Soft error robust storage SRAM cells and flip-flops

The proposed Eight Transistor Storage Cell topology offers improved soft error robustness compared to the standard 4T dual-inverter storage cell commonly used in 6T SRAM cells and static flip-flops. This application claims the benefit of priority of U.S. Provisional Patent Application No. 61/202,707 filed Mar. 30, 2009 which is incorporated herein by reference in its entirety.

SRAMs are one of the most popular ways to store data in electronic systems. Similarly, embedded SRAMs are a vital building block in integrated circuits. SRAMs are a preferred method of implementing embedded memories owing to higher speed, robust design, and ease of integration. SRAMs, in general, occupy a significantly large portion of the chip's die area, making it an important block in terms of area, yield, reliability and power consumption. With increasing demand for highly integrated System on Chip (SoC), improving various aspects of embedded SRAMs have received a significant interest.

The six transistor SRAM cell is the most popular configuration because of its high speed and robustness. However the 6T SRAM cell, shown in FIG. 1 , suffers from a relatively high vulnerability to soft errors. At the heart of the 6T SRAM cell is the 4T dual-inverter storage cell shown in FIG. 2 . The inherent feedback of the back-to-back inverters means that if sufficient charge is collected on one node, the feedback will cause the second node to also switch. This results in an erroneous state being stored, and the cell cannot recover.

Flip-flops are critical components in integrated circuits. Flip-flops enable sequential logic by storing the result of combinational logic. A flip-flop stores data by sampling the input data signal with a clock signal at a particular instant of time, typically at the edge of the clock. Thus, the output of the flip-flop is sensitive to the input data signal only at the clock edge. At all other times, the flip-flop's output is constant and does not respond to changes in input data signal. A flip-flop can be realized in a variety of ways. A typical way of realizing a flip-flop is to use two series connected latches called master and slave. This architecture of flip-flop is called the master-slave D-flip-flop (DFF), a schematic of which is given in FIG. 3 . Whatever the architecture every flip-flop must have a storage element. Most static flip-flops use a dual-inverter structure, shown in FIG. 2 .

The scaling of CMOS processes has resulted in lower node capacitance and lower supply voltages in integrated circuits. The result of these effects is that less charge is used to store data in a storage cell. The increased density enabled by the smaller geometries of scaled processes has also enabled the integration of large SRAM arrays and large numbers of flip-flops. These conditions resulted in an increased sensitivity to transients induced by ionizing radiation consisting of energetic cosmic neutrons and alpha particles. When these particles strike the chip they generate a large number of electron hole pairs. Depending on the location of the particle strike the deposited charge may be collected by a node in a storage cell. If sufficient charge is collected the storage cell can switch its logical state, which is called a soft error. In view of this vulnerability numerous methods have been proposed to improve the robustness of the storage cell. More effort has been put into improving the soft-error robustness of SRAM cells, however the SER of flip-flops has been continually degrading to the point where it has become a significant reliability consideration. Both the 6T SRAM and the master-slave static CMOS DFF use the dual-inverter storage cell to hold a logical state. For both circuits the data value stored in this structure can be susceptible to soft errors. If a soft error transient is able to change the state of one node, the feedback of the dual-inverter storage cell will change the second node and result in erroneous data being stored.

FIG. 4 shows the schematic diagram of an SRAM cell which has an improved robustness to radiation induced soft errors. In the proposed circuit the SRAM cell has the same dual-inverter storage cell as in the traditional case, however it also includes a coupling capacitor between the two nodes where the data is stored. The effect of adding the coupling capacitor is that the critical charge (Qc) of the cell is significantly increased. The critical charge is the minimum amount of charge which will result in a soft error. However, the addition of a large, area efficient coupling capacitor requires a special semiconductor manufacturing process. Therefore this SRAM cell is not easily integrated into common Complementary Metal-Oxide-Semiconductor (CMOS) processes. As a result, for Application Specific Integrated Circuits (ASICs) where embedded SRAM cells are widely realized using standard CMOS process, this approach is costly to implement. Another problem with this approach is that while the capacitor increases the critical charge, it also makes the cell slower to write to. For data to be written into this SRAM cell the coupling capacitor will have to be charged/discharged, which will take more time than for a standard 6T SRAM cell. This results in a lower frequency of operation, which is again not desirable. Another drawback of this cell is that the extra capacitance which must be charged and discharged results in larger power consumption compared to the standard 6T SRAM cell.

