Low-cost dual-stage offset-cancelled sense amplifier with hybrid read reference generator for improved read performance of RRAM at advanced technology nodes

Qiao Wang; Donglin Zhang; Yulin Zhao; Chao Liu; Xiaoxin Xu; Jianguo Yang; Hangbing Lv

doi:10.1088/1674-4926/42/8/082401

Abstract

In this work, two process-variation-tolerant schemes for a current-mode sense amplifier (CSA) of RRAM were proposed: (1) hybrid read reference generator (HRRG) that tracks process-voltage-temperature (PVT) variations and solve the nonlinear issue of the RRAM cells; (2) a two-stage offset-cancelled current sense amplifier (TSOCC-SA) with only two capacitors achieves a double sensing margin and a high tolerance of device mismatch. The simulation results in 28 nm CMOS technology show that the HRRG can provide a read reference that tracks PVT variations and solves the nonlinear issue of the RRAM cells. The proposed TSOCC-SA can tolerate over 64% device mismatch.

1. Introduction

Resistive random access memory (RRAM) has been studied as a promising candidate for the next generation of embedded non-volatile memory due to its advantages such as high scalability, low power consumption, high speed and non-volatility^{[1, 2]}. However, the practical application of RRAM at advanced technology nodes is challenged by a limited sensing margin^[3]. The sense amplifier (SA) needs to tolerate the increasing mismatch of devices. A read reference that tracks process and temperature variations is also essential^[4-6]. Several offset-cancellation techniques have been proposed to reduce the device mismatch and improve the sensing margin^[7-9]. However, only the first stage (current-sampling) or the second stage (latch) of the SA was calibrated. In addition, multi capacitors in the offset-cancellation circuits also incur a remarkable area overhead. Moreover, the nonlinearity of RRAM cells has not been considered in previous dummy-cell read references.

In this work, a two-stage offset-cancelled current sense amplifier (TSOCC-SA) with only two capacitors is proposed. Two capacitors are used to maximize the sensing margin of RRAM and the tolerance of device mismatch. In addition, a hybrid read reference generator (HRRG) is designed by combining RRAM cells and linear resistors to provide a read reference that can track PVT variations and solve the nonlinear issue of the RRAM cells.

2. RRAM cell and its nonlinearity

The structure of the 1T1R RRAM cell used in this study includes an NMOS switch transistor and a TaO_x-based bipolar resistor memory device. Fig. 1 shows the cross-section transmission electron microscope (TEM) images of the fabricated RRAM cells in the 28 nm CMOS process. The RRAM is fabricated at the back end of line (BEOL) process and is fully compatible with the logic process^[10].

Figure 1.The TEM images of 1T1R RRAM cells.

The schematic of the 1T1R RRAM cell is shown in Fig. 2, including the word line (WL) used to select the cell, and the bit line (BL) and source line (SL) used to apply the corresponding operation voltage. A fresh RRAM cell is in an initial state with extremely high resistance^[11] and requires a one-time forming process to form a large number of conductive filaments (CFs) as shown in Fig. 2(a). After the forming process, the RRAM cell is in a low resistance state (LRS) as shown in Fig. 2(b). The reset operation is shown in Fig. 2(c). A reset voltage (V_Reset) is applied to the RRAM cell from the SL to BL. The reset operation ruptures the formed CFs, making the RRAM cell from the LRS to the high resistive state (HRS). In the set operation, a set voltage (V_Set) is applied to the RRAM cell from the BL to SL and CFs form again to make the cell into an LRS^[12].

Figure 2.(Color online) RRAM cell basic operations: CFs forming, Reset and Set.

In the read operation, a small read voltage (V_read is much lower than set voltage) is applied to the BL and then the CSA detects the current in the read path to determine the state of the cell resistance. The V_read applied to the BL will cause the migration of a small number of oxygen atoms. Then the distribution of oxygen vacancies inside the RRAM cell is changed and the resistance of RRAM cells exhibits nonlinearity. As shown in Fig. 3, the ideal linear I–V curve of HRS and LRS is not exactly consistent with the measured I–V curve.

