into the partial access mode, the semiconductor device 10 repeats periodically entry to and exit from the self refresh mode. The semiconductor device 10 repeats partial access period and self refresh period. The first entry to the self refresh mode may be performed upon completion of the copy operation to be performed at the entry to the partial access mode. During the partial access mode period, the semiconductor device 10 holds one bit data in the plurality of memory cells, so that retention time t_REF is longer than that of the related semiconductor device. For example, the retention time t_REF during the partial access mode is 512 ms, and the retention time t_REF during the self refresh mode is 6.0 s. Upon exit from the self refresh mode, all bank refresh is not performed. As a result, for example, the average power consumption is a 61.4 μA.

When the portable phone exits the low power consumption mode, semiconductor device 10 exits from the partial access mode. At that time, all bank refresh is performed for all the memory cell arrays 18. At this time, if there is being in the self refresh mode, the phone exits from the partial access mode after exiting from the self refresh mode. In this way, all bank refresh can be eliminated by performing all bank refresh when the phone exits from the partial access mode.

FIG. 6D shows a time variation of the power consumption in the low power consumption mode. It is assumed that the mobile phone communicates with the base station, by communication access period 280 μs and communication cycle 2.0 s.

The semiconductor device 10 performs the 64M bit partial access, the retention time t_REF in self refresh mode at the time is 6.0 s, and the refresh number (equal to the target word line number) is 1 Kref. Therefore, the refresh operation interval may be 6.0 s/1 K=6.0 ms or less. In this case, 5.86 ms is used for the refresh operation interval. The retention time t_REF in the partial access time is 512 ms, and refresh times as necessary in the meantime is a 1 Kref same. Therefore, the refresh operation interval may be any 512 ms/1K=512 μs or less. Here, a period in which communication with the base station is performed is 280 μs. Since the period is shorter than the refresh operation interval, no refresh (interrupt) is needed.

Power consumption may be increased in copying in the entry to the partial access mode and all bank refresh in exiting from the partial access mode. However, compared with the effect of reducing the power consumption of partial access mode, such increase can be negligible compared with the power consumption effect by the partial access mode.

FIG. 6E shows an example of the relationship between the refresh operation interval and the number of memory cells for one bit data (index N means that one bit data is held by 2^N memory cells). It is assumed that refresh operation interval=refresh cycle/access number of words=t_REF/(8K/2^N). In the case that the access period is 280 μs, when the N>=3, access period <refresh operation interval, resulting in that refresh operation is unnecessary. In self refresh period, self refresh period is longer than the refresh period access period.

FIG. 6F shows an example of the relationship between the standby current and the storage capacity of the holding area (access capacity). It is possible that the smaller the storage capacity of the holding area is, the longer the refresh period (t_REF) can be, resulting in reducing the refresh current consumption. As a result, the smaller the storage capacity of the holding area is, standby current can be also reduced.

FIG. 6G shows a relationship between a standby current and storage capacity of the data holding area (access capacity) of FIG. 6F as the standby current and the number of memory cells of one bit per (index N). As can be seen, when N>=3, the refresh current is negligibly small relative to the steady-state current (DC current).

Referring now to FIG. 7, the data set for the partial access mode is stored in the mode register 14. The control circuit 13 controls each unit of the semiconductor device 10 and implements the operation of the partial access mode based on the stored data set.

The data set for the partial access mode defines the copy area and the data holding area. The copy area and the data holding area are set by storing (changing) the data set for the partial access mode in the mode register 14.

As shown in this figure, the data set is defined such that the data holding area can be selectable by a volume unit. In particular, the data holding area can be selected from among all the banks (256M), 128M, 64M, 32M and 16M bit. Such selection can be implemented using signals XA0 (X0), XA1 (X1), XA2 (X2), XA3 (X3) . . . row address.

The copy area is set so that the memory cells in the copy area are used for retaining the data held by memory cells in the data holding area. Where the data holding area is set ½^N for all banks, the number of memory cells holding the same data is 2^N, that is, one memory cell in the holding area and 2^N−1 memory cells in the copy region. By this, the data retention time (retention time) t_REF can be increased while the storage capacity of the holding area is decreased. The example in this figure shows 8 Kref, 4 Kref, 2 Kref, 1 Kref, 512 ref. By this, the refresh interval of the refresh operation is 4 times when the storage area becomes a half, and 16 times when the storage area becomes ¼. Thus, it is possible to reduce its power consumption by lengthening the refresh interval operation, and reducing the number of refresh operations.

PAM Specification

FIG. 6A shows the block diagram for the PAM according to some embodiments, and directly explains the elimination of Problem 1 and Problem 2 discussed for PASR previously. The PAM enables a reduction in the refresh operation frequency by two effects: t_ret extension and reduction in the memory capacity. This is summarized in FIG. 7 as the PAM specification. Conversion from 1 cell/bit to 2^N cells/bit expands t_ret to 2^N×64 ms at least. The total number of refresh operations to restore all memory cells is 8-K/2^N owing to the reduction in the memory capacity to 256-M/2^N. This dual effect changes the refresh operation interval t_REF from 7.8 μs to 2^N×2^N×7.8 μs. t_ret of 2^N cells/bit is analyzed using the measured data presented previously and it is shown that t_ret can be one order of magnitude greater than 2^N×64 ms. Here, it can be stated that t_ret can be set to 2^N×64 ms.