FIG. 5 shows the schematic diagram of a storage cell which has four extra transistors in order to improve its robustness to soft errors. This circuit is known as the DICE cell, which stands for dual-interlocking storage cell. In the proposed solution, the data is stored on multiple nodes and the storage cell is immune to single node upsets. However, the SRAM cell based on the DICE storage cell requires twice as many transistors to implement compared with the standard 6T SRAM circuit, making it expensive. Also, in order for this cell to be used in a flip-flop, at least three nodes must be accessed. The result is that flip-flops which implement this storage cell require a greater number of transistors and use more power.

FIG. 6 shows the schematic diagram of a third storage cell which has an improved robustness to soft errors. This circuit consists of eight transistors and is similar to the DICE cell in that data is stored at multiple nodes. However, in this storage cell data is stored on four nodes which are interconnected using four half-latches. This SRAM cell is almost completely immune to soft-error induced upsets. The SRAM circuit based on this storage cell requires only ten transistors, which is more than the 6T SRAM cell, but less than the DICE SRAM cell. This storage cell can also be used to implement flip-flops. This data on this storage cell can be changed by accessing only two nodes, and as such fewer transistors are required as compared to the DICE based flip-flop.

A new storage cell configuration is introduced. The storage cell configuration allows for logical states to be stored with robustness against noise events such as soft errors. Various types of circuits can be implemented with this storage cell, including SRAM cells and flip-flops.

In accordance with an aspect of the present invention there is provided a . . .

An embodiment of the present invention will now be described by way of example only with reference to the following drawings in which:

FIG. 1 is a schematic showing the prior art of the six transistor SRAM cell.

FIG. 2 is a schematic showing the prior art of the dual-inverter storage cell.

FIG. 3 is a schematic showing the prior art of a static CMOS MS DFF.

FIG. 4 is a schematic showing the prior art of a six transistor SRAM cell with improved SER robustness.

FIG. 5 is a schematic showing the prior art of an 8T SER robust DICE storage cell.

FIG. 6 is a schematic showing the prior art of an 8T SER robust storage cell.

FIG. 7 is a schematic diagram showing the architecture of the proposed storage cell.

FIG. 8 is the proposed storage cell storing 1010 .

FIG. 9 is a schematic of an SER robust SRAM cell utilizing the proposed storage cell.

FIG. 10 is a diagram of an SRAM array implemented with SER robust SRAM cells using the proposed SER robust storage cell.

FIG. 11 is a flowchart which describes the ead operation on a SER robust SRAM cell using the proposed storage cell.

FIG. 12 shows simulation waveforms illustrating a read operation on a SER robust SRAM cell using the proposed storage cell.

FIG. 13 shows simulation waveforms illustrating a half-selected SER robust SRAM cell using the proposed storage cell.

FIG. 14 is a flowchart which describes the write operation on a SER robust SRAM cell using the proposed storage cell.

FIG. 15 shows simulation waveforms illustrating a write operation on a SER robust SRAM cell using the proposed storage cell.

FIG. 16 is a schematic of an alternate SER robust SRAM cell utilizing the proposed storage cell but connecting to nodes C and D.

FIG. 17 is a schematic of an alternate SER robust SRAM cell utilizing the proposed storage cell however connecting to nodes A and B via PMOS access transistors.

FIG. 18 is a schematic of an alternate SER robust SRAM cell utilizing the proposed storage cell however connecting to nodes C and D via PMOS access transistors.

FIG. 19 is an illustration of the response of the proposed storage cell to a soft error induced upset on node A.

FIG. 20 is a schematic of a register file memory cell with one write port and one read port which utilizes the proposed storage cell.

FIG. 21 is a schematic diagrams showing an alternate embodiment of the storage cell with PMOS inner transistors and NMOS restorative transistors.

FIG. 22 shows schematic diagrams of three possible flip-flop circuits utilizing the proposed SER robust storage cell.

For convenience, like structures in drawings will be referenced by like numerals in the description. The following describes an eight transistor storage cell which can be used in soft error robust SRAM cell or in soft error robust flip-flops.

