Dram and access method

Abstract

A DRAM and a method of accessing a DRAM capable of obtaining a high data rate in a random row access. By selecting one of a plurality of main word lines (14), eight sub-word lines (16) are selected from 512 sub-word lines (16), and one sub-word line (16) is selected by a sub-word line (16) selecting signal and an enable signal.

Description

TECHNICAL FIELD

[0001] The present invention relates to a DRAM (Dynamic Random Access Memory) and access method thereof, and more particularly to a DRAM and access method thereof for achieving a high data rate in random row access. According to the present invention, an access is performed as if there were a large number of banks by limiting an access to multiple divided blocks, wherein latching of its row address, activation of a sense amplifier, and precharge after write-back are all controlled by signals of circuit within this block.

BACKGROUND

[0002] Since the speed of DRAM is slower compared with MPU (Micro Processor Unit), it has been the bottleneck to improve the performance of computers. Particularly in random row access where the row address changes continuously, since the access time is long in addition to the fact that precharge for the previous access also takes some time, operations of DRAM become very slow. In order to prevent random row access as much as possible, recent high-performance DRAM's such as SDRAM (synchronous DRAM), SDRAM DDR (SDRAM Double Data Rate), Rambus, etc., are all provided with banks. Furthermore, for speeding up for DRAM, every efforts are made to a program or memory mapping such that a column address in the same page would be accessed.

[0003] However, this is not possible between individual programs and consequently an access to another row address is bound to occur. Given that the banks are provided, an access to the next row address can occur without precharging the row currently accessed if the next row address is in a different bank than the one currently accessed. Therefore, the next burst takes place just when the previous burst ends, which allows the fast processing without temporal empty periods on the data bus.

[0004] In order to move to the other bank one after another without bank conflicts, it is necessary to increase the number of banks considerably. There is needed an active, precharge, read, and write signals and control for these signals for each bank. Since provision of a number of banks brings an increase of chip size, realistically only four banks are able to be provided for SDRAM.

[0005] For Rambus, there is provided only 16 banks of 72 Mb each or 32 banks of 144 Mb each at a large chip size penalty. In the case of Rambus, the cycle time is long and three banks are occupied for one row access because banks are configured across shared sense amplifiers. Thus, there will be a small gain in performance even if the number of banks is increased to 32 with further chip size penalty.

[0006] Therefore, all of the prior-art DRAMs can not show a high data rate for random row access.

[0007] Now the above content will be described in detail taking the same hardware of 128 Mb (8 Mb×16) as an example. AS shown in FIG. 3, a conventional SDRAM 40 has four banks of 32 Mb each, which are independent blocks. One row access activates 8-K sense amplifiers. One bit line is traversed by 512 word lines orthogonally. Therefore, this means that a block comprised of a cell array of 4 Mb (=512×8 K) is activated. Namely, a 32 Mb bank is comprised of 8 blocks. According to the design, other row addresses in the same bank can not be accessed.

[0008] However, practically, row addresses that can never be accessed are only 511 other word lines in the same 4 Mb bank. Word lines of other seven blocks that do not share sense amplifiers are realistically accessible even in the same bank. Nevertheless, the reason for not making this block unit be a bank is that the chip size increases because of the increase of complexity for controlling banks and the increase of the number of signal lines if the number of banks increases. Since 16 data lines from each bank are connected to 16 data I/O pads, the number of wire connections to I/O pads increases when the number of banks increases.

[0009] On the other hand, for Rambus 42, banks are composed as a 4 Mb block that is surrounded by 512 word lines sharing sense amplifiers, wherein 32 banks are provided in a whole chip, which is more than SDRAM. In order to avoid the congestion, it is configured as a vertical stack structure shown in FIG. 4, however, the increase of chip size is inevitable due to a large number of control signals to each bank. Furthermore, since the length of pages becomes short and the probability of page misses increases when increasing the number of banks, it is necessary to increase the number of banks to be activated.

[0010] However, since the number of banks that can be used for a new row access decreases, resulting in the increase of probability of bank conflicts, the purpose of increasing the number of banks, that is, decreasing the number of bank conflicts, can not be accomplished. Therefore, though the probability of bank conflicts is reduced by increasing the number of banks while maintaining most banks not activated, the page hit rate may be decreased. In this manner, expecting a high hit rate in the page mode and attempting to avoid bank conflicts contradict each other, thus the performance is not enhanced very much regardless to the number of banks. As a result, it is essentially impossible to solve two problems of page misses and bank conflicts by the number of banks.

PROBLEMS TO BE SOLVED BY THE INVENTION

[0011] It is therefore an object of the present invention to provide a DRAM and access method for DRAM with a high data rate in a random row access.