Figure 3.(Color online) Measured I–V curve and ideal I–V curve of RRAM cell. Ideal I_MP is the mid-point current of the ideal value of I_HRS and I_LRS. Actual I_MP is the mid-point current of the measured value of I_HRS and I_LRS. The actual I_MP is 18.4% lower than the ideal I_MP.

Table 1 shows the operating conditions of the 1T1R RRAM cell used in this study. The RRAM cell has a large resistance variation (R-variation). R-variation leads to the wide distribution of cell current (I_cell) and references current (I_ref) that will cause sensing yield issues^[13]. Therefore, for an SA, a read reference that tracks process and temperature variations is essential.

Table Infomation Is Not Enable

3. Hybrid read reference generator

The previous dummy reference cell structures of the Serial-Parallel (SP) scheme^[14], Parallel-Series Reference Cell (PSRC) scheme^[15], and the proposed hybrid read reference generator (HRRG) are shown in Fig. 4. The HRRG adopts the Parallel-Series structure like the PSRC scheme. Besides, several dynamic-reference sensing schemes have been proposed to maximize the sensing margin^[16-18]. Nevertheless, the dynamic-reference sensing schemes are always accompanied by higher power consumption and area overhead.

Figure 4.(Color online) Reference cell structure diagrams for SP scheme, PSRC scheme and HRRG scheme.

The average of HRS RRAM cell and LRS RRAM cell is used as the reference in the SP scheme, and the I_sp is as Eq. (1).

${I_{\rm{sp}}} = \frac{{{V_{\rm{read}}}}}{{{R_{\rm{sp}}}}} = \frac{{{V_{\rm{read}}}}}{{\left( {{R_{\rm{L}}} + {R_{\rm{H}}}} \right){\rm{||}}\left( {{R_{\rm{L}}} + {R_{\rm{H}}}} \right)}}{\rm{ = }}\frac{{2{V_{\rm{read}}}}}{{ {{R_{\rm{L}}} + {R_{\rm{H}}}} }}.$ (1)

Fig. 5 shows the simulated reference current (I_ref) distributions of conventional reference schemes and the proposed hybrid read reference, not considering the nonlinearity of the RRAM cell. The ideal reference for CSA should be the mid-point current of the two neighboring memory states (I_MP). The reference current generated by SP schemes is closer to the current of HRS (I_HRS) and the existing overlap may cause sensing failure with PVT variations. The reference current generated by PSRC (I_PSRC) and the proposed HRRG (I_HRRG) is in the middle of I_HRS and I_LRS, as shown in Eq. (2).

Figure 5.(Color online) The distributions of cell current and the reference current.

${I_{\rm{ps}}} = \frac{{{V_{\rm{read}}}}}{\left( {{R_{\rm{L}}}{\rm{||}}{R_{\rm{H}}}} \right) + \left( {R_{\rm{L}}}{\rm{||}}{R_{\rm{H}}} \right)} = \frac{{{V_{\rm{read}}}\left( {{R_{\rm{L}}} + {R_{\rm{H}}}} \right)}}{{2{R_{\rm{L}}} {R_{\rm{H}}}}} = \frac{{{I_{\rm{LRS}}} + {I_{\rm{HRS}}}}}{2}.$ (2)

This is very close to the ideal reference for the CSA. In addition, as shown in Fig. 5, the HRRG can achieve a narrower distribution compared with the PSRC. Specifically, the HRRG reduces the σ-I_REF by 49% compared with the PSRC scheme. The difference between the proposed HRRG and the previous PSRC scheme is that the pair of parallel RRAM cells in the PSRC scheme is replaced by linear resistors to solve the nonlinear issue of RRAM cells, as shown in Fig. 4. The two linear resistors are trimmed close to HRS and LRS of RRAM cells, respectively. Assuming that the four linear resistors are used in HRRG, the reference current with the most narrower distribution range can be obtained. However, it will not be able to track the PVT variations of the RRAM cell.

