Method for hiding a refresh in a pseudo-static memory

Abstract

A method and device for hiding refresh operations during accesses to sub-arrays of a pseudo-static memory device. By refreshing sub-arrayj while filling the row Ri (where i≠j) corresponding to sub-arrayi, refresh operations can be performed without risking that a read request will interrupt the refresh operation. Additional refresh operations of sub-arrayi can be performed serially with the operations of filling the row Ri without delaying the overall burst read or write operation if the burst is made sufficiently long.

Description

FIELD OF INVENTION

A method and system for hiding, from an external memory controller, a refresh operation (or plural refresh operations) in a pseudo-static memory device.

BACKGROUND OF THE INVENTION

A variety of memory devices are presently available with varying characteristics. Dynamic Random Access Memory (DRAM) has the advantage that the number of gates per cell is small and the density is generally quite high. On the other hand, DRAM is disadvantageously prone to data loss if the individual rows of data are not periodically refreshed. Accordingly, known systems have used external or internal refresh circuitry to prevent data loss. External circuitry complicates the design of an external memory controller and may therefore be disadvantageous. DRAMs disadvantageously have relatively long array cycle times as compared to other memory devices (e.g., static memories) and therefore may act as a bottleneck for a processor that requests memory more quickly than the DRAM can provide it.

As an alternative, Static Random Access Memory (SRAM) devices utilize a greater number of transistors per memory cell and, as a result, do not require refreshing. Moreover, the transistor interconnections allow data to be read from and written to the device significantly more quickly than DRAMs. Unfortunately, the cost of SRAMs per bit is significantly more expensive than the cost of DRAMs per bit. Accordingly, it is often prohibitively expensive to use SRAM for a computer's main memory, and instead a small amount of SRAM is used only for a memory cache between the processor and a larger amount of DRAM.

As an alternative to both DRAM and SRAM designs, hybrid memories have been introduced that have some of the characteristics of both DRAM devices and SRAM devices. One such device is known as an “Enhanced DRAM” (EDRAM) and is described in U.S. Pat. Nos. 5,887,272, 5,721,862, and 5,699,317 (hereinafter “the '272 patent,” “the '862 patent”, and “the '317 patent,” respectively), each naming Sartore et al. as inventors. (Those inventions have been assigned to the parent company of the assignee of the present invention.) The EDRAM devices disclosed therein provide increased data throughput by providing at least one row register (i.e., a read cache) associated with each DRAM sub-array or with each group of DRAM sub-arrays. Although an EDRAM device can achieve the higher data rate, resembling SRAM access speed, for accesses to data stored in the row register, an EDRAM device still requires externally initiated refresh operations. Column 15, lines 42-56, of the '272 patent discloses that an external signal labeled “/F” controls a refresh cycle. This characteristic requires the use of a memory controller that understands the operation of EDRAM devices and, thus, disadvantageously includes the additional circuitry for initiating those refresh cycles.

SUMMARY OF THE INVENTION

It is an object of the present invention to provide a memory device with faster access speed than a conventional DRAM device but without (1) the cost of a pure SRAM device and (2) the need for an external memory controller to initiate refresh cycles.

This object and other advantageous effects are achieved using a pseudo-static memory device that substantially hides all refresh operations from a memory controller coupled to the pseudo-static memory device. In a first embodiment, such a device internally includes a set of DRAM sub-arrays (matrices) each having plural array rows, such that at least one row in one sub-array can be refreshed while responding to an access request in another sub-array. In a second embodiment, at least one non-array row (e.g., an SRAM row or a register), external to a set of DRAM sub-arrays, is used to allow a refresh of at least one row of its corresponding sub-array when the internal controller, as opposed to an external controller, determines that a refresh would be appropriate. The at least one non-array row may either be either direct mapped or fully associative, and the size of the non-array row may be smaller than, larger than, or equal to the size of each of the plural array rows.