PAM Circuit Implementation

FIG. 8A shows a conventional 256-Mb DRAM has X decoding circuits that select one WL from X addresses <X0T-X12>. The X decoding circuits produce RX signals RX0T-511T and FX signals based on X address signals X0T-12T.

FIG. 8B shows an exemplary circuit to implement simultaneous selection of 2^N WLs. Two 4-bit decoders in FIG. 8A are modified to realize 2^N cells/bit (N=1, 2, 3, 4) access mode. One 4-bit decoder selects one output signal of RF0T<0-3> from two X addresses <X0T, X1T>. The other 4-bit decoder selects one output of RF2T<0-3> from two X addresses <X2T, X3T>. The code flag N1 is prepared for 2 WL selections, RF0T<0,1>, which is enabled by simultaneous choice of X0T and X0B. In the same way, N2 is prepared for 4 WL selections, RF0T<0-3>, enabled by simultaneous choice of two pairs, X0T/B and X1T/B. N3 is prepared for 8 WL selections, RF0T<0-3> and RF2T<0,1>, by three pairs X0T/B, X1T/B, and X2T/B. N4 is prepared for 16 WL selections, RF0T<0-3> and RF2T<0-3>, by four pairs X0T/B, X1T/B, X2T/B, and X3T/B.

A delay circuit with copy flag CP in FIG. 8B implements the copy operation of partial access conversion in FIGS. 5A and 5B. This circuit enables delayed selection of X0B for N1, which sets the copy operation from one cell to another. The address specification X0T=L in N1 (2 cells/bit mode) forms a delayed selection X0B from X0T.

In the same way, the address specification (X address signals) X0T=X1T=L in N2 (4 cells/bit mode) forms a delayed selection X0B and X1T/B from X0T. X0T=X1T=X2T=L in N3 forms a delayed selection, X0B, X1T/B, and X2T/B. X0T=X1T=X2T=X3T=L in N4 forms a delayed selection, X0B, X1T/B, X2T/B, and X3T/B.

The circuit shown in FIG. 8B can perform Partial Access Copy operation for N=2 (4 cells/bit). The operation includes pre-charging, multi selection and copy operation.

FIG. 9A shows the waveforms of the PAM X-decoder when N2=H. Referring to FIG. 8B, Multi select decoder for two outputs (N1=H), selects no output in case of X0T=H. PRE=H realizes the pre-charging operation in the same way with N1=X0T=H.

FIG. 8C shows an exemplary conventional DRAM with modifications to enable PAM. The left panel of FIG. 8C is refresh control including the modification for PAM ENTRY and PAM EXIT. When the DRAM receives PAM ENTRY, command decoder generates PA, a one shot pulse, which enables receipt of a PAM code. A0-A2 for PAM code enters PAM code decoder, which outputs N0-N4 signals to keep composed cell number n until the DRAM receives the next PAM EXIT command. PAM code decoder outputs CREF=H and CP=H (set by A3) those decide the continuous copies operation. CREF=H selects CREF_PLS as input signal into refresh counter, which sends the pulses for a continuous copies operation in 100 ns interval repeatedly until compare circuit outputs CREF_STP=H which stops continuous copies at the point of the full array copy completion. The reset operation changes CREF to L and CP to L. FIG. 9B shows these signal transitions. In PAM ENTRY operation N0-N4 signal selects the lowest bit of refresh counter among X0T-X4T.

In PAM, AREF command interval is defined in FIG. 7 and the refresh operation interval in SELF is determined by measured t_ret data in wafer condition which is estimated in the following discussion. In both cases output signals REFX0-12 in FIG. 8C enables refresh operation selected by ACS_REF as X0T-12T. If PAM is selected, for example, when a mobile phone system is in a call waiting state, both refresh operation intervals of normal and SELF can be extended.

When the DRAM receives PAM EXIT, the circuit operation about CREF is same. But PAM code decoder keeps CP to L by A3. The A0-A2 code in PAM code decoder is reset at the point of the continuous refreshes completion by the signal, CREF_STP=PACR=H. PACR=H selects NO through PAM code decoder.

CP and N1-N4 become input signals of X-decoder circuit, two 4-bit decoders for X0T/X1T and X2T/X3T in the right panel of FIG. 8C. This circuit enables delayed n−1 word lines selections in case of CP=L, PAM ENTRY in FIG. 5B, and n simultaneous selection in case of CP=L, PAM EXIT in FIG. 5B. N1-N4 decide composed cell number n and CP decides whether operations become continuous copies or continuous refreshes. The address specification X0T=X1T=L in case of N2 (4 cells/bit mode) makes a delay of a few nanoseconds of X0L=X01L=L to the next inverters. This operation generates the delayed selections of RF0T<1-3> from RF0T<0>, which correspond to the delayed selections of WL<L-3> from WL<0>, for RF0T<0-3> and RF2T<0-3> are input signals of the 16-bit decoder for 16 WLs.

