Dual-port SRAM memory using single-port memory cell

[0001] The present invention relates to semiconductor memory systems, such as static random access memory (SRAM) systems or dynamic random access memory (DRAM) systems. In particular, the present invention relates to a method and apparatus of implementing a dual-port memory using a single-port memory cell. In addition, the present invention relates to dual-port memories in which accesses on the two ports can be asynchronous to each other.

[0002] In the past, dual-port semiconductor memories have been implemented with conventional dual-port memory cells, such as dual-port static random access memory (SRAM) cells and dual-port dynamic random access memory (DRAM) cells.

[0003] FIG. 1 is a circuit diagram illustrating a conventional dual-port SRAM cell 100. Dual-port SRAM cell 100 includes eight transistors 101-108, and requires two word lines WL_P1-WL_P2 and four bit lines BL_P1, BL_P1#, BL_P2 and BL_P2#. The first port of SRAM cell 100 (port 1) is accessed via word line WL_P1 and bit lines BL_P1-BL_P1#. The second port of SRAM cell 100 (port 2) is accessed via word line WL_P2 and bit lines BL_P2-BL_P2#.

[0004] FIG. 2 is a circuit diagram illustrating a conventional dual-port DRAM cell 200. Dual-port DRAM cell 200 includes two transistors 201-202 and a capacitor 203, and requires two word lines WL₁-WL₁ and two bit lines BL₁-BL₁. The first port of DRAM cell 200 is accessed via word line WL₁ and bit line BL₁. The second port of DRAM cell 200 is accessed via word line WL₁ and bit line BL₁. Dual-port DRAM cell 200 is described in more detail in U.S. Pat. No. 5,923,593, entitled “Multi-port DRAM Cell and Memory System Using Same”.

[0005] FIGS. 3 and 4 are circuit diagrams illustrating a conventional single-port SRAM cell 300 and a conventional single-port DRAM cell 400, respectively. Single-port SRAM cell 300 includes six transistors 301-306, and requires two bit lines BL and BL# and one word line WL. Single-port DRAM cell 400 includes one transistor 401, one capacitor 402 and requires one bit line BL and one word line WL. Because the dual-port cells 100 and 200 include significantly more circuit elements than the corresponding single-port cells 300 and 400, the dual-port cells 100 and 200 occupy significantly more layout area than the corresponding single-port cells 300 and 400.

[0006] In addition, when both ports of dual-port SRAM cell 100 (or both ports of dual-port DRAM cell 200) are accessed at the same time, intermittent data usually results. That is, when a write operation is performed on one port of a dual-port cell, and a write operation is performed on the other port of the dual-port cell, the write and read operations may interfere with one another. In these cases, erroneous (or indefinite) data can be read from or written to the dual-port cell. Note that the intermittent data issue does not exist in single-port cells, such as SRAM cell 300 and DRAM cell 400.

[0007] It would be desirable to be able to use single-port cells to perform a dual-port function, thereby minimizing cell layout area and eliminating intermittent data.

[0008] Accordingly, the present invention provides a dual-port memory system implemented using single-port memory cells. More specifically, the present invention includes an array of single-port memory cells, which may be, for example, SRAM or DRAM cells. A first set of registers store data, address and control values associated with a first port, and a second set of registers store data, address and control values associated with a second port. An access arbiter is configured to receive first and second port enable signals associated with the first and second ports, respectively. If the single-port memory cells require refresh, then the access arbiter also receives a refresh request signal.

[0009] The access arbiter includes circuitry for activating a memory request signal when one or more of the first port enable signal, the second port enable signal or the refresh request signal is activated. The activated memory request signal causes a memory sequencer to sequentially activate and de-activate the control signals necessary to access the memory array. In one embodiment, one of these control signals is a row access signal (RAS).

[0010] The access arbiter also includes a synchronization circuit having a first stage and a second stage. The first and second port enable signals and the refresh request signal are all latched into the first stage of the synchronization circuit in response to the row access signal. A delay circuit introduces a delay T to the row access signal, thereby creating a delayed row access signal. The contents of the first stage of the synchronization circuit are latched into the second stage of the synchronization circuit in response to the delayed row access signal. The delay T is selected to control the expected mean-time-before-failure.

[0011] The contents of the second stage of the synchronization circuit are provided to a prioritization circuit, which generates a first select signal associated with the first port, a second select signal associated with the second port, and a third select signal associated with the refresh request. The prioritization circuit ensures that only the highest priority signal of the first, second and third select signals is activated at any given time. When activated, the first, second and third select signals enable an access on the first port, an access on the second port and a refresh access, respectively.

[0012] The first, second and third select signals are used in combination with the row access signal to de-activate the memory request signal.

[0013] In accordance with one embodiment, one of the port enable signals is activated and processed while a clock signal has a first logic state. The refresh request signal can then be activated and processed while the clock signal has a second logic state. This embodiment is useful for implementing a display buffer.