BRIEF DESCRIPTION OF THE DRAWINGS

A more complete appreciation of the invention and many of the attendant advantages thereof will be readily obtained as the same becomes better understood by reference to the following detailed description when considered in connection with the accompanying drawings, wherein:

FIG. 1A is a schematic illustration of a pseudo-static memory architecture according to a first aspect of the present invention;

FIG. 1B is a schematic illustration of the pseudo-static memory architecture of FIG. 1A in which row registers and tag registers have been added;

FIG. 2 is a timing diagram showing parallel read and refresh operations according to a first embodiment of the present invention;

FIG. 3 is a timing diagram showing parallel read and refresh operations according to a second embodiment of the present invention;

FIG. 4 is a timing diagram showing parallel read and refresh operations according to a third embodiment of the present invention;

FIG. 5 is a schematic illustration of a pseudo-static memory architecture according to a second aspect of the present invention;

FIG. 6 is a timing diagram showing parallel read and refresh operations according to a fourth embodiment of the present invention;

FIG. 7 is a schematic illustration of a pseudo-static memory architecture according to a third aspect of the present invention; and

FIG. 8 is a timing diagram showing parallel read and refresh operations according to a sixth embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT

Referring now to the drawings, in which like reference numerals designate identical or corresponding parts throughout the several views, FIG. 1A is an illustration of a pseudo-static memory architecture according to a first aspect of a first embodiment of the present invention. Generally, the present invention includes control circuitry for performing a refresh operation in at least a first DRAM sub-array when a second DRAM sub-array is being accessed (for reading or writing).

Turning now to FIG. 1B, in the second aspect, like the memory architecture shown in FIG. 7 of the '317 patent, a pseudo-static memory device includes both plural DRAM sub-arrays and at least one row (e.g., an SRAM row) coupled to the plural DRAM sub-arrays. The at least one row can act as an intermediate storage area for at least one of the plural DRAM sub-arrays when reading from or writing to that sub-array while contemporaneously refreshing a different DRAM sub-array.

For example, as illustrated in FIG. 2, a read operation corresponding to a memory location in the second row of the first DRAM sub-array (designated “DSA₁”) creates a row miss since the last read row for DSA₁ (labeled “LRR₁”) indicates that the last read row was row zero. Since the data is not already loaded into the row corresponding to DSA₁, the control circuitry of the first embodiment begins the process of loading the row corresponding to DSA₁ with the data of row zero. In parallel, the control circuitry of the first embodiment begins to initiate a refresh operation corresponding to at least one row in the remaining DRAM sub-arrays (i.e., in either DSA₂ or DSA₃ or both). By running a refresh cycle concurrently with an access cycle, it can be insured that the refresh cycle is not interrupted by an access to the same DRAM sub-array.

In this first embodiment, the address of the row to be refreshed is supplied from a standard refresh counter, labeled RC_(global) in FIG. 2, during every DRAM access cycle that causes a read (or write) miss. As shown, when DSA₁ is processing a row read miss and filling the first row, R₁, then the contents of the refresh counter, RC, are read, indicating that row 6 should be refreshed in parallel. At the end of the combined read/refresh cycle, the first row, R₁, contains the data corresponding to row 2, and the refresh counter, RC, contains the address of the next row to be refreshed (i.e., row 7).

The approach illustrated in FIG. 2, however, may adversely waste power by running refreshes excessively often. Consequently, in a second embodiment of the present invention shown in FIG. 3, the pseudo-static memory device determines whether or not an appropriate amount of time has elapsed since the last refresh cycle. A refresh cycle is run behind the given access cycle only if the time has expired. As shown in FIG. 3, consecutive read misses occur for rows 1-4 of DSA₁, thereby updating R₁ and LRR₁ after each miss. Unlike the timing diagram of FIG. 2, on the first read miss, a refresh cycle of at least one other sub-array does not occur since the timer has not yet decremented to zero. Instead, each clock cycle (either internal or external) the timer is decremented, thus RC_(global) remains constant until the timer hits zero. In one implementation, the rollover of the counter sets a “refresh pending” flag, and the timer is allowed to decrement again immediately. In an alternate implementation, the timer is held at zero until the refresh occurs. In either case, the timer is eventually reset (e.g., reset to its maximum value if the timer is a down counter) to signal the length of the next period to wait.)

In FIG. 3, it is not until the fourth row miss that the timer equals zero and a refresh cycle is initiated in parallel with a next row (e.g., read) miss. After performing the refresh, RC_(global) is incremented to its next value (i.e., 7).

As shown in FIG. 4, a third embodiment of the present invention includes both a counter and a timer. The counter acts as a running total of the number of refresh cycles that were missed because a hidden refresh did not occur recently and is referred to herein as a “missed refresh counter” (hereinafter “MRC”). In that configuration, a timer is coupled to an internal or external clock signal, and the value stored in MRC is incremented every time the timer resets. While the value stored in MRC is zero, the third embodiment acts like the embodiment of FIG. 3. However, after MRC accumulates at least one missed refresh cycle, the control circuitry provides a hidden refresh cycle during the next opportunity, to a block other than the one requested by the refresh address, regardless of the value of the timer. After performing the refresh cycle, the value in MRC is decremented to indicate that the control circuitry has “caught up” by one refresh cycle. Subsequent hidden refreshes occur again when possible until the value in MRC is decremented to zero. In a variation of the third embodiment, MRC is used to implement a burst of refreshes, as needed, upon entry into any device mode when internally initiated refreshes are allowed (e.g., a sleep mode).