As described above, by setting the copy region and the holding region, restricting accessible region into the data holding area, and copying data of memory cells in the data holding area to the memory cells in the copy region, it is possible to expand the refresh interval, reduce the number of refresh operations, and reduce power consumption consumed by auto refresh as well as self refresh.

The refresh interval can be changed by changing the number of memory cells holding the same one bit data. Therefore, by increasing the number of memory cells holding the same one bit data, it is possible to eliminate interruption for auto refresh in short (small data) communication. Further, by performing all bank refresh in exit from the partial access mode and by not performing all bank refresh in exiting from the self refresh, it is possible to reduce more power consumption. Any specific circuit is not needed to avoid all bank refresh in exiting from the self refresh. Further, even if the memory capacity is larger, it can suppress the increase in power consumption by increasing the number of word line as access unit.

Embodiment 2

In the PAM of Embodiment 1, 2^N memory cells compensate for each other's storage charge by selecting 2^N WLs at the same time in both the normal and the SELF modes, as shown in FIG. 5. In the PAM of Embodiment 2, 2^N memory cells compensate for each other's storage charge by selecting 1 WL in normal mode and 2^N WLs in the SELF mode for a first PAM mode, and by selecting 2^N WLs at the same time in both the normal and the SELF modes for a second PAM mode, as will be discussed in more detail below.

Partial access mode may reduce the refresh operation frequency by a factor of 2^N×2^N. The load of one refresh operation for n cells/bit has two increasing factors: the additional n−1 storage capacitors connected to each BL, and the additional n−1 WLs to drive at once. These two effects are estimated as follows:

(1) Storage capacitance effect: The storage capacitance C_s is 32 fF from the measured data, and VARY is 1.8V. The average number of driven capacitors is 8K×4 banks×½=16K. The total current charging all storage capacitors for t_REF=7.8 μs is 32 fF×1.8V×16K/7.8 μs=121 μA.

(2) Effect of driven WLs: The WL capacitance is 136.5 fF per WL from the design parameters. The number of driven WLs is 16 WLs per 1 bank. X-decoder must drive the signal to select SWD (called main word line), which has 1053 fF per 1 bank. VPP is 3.6V and generated from VDD=2.5V at ⅓ efficiency. The total current for driving all WLs at once for t_REF=7.8 μs is (136.5×16+1053)[fF]×4 banks×3.6V×3/7.8 [μs]=18 μA.

The n cells/bit refresh operation current is estimated by summation of the measured refresh current adjusted by t_REF, n−1 times the charging current to cells, and n−1 times the driving current to WLs.I_SELF[μA]=(7.8/t_REF)×[1238+139×n]+81.

2^N cells/bit access mode in SELF changes the current consumption to 100 μA, which is almost equal to the DC current 81 μA. This value is expressed in FIG. 10B.

The normal operation current I_normal is composed of the X-addressing current, the refresh operation current including the DC current (81 μA), and the current for READ or WRITE data transfer. We define their time-averaged values, which depend on the operating frequency, as I_ACT, I_AREF, and I_DATA.

Here, we assume that I_DATA is constant.

I_normal in the following equation is the function of composed cell number n=2^N and X-addressing interval t_ACT:I_normal(N,t_ACT)=I_ACT(N,t_ACT)+I_AREF(N)+_DATA,

where I_ACT(N, t_ACT) and I_AREF(N) are given by the next two equations, where the bank number: N_BANK=4 and t_REF=7.8 [μs]×2^N×2^N:I_ACT(N,t_ACT)[μA]=[1244+133×2N]/(N_BANK)×(7.8)/(t_ACT[μs]),andI_AREF(N)[μA]=[1244+133×2N]×(7.8)/(t_REF[μs])+81.

FIG. 10A shows I_normal with the exception of I_DATA.

I_ACT has a t_ACT dependency that decides the features of I_normal, where the minimum t_ACT is 100 ns, decided by the minimum X-addressing interval.

In the area of t_ACT>100 μs, I_normal depends on composed cell number n=2^N of I_AREF.

A system operation to the DRAMs decides t_ACT, which increases I_normal above to 100 ns owning to the increasing current factor.

From the power viewpoint, the accessed memory capacity may be restricted when the full capacity is not used, since total refresh operation times are decreased. But whether 1 cell/bit is kept or n cells/bit is composed may be dependent on the X-addressing frequency.

When a system condition has high X-addressing frequency in the random X-addressing case, a composed memory cell may be constructed in the SELF mode to realize the power reduction of ½^N×½^N.

PAM Specification

The specifications of the PAM of Embodiment 2 is illustrated as PAM1 in FIG. 11, the standardized retention time is constant t_ref1(=64 ms). The refresh operation interval in normal operation is extended to t_REF=2^N×7.8 μs by the accessed capacity reduction. This extension causes the current reduction in any case of t_ACT, as shown in FIG. 13A.