[0014] The present invention will be more fully understood in view of the following description and drawings.

[0024] The embodiments described herein are provided for illustrative purposes only and are not meant to serve as limits to the present invention. In a first embodiment, both ports are accessed with no continuous running clock. In a second embodiment, one port has an access interface with a continuous running clock, but the other port does not. The memory of the second embodiment is particularly useful as a display buffer, wherein the memory stores display data for an LCD panel or display monitor. The display buffer has a first (display) port from which the display data is read in regular fashion to refresh the display screen. This display port is usually synchronous with a continuous running pixel clock or row display clock. The display buffer may also have a second (screen update) port to allow memory accesses by a graphics controller or a processing unit (which puts new display data in the display buffer for screen update). The screen update port usually works without a continuous running clock signal. Rather, a memory access signal is used to access the screen update port. The activated memory access signal indicates the beginning of an external memory access.

[0025] In the dual-port memory systems described below, single-port Dynamic Random Accessible Memory (DRAM) cells are used implement a dual-port memory, wherein the refreshing of the DRAM cells is completely hidden (such that the memory behaves like a dual-port SRAM). Memory accesses from the two ports can be synchronous or asynchronous with respect to each other. Memory accesses from the two ports are scheduled according to their order of arrival, wherein the earlier arriving memory requests are served first. When both ports receive access requests at substantially the same time, an access arbiter determines which access is processed first. The access arbiter also arbitrates refresh access requests.

[0026] Each port includes a transaction buffer, which stores the access information (e.g., address, data, read-write indicator and byte-write indicator) of a transaction that is pending while the memory is busy serving another transaction or refresh. Because the memory array handles only one access at a time, simultaneous accesses from the two ports to the same memory location does not cause the memory content or the read data to be corrupted. However, the memory can be overrun if the access frequency of the two ports together with refresh exceeds the maximum operating frequency of the memory array. To avoid the memory overrun situation the memory array is designed so that it can operate at frequencies equal to or greater than the sum of the maximum access frequencies of the two ports and internal refreshes.

[0027] A refresh controller similar to the one described in U.S. Pat. No. 6,504,780B2, “Method and Apparatus For Completely Hiding Refresh Operations In A DRAM Device Using Clock Division” can be incorporated to generate refresh accesses periodically to refresh the memory contents. In one embodiment, the clock to the refresh controller is generated inside the memory using a ring oscillator or a bi-stable vibrator. Such circuit is well known in the art and is not described further in this patent application. In this case, the refresh controller operation is not synchronous with either of the two ports and the two ports are not synchronous with each other. Therefore, the three memory accesses, refresh, and access from the two ports, are asynchronous with each other. In the second embodiment described below, the clock input of the refresh controller is connected to the clock signal of the display port. Therefore, the memory refresh is synchronous with the display clock.

[0028] FIG. 5 is a block diagram of a dual-port memory system 500, which is implemented using single-port memory cells in accordance with one embodiment of the present invention. As described in more detail below, refresh operations and external accesses from the two ports of memory system 500 are asynchronous with respect to one another.

[0029] Dual-port memory system 500 includes memory array 501, 3-to-1 multiplexer 502, memory sequencer 503, access arbiter 505, a first transaction buffer 510 associated with a first port (port A), a second transaction buffer 520 associated with a second port (port B), a refresh controller 530 and logical AND gates 541-542. The first transaction buffer 510 includes input data register 511, output data register 512 and address/control register 513. Similarly, the second transaction buffer 520 includes input data register 521, output data register 522 and address/control register 523. The first transaction buffer 510 stores an input data word LD_A[31:0] (in input data register 511), an output data word Q_A[31:0] (in output data register 512) and an access address LA_A[14:0] and a read-write control signal LRW_A# (in address/control register 513). Similarly, the second transaction buffer 520 stores an input data word LD_B[31:0] (in input data register 521), and an output data word Q_B[31:0] (in output data register 522) and an access address LA_B[14:0] and a read-write control signal LRW_B# (in address/control register 523). The first and second transaction buffers 510 and 520 store access information associated with port A and port B, respectively, in case memory system 500 is busy serving the other port.

[0030] Memory array 501 can include one or more memory banks. In the described examples, memory array 501 is implemented using ordinary DRAM cells or embedded DRAM cells, such as those described in U.S. Pat. No. 6,075,720, entitled “Memory Cell For DRAM Embedded In Logic”. Memory array 501 can alternately be implemented using other volatile memory cells that require periodic refreshing. Memory array 501 can also be implemented using SRAM cells. If memory array 501 is implemented using SRAM cells, the refresh operations of the embodiments described below will not exist, and may therefore be ignored.