As shown in Fig. 3, the resistance of the RRAM cell is nonlinear. The voltage falling on two pairs of parallel resistors is about V_read/2 according to Ohm’s law. Therefore, the expression of the actual currents of the PSRC scheme and the HRRG scheme can be described as Eq. (3).

${I_{\rm{ps\_real}}} = {I_{\rm{LRS}}}{|_{{{{V_{\rm{read}}}}}/{2}}} + {I_{\rm{HRS}}}{|_{{{{V_\rm{read}}}}/{2}}}.$ (3)

Taking into account the nonlinearity of the RRAM cell, Fig. 3 shows the measured I_LRS and I_HRS which are 25% and 70% lower than the ideal value at a read voltage of 0.3 V, respectively. Hence, the actual value of the I_MP is 18.4% lower than the ideal value of the I_MP. The I_PSRC is 14.86% higher than the actual value of the I_MP calculated from the I–V curve in Fig. 3. By contrast, the proposed I_HRRG is only 0.49% higher than the actual value of the I_MP.

Using the proposed HRRG scheme, the reference current shows a more concentrated distribution and is more agreed with the true value of I_MP. Hence, the sensing margin is maximized and the accuracy of the reading is improved. Moreover, the latency of the CSA with HRRG is minimum compared with the CSA with SP or PSRC, as shown in Fig. 6.

Figure 6.(Color online) The maximum latency of the CSA with different reference cells.

4. Two-stage offset-cancelled current sense amplifier

A two-stage offset-cancelled current sense amplifier (TSOCC-SA) with only two capacitors is proposed to improve the sensing margin. Fig. 7 shows the schematic of TSOCC-SA. The TSOCC-SA is comprised of a current-sampling and a latch. Compared to the previous OCCS-SA in Ref. [7], the proposed TSOCC-SA uses two cross switches (S1 and S2) for offset-cancellation. The switches (S3 and S4) are used to enlarge the sensing margin. Combined with the use of switches S1, S2, S3, and S4, only one capacitor can realize the function in Ref. [7], which significantly reduces the area of the circuit. The offset cancellation of the second stage of TSOCC-SA is realized by using a similar scheme proposed by Dong^[8].

Figure 7.(Color online) Schematic diagram of TSOCC-SA.

Fig. 8 shows the timing of TSOCC-SA. In P1, the outputs of the two inverters are connected to their inputs, respectively. The difference between the trip voltages of the two inverters (V_TR – V_TL) is sampled by capacitor C2. In P2, the outputs of the two inverters are reset to “0”. And, the input of the right inverter becomes “V_TR – V_TL” while that of the left inverter remains “0”. In P3, two diode-connected transistors (M1 and M2) supply pre-charge currents (I_pre1 and I_pre2) to A and B nodes. After a sufficient pre-charge time, the currents of M1 and M2 (I_M1 and I_M2) decrease to nearI_ref and I_cell, regardless of the device mismatch in M1 (M2) as long as I_cell and I_ref are constant in P3. Finally, the gate voltages (V_G1 and V_G2) of M1 and M2 are stored at the left and right ends of capacitor C1, respectively. In P4, the four switches (S5–S8) are turned off and the two switches (S1 and S2) are turned on. The A (B) node has a current path to GND through S3 (S4), resulting in strong positive feedback. In P5, the latch comparator starts to work and outputs “1” or “0”.

Figure 8.(Color online) The timing of TSOCC-SA.