SUMMARY OF THE INVENTION

[0012] In a first aspect of the present invention, there is provided a DRAM comprising: a block composed of a plurality of segments storing data; main word lines for selecting a predetermined number of sub-word lines from a plurality of sub-word lines; and corner blocks for selecting one of the plurality of segments and one of the predetermined number of sub-word lines, wherein the corner block comprises: a plurality of global Z-lines to which a signal is sent for selecting the one of the predetermined number of sub-word lines; segment select lines to which a signal is sent for selecting the segment; a plurality of NAND circuits to which the sub-word lines and segment select lines are connected; a plurality of latch circuits operated by predetermined signals from the NAND circuits.

[0013] In another aspect of the invention, there is provided a method for accessing a DRAM where one block is divided into multiple segments by a segment decoder and a desired sub-word line is selected from a plurality of sub-word lines in the block, the method comprising the steps of: activating main word lines to select a predetermined number of sub-word lines from the plurality of sub-word lines; activating a desired global Z-line among a plurality of global Z-lines contained in a corner block and segment select lines to activate a local Z-line corresponding to the desired global Z-line among the plurality of local Z-lines; and selecting a sub-word line corresponding to the activated local Z-line from the predetermined number of sub-word lines.

BRIEF DESCRIPTION OF THE DRAWINGS

[0014] FIG. 1 depicts a schematic diagram of a DRAM of the present invention.

[0015] FIG. 2 depicts a structure of DRAM shown in FIG. 1 in detail using four detailed drawings.

[0016] FIG. 3 depicts a schematic diagram of a conventional SDRAM.

[0017] FIG. 4 depicts a schematic diagram of conventional Rambus.

DETAILED DESCRIPTION OF THE INVENTION

[0018] Now a DRAM and access method of the present invention will be described with reference to the accompanying drawings.

[0019] A DRAM 10 of the present invention shown in FIG. 1 is based on a memory that attempts seamless operations in random row access, which needs no page length since it does not use a page mode. Ideally, it may have enough number of sense amplifiers (SA) for the burst length. Therefore, it is able to have activation of an extremely small block 12 and a large number of blocks, thus the performance of random row access may be improved by increasing the number of blocks 12 as much as possible in order to reduce bank conflicts without considering page hits and misses.

[0020] In order to avoid congestion of data lines, DRAM 10 is configured such that 16 I/O's are divided into four, wherein four I/O's are drawn from each 32 Mb divided array 15. Block 12 of DRAM 10 is composed of a matrix comprising 512 word lines and 512 bit line pairs, wherein data is stored in a cell located at each lattice. Namely, block 12 can store 256 Kb (=512×512) of data. Since array 15 is 32 Mb, the number of blocks contained in an array 15 is 128.

[0021] FIG. 2 depicts a structure of block 12 of 256 Kb. Block 12 is divided into four segments 21 by segment row decoder 20. 512 Bit line pairs 18 contained in block 12 are divided into four by segment row decoder 20. A sub-word line 16 comprised of eight polycides is selected from one main word line 14 that crosses 1024 bit line pairs 18, wherein one of these eight polycides is to be selected. There are 64 main word lines 14 in a block 12. Sub-word line 16 comprised of polycides are arranged across 256 bit line pairs 18 on both sides of segment row decoder 20. One of four sub-word lines 16 is selected in a segment row decoder 20.

[0022] On the far right side of FIG. 2, there are depicted segment row decoder 20 and corner block (CB) 24. Corner block 24 comprises global Z-lines 26 to which a signal is sent to select a sub-word line 16 from four sub-word lines 16, segment select lines 28 to which an enable signal is sent to select a segment 21, four NAND circuits 30, and latch circuits 32. In addition, a reset line 36 is provided where a reset signal is input.

[0023] Segment row decoder 20 comprises four local Z-lines 34 connected to latch circuits 32. By activating one of four local Z-lines 34, sub-word lines 16 connected to that local Z-lines 34 are selected. There is provided a sense amplifier (SA) 23 in each segment 21.

[0024] Then, it will be described about a method for selecting sub-word lines 16, that is, a method for accessing DRAM 10 of the present invention. Selecting main word line 14 allows selection of eight sub-word lines 16. How to select main word line 14 will be described later in detail.

[0025] From beneath the 32 Mb divided array 15, a low level output signal of a pre-decoder is transmitted to one of four global Z-lines 26 through data line pairs 22. Namely, a signal for selecting sub-word line 16 is supplied to one of four global Z-lines 26. Data line pair 22 comprises 16 lines in one block. Segment select line 28 in corner block 24 is also supplied with an enable signal in the same manner as the output signal mentioned above. The signal for selecting sub-word line 16 and the enable signal are both a high pulse, which makes one of four NAND circuits 30 be low. Furthermore, this pulse also serves as a signal for starting a series of timing chains.