[0031] The memory cells of array 501 can be accessed through either of the dual memory ports A or B. In the described example, memory array 501 includes memory cells organized into 1k rows by 1k columns (although this size is not necessary to practice the present invention). In addition, memory array 501 includes sense amplifiers, word line drivers, read data amplifiers, write data drivers, and row and column address decoders. An example of such an array is described in more detail in U.S. Pat. No. 6,028,804, entitled “Method and Apparatus For 1-T SRAM Compatible Memory”.

[0032] The address MA[14:0] provided to memory array 501 is driven by 3-to-1 multiplexer 502, which is controlled to select the latched port A access address LA_A[14:0], the latched port B access address LA_B[14:0], or the refresh address RFA[9:0] provided by refresh controller 530. Because memory array 501 is refreshed one row at a time, and memory array 501 has only 1024 rows, only the 10-bit row address MA[14:5] is required during a refresh operation.

[0033] Multiplexer 502 also passes the read-write indicator signal associated with the selected port (i.e., LRW_A# for port A or LRW_B# for port B) to memory array 501 as the read-write indicator signal MRW#. During a refresh operation (i.e., when the refresh control signal REF is activated high), the logic value of the read-write indicator signal MRW# is ignored.

[0034] Multiplexer 502 also passes the latched input data associated with the selected port (i.e., LD_A[31:0] for port A or LD_B[31:0] for port B) to memory array 501 as the write data value MD[31:0].

[0035] The internal operation of memory array 501 is controlled by a row access (RAS) signal, a sense amplifier enable (SENSE) signal, a column access (CAS) signal, and a pre-charge (PRC) signal. In general, the RAS, SENSE, CAS and PRC signals can be referred to as memory access control signals. The RAS, SENSE, CAS and PRC signals are all generated by memory sequencer 503 in response to a memory request signal (MRQ) and a refresh request signal (REF), which are generated within access arbiter 505. In one embodiment, memory sequencer 503 and the associated memory control signals RAS, SENSE, CAS and PRC can be implemented in a manner similar to that described in U.S. Pat. No. 6,222,785 B1, entitled “Method and Apparatus for Refreshing a Semiconductor Memory Using Idle Memory Cycles”. In general, a memory access cycle starts with the activation of the RAS signal and ends with the deactivation of the RAS signal. Memory sequencer 503 initiates a memory access cycle each time the received memory request signal MRQ is activated to a logic high state. The memory request signal MRQ is driven by memory access arbiter 505, which arbitrates the access requests associated with ports A and B and memory refresh controller 530.

[0036] Each port interface consists of data output bus Q_N[31:0], input data bus D_N[31:0], input address bus A_N[14:0], read-write indicator line RW_N#, and port enable line P_NE, where the ‘N’ in each signal name identifies either port A or port B.

[0037] An external device (not shown) initiates an access on one of the ports by first driving the address on the address bus A_N[14:0], the read write indicator RW_N# in the correct state (‘0’ for write and ‘1’ for read), and in case of a write operation, the write data on the input data bus D_N[31:0]. Subsequently, the external device drives the port enable signal P_NE to a logic high state. As a result, the address A_N[14:0] and the read write indicator RW_N# are latched into the associated address/control register (i.e., 513 or 523), and in the case of a write operation, the write data D_N[31:0] is latched into the associated input data register (i.e., 511 or 521).

[0038] Refresh controller 530 initiates refresh accesses within memory system 500. Refresh controller 530 periodically activates a refresh request signal RREQ to a logic high state each time a row of memory array 501 is to be refreshed. The period of the refresh request signal RREQ is selected to ensure that all memory cells of array 501 are refreshed in a timely manner (i.e., there is no loss of data). Refresh controller 530 also provides the row address RFA[9:0] of the row to be refreshed. The refresh row address RFA[9:0] is updated each time a refresh operation is performed.

[0039] Memory arbiter 505 arbitrates access requests received from port A, port B and refresh controller 530. Access to memory array 501 is granted on a first-come-first-served basis. If all three access requests arrive substantially simultaneously, then priority is given, in order, to the port A access request, the port B access request, and then the refresh request. The same priority is given if two of the three access requests arrive substantially simultaneously.