Fig. 9 shows the simulated waveforms of the CSB-SA^[9] and the proposed TSOCC-SA. TSOCC-SA has a substantial influence on the sensing margin and performance. In P3, the operations of the CSB-SA and the TSOCC-SA are exactly the same. Four switches (S1–S4) are turned on at the beginning of P3, and the CSB-SA samples I_cell and I_ref through these switches. Similarly, the TSOCC-SA samples I_cell and I_ref through six switches (S3–S8). In P4, the operations of the CSB-SA and the TSOCC-SA are different. In the CSB-SA, the four switches (S1–S4) are turned off at the beginning of P4. However, in the TSOCC-SA, the double sensing margin switches (S1–S4) are turned on, and the four switches (S5–S8) are turned off. In the CSB-SA, without discharge path, both V_A and V_B in P4 monotonically increase until one of them reaches VDD. In the TSOCC-SA, the M1 (M2) charges the A (B) node with the sampled current I_cell (I_ref), while the S3 (S4) discharges the A (B) node with the current I_ref (I_cell). Thus, the current difference between the A and B nodes in the TSOCC-SA is twice that of the CSB-SA.

Figure 9.(Color online) CSB-SA and proposed TSOCC-SA. Bit-cell state 1 (I_cell > I_ref) is assumed in the V_A and V_B waveforms.

Fig. 10(a) shows the sensing margin development with V_TH mismatch between M1 and M2. For TSOCC-SA, when the device mismatch reaches 64% of the threshold voltage, the sensing margin is still enough, and the sense amplifier is valid. It clearly shows that the TSOCC-SA has higher offset-tolerance than the CSB-SA in Ref. [9]. And the TSOCC-SA can work at a lower operation voltage to reduce power consumption by introducing offset-cancellation techniques, as shown in Fig. 10(b). In addition, when the mismatch between M1 and M2 exceeds 225 mV, the swing between A and B is lower or even only tens of millivolts. Then, the mismatch between M4 and M5, M6 and M7 in the latch comparator will lead to the invalidation of SA. Hence, it is extremely necessary to introduce the cross-coupling capacitor C2 in the latch comparator to cancel these device mismatches, which can further improve the accuracy of SA. Dong et al. have proved that the standard deviation of the mismatches is reduced by more than 60% by C2^[8].

Figure 10.(Color online) (a) Simulated ∆V [= min |V_A –V_B|] and ∆I [= min |I_A –I_B|] vs. V_TH mismatch between transistors M1 and M2. (b) The maximum V_TH mismatch that can be tolerated by TSOCC-SA at different operation voltage.

Moreover, the TSOCC-SA avoids excessive area overhead. In the layout, the area of the coupling capacitor is equivalent to that of a transistor. Fig. 11 shows the sensing margin of several offset-cancellation techniques (CSB-SA^[9], OCCS-SA^[7], SCOC-SA^[8], and TSOCC-SA) and the extra area overhead caused by the coupling capacitor.

Figure 11.(Color online) The extra area overhead and sensing margin of several offset-cancellation techniques.

Table 2 shows the performance comparison of several offset-cancellation techniques. The area is characterized by the number of coupling capacitors (Caps). Stage indicates that the mismatch of how many stages of SA can be cancelled. Cancellation ability indicates the percentage of the mismatch voltage to the threshold voltage that can be tolerated by SA.

Table Infomation Is Not Enable

5. Conclusion

In this paper, we proposed a two-stage offset-cancelled current sense amplifier (TSOCC-SA) with only two capacitors for RRAM at an advanced technology node. A hybrid read reference generator (HRRG) is designed by combining RRAM cells and resistors to provide a read reference that tracking PVT variations and to solve the nonlinear issue of the RRAM cells. The simulation results in the 28 nm CMOS process have proved the effectiveness of HRRG and the reliability of the TSOCC-SA.

Acknowledgements

This work was supported in part by the National Key R&D Program of China under Grant No. 2019YFB2204800 and in part by the Major Scientific Research Project of Zhejiang Lab (Grant No. 2019KC0AD02) and in part by the National Natural Science Foundation of China under Grants 61904200 and the Strategic Priority Research Program of the Chinese Academy of Sciences under Grant No. XDB44000000.

