A method and system for hiding, from an external memory controller, a refresh operation (or plural refresh operations) in a pseudo-static memory device.
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.
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.
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:
Referring now to the drawings, in which like reference numerals designate identical or corresponding parts throughout the several views,
Turning now to
For example, as illustrated in
In this first embodiment, the address of the row to be refreshed is supplied from a standard refresh counter, labeled RC(global) in
The approach illustrated in
In
As shown in
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
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
In the most complex refreshing technique of a second-order hierarchy, each sub-array of each group includes a refresh counter (RCi,j) and a corresponding missed refresh counter (MRCi,j). Thus, in the embodiment illustrated in
As shown in
Moreover, a burst write operation may likewise provide an opportunity for serial refresh operations. While a row Ri is being filled with data corresponding to a burst write, at least one row (depending on the size of the burst) of DSAi 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 Ri is being filled by DSAi, a maximum of n−1 refresh operations can be performed in parallel for DSAj 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 Ri, 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 Ri, then a maximum of n−1 refresh operations at a time could be performed in parallel with writing the data from Ri to DSAi.
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.
This is a continuation application of U.S. patent application Ser. No. 09/828,283 filed Apr. 5, 2001, now U.S. Pat. No. 7,085,186 issued Aug. 1, 2006, which is herein incorporated for all purposes.
Number | Name | Date | Kind |
---|---|---|---|
4888733 | Mobley | Dec 1989 | A |
5134310 | Mobley | Jul 1992 | A |
5226147 | Fujishima et al. | Jul 1993 | A |
5446696 | Ware et al. | Aug 1995 | A |
5559750 | Dosaka et al. | Sep 1996 | A |
5577223 | Tanoi et al. | Nov 1996 | A |
5619468 | Ghosh et al. | Apr 1997 | A |
5699317 | Sartore et al. | Dec 1997 | A |
5721862 | Sartore et al. | Feb 1998 | A |
5778436 | Kedem et al. | Jul 1998 | A |
5887272 | Sartore et al. | Mar 1999 | A |
5890195 | Rao | Mar 1999 | A |
5991851 | Alwais et al. | Nov 1999 | A |
5999474 | Leung et al. | Dec 1999 | A |
6023745 | Lu | Feb 2000 | A |
6055192 | Mobley | Apr 2000 | A |
6064620 | Mobley | May 2000 | A |
6104658 | Lu | Aug 2000 | A |
6134167 | Atkinson | Oct 2000 | A |
6141281 | Mobley | Oct 2000 | A |
6147921 | Novak et al. | Nov 2000 | A |
6151236 | Bondurant et al. | Nov 2000 | A |
6208577 | Mullarkey | Mar 2001 | B1 |
6215714 | Takemae et al. | Apr 2001 | B1 |
6324626 | Uenoyama et al. | Nov 2001 | B1 |
6392948 | Lee | May 2002 | B1 |
7085186 | Mobley | Aug 2006 | B2 |
20010033245 | Campanale et al. | Oct 2001 | A1 |
20020003741 | Maesako et al. | Jan 2002 | A1 |
Number | Date | Country | |
---|---|---|---|
Parent | 09828283 | Apr 2001 | US |
Child | 11238182 | US |