[0040] FIG. 6 is a circuit diagram of access arbiter 505 in accordance with a first embodiment of the present invention. Access arbiter 505 includes set-reset (S-R) flip-flops 601-603, logical NOR gates 610-611, synchronizer 620, priority encoder 630, logical AND gates 641-643 and delay circuit 650. Synchronizer 620 includes first stage flip-flops 621-623 and second stage flip-flops 625-627. Priority encoder 630 includes logical NOR gates 631-632

[0041] The access request signals P_AE, P_BE and RREQ are applied to set input terminals of S-R flip-flops 601, 602 and 603, respectively. When any one of these access request signals P_AE, P_BE, or RREQ is activated high, the corresponding S-R flip-flop 601, 602 or 603 is set to provide a logic high output signal. The latched output signals provided by S-R flip-flops 601, 602 and 603, are labeled LPE_A, LPE_B and RFRQ, respectively. The LPE_A, LPE_B and RFRQ signals are provided to the input terminals of NOR gate 610. If any one of the LPE_A, LPE_B and RFRQ signals has a logic high state, NOR gate 610 provides a logic low output signal to one input terminal of NOR gate 611. The other input terminal of NOR gate 611 is coupled to receive the row access signal RAS. As long as memory array 501 is not currently being accessed (i.e., the RAS signal is de-activated low), NOR gate 611 provides a logic high memory request signal MRQ in response to the logic low output of NOR gate 610. The logic high state of the memory request signal MRQ activates memory sequencer 503 to generate the memory control signals RAS, SENSE, CAS and PRC. The activation of the RAS signal starts a memory access cycle.

[0042] Within arbiter 505, the activation of the RAS signal causes NOR gate 611 to de-activate the memory request signal MRQ to a logic low state. The RAS signal is also provided to the clock terminals of first stage flip-flops 621-623 and the input terminal of delay circuit 650. The logic high state of the RAS signal causes the LPE_A, LPE_B and RFRQ signals to be latched into first stage flip-flops 621, 622 and 623, respectively. In addition, the logic high state of the RAS signal propagates through delay circuit 650. Delay circuit 650 is designed to have a delay time of T nanoseconds (ns). As a result, delay circuit 650 provides an output signal (D_RAS) that is delayed by T ns with respect to the RAS signal. The delayed RAS signal (D_RAS) is applied to the clock terminals of flip-flops 625-627. When the D_RAS signal transitions to a logic high state (i.e., T ns after the RAS signal is activated), the output signals provided by first stage flip-flops 621-623 are latched into second stage flip-flops 625-627, respectively, and are provided as the output signals SEL_A, SEL_B and REF, respectively.

[0043] First stage flip-flops 621-623 and second stage flip-flops 625-627 serve as a dual-stage synchronizer for the asynchronous signals LPE_A, LPE_B and RFRQ. The use of multi-stage flip-flops for synchronization of asynchronous events has been documented, for example, in “Digital Logic MetaStability and Flip Flop MTBF Calculation”, http://www.interfacebus.com/Design MetaStable.html and “The behavior of Flip-Flops Used as Synchronizers and Prediction of Their Failure Rate”, by Harry J. M. Veendrick, JSSC vol. SC-15, No. 2, 1980, pp. 169-176. The mean-time-before-failure (MTBF) of a synchronizer is an exponential function of the delay between triggering the first stage of the synchronizer (e.g., flip-flops 621-623) and triggering the second stage of the synchronizer (e.g., flip-flops 625-627). Thus, the delay T introduced by delay circuit 650 is chosen so that the MTBF is tolerable in memory system 500.

[0044] NOR gates 631 and 632 are connected to form an access priority encoder 630, which generates the output signals SEL_A, SEL_B and REF. If the SEL_A signal (i.e., the contents of second stage flip-flop 625) is activated high, NOR gates 631 and 632 will de-activate the SEL_B and REF signals, respectively, to logic low states. This effectively provides the SEL_A signal with the highest priority. As described in more detail below, the logic high state of the SEL_A signal enables an access operation to be performed on port A of memory system 500.

[0045] Logic AND gate 641 is configured to receive the SEL_A signal and the RAS signal. During an access operation on port A, both the SEL_A and RAS signals will be activated high. Under these conditions, AND gate 641 provides a logic high output signal to the reset terminal of S-R flip-flop 601. In response, S-R flip-flop 601 is reset to provide a logic low output signal LPE_A. As long as the next access to memory array 501 is not on port A, the logic low LPE_A signal is transferred through synchronizer 660 in response to the next rising edge of the RAS signal (i.e., the next access of memory array 501), thereby returning the SEL_A signal to a logic low state after the access on port A is complete. Note that if the next access to memory array 501 occurs on port A, then the LPE_A signal will transition to a logic high state prior to the next rising edge of the RAS signal (thereby enabling the SEL_A signal to remain activated high).

[0046] If the SEL_A signal is de-activated low and the Q output of second stage flip-flop 626 is activated high, then NOR gate 631 activates the SEL_B signal to a logic high state, and NOR gate 632 deactivates the REF signal to a logic low state. This effectively provides the SEL_B signal with the second highest priority. As described in more detail below, the logic high state of the SEL_B signal enables an access operation to be performed on port B of memory system 500.