[0026] In corner block 24, latch circuit 32 that receives a low signal from NAND circuit 30 makes local Z-line 34 connected therewith be high. After being latched, the pulse of the enable signal becomes low. In this manner, latch circuit 32 operates with a low signal, whereby one of four local Z-lines 34 is selected. As a result, sub-word line 16 is activated which is connected to local Z-line 34 that is made high in segment row decoder 20. In other words, one of sub-word lines 16 is to be selected from eight sub-word lines 16 that are selected by main word line 14.

[0027] In summary, main word line 14 selects a predetermined number of (i.e., eight) sub-word lines 16 among 512 sub-word lines 16, and then one of sub-word lines 16 is selected by a signal for selecting sub-word line 16 and an enable signal.

[0028] Likewise, main word line 14 is also latched in the same manner by main row decoder 38 shown in FIG. 2, whereby two polycide sub-word lines 16 are activated across 256 bit line pairs 18 in 256 Kb block 12. Namely, 512 bit line pairs 18 are activated by two sub-word lines 16.

[0029] There is also provided in corner block 24 a circuit for driving set nodes of the sense amplifier. Further provided in corner block 24 is a circuit for, in receipt of a drive signal from the drive circuit, generating a reset signal indicating that rewriting has been completed, which is then supplied to latch circuit 32. That is, the reset signal is the one that indicates completion of data rewriting. Using this signal for releasing the latch for local Z-line 34, sub-word line 16 can be restored. The reset signal is also used for restoring the set node of the sense amplifier after some delay of time.

[0030] In this manner, a series of operations from activation of sub-word line 16 to precharge together with the reading and the writing is sequentially processed by signals generated in corner block 24. Even if a next low access occurs for the other 256 Kb block 12 before this series of operations completes, block 12 that was accessed first won't be affected by the change of global Z-lines 26 and continues its processing. Besides, processing of the next 256 Kb block 12 proceeds concurrently and the cycle will be able to be completed using its own signals. In this manner, block 12 can operate as if it were a bank without requiring global signals for a specific bank, thus it is called virtual bank (VB). The virtual bank structure allows to have a large number of banks without bank control signals.

[0031] For a memory that attempts seamless row-to-row operations by accessing row addresses in a short-cycle pipeline with automatic precharge after prefetching full-burst length data or writing preloaded full-burst length data simultaneously, the intention can provide a substantially large number of banks without bank control and thus can improve the data rate.

[0032] The present invention abolishes a page mode, thus the concept of page hits and misses is unnecessary, wherein seamless operations are allowed even in a row-to-row access unless bank conflicts exist, so that the data rate is always maintained at a peak rate. In view of the above, the attached table 1 represents the comparison of performance between the background art and the present invention in terms of the probability of bank conflicts.

[0033] The comparison assumes that the memory system bus is 64 bits in width (while Rambus is 16 bits in width) and 32 bytes of cache line is provided. One more important item in terms of bank conflicts is the ratio of the cycle time to the time required for the burst. For example, when processing 32 bytes cache line with 64 bit width, 4-bit burst is required. Since Rambus is 16 bits in width, 16-bit burst is required. Table 1 shows the probability of random row access becoming seamless.

[0034] First of all, when the cycle time is shorter than or equal to the burst time, any address can be accessed seamlessly in the row-to-row access even if there is no bank, thus the probability is 100%. When the cycle time is twice the burst time, the currently accessed bank will be able to be used in the row access after next. Accordingly, if there are two banks, the other bank can be used in the next cycle, resulting in the probability being 50%. The ratio of the cycle time to the burst time is 1 or 2 for memories that attempt seamless operations in a random row access, therefore, the success ratio of seamless operations is high with less bank conflicts occurring.

[0035] On the contrary, conventional SDRAM DDR and Rambus have a long cycle time, wherein the ratio may be very large such as four to five. In this case, it is necessary to avoid bank conflicts during four to five cycles continuously in order to succeed in seamless access. Accordingly, it is inevitably impossible to obtain a high probability to avoid bank conflict. For example, for SDRAM DDR whose ratio is four and has four banks, the first row access can use any bank, while the second can use three banks among four, the third can use two banks, and the fourth can use only one remaining bank because the first bank has not completed its precharge yet, wherein the probability of all four cycles succeeding in seamless access without bank conflicts would be 9.375% (=(1×3/4×2/4×1/4)×100). In general, assuming that the number of banks be B and the ratio of cycle time to burst time be R, the success rate is given by the following equation. 1$\begin{array}{cc}\left[\mathrm{Equation}1\right]\frac{B-1}{B}\times \frac{B-2}{B}\times \cdots \times \frac{B-\left(R-1\right)}{B}\times 100& \left(1\right)\end{array}$