It should be noted that the ability to hide a refresh cycle behind an access is contingent upon the ability to have the refresh cycle occur for an address in a different DRAM sub-array than the sub-array where the access is occurring. However, if a series of accesses are directed to a first DRAM sub-array, then the other sub-arrays are being refreshed properly, but the first sub-array is not. Accordingly, as shown in FIG. 5, a fourth embodiment of the present invention includes separate refresh counters for each sub-array. This prevents a row in the first sub-array from being “skipped” when the corresponding row is refreshed in the other sub-arrays. As shown in FIG. 6, when a read miss occurs in DSA₁, refreshes in DSA₂ and DSA₃ are performed and their corresponding refresh counters (i.e., RC₂ and RC₃) are updated. However, the value stored in RC, is left unchanged, so that no rows in DSA₁ are skipped for refresh.

In a fifth embodiment of the present invention, in addition to having a dedicated refresh counter per sub-array, each sub-array further includes a missed refresh counter (MRC). By tracking how far behind any sub-array is, each sub-array can “catch up” on missed refresh cycles without refreshing a row in all the sub-arrays, which would otherwise expend excess energy. For example, if a first sub-array (of four sub-arrays) is behind three cycles, a second sub-array is behind two cycles, and a third sub-array is behind one cycle, all three sub-arrays can “catch up” in parallel if the next read miss is to the sub-array that is not behind. After decrementing the corresponding MRCs, only two sub-arrays will be behind. Those remaining two sub-arrays can be updated when possible, and, if it was possible to update those two sub-arrays simultaneously, the last sub-array is updated by itself subsequently.

The embodiments above described can logically be considered an n-way grouping (e.g., 3-way grouping in FIGS. 1A and 1B) with a single DRAM sub-array per group. However, the method and structure of the present invention can be expanded to include groupings with plural sub-arrays per group, as shown in FIG. 7. In such a hierarchical embodiment, group refreshes can be performed (1) one sub-array per group per refresh cycle, (2) plural, but not all, sub-arrays per group per refresh cycle, or (3) all sub-arrays per group per refresh cycle. It is also possible to select a first one of three group refreshing techniques for one group while selecting a second group refreshing technique for a second group. Moreover, the groups need not be refreshed at the same time. It is possible to either refresh all groups simultaneously (with each group refreshing at least one sub-array), or refresh individual groups individually based on the needs of that group. It is to be understood that attempting to refresh all sub-arrays simultaneously may lead to excessive current draw. As a result, the preferred embodiment does not refresh all possible sub-arrays of all groups simultaneously. Moreover, the hierarchy is not limited to a second-order hierarchy (i.e., a hierarchy of groups and sub-arrays). Any number of sub-arrays can be formed into a sub-group, and plural sub-groups can be formed into a group, with plural groups being formed into a super-group, etc.

In the most complex refreshing technique of a second-order hierarchy, each sub-array of each group includes a refresh counter (RC_i,j) and a corresponding missed refresh counter (MRC_i,j). Thus, in the embodiment illustrated in FIG. 7, on two consecutive row read misses corresponding to DSA_4,1, it is possible to perform two consecutive sets of refresh operations. In parallel with the first row read miss, the following rows are illustratively refreshed: the row referenced by RC_1,1 of DSA_1,1, the row referenced by RC_2,2 of DSA_2,2, both of the rows referenced by RC_3,1 of DSA_3,1 and RC_3,2 of DSA_3,2, and the row referenced by RC_4,2 of DSA_4,2, (resulting in a value of zero in MRC_4,2). For each of the RC_i,j's that were accessed, their values are incremented by one. Then, in parallel with the second row read miss, the following rows are illustratively refreshed: the row referenced by RC_1,2 of DSA_1,2, the row referenced by RC_2,1 of DSA_2,1, and both of the rows referenced by RC_3,1 of DSA_3,1 and RC_3,2 of DSA_3,2. It is noted that the row referenced by RC_4,2 of DSA_4,2 is not refreshed unless the timer is zero since DSA_4,2 has caught up to all missed refreshes.