[0047] Logic AND gate 642 is configured to receive the SEL_B signal and the RAS signal. During an access operation on port B, both the SEL_B and RAS signals will be activated high. Under these conditions, AND gate 642 provides a logic high output signal to the reset terminal of S-R flip-flop 602. In response, S-R flip-flop 602 is reset to provide a logic low output signal LPE_B. As long as the next access to memory array 501 is not on port B, the logic low LPE_B signal is transferred through synchronizer 660 in response to the next rising edge of the RAS signal (i.e., the next access of memory array 501), thereby returning the SEL_B signal to a logic low state after the access on port B is complete. Note that if the next access to memory array 501 occurs on port B, then the LPE_B signal will transition to a logic high state prior to the next rising edge of the RAS signal.

[0048] If the SEL_A signal and the Q output of second stage flip-flop 626 are both deactivated low, then NOR gate 631 provides a SEL_B signal that is deactivated low. If the Q output of flip-flop 627 is activated high under these conditions, then NOR gate 632 activates the REF signal to a logic high state. This effectively provides the REF signal with the lowest priority. As described in more detail below, the logic high state of the REF signal enables a refresh operation to be performed within memory array 501.

[0049] Logic AND gate 643 is configured to receive the REF signal and the RAS signal. During a refresh access to memory array 501, both the REF and RAS signals will be activated high. Under these conditions, AND gate 643 provides a logic high output signal to the reset terminal of S-R flip-flop 603. In response, S-R flip-flop 603 is reset to provide a logic low output signal RFRQ. As long as the next access to memory array 501 is not a refresh access, the logic low RFRQ signal is transferred through synchronizer 660 in response to the next rising edge of the RAS signal (i.e., the next access of memory array 501), thereby returning the REF signal to a logic low state after the refresh access is complete. Note that if the next access to memory array 501 is a refresh access, then the RFRQ signal will transition to a logic high state prior to the next rising edge of the RAS signal.

[0050] The SEL_A and SEL_B signals are provided to select input terminals of 3-to-1 multiplexer 502 to select the address used to access memory array 501. As described above, the priority encoder 630 formed by NOR gates 630 and 631 allows only one of the outputs SEL_A, SEL_B and REF to be activated at a time.

[0051] When the SEL_A signal is activated high (and the SEL_B and REF signals are de-activated low), multiplexer 502 routes the latched address signal LA_A[14:0], the latched read write signal LRW_A# and the latched data signal LD_A[31:0] associated with port A to memory array 501.

[0052] When the SEL_B signal is activated high (and the SEL_A and REF signals are de-activated low), multiplexer 502 routes the latched address signal LA_B[14:0], the latched read write signal LRW_B# and the latched data signal LD_B[31:0] associated with port B to memory array 501.

[0053] When both the SEL_A and SEL_B signals are deactivated low, multiplexer 502 routes the refresh address RFA[9:0] to memory array 501. Arbiter 505 routes the refresh signal REF directly to memory array 501. If the refresh signal REF is activated high (the SEL_A and SEL_B signals are de-activated low), then memory array 501 performs a refresh operation to the row identified by refresh address RFA[9:0]. The read/write control signal MRW# provided by multiplexer 502 is ignored when the refresh signal REF is activated.

[0054] The SEL_A signal is also provided to AND gate 541 (FIG. 5), along with the pre-charge signal PRC, and the latched port A read/write signal LRW_A#. In response, AND gate 541 provides the port A output enable signal P_AO. The port A output enable signal P_AO is activated high at the end of a read access to port A (i.e., LRW_A#=1 (read), SEL_A=1 (access to port A), PRC=1 (end of access)). The activated port A output enable signal P_AO is used to latch the read data value MQ[31:0] into data output register 512.

[0055] Similarly, the SEL_B signal is also provided to AND gate 542, along with the pre-charge signal PRC, and the latched port B read/write signal LRW_B#. In response, AND gate 542 provides the port B output enable signal P_BO. The port B output enable signal P_BO is activated high at the end of a read access to port B (i.e., LRW_B#=1 (read), SEL_B=1 (access to port B), PRC=1 (end of access)). The activated port B output enable signal P_BO is used to latch the read data value MQ[31:0] into data output register 522.

[0056] FIG. 7 is a waveform diagram 700 illustrating the signals in arbiter 505 and memory sequencer 503 for the worst case operating conditions of memory system 500, wherein access requests on ports A and B and a refresh request occur simultaneously.

[0057] Prior to time T0, an external accessing circuit (not shown) provides the read/write control signal RW_A#, the address value A_A[14:0] and data value D_A[31:0] associated with an access on port A. Similarly, an external accessing circuit provides the read/write control signal RW_B#, the address value A_B[14:0] and the data value D_B[31:0] associated with an access on port B. Note that data values D_A[31:0] and D_B[31:0] are not provided for read accesses.