In case of PAM1, SELF is formed through continuous copy (partial access conversion: PAC) and continuous refresh (partial access re-conversion: PAR). The refresh operation interval is extended by the dual effect in the SELF mode. The power consumption is shown in FIG. 13C. Functional waveform is shown in FIG. 12A.

The PAM of Embodiment 1 is defined as PAM2 in FIG. 11. The standardized retention time is extended (t_ref2=2^N×64 ms) and the refresh operation interval is extended (t_REF2=2^N×2^N×7.8 μs) by the dual effect of retention time extension and accessed capacity reduction. Both are defined in FIG. 11. In FIG. 13B, the extension of PAM2 causes a current reduction in the area of t_ACT>100 s, which is larger than that of PAM1 in FIG. 13A. But I_normal increases with the N increment in the area of t_ACT<100 μs due to the current increment of SA activation, which is included in AREF and X-addressing. The functional waveform is shown in FIG. 12B.

The table of FIG. 13 is the summation of PAM1 and PAM2. In case of PAM1, the DRAM is operated with 1 cell/bit mode (normal) and 2^N cells/bit mode (SELF). By contrast in case of PAM2, 2^N cells/bit access mode is selected in both normal and SELF. Moreover, 1 cell/bit access mode is included as PAM0 in the table of FIG. 13. PAM0 may provide ½^N times lower frequency of refresh disturbance and ½^N times lower refresh power consumption, when it is compared to the conventional DRAM with SELF.

Both of the current consumptions are expressed in FIG. 14A and FIG. 14B. PAM1 and PAM2 both enable the lower power consumption to PAM0.

PAM Circuit Implementation

A DRAM becomes selectable for PAM1 and PAM2 by a few modifications. A first modification is the common change for PAM1 and PAM2, which is implementation of a code set. A system issues the code of PAM, for example, A0-A2 in FIG. 11.

FIG. 15 illustrates a partial access configuration circuit according to some embodiments of the invention. The circuit on the right is a partial access X-decoder in accordance with an embodiment of the invention, for example, the partial access X-decoder shown in FIG. 8B and previously described with the PAM circuit implementation of Embodiment 1. Operation of the partial access X-decoder was previously described. The circuit on the right of FIG. 15 is refresh control including modifications for PAM ENTRY and PAM EXIT for Embodiment 2.

The DRAM changes the lowest bit for the refresh operation by N0-N4, which is the output of the PAM code decoder in FIG. 15. This brings t_REF1 extension and refresh power reduction instead of capacity reduction.

In the case of PAM1 external commands, PAC and PAR do not exist. FIG. 12A shows that the SELF ENTRY includes PAC and SELF EXIT includes PAR. In contrast, external commands PAC and PAR after PAM execute the operations for the composed memory cell in the case of PAM2, which is expressed in FIG. 12B. All that is needed for the mode register set (MRS) concerning PAM½ are the selection of SELF ENTRY and external PAC, and the selection of SELF EXIT and external PAR, as shown in FIG. 15.

In PAM1, A4=L for the MRS. The output signal L of the A4 register selects SELFA through CREF-SW as the flag for a continuous refresh for the composed memory cell, and also selects SELF through CP-SW for a continuous copy. The L of the A4 register sets the configuration of the composed memory cell only in the SELF mode through N-SW.

FIG. 12A shows these signal transitions in the case of (N=0→N=2→N=0). When the DRAM receives SELF ENTRY, the command decoder generates SELFA, a one-shot pulse, which enables CREF=H and CP=H (set by SELF) to decide the continuous copy operation. CREF=H selects CREF_PLS as the input signal in the refresh counter, which sends the pulses for the continuous copy operation in 100 ns interval repeatedly until the compare circuit outputs CREF_STP=H, which stops continuous copies at the point of the full array copy completion. The reset operation changes CREF to L and CP to L.

In the SELF ENTRY operation, the N0-N4 signal selects not only the lowest bit of the refresh counter among X0T-X4T, but also selects the right panel X-decoder circuit that enables the simultaneous WL selection for n=2^N cells/bit of the composed memory cell.

When the DRAM receives the SELF EXIT, the circuit operation about CREF is the same. But the SELF ENTRY register keeps CP to L by SELF=L.

In PAM2, A4=H for the MRS. The H of the A4 register selects PA through CREF-SW as the flag of the continuous refresh, and selects the A3 register output through CP-SW as the flag of the continuous copy. The H of the A4 register sets the configuration of the composed memory cell in both the normal and the SELF modes through N-SW.

FIG. 12B shows these signal transitions in the case of (N=0→N=2→N=0). When the DRAM receives the PAC, the command decoder generates PA, a one-shot pulse, which enables receipt of the identification code A3. PA connects to the CREF register, whose output signal CREF=H decides the continuous refresh operation. The A3 register outputs CP=H, which decides the continuous copy operation. CREF=H selects CREF_PLS as the input signal into the refresh counter, which sends the pulses for continuous copy operation in the 100 ns interval repeatedly until the compare circuit outputs CREF_STP=H, which stops continuous copies at the point of full array copy completion. The reset operation of the PA register changes CREF to L and that of the A3 register changes CP to L.