Referring to FIG. 7 , an eight transistor storage cell in accordance with an embodiment of the present invention is illustrated generally by the numeral 700 . The eight transistor storage cell 700 comprises four n-channel transistors N 1 , N 2 , N 3 and N 4 , four p-channel transistors P 1 , P 2 , P 3 and P 4 and four internal nodes A, B, C and D. Transistors N 1 and N 3 are given the name of drive transistors. Transistors P 1 and P 3 are given the name of load transistors. Transistors N 2 and N 4 are given the name of inner transistors. Transistors P 2 and P 4 are given the name of restorative transistors. Nodes A and B are given the name of lower storage nodes and are located at the drains of the drive transistors. Nodes C and D are given the name of upper storage nodes and are located at the drains of the load transistors.

In the storage cell 700 one logical value will be stored on nodes A and C and a complementary logical value will be stored on nodes B and D. Referring to FIG. 8 the storage cell 700 is illustrated storing a logical 0 on nodes A and C and a logical 1 on nodes B and D.

Transistors N 1 , N 2 , N 3 , N 4 , P 1 , P 2 , P 3 and P 4 in the storage cell 700 are connected in such a way that the four storage nodes A, B, C, D interlock to work against a change in the state of the storage cell due to a single ended disturbance such as a soft error. The drains of the load transistors P 1 and P 3 are connected to the drains of the inner transistors N 2 and N 4 . The drains of the drive transistors N 1 and N 3 are connected to both the drains of the restorative transistors P 2 and P 4 and the sources of the inner transistors N 2 and N 4 .

The drive transistors, load transistors and restorative transistors of the storage cell 700 selectively connect the storage cell to the supply voltages. The sources of these transistors are connected to either the logical high supply voltage VDD or the logical low supply voltage GND. The sources of the load transistors P 1 and P 3 and the sources of the restorative transistors P 2 and P 4 are connected to VDD and the sources of the drive transistors N 1 and N 3 are connected to GND. Referring to FIG. 8 the situation where the initial logical state at nodes A and C is 1 and the initial logical state at nodes B and D is 0 is illustrated. Initially transistors P 1 , P 2 , N 3 and N 4 are on and transistors N 1 , N 2 , P 3 and P 4 are off.

Therefore, the storage cell 700 is able to retain logic data as long as it is powered.

Referring to FIG. 9 , a sample SRAM cell which utilizes the storage cell 700 is illustrated generally by numeral 900 . Two NMOS transistors N 5 and N 6 are added in order to selectively couple the storage cell 700 to bitlines BL and BLB. Transistors N 5 and N 6 are called access transistors. The gates of the access transistors are couples at their gates to a wordline WL node. The drains of access transistors N 5 and N 6 are coupled to bitlines BL and BLB respectively. The sources of the access transistors N 5 and N 6 coupled to nodes A and B of the storage cell 700 . The nodes WL, BL and BLB are controlled in order to enable read and write operations on the SRAM cell 900 .

Referring to FIG. 10 , a sample array of SRAM cells 900 is illustrated generally by numeral 1000 . The array 1000 comprises M rows and N columns of SRAM cells 900 . A bitline pair BL and BLB is shared among the cells located in a given column. A wordline WL is shared among all cells in a given row. In addition to the array 1000 , a memory will also contain blocks such as address decoders, timing and control, sensing and column drivers. These blocks are similar to those found in state-of-the-art SRAM configurations, and therefore are not described in detail. Alternatively, the bitline pair BL and BLB may be shared among the cells located in a given row and the wordline WL may be shared among all cells in a given column.

Referring to FIG. 11 the read operation on the 10T SRAM cell 900 is illustrated generally by numeral 1100 . In the present embodiment, it is assumed that the supply voltage VDD is 1V, the initial voltage at nodes A and C is 1V value and the initial voltage at nodes B and D is 0V. Thus, the cell 1100 stores a logic 1 .

At step 1101 , the differential bitline pair BL and BLB are pre-charged to 1V and the allowed to float.

At step 1102 , the wordline WL voltage is changed so as to turn on the access transistors. In this embodiment using SRAM cell 900 the WL is raised to logic high so as to enable the NMOS access transistors.