[0058] At time T₀, refresh controller 530 activates the refresh request signal RREQ, and port enable signals P_AE and P_BE are activated. That is, the access request signals P_AE, P_BE, and RREQ all go high at the same time. Note that the refresh request signal RREQ, the port enable signal P_AE and the port enable signal P_BE are asynchronous with respect to each other. That is, there is no common clock signal controlling the activation of the RREQ, P_AE and P_BE signals.

[0059] Within access arbiter 505, the logic high states of the P_AE, P_BE and RREQ signals simultaneously set S-R flip-flops 601, 602 and 603, respectively, thereby causing the corresponding latched output signals LPE_A, LPE_B and RFRQ to transition to logic high states. In response, NOR gate 610 provides a logic low output signal, which in combination with the logic low state of the RAS signal, causes NOR gate 611 to drive the memory request signal MRQ to a logic high state. The logic high state of the memory request signal MRQ activates memory sequencer 503, such that memory sequencer 503 activates and de-activates the memory control signals (RAS, SENSE, CAS and PRC) according to the waveform pattern 701 illustrated in FIG. 7.

[0060] When the RAS signal is activated high, NOR gate 611 drives the memory request signal MRQ to a logic low state. In this manner, the RAS and MRQ signals form a handshake pair for requesting and acknowledging a memory access. Within arbiter 505, the rising edge of the RAS signal also causes first stage registers 621-623 to latch the logic high values of the LPE_A, LPE_B and RFRQ signals, respectively.

[0061] After a delay of T ns, the rising edge of the RAS signal propagates through delay circuit 650 as the delayed signal D_RAS. The rising edge of the D_RAS signal causes second stage flip-flops 625-627 to latch the logic high values provided at the outputs of first stage flip-flops 621-623, respectively, thereby completing the synchronization between the RAS signal and the latched request signals LPE_A, LPE_B and RFRQ. The synchronized LPE_A, LPE_B and RFRQ signals appear at the outputs of second stage flip-flops 625-627, respectively.

[0062] The synchronized LPE_A signal at the output of flip-flop 625 forms the SEL_A signal. The high state of the SEL_A signal causes NOR gates 631 and 632 to drive the SEL_B and REF signals to logic low states. The high states of the SEL_A and RAS signals cause AND gate 641 to provide a logic high output, thereby resetting the output of S-R flip-flop 601 (i.e., LPE_A) to a logic low state. However, the LPE_B and RFRQ signals provided by S-R flip-flops 602 and 603 still have logic high states. As a result, NOR gate 610 still provides a logic low output signal to NOR gate 611.

[0063] The high state of the SEL_A signal and the low states of the SEL_B and REF signals cause multiplexer 502 to drive the latched memory address LA_A[14:0] from address/control register 513 to memory array 501 as memory address MA[14:0]. Multiplexer 502 also drives the latched read/write control value LRW_A# to memory array 501 as memory array read/write enable signal MRW#. In addition, multiplexer 502 drives the latched input data value LD_A[31:0] from data input register 511 to memory array 501 as input memory data value MD[31:0].

[0064] If the access on port A is a write operation, the read/write control signal MRW# is low, and the input memory data value MD[31:0] is written to the memory location identified by memory address MA[14:0]. If the access on port A is a read operation, the read/write control signal MRW# is high, and a read data value MQ[31:0] is retrieved from the memory location identified by memory address MA[14:0]. When the pre-charge signal PRC is activated high, the port A output enable signal P_AO is activated high, thereby latching the read data value MQ[31:0] into data output register 512. The data output register 512 provides this read data value on the output data bus of port A as the output data value Q_A[31:0].

[0065] The activation of the PRC signal causes the RAS signal to transition to a logic low state. Within memory sequencer 503, the logic low state of the RAS signal, coupled with the logic low signal provided by NOR gate 610, causes NOR gate 611 to drive the memory request signal MRQ to a logic high state, thereby initiating another memory access. That is, memory sequencer 503 again activates the RAS, SENSE, CAS and PRC signals in the same manner defined by waveform pattern 701. When the RAS signal is activated high, the logic low state of the LPE_A signal is stored in flip-flop 621. The rising edge of the D_RAS signal subsequently causes the low state of the LPE_A signal to be stored in flip-flop 625, such that this flip-flop provides a SEL_A signal having a logic low state. In response, NOR gate 631 provides a SEL_B signal having a logic high state, thereby causing multiplexer 502 to drive the latched memory address LA_B[14:0], the latched read/write control value LRW_B# the latched input data value LD_B[31:0] to memory array 501. In response, a read or write operation (as identified by the LRW_B# signal) is performed by memory array 501.

[0066] In the case of a write operation, data value D_B[31:0] is written to the memory location selected by address A_B[14:0]. In the case of a read operation, data from the memory location associated with address A_B[14:0] is driven as the output signal MQ[31:0], and is subsequently latched in output data register 522 at the rising edge of the memory pre-charge signal PRC. Output data register 522 drives the memory read data on the read data bus of port B Q_B[31:0]

[0067] The logic high states of the SEL_B and RAS signals cause AND gate 642 to provide a logic high output, thereby resetting the output of S-R flip-flop 602 (i.e., LPE_B) to a logic low state. However, the RFRQ signal provided by S-R flip-flop 603 still has a logic high state. Thus, NOR gate 610 continues to provide a logic low output signal to NOR gate 611.