When the DRAM receives the PAR, the circuit operation about CREF is the same. But the A3 register decides CP=L.

In summary, the 1 cell/bit access mode PAM0 realizes ½^N times lower frequency of refresh disturbance and ½^N times lower refresh power consumption, when it is compared to the conventional DRAM with SELF. PAM1 enables the minimum power consumption and the minimum refresh interruption through 1 cell/bit access mode in normal and 2^N cells/bit access mode in SELF, even if X-addressing interval is short (t_ACT<100 μs). PAM2 enables the minimum power consumption and the minimum refresh interruption through 2^N cells/bit access mode in normal and SELF, when X-addressing interval is long enough (t_ACT>100 μs).

FIG. 16 illustrates a PAM1 example for a memory module according to the present disclosure.

In a personal computer system, the demand from the system frequently changes the used memory capacity. It is reasonable to use DRAMs in the PAM1 mode from the viewpoints of both power consumption and system performance. The DRAM is typically used in the set of 8 DRAMs in a module. The memory module of 8 DRAMs has 64 I/O pins, which is the sum of 8 I/O pins in each DRAM. The address pins are shared between the 8 DRAMs. This example is explained in FIG. 16B, which is composed of the DRAMs shown in FIG. 16A. FIG. 16A illustrates a memory module according to an embodiment of the disclosure. FIG. 16B illustrates memory capacity allocation according to an embodiment of the disclosure. We assume one set of modules and omit the chip select (CS) pin, which identifies the one DRAM module from multiple modules. The memory capacity of FIG. 16B is 8 times larger than that of the one chip DRAM. If a system uses PAM1 in the form of a module, the memory capacity reduction and current reduction effect are 8 times that of the one DRAM case of FIGS. 13A-C. The standby current of the memory module can be lowered to just the sum of the DRAM's DC current, if the SELF of PAM1 is used as N=2, 3, 4.

FIG. 17 illustrates a PAM2 example in a wireless system according to an embodiment of the disclosure. Mobile phones are now widely used all over the world. DRAM is the main memory in each phone. It is desirable for the system to reduce power consumption for the wireless condition, even though the system must detect all in-coming phone calls, and ideally, the power consumption for detection should be lowered as much as possible. As previously described, each mobile phone communicates with a radio station for 280 μs with an interval of 2 s. The communication interval 2 s and the communication time 280 μs are defined by the specification of the wireless system. The system flow is expressed in FIG. 17. The mobile phone judges whether a phone call to itself can come in communication time 280 μs.

A small memory capacity of the DRAM is used to detect an in-coming phone call. It is reasonable to use DRAMs in the PAM2 mode. Ten percent of 256 Mb memory capacity is large enough to be accessed in the call-waiting state. The t_REF2 of AREF (PAM) at the point of N=3 is 500 μs in FIG. 11, which is larger than the communication time 280 μs. This means the mobile phone system does not have to issue the AREF command in every communication time of 280 μs.

Effect of Improvement: Half Pre-Charge CMOS Sense Amplifier

A CMOS sense amplifier in a DRAM is a circuit that is used to amplify the signal, ΔV, the difference between the voltage on the BLB and BLT. One voltage is connected to the cell capacitor, and the other is connected to the reference voltage. The operation of sense amplifiers was expressed previously with reference to FIGS. 2C and 2D. In a sense amplifier, the pre-charge voltage is VARY/2, which enables correct amplification even though sense amplifiers in a DRAM have threshold voltage variation, which form a normal distribution.

The signal amounts of H/L data are given by the charge conservation law in FIG. 18A: ΔV=+(VARY/2)×(C_s/(C_b+C_s)) for H data, and ΔV=−(VARY/2)×(C_s(C_b+C_s)) for L data. This relationship may be referred to as identity 1. In this circumstance, C_b and C_s are the capacitances of the BL and storage capacitor, respectively.

When the pause time is infinite, the memory node voltage V_S converges to the p-well voltage V_B in FIG. 18A. H data is broken when V_S is equal to VARY/2, the pre-charge level for the reference voltage. There is no destruction of L data, because V_S for the L data is always within V_B<V_S<VSS and lower than VARY/2.

Effect of Improvement: Pause Time Dependency of Accumulated Charge (H Data)

The following disclosure examines the pause time dependency for an amount of signal for an ACH pause to estimate t_ret and the power consumption for the refresh operation in SELF of the PAM. The preceding disclosure explained that the destruction of the memory node voltage H restored in the DRAM memory cell is caused by the discharge of the storage capacitor via leakage current. Destruction itself means that the BL voltage is amplified to a different side from the written side because of storage charge reduction.

The leakage current is represented by a current source I_L that has strong dependence on the voltage difference, V_S−V_B for the pn-junction's leakage, V_S−VSS for GIDL and V_S−VARY12 for sub-threshold current. Generally, the dependence is strong when the voltage difference is small. It can be approximated that h is constant in the region VARY>=V_S>=VARY12, where the storage charge is destroyed from H. This approximation is not appropriate for the sub-threshold current but acceptable for pn-junction's leakage and GIDL.