As shown in FIG. 8, a sixth embodiment of the present invention analyzes the type of command to be performed to further hide at least one refresh cycle for the sub-array being addressed. In a burst read operation that produces a row read miss, data is read from a DRAM sub-array (DSA_i) and loaded into a corresponding row (R_i). Since the burst read is guaranteed to perform plural consecutive reads, an interruption-free refresh operation can be performed on DSA₁ since R_i is providing the data for the burst. In the illustrated example, the four data words (or bits) are output in parallel with a refresh of the next row in DSA_i, but the refresh is performed serially as compared to the row filling operation. (This type of refresh operation is referred to hereinafter as a “serial refresh operation.”) By extending the number of words (or bits) output in any burst, the number of rows that can be serially refreshed per burst increases. For example, it may be possible to hide at least two serial refresh cycles behind a burst read of 8 words (or bits).

Moreover, a burst write operation may likewise provide an opportunity for serial refresh operations. While a row R_i is being filled with data corresponding to a burst write, at least one row (depending on the size of the burst) of DSA_i can be refreshed.

In a seventh embodiment of the present invention, the control circuitry for performing parallel refreshes and the control circuitry for performing serial refreshes is combined to provide a high number of refreshed rows in burst accesses. In addition to the refresh operations that can be performed in the first through fifth embodiments, additional refresh operations of non-accessed sub-arrays can be performed in parallel with the serial refresh operations. For example, in an n-way grouping with one sub-array per group, while R_i is being filled by DSA_i, a maximum of n−1 refresh operations can be performed in parallel for DSA_j where i≠j. Then, during the burst read, one row for, at most, each of the n sub-arrays at a time can be refreshed while data is read from R_i, but fewer rows can be refreshed if a power consumption checking circuit determines that a maximum power (or current) draw would be exceeded.

If a write row is provided as well, the opposite refreshing schedule would be used during a burst write operation. That is, a maximum of n rows could be refreshed while the data is being written to R_i, then a maximum of n−1 refresh operations at a time could be performed in parallel with writing the data from R_i to DSA_i.

In order to reduce the possibility that any one sub-array will become too far behind in its refreshes, in a variation of the first through seventh embodiments disclosed herein, the address decoders are configured to put logically adjacent rows in different sub-arrays. For example, in a row of 128 bits, binary addresses XXX00XXXXXXX are stored in a first sub-array, binary addresses XXX01XXXXXXX are stored in a second sub-array, binary addresses XXX10XXXXXXX are stored in a third sub-array, and binary addresses XXX11XXXXXXX are stored in a fourth sub-array, where “X” represents a “don't care” in the address. Thus, by consecutively referencing all memory locations between 000000000 and 111111111111, the memory device will cycle between sub-arrays 0, 1, 2, and 3 and then return to zero eight times. Using a linear addressing scheme, 1024 consecutive operations would occur for DSA₁ followed by 1024 consecutive operations for DSA₂, etc., and data could be lost during those bursts.

The present invention can utilize a number of different conditions to trigger an internal examination as to whether a refresh cycle should be performed. The memory device receives commands (e.g., in the form of an explicit command sent to the memory or in the form of the assertion or de-assertion of a signal on an external pin) that are decoded internally by the memory device. For example, a read command issued to the memory device can be decoded to determine if, for example, (1) if it is time to hide a refresh behind a read from a non-array row or (2) if it is time to hide a refresh of an array row of a first sub-array behind an access to an array row of a second sub-array. Similarly, upon receipt of a command in the form of a refresh signal on a refresh pin, the memory device may decide internally that an insufficient amount of time has passed (or that power considerations would make it inadvisable to perform a refresh at the requested time), and the memory device blocks a refresh from occurring in response to that command. Other commands that can be decoded include, but are not limited to, write requests and chip deselects.

Obviously, numerous modifications and variations of the present invention are possible in light of the above teachings. It is therefore to be understood that, within the scope of the appended claims, the invention may be practiced otherwise than as specifically described herein.

Claims

1. A device, comprising: a plurality of DRAM sub-arrays, wherein each DRAM sub-array includes a plurality of array rows;an address decoder configured to decode an address of a memory access request, wherein the decoded address indicates an accessed DRAM sub-array from the plurality of DRAM sub-arrays; andrefresh circuitry configured to perform a refresh of a plurality of refreshed array rows for at least one refreshed DRAM sub-array, wherein the at least one refreshed DRAM sub-array is not the accessed DRAM sub-array, and wherein the plurality of refreshed array rows is determined based on power consumption;wherein each DRAM sub-array further comprises a missed refresh counter, and wherein each missed refresh counter is configured to track a number of refreshes missed by at least one of the plurality of DRAM sub-arrays.