[0068] When the RAS signal is de-activated low at the end of the access, NOR gate 611 in arbiter 505 again drives the MRQ signal to a logic high state, thereby requesting another memory access (i.e., a refresh access). In response, memory sequencer 503 drives RAS, SENSE and PRC signals in accordance with the waveform pattern 701. The rising edges of the RAS and D_RAS signals cause the logic low state of the LPE_B signal to be transmitted through flip-flops 622 and 626, respectively. In response, NOR gate 631 provides a logic low SEL_B signal and NOR gate 632 provides a logic high REF signal. The high state of the REF signal designates the current memory cycle a refresh access.

[0069] The logic low states of the SEL_A and SEL_B signals cause multiplexer 502 to route the refresh row address RFA[9:0] from refresh controller 530 to memory array 501 as the address signal MA[14:5]. The high state of the REF signal suppresses the generation of the column access signal CAS in the memory sequencer, since there is no column access in a refresh operation. The high state of REF signal also inhibits any read or write operation in memory array 501. Within arbiter 505, the logic high states of the REF and RAS signals cause AND gate 643 to provide a logic high output, thereby resetting the RFRQ signal provided by S-R flip-flop 603 to a logic low state. When the pre-charge signal transitions to a logic high state and the RAS signal transitions to a logic low state, memory system 500 is ready to accept another access request.

[0070] FIG. 8 is a block diagram of a dual-port memory system 800, which is implemented using single-port memory cells in accordance with an alternate embodiment of the present invention. Because dual-port memory system 800 is similar to dual-port memory system 500, similar elements in FIGS. 8 and 5 are labeled with similar reference numbers. Within dual-port memory system 800, port B is a read-only port. In addition, the port enable signal P_BE of memory system 500 is replaced with a clock signal SCLK, and the read/write signal RW_B# of memory system 500 is replaced with a read enable signal REN_B in dual-port memory system 800. Thus, dual-port memory system 800 includes an access arbiter 805 which operates in a different manner than the access arbiter 505 of dual-port memory system 500.

[0071] FIG. 9 is a circuit diagram of access arbiter 805 in accordance with one embodiment of the present invention. Access arbiter 805 includes S-R flip-flop 601, data flip-flops 902-903, NOR gates 610-611, synchronizer 920, priority encoder 930, AND gates 641-643 and delay circuit 650. Synchronizer 920 includes OR gate 921, first stage flip-flops 621-622 and second stage flip-flops 625-626. S-R flip-flop 601, NOR gates 610-611, first stage flip-flops 621-622, second stage flip-flops 625-626, delay circuit 650 and logical AND gates 641-643 have been described above in connection with access arbiter 605 (FIG. 6). Priority encoder 930 includes logical AND gates 931-932 and NOR gate 933.

[0072] The data (D) input terminals of flip-flops 902-903 are coupled to receive the port B read enable signal REN_B and the refresh request signal RREQ, respectively. The reset (R) terminals of flip-flops 902-903 are coupled to receive the output signals provided by AND gates 642 and 643, respectively. Flip-flops 902 and 903 include a clock input terminal and an inverting clock input terminal, respectively, which are coupled to receive the SCLK signal. Flip-flops 601, 902 and 903 provide latched access enable signals LPE_A, LPE_B and RFRQ, respectively, to NOR gate 610. NOR gate 610 is coupled to NOR gate 611, which generates the MRQ signal in the manner described above in connection with FIG. 6. The LPE_B and RFRQ signals are also provided to OR gate 921 within synchronizer 920. If either the LPE_B signal or the RFRQ signal has a logic high state, then a logic high signal is provided to the D input terminal of first stage flip-flop 622. Otherwise, a logic low signal is provided to the D input terminal of first stage flip-flop 622. First stage flip-flops 621 and 622 are clocked by the RAS signal, and second stage flip-flops 625 and 626 are clocked by the D_RAS signal provided by delay circuit 650.

[0073] The port A select signal SEL_A is provided at the output of second stage flip-flop 625. NOR gate 933 has a non-inverting terminal coupled to receive the SEL_A signal, and an inverting input terminal coupled to receive the Q output signal of second stage flip-flop 626. In response, NOR gate 933 provides an output signal to AND gates 931 and 932. The other input terminal of AND gate 931 is coupled to receive the SCLK signal, and the other input terminal of AND gate 932 is coupled to receive the inverse of the SCLK signal. In response, AND gates 931 and 932 provide the SEL_B and REF signals, respectively. In general, the SEL_B and REF signals are deactivated low if the SEL_A signal is activated high. If the SEL_A signal is de-activated low, then the SEL_B and REF signals can be activated high during alternate half cycles of the SCLK signal.