FIGS. 18B and 18C show how the signal amount depends on the pause time for the measurement of a 256-Mb DRAM. These data show that the signal amount decreases approximately in proportion to the pause time t for an ACH pause, indicating that I_L is nearly constant. It can be assumed for simplicity that the retention time is given by the time when the signal amount is zero, i.e. ΔV(t_ret)=0. In this simple approximation, the amount of signal ΔV(t) is given by: ΔV(t)=(VARY/2)×(C_s(C_b+C_s))×(1−t/t_ret). This relationship may be referred to as identity 2. This means that the storage charge Q(t) is given by: Q(t)=(VARY/2)×C_S×(1−t/t_ret), for t<=2.3×t_ret. This relationship may be referred to as identity 3. From the viewpoint of Q(t), FIGS. 18B and 18C show that Q(t) decreases in proportion to the pause time from VARY×(C_s/2) at t=0 to −VARY×(C_s/2) at t=2t_ret. For t>2.3t_ret, Q(t) is constant at −(1.3×VARY×C_s)/2 because the memory node voltage has settled to the lowest voltage V_B.

FIG. 21A shows a delay circuit with copy flag CP, which implements the copy operation. This circuit enables delayed selection of X0B for N1, which sets the copy operation from one cell to another. The address specification X0T=L in N1 (2 cells/bit mode) forms a delayed selection X0B from X0T. In the same way, the address specifications in N2, N3, and N4 form their delayed selections.

N2=H (4 cells/bit) sets N1234=N234=H. PRE transition to L makes a delay of a few nanoseconds of X0L=X01L=X012L=L to the next inverters when addresses are specified: <X0T, X1T>=<0, 0>. This makes the delayed selections of RF0T<1-3> from RF0T<0>, which correspond to the delayed selections of WL<1-3> from WL<0>. This operation is expressed where <X2T, X3T> is periodically incremented: <X2T, X3T>=<0, 0>, <1, 0>, <0, 1>.

Effect of Improvement: Data Retention Time for 2, 4, 8, and 16 Cells/Bit

The following analysis leads the data retention time of the composed memory cell from identity 3, the cumulative distribution of 1 cell/bit measured error ratio in FIG. 19B, and the measured random fail-bit map in FIG. 19A. The amount of signal for the case of 2^N cells/bit (N>=1) can be obtained by the charge conservation law written with identity 3's Q(t) as in identity 1: This is given as follows: ΔV(j₁, . . . , j_n, t)=(1/(C_b+n×C_s))×SUM (i=1 . . . n) Q(j_i, t). This relationship may be referred to as identity 4. In this case, n is the composed cell number (n=2^N), and j is one combination of n memory cells. When ΔV(j₁, . . . , j_n, t)>0, the composed cell j is judged as pass. When ΔV(j₁, . . . , j_n, t)<=0, the composed cell j is judged as fail.

The expected value of the fail-bit count of 256/n-Mb DRAM is expressed by the following identity 5, where t is the pause time, θ is the function defined by the following identity 6, and p_j1, . . . , p_jn are the probabilities that each cell has a certain pause time. Identity 5 is: E⁽ⁿ⁾(t)=(256 M/n)×SUM(j1) . . . SUM(jn) θ(j₁, . . . , j_n, t) p_j1, . . . , p_jn. Identity 6 is: θ(j₁, . . . , j_n, t)=1 for V(j₁, . . . , j_n, t)<=0, and θ(j₁, . . . , j_n, t)=0 for V(j₁, . . . , j_n, t)>0.

For the case of N=1 (2 cells/bit), the expected value E⁽ⁿ⁾(t) is calculated by identity 5 from the probabilities pt_ret1, pt_ret2 from ΔN_fail(t_ret)/256M. ΔN_fail(t_ret) is the difference between two values side-by-side in the measured 1 cell/bit distribution of FIG. 19B when t_ret is the center pause time.

For 2^N cells/bit (N>=2), the number of calculations is enormous. Here will be used an approximation, the majority decision rule, that a composed cell with 2^N cells/bit fails if half the memory cells in 2^N cells fail, based on the mechanism of data loss. This approximation means that the memory node voltage of each cell takes either H or L and each cell charge compensates for the charges of the other memory cells. Another assumption is the random formation of the fail-bit distribution. This is the same as in the case of 2 cells/bit. These two assumptions yield the expected value of the fail-bit count of composed cells. The case of composed cells with 2^N cells/bit will be discussed using the following notation: n is the number of cells in a composed cell, q is the number of fail cells in a composed cell, and r is the number of fails in 256-Mb memory cells.