There is now no voltage difference across transistor N 5 in the SRAM cell 900 and as such no current will flow. However, there is now a voltage difference across transistor N 6 in the SRAM cell 900 which is turned on. The voltage on BLB is VDD whereas the voltage at node B is GND. As such at step 1103 the voltage on BLB will begin to discharge through node B in the storage cell 700 .

At step 1104 , either the current resulting from the read operation, or the resulting differential voltage across the bitlines, is sensed by a sense amplifier (not shown).

Since the cell is differential, the read operation will be similar when the stored value is reversed and the initial voltage at nodes A and C is 0V value and the initial voltage at nodes B and D is 1V. However, in such an embodiment node the voltage on BL would be zero and as such the BL would be discharged, as opposed to BLB.

It will be appreciated that the logic state of a cell can be determined by examining the bitline pair BL or BLB is at the end of the read operation.

An example of a read operation when a logic 1 is stored in the SRAM cell 900 will be described with reference to the timing diagrams illustrated in FIG. 12 . As illustrated in FIG. 12 d , the read operation begins when the signal WL is raised from GND to VDD. As illustrated in FIG. 12 c , the read operation results in the voltage BLB decreasing as it is discharged through the storage cell creating the differential voltage across the bitline pair BL and BLB detected by the sense amplifier. As illustrated in FIG. 12 a and FIG. 12 b , the internal nodes of the storage cell 700 do not change state, hence the read operation is non-destructive.

Referring to FIG. 13 an example of a SRAM cell 900 which is not being read, but which connected to the WL which is enabled. This causes the SRAM cell 900 to be half-selected. As illustrated in FIG. 13 d the wordline WL is raised from GND to VDD for a period of time. However, as illustrated in FIG. 13 c the bitlines BL and BLB are not left to float, but rather are held to VDD. As illustrated in FIG. 13 a and FIG. 13 b the internal nodes of the storage cell 700 are perturbed, but do not change stage.

Referring to FIG. 14 , a flow chart illustrating the steps for writing a logic 0 to the SRAM cell 900 is illustrated by numeral 1400 . In this example, at step 1401 the initial state of the SRAM cell 900 is logic 1 so the initial voltage at nodes A and C is 1V and that the initial voltage at nodes B and D is 0V.

At step 1402 the bitline pair is set so that bitline BL is set to 0V and bitline BLB is set to 1V. At step 1403 , the voltage on the wordline WL is changed to a voltage which turns on the access transistors in the SRAM cell 900 .

At step 1404 , the voltage at node A is logically high, however the voltage on the bitlines BL connected to node A via access transistor N 5 is low. As such node A is discharged through the access transistors N 5 . This results in the outer transistor N 3 turning off and outer transistor P 3 turning on.

At step 1405 , Node D is charged up through outer transistor P 3 , turning off core transistor P 2 and turning on core transistor N 2 .

At step 1406 , Node B is partially charged up through access transistor N 6 and partially through transistors P 3 and P 4 . Finally at step 1407 Node C is discharged through outer transistor N 1 .

At step 1408 , the internal nodes of storage cell 700 are stable and the storage cell now stores a logic 0 state. At step 1409 , the wordline WL is changed to a state which turns off the access transistor and isolates the storage cell 700 from the bitlines BL and BLB.

Since the cell is differential, writing a logic 1 to an SRAM cell 900 storing a logic 0 operates in a similar fashion to that described with reference to FIG. 14 . However, in this example, the initial voltage at nodes A and C is 0V and that the initial voltage at nodes B and D is 1V. Accordingly, in order to write a logic 1 the bitline pair is set so that bitline BL is set to 1V and bitline BLB is set to 0V. Thus, when the voltage of the wordline WL is changed so as to turn on the access transistor the voltages of the storage cell will flip and the cell will store a logic 1 state.

An example of a write operation on the SRAM cell 900 will be described with reference to the timing diagrams illustrated in FIG. 15 . The initial state of the cell is logic 1 . First a logic 0 will be written into the cell and then a logic 1 will be written into the cell. As illustrated in FIG. 15 d , the state written into the storage cell depends on the complementary values on the bitlines BL and BLB. As illustrated in FIG. 15 c , the write operation begins when the signal WL is raised from GND to VDD. As illustrated in FIG. 12 a and Figure b, the internal storage nodes of the storage cell are changed during a write operation in such a way that the cell stores the desired state.