2. The device of claim 1, wherein the refresh circuitry comprises a timer configured to establish a time period before a next array row is refreshed from the at least one refreshed DRAM sub-array.

3. The device of claim 1, further comprising a command decoder configured to internally determine when a refresh cycle can be hidden.

4. The device of claim 1, wherein the memory access request is a burst read operation to be performed on an array row of the accessed DRAM sub-array, and wherein the refresh circuitry is further configured to refresh, while the burst read operation is performed, at least one array row consecutive with the array row of the accessed DRAM sub-array.

5. The device of claim 4, wherein the refresh circuitry is further configured to serially refresh at least two array rows consecutive with the array row of the accessed DRAM sub-array.

6. The device of claim 1, wherein the plurality of DRAM sub-arrays are organized into a plurality of DRAM sub-array groups, and wherein the refresh circuitry is further configured to refresh at least one DRAM sub-array group in a single refresh cycle.

7. The device of claim 6, wherein the refresh is configured to be managed by refreshing exactly one DRAM sub-array per DRAM sub-array group per refresh cycle.

8. The device of claim 6, wherein the refresh is configured to be managed by refreshing all of the DRAM sub-arrays per DRAM sub-array group for the at least one DRAM sub-array group.

9. The device of claim 6, wherein the refresh is configured to be managed by refreshing a plurality of, but not all, DRAM sub-arrays per DRAM sub-array group for the at least one DRAM sub-array group.

10. The device of claim 6, wherein the refresh is performed using at least one group refresh technique, and wherein each DRAM sub-array group from the plurality of DRAM sub-array groups is refreshed independently.

11. The device of claim 10, wherein the at least one group refresh technique comprises a refresh technique configured to refresh a DRAM sub-array group selected from the group consisting of: (i) refreshing exactly one DRAM sub-array for the DRAM sub-array group, (ii) refreshing a plurality of, but not all, DRAM sub-arrays for the DRAM sub-array group and (iii) refreshing all DRAM sub-arrays for the DRAM sub-array group.

12. The device of claim 1, further comprising a refresh counter for each DRAM sub-array, wherein each refresh counter is configured to track a next array row to be refreshed for at least one DRAM sub-array.

13. The device of claim 1, wherein the refresh circuitry comprises circuitry that, in response to refreshing an array row in a DRAM sub-array, is configured to increment the missed refresh counter for the DRAM sub-array.

14. The device of claim 1, wherein the refresh circuitry comprises circuitry that, in response to refreshing an array row in a DRAM sub-array, is configured to decrement the missed refresh counter for the DRAM sub-array if the dedicated sub-array missed refresh counter exceeds zero.

15. The device of claim 1, wherein the refresh circuitry comprises circuitry that, in response to not refreshing an array row in a DRAM sub-array, is configured to increment the missed refresh counter for the DRAM sub-array.

16. A method of refreshing a memory device having a plurality of DRAM sub-arrays, wherein each DRAM sub-array has a plurality of array rows, the method comprising: decoding an address of a first memory access request;performing a first refresh of at least one array row in at least one DRAM sub-array;executing the first memory access request contemporaneously with the first refresh;after performing the first refresh, determining a time period before performing a second refresh;blocking a command to perform the second refresh until the time period has elapsed; andtracking a number of refreshes missed by at least one of the plurality of DRAM sub-arrays.

17. The method of claim 16, further comprising: determining which of the DRAM sub-arrays are referenced by a second memory access request and which of the plurality of array rows in the referenced DRAM sub-arrays are referenced;in response to the second memory access request being a read request, determining that a third refresh for a refreshed array row of a refreshed DRAM sub-array can be hidden behind the second memory access request because the second memory request is either (a) a read request corresponding to an array row that is not the refreshed array row or (b) a read request corresponding to a DRAM sub-array that is not the refreshed DRAM sub-array; andin response to the determination that the third refresh can be hidden, performing the third refresh simultaneously with the second memory access request.

18. The method of claim 16, wherein the plurality of DRAM sub-arrays are organized into a plurality of DRAM sub-array groups, and wherein the method further comprises: refreshing the memory device by using at least one group refresh technique, wherein each DRAM sub-array group from the plurality of DRAM sub-array groups is refreshed independently.

19. The method of claim 16, further comprising receiving, into a non-array row external to the plurality of DRAM sub-arrays, from the DRAM sub-array referenced by the address of the first memory access request at least a portion of an array row corresponding to the address of the first memory access request, wherein said receiving comprises receiving the portion into a static random access memory (SRAM) row.