[0074] In accordance with one embodiment, access arbiter 805 is used to implement a display buffer, wherein port A is used by a graphics controller to implement a screen update function, and port B is used to implement a screen refresh function. Port A (i.e., the screen update port) is a read-write port, which has priority over port B (i.e., the screen refresh port) and the refresh mechanism. Port A operates in a manner similar to that described above in connection with FIGS. 6-7. That is, when the port enable signal P_AE is activated high, the latched port enable signal LPE_A and the memory request signal MRQ are both activated high. The RAS signal is activated in response to the activated MRQ signal. The delayed RAS signals D_RAS is activated in response to the activated RAS signal. The SEL_A signal is then activated in a synchronized manner in response to the RAS and D_RAS signals. The activated SEL_A and RAS signals reset S-R flip-flop 601 prior to the next access on port A.

[0075] The screen refresh operation is performed on port B on a regular basis, and is synchronized using the SCLK signal. The port B enable signal REN_B is activated high to indicate that a read transaction is arriving at screen refresh port B. The continuously running SCLK signal is also input to refresh controller 530 to generate the refresh request signal RREQ, as well as the refresh address RFA[9:0]. With this arrangement, the refresh request signal RREQ is synchronous with the SCLK signal, such that refresh requests and read requests from port B are synchronous with respect to each other. In accordance with one embodiment, the refresh requests are controlled to occur while the SCLK signal has a logic low state, and read requests from port B are controlled to occur while the SCLK signal has a logic high state (or vice versa), thereby eliminating the need to arbitrate between read accesses on port B and refresh accesses. Using this configuration simplifies access arbiter 805 (with respect to access arbiter 505) by reducing the required number of synchronizers from three to two.

[0076] Returning now to FIG. 9, the read request signal REN_B and the refresh request signal RREQ are latched by data flip-flops 902 and 903, respectively, at the rising and falling edges of the SCLK signal, respectively. The latched memory request signals LPE_A, LPE_B and RFRQ are logically OR'ed to generate the memory request signal MRQ. Because the latched access request signal LPE_B and the latched refresh request signal RFRQ are synchronous signals, OR gate 921 performs a logical OR function on these signals prior to synchronization with the RAS signal.

[0077] Synchronization is carried out by the multi-stage synchronizer formed by data registers 622 and 626. NOR gate 933 provides access priority to port A, such that the SEL_B and REF signals are deactivated low if the SEL_A signal is activated high. AND gates 931-932 (along with the SCLK signal) resolve the access priority between a read access to port B and a refresh access. When the output of NOR gate 933 has a logic high state and the SCLK signal has a logic high state, a read access is performed on port B. When the output of NOR gate 933 has a logic high state and the SCLK signal has a logic low state, a refresh access is performed within memory array 501.

[0078] Even though three multi-stage synchronizers are used in access arbiter 505, and two multi-stage synchronizers are used in access arbiter 805, only two signals need to be synchronized in arbiter 505, and only one signal needs to be synchronized in arbiter 805. This is because the RAS signal is activated by the MRQ signal, which in turn is activated by the high state of at least one of the LPE_A, LPE_B or RFRQ signals. Each path of the synchronizer can readily resolve those signals (LPE_A, LPE_B or RFRQ) activated before the activation of the RAS signal and having sufficient setup time at the first stage flip-flop. Such signals are considered synchronized with the RAS signal. Thus, the RAS signal is synchronized with at least one of the three requests signals (LPE_A, LPE_B or RFRQ) every time the RAS signal is activated. The net MTBF of the all the synchronizer paths is equal to the number of signals that need to be synchronous, times the MTFB of a single synchronizer path. Therefore, the net MTBF of arbiter 805 is twice as long as the net MTBF of arbiter 505. This is the result of using the continuous SCLK signal to generate both the refresh request signal RREQ and the port B read access enable signal REN_B, such that refresh accesses and port B read accesses are synchronous, thereby eliminating one signal that needs to be synchronized.

[0079] Although the present invention has been described in connection with several embodiments, it is understood that this invention is not limited to the embodiments disclosed, but is capable of various modifications and embodiments which would be apparent to one of ordinary skill in the art. For example, although the disclosed embodiments have concentrated on a memory system that implements single-port DRAM cells, with minor modifications by ordinary skill memory designers, the present invention can be practiced using single-port SRAM cells. When using single-port memory cells that do not require periodic refresh, refresh controller 530 can be eliminated, and the refresh request signal RREQ can be grounded. It is, therefore, contemplated that the appended claims will cover any such modifications or embodiments as falling within the true scope of the invention.