Although the previous embodiments have been described with a particular configuration of storage node voltages for logic 1 and complementary voltages for logic 0 , it will be appreciated that the inverse may also be the case. That is, a storage node configuration described as logic 1 could, instead, be defined as logic 0 , and vice versa.

Further, although the previous embodiment of the SRAM cell 900 had NMOS access transistors accessing nodes A and B of a storage cell 700 it will be appreciated that alternative embodiments are possible. Referring to FIG. 16 one alternate SRAM cell embodiment is illustrated generally by numeral 1600 . In this embodiment an SRAM cell is created by accessing a storage cell 700 using two NMOS access transistors which couple storage cell 700 nodes C and D to BL and BLB respectively. Referring to FIG. 17 a second alternate SRAM cell embodiment is illustrated generally by numeral 1700 . In this embodiment an SRAM cell is created by accessing a storage cell 700 using two PMOS access transistors which couple nodes A and B of storage cell 700 to BL and BLB respectively. Referring to FIG. 18 a third alternate SRAM cell embodiment is illustrated generally by numeral 1800 . In this embodiment an SRAM cell is created by accessing a storage cell 700 using two PMOS access transistors which couple nodes C and D of storage cell 700 to BL and BLB respectively. It will be understood by those skilled in the art that these SRAM cells perform the same functions as described in the first embodiment SRAM cell 900 .

While the SRAM cell 900 can be written into via differential access transistors, the storage cell 700 resists single ended stresses, and as such is robust to soft-errors. The response of the proposed storage cell 700 to a particle strike on node A is illustrated in FIG. 19 . The initial logical state at nodes A and B is 1 and the initial logical state at nodes C and D is 0. Initially transistors P 1 , P 2 , N 3 and N 4 are on, and transistors N 1 , N 2 , P 3 and P 4 are off. The charge collected by the particle strike causes the state of node B to flip, such that it is now at a logical 0 state. This has the effect of turning off transistor N 4 and turning on transistor P 4 . However, transistor N 3 can sink the current through P 4 , and as such the particle strike does not result in any other node changing state. The current from P 1 will eventually charge Node B to a logical state 1 returning the storage cell to its original state.

The proposed storage cell can be used to implement a register file memory cell. Referring to FIG. 20 , a sample register file memory cell which utilizes the storage cell 700 is illustrated generally by numeral 2000 . The addition of transistors N 7 and N 8 along with signals RBL and RWL create a single read port. It will be understood by those skilled in the art that it is trivial to increase the number of read or write ports to register file memory cell 2000 .

Referring to FIG. 7 , it will be understood by those skilled in the art that while the storage cell has been described with the inner transistors as NMOS and the restorative transistors as PMOS, it is trivial to create an alternate embodiment of the cell with PMOS inner transistors and NMOS restorative transistors. This embodiment is illustrated generally by numeral 2100 and is shown in FIG. 21 . In this embodiment the sources of the restorative transistors N 2 and N 4 are connected to GND. This storage cell is able to perform the same functions as the storage cell embodiment 700 .

In logic storage cells are an integral part of flip-flops, and as such the proposed storage cell 700 can be used in a variety of flip-flop designs. Referring to FIG. 21 several sample embodiments of flip-flops are shown which are implemented using the proposed storage cell 700 . FIG. 22 a illustrates a flip-flop with a master-slave DFF architecture which utilizes the storage cell 700 . FIG. 22 b illustrates a flip-flop with a C2MOS architecture which utilizes the storage cell 700 . FIG. 22 c illustrates a flip-flop with a pulsed-latch architecture which utilizes the storage cell 700 . While the figures show the various transfer gates accessing the storage cell 700 via nodes C and D, the storage cell 700 could also be access via nodes A and B instead.

It will be understood by those skilled in the art that these are not an exhaustive list of flip-flops which can be implemented using the storage cell 700 . A flip-flop can be created using any transfer gate which passes the input data signals to the storage cell 700 gated appropriately by a clock signal.

Further, although preferred embodiments of the invention have been described herein, it will be understood by those skilled in the art that variations may be made thereto without departing from the spirit of the invention or the scope of the appended claims.