The number of combinations needed to select q failed cells in n cells in a composed cell is given by _nC_q. The number of combination to select r−q failed cells in the other 256M−n cells is given by _256M-nC_q. Thus, the total number of combinations is given by _256M-nC_r-q. However, the number of combinations to select r failed cells in 256M cells is given by _256MC_r. Thus, the probability for the above condition is given by (_nC_q×_256M-nC_r-q)/_256MC_r. By the majority decision rule, failure of composed cell occurs when q>=n/2. Therefore, the expected value of fail bits of the composed cell is given by: (256M/n)×SUM(q=n/2 to n) ((_nC_q×_256M-nC_r-q)/(_256MC_r) for (N>=2). This expression can be referred to as identity 7. Error ratios using identity 7 are plotted in FIG. 19B for t⁽ⁿ⁾_ret, n=4, 8, and 16. These results show that t_ret improves from t_ret=100 ms (1 cell/bit) to t_ret=3 s (8 cells/bit). Increasing n reduces the variation in t_ret and eliminates the tail distribution. Therefore, the error ratio in FIG. 19B increases steeply with an increase in n when the pause time is 7 s (50% error ratio). Identity 7 for n(>=4) derives the feature. This estimation can be applied to an ACH disturb because the added electric field in the BL direction brings no difference to the data destruction mechanism and random formation of fail bits.

The improvement of retention time must be examined, including the variation of cumulative distribution among DRAM chips. The minimum retention time t_ret,min, is around 100 ms which is longer than the standardized specification t_ref (=64 ms). The process development must achieve this target. The minimum 2 cells/bit retention time is the twice of 1 cell/bit retention time, the compensated result of the t_ret,min memory cell which is expressed in identity 4 and FIG. 19B. The tail distribution slopes in N>=2 are steeper than that of N=1. The retention time improvement of FIG. 19B is ensured in spite of the process variation across the DRAM chip.

Effect of Improvement: Current Consumption

The following disclosure discusses the current consumption of the SELF and normal operation modes. In the SELF mode, the current for the refresh operation and the DC current comprise the total current consumption. In this case, the DC current is the sum of the OFF current of all transistors, operating current for the internal voltage generator, and current for the refresh period oscillator circuit. The measured DC current of the 256-Mb DRAM is around 81 μA. FIG. 20A shows the SELF current comparison of the PAM and PASR. The conventional method PASR always consumes two types of currents in the SELF mode: a refresh operation current of 880 μA derived in identity 8 from an extended t_ref=12.2 μA and a DC current of 81 μA. PASR reduces the refresh operation current from 880 μA to 880 μA/2^N. This can be attributed to the reduction in the preserved density. This gives identity 8 as: I_SELF(PASR)[μA]=(880/2^N)+81.

As previously described, the PAM may reduce the refresh operation frequency by a factor of 2^N×2^N. The load of one refresh operation for n cells/bit has two increasing factors: the additional n−1 storage capacitors connected to each BL and the additional n−1 WLs to drive at once. These two effects were estimated to result in a current consumption of 121 μA for the storage capacitance effect and 18 μA for the effect of driven WLs.

As also previously described, the n cells/bit refresh operation current may be estimated by summation of the measured refresh current adjusted by t_REF, n−1 times the charging current to cells, and n−1 times the driving current to WLs. This gives identity 9 as: I_SELF(PAM)[μA]=(1238+139×n)×(7.8/t_REF[μs])+81. Partial access conversion from 1 cell/bit to 4 cells/bit changes the current consumption to 100 μA, which is almost equal to the DC current; however 1-bank-PASR (64-Mb PASR) consumer 300 μA of current in the SELF mode. These values are expressed in FIG. 20A.

The present disclosure now discusses the normal operation current, which can be classified into SA activation with X-address selection, data transfer for READ or WRITE operation, AREF operation, and a DC current. Their time-averaged values can be defined I_ACT, I_DATA, and I_AREF, which depend on the operation frequency. The DC current of 81 μA is the same as that in the SELF mode. The common equation for PASR and PAM is expressed by: I_normal[μA]=I_ACT+I_DATA+I_REF+81. This relationship can be referred to as identity 10.

For PASR, I_ACT(PASR)[μA]=(1377/N_BANK)×(7.8/t_ACT[μs]). For PASR, I_AREF(PASR)[μA]=1377×(7.8/t_REF[μs]). For PASR, t_REF 7.8 μs. This collection of relationships can be referred to as identity 11.

For PAM, I_ACT(PAM)[μA]=((1238+139×n)/N_BANK)×(7.8/t_ACT[μs]). For PAM, I_AREF(PASR)[μA]=[1238+139×n]×(7.8/t_REF[μs]). For PAM, t_REF=7.8 [μs]×2^N+×2^N. This collection of relationships can be referred to as identity 12.

The number of banks, N_BANK in identities 11 and 12 is four for a 256-Mb DRAM. One bank is activated, contrary to the AREF operation for all banks. Parameter t_ACT in identities 11 and 12 depends on the condition of the system. Below is considered the normal operation current for two typical conditions, minimum and maximum t_ACT at maximum data transfer rates.

Random Addressing: 256-Mb DRAM is activated repeatedly in a 100 ns interval, and t_ACT=100 ns. In this case, t_ACT is much greater than I_AREF, and the PAM (N>0) increases the operation current as shown in FIG. 20B.