[0036] For Rambus, when one of the banks is used by the shared sense amplifier, its upper and lower banks are also used by the sense amplifier, thus a total of three banks can be used, whereby the success rate is given by the following equation, resulting in low with respect to the number of banks. 2$\begin{array}{cc}\left[\mathrm{Equation}2\right]\frac{B-3}{B}\times \frac{B-6}{B}\times \frac{B-9}{B}\times 100& \left(2\right)\end{array}$

[0037] On the contrary, the present invention provides a memory that attempts seamless operations in random row access where R is 1 or 2, wherein the success rate is 50% when only two banks are provided and it increases rapidly as the number of banks increases. The success rate reaches 94% when 16 banks are provided, while it reaches 99.3% for 128 banks, which is feasible for 128 Mb, which means that a row-to-row access for any address is almost always performed seamlessly. It is seen that for a shared sense amplifier such as Rambus, the success rate reaches 81% using 16 banks.

[0038] In addition, the virtual bank scheme of the present invention can achieve cost reduction by greatly reducing test time. With a bank architecture of conventional SDRAM or Rambus, an access to the other bank is made while maintaining activation of a bank, a variety of tests are required combining these accesses. As the number of banks increases, the number of combinations also becomes huge, thereby greatly increasing the cost for testing. For this reason, the number of banks can not be increased. In fact, for Rambus having 16 or 32 banks, the number of banks that can be activated simultaneously is limited to four. This means reducing the page length, which in turn becomes the reason for not being able to improve a page hit rate. On the other hand, since the present invention needs no bank control and further never accesses other banks while maintaining the sense amplifier activated, such complicated combination tests are unnecessary. For DRAM 10 of the present invention, normal random row accesses just have to be performed with a short period of time.

[0039] Moreover, the present invention provides a method for processing low access operations from selection of a word line to precharge by using signals generated in the corner block, wherein a great effect is brought out by combining (1) a scheme used in a memory that attempts seamless operations in random row access for automatically performing precharge after prefetch or collective writing, and (2) a small ratio of cycle time to burst time such as two used in a memory that attempts seamless operations in random row access.

[0040] The present invention has been described in relation to the preferred embodiments, however, it should be construed that the present invention is not limited to those embodiments. For example, latch circuit 32 in corner block 24 may be operated with a high level signal instead of low signal. In this case, NAND circuit 30 is replaced with AND circuit.

[0041] In addition, the present invention will be improved, modified or changed by those skilled in the art without departing from the spirit and scope of the present invention.

ADVANTAGES OF THE INVENTION

[0042] As mentioned above, according to a DRAM and access method of the present invention, it becomes possible to overcome the idea of the background art where banks are provided in consideration of page misses due to the heavy use of page mode. Namely, performance of random row access, as such, is improved by abolishing the page mode and greatly reducing the cycle time through activation of blocks. In summary, the present invention makes good use of the activation of blocks and substantially allows an operation as if there were a large number of banks without conventional bank control, thereby achieving a high data rate of about 80%.

Claims

1. A DRAM comprising: a block composed of a plurality of segments storing data; main word lines for selecting a predetermined number of sub-word lines from a plurality of sub-word lines; and corner blocks for selecting one of said plurality of segments and one of said predetermined number of sub-word lines, wherein said corner block comprises: a plurality of global Z-lines to which a signal is sent for selecting the one of said predetermined number of sub-word lines; segment select lines to which a signal is sent for selecting the segment; a plurality of NAND circuits to which the sub-word lines and segment select lines are connected; a plurality of latch circuits operated by predetermined signals from said NAND circuits.

2. The DRAM according to claim 1, further comprising local Z-lines connected to said larch circuits, said local Z-lines being activated by said latch circuits.

3. The DRAM according to claim 1 or 2, wherein the corner block comprises a circuit for generating a signal for driving set nodes of a sense amplifier.

4. The DRAM according to claim 3, wherein the corner block comprises a circuit for, in receipt of said signal for driving said set nodes, generating a reset signal indicating that rewriting of data has been completed.

5. A method for accessing a DRAM where one block is divided into multiple segments by a segment decoder and a desired sub-word line is selected from a plurality of sub-word lines in the block, the method comprising the steps of: activating main word lines to select a predetermined number of sub-word lines from said plurality of sub-word lines; activating a desired global Z-line among a plurality of global Z-lines contained in a corner block and segment select lines to activate a local Z-line corresponding to said desired global Z-line among said plurality of local Z-lines; and selecting a sub-word line corresponding to said activated local Z-line from said predetermined number of sub-word lines.

6. The method according to claim 5, wherein a sense amplifier is provided for each said block, further comprising the step of transmitting a signal from said corner block for driving set nodes of said sense amplifier.

7. The method according to claim 5 or 6, further