Continuous Addressing: consider a 100-MHz clock synchronized operation as an example. If a DRAM is accessed continuously in the Y-address direction, each 8-bit data are accessed in a 10-ns interval repeatedly, an 8-K bit page is fully accessed in 10 μs, and t_ACT=10 μs. In this case, PAM (N>0) can reduce the total current, as shown in FIG. 20B. Moreover, the AREF disturbance frequency decreases by t_REF=7.8 μs×2^N×2^N. This means that the interruption frequency is reduced if the PAM (N>0) is used for Y-continuous access. t_ACT depends on the data access interval determined by the clock frequency. In 2012, a 1-GHz operation frequency is widely used. A higher frequency shortens t_ACT and lowers the current reduction effect by the PAM. There is a transition point in t_ACT,tran where the PAM (N>0) can or cannot reduce the current. The transition point for N=1 is 260 ns. For t_ACT<260 ns, full array access (N=0) is more useful not only from a memory view point but also from the current consumption view point.

While the invention may be susceptible to various modifications and alternative forms, specific embodiments have been shown by way of example in the drawings and have been described in detail herein. However, it should be understood that the invention is not intended to be limited to the particular forms disclosed. Rather, the invention is to cover all modifications, equivalents, and alternatives falling within the spirit and scope of the invention as defined by the following appended claims.

The construction and arrangement of the elements as shown in the exemplary embodiments are illustrative only. Although only a few embodiments of the present disclosure have been described in detail, those skilled in the art who review this disclosure will readily appreciate that many modifications are possible (e.g., variations in sizes, dimensions, structures, shapes and proportions of the various elements, values of parameters, mounting arrangements, use of materials, colors, orientations, etc.) without materially departing from the novel teachings and advantages of the subject matter recited. For example, elements shown as integrally formed may be constructed of multiple parts or elements. The elements and assemblies may be constructed from any of a wide variety of materials that provide sufficient strength or durability, in any of a wide variety of colors, textures, and combinations. Additionally, in the subject description, the word “exemplary” is used to mean serving as an example, instance, or illustration. Any embodiment or design described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments or designs. Rather, use of the word “exemplary” is intended to present concepts in a concrete manner. Accordingly, all such modifications are intended to be included within the scope of the present disclosure. Other substitutions, modifications, changes, and omissions may be made in the design, operating conditions, and arrangement of the preferred and other exemplary embodiments without departing from the scope of the appended claims.

Claims

1. A device comprising: a plurality of memory cells;a controller configured to control storage of data in the plurality of memory cells, the controller configured to operate a first access mode which stores one bit data into 1 cell of the memory cells, and a second access mode which stores one bit data into 2N cells of the memory cells, wherein N is a natural number; andan address decoder coupled to the controller and configured to simultaneously select one or more word lines coupled to one or more memory cells of the plurality of memory cells to store a data in a self refresh operation mode, the address decoder comprising a logic circuit configured to receive a plurality of code flag signals and a plurality of switches coupled to the logic circuit, andwherein the controller is configured to operate in the first access mode which activates a word line in a mat to access the one bit data and further configured to operate in the second access mode which simultaneously activates at least one other word line in the mat to access the one bit data, the second access mode being associated with a self refresh operation mode, andwherein the logic circuit is configured to activate one or more switches of the plurality of switches to select the one or more word lines responsive to the plurality of code flag signals in the second access mode.

2. The device of claim 1, wherein the first access mode is associated with a normal operation mode.

3. The device of claim 1, wherein the controller is configured to operate in the first access mode which activates a word line to access the one bit data and in the second access mode which activates 2N word lines to access the one bit data.

4. The device of claim 1, wherein the controller is configured in a first partial access mode to operate the first access mode in a normal operation mode and the second access mode in the self refresh operation mode, and further configured in a second partial access mode to operate the second access mode in both the normal operation mode and the self refresh operation mode.

5. The device of claim 4, wherein the controller is configured in a third partial access mode to operate the first access mode in both the normal operation mode and the self refresh operation mode.

6. The device of claim 1, wherein the controller storing causes the power consumption of the device to be a first reduced amount for the first access mode, and causes the power consumption of the device to a second reduced amount for the second access mode, the second reduced amount less than the first reduced amount.

7. A device comprising: a plurality of memory cells;a controller controlling storage of data in the plurality of memory cells, the controller configured to store data in the plurality of memory cells as 1 data bit/cell for a normal operation mode and store data in the plurality of memory cells as 1 data bit/2N cells for a self refresh operation mode in a first partial access mode, and further configured to store data in the plurality of memory cells as 1 data bit/2N cells for both a normal operation mode and a self refresh operation mode in a second partial access mode, andan address decoder coupled to the controller and configured to simultaneously select word lines coupled to the plurality of memory cells to store the data in the plurality of memory cells in the self refresh operation mode data, the address decoder comprising a logic circuit and a plurality of switches coupled to the logic circuit,wherein the logic circuit is configured to receive a plurality of code flag signals and further configured to provide one or more activation signals to one or more switches to select the word lines, andwherein the plurality of switches is configured to provide a delay in selecting the word lines responsive to the one or more activation signals.

8. The device of claim 7 wherein the address decoder is configured to handle bits of a row address as “don't care” according to the first or second partial access mode.

9. The device of claim 7, wherein the plurality of memory cells are included in memories of a memory module.