References

[1] C W S Yeh, S S Wong. Compact one-transistor-N-RRAM array architecture for advanced CMOS technology. IEEE J Solid-State Circuits, 50, 1299(2015).

[2] J Woo, S M Yu. Two-step read scheme in one-selector and one-RRAM crossbar-based neural network for improved inference robustness. IEEE Trans Electron Devices, 65, 5549(2018).

[3] M F Chang, C W Wu, C C Kuo et al. A 0.5 V 4 Mb logic-process compatible embedded resistive RAM (ReRAM) in 65 nm CMOS using low-voltage current-mode sensing scheme with 45 ns random read time. 2012 IEEE International Solid-State Circuits Conference, 434(2012).

[4] P Jain, U Arslan, M Sekhar et al. 13.2 A 3.6 Mb 10.1Mb/mm² embedded non-volatile ReRAM macro in 22 nm FinFET technology with adaptive forming/set/reset schemes yielding down to 0.5 V with sensing time of 5 ns at 0.7 V. 2019 IEEE International Solid-State Circuits Conference, 212(2019).

[5] W H Chen, C M Dou, K X Li et al. CMOS-integrated memristive non-volatile computing-in-memory for AI edge processors. Nat Electron, 2, 420(2019).

[6] C C Chou, Z J Lin, C A Lai et al. A 22nm 96K × 144 RRAM macro with a self-tracking reference and a low ripple charge pump to achieve a configurable read window and a wide operating voltage range. IEEE Symposium on VLSI Circuits, 1(2020).

[7] T Na, B Song, J P Kim et al. Offset-canceling current-sampling sense amplifier for resistive nonvolatile memory in 65 nm CMOS. IEEE J Solid-State Circuits, 52, 496(2017).

[8] Q Dong, Z H Wang, J Lim et al. A 1Mb 28nm STT-MRAM with 2.8ns read access time at 1.2V VDD using single-cap offset-cancelled sense amplifier and in situ self-write-termination. IEEE International Solid-State Circuits Conference, 480(2018).

[9] M F Chang, S J Shen, C C Liu et al. An offset-tolerant current-sampling-based sense amplifier for sub-100nA-cell-current nonvolatile memory. IEEE International Solid-State Circuits Conference, 206(2011).

[10] Y L Huang, R Huang, Y M Cai et al. A TaO_x based threshold switching selector for the RRAM crossbar array memory. 12th Annual Non-Volatile Memory Technology Symposium Proceedings, 85(2012).

[11] C C Chou, Z J Lin, P L Tseng et al. An N40 256K × 44 embedded RRAM macro with SL-precharge SA and low-voltage current limiter to improve read and write performance. IEEE International Solid-State Circuits Conference, 478(2018).

[12] S Chakrabarti, D Jana, M Dutta et al. Impact of AlO_x interfacial layer and switching mechanism in W/AlO_x/TaO_x/TiN RRAMs. 2014 IEEE 6th Int Mem Work IMW, 1(2014).

[13] T Na, J Kim, J P Kim et al. Reference-scheme study and novel reference scheme for deep submicrometer STT-RAM. IEEE Trans Circuits Syst I, 61, 3376(2014).

[14] M Durlam, P J Naji, A Omair et al. A 1-Mbit MRAM based on 1T1MTJ bit cell integrated with copper interconnects. IEEE J Solid-State Circuits, 38, 769(2003).

[15] M F Chang, S S Sheu, K F Lin et al. A high-speed 7.2-ns read-write random access 4-mb embedded resistive RAM (ReRAM) macro using process-variation-tolerant current-mode read schemes. IEEE J Solid-State Circuits, 48, 878(2013).

[16] Q K Trinh, S Ruocco, M Alioto. Dynamic reference voltage sensing scheme for read margin improvement in STT-MRAMs. IEEE Trans Circuits Syst I, 65, 1269(2018).

[17] W Kang, T T Pang, W Lv et al. Dynamic dual-reference sensing scheme for deep submicrometer STT-MRAM. IEEE Trans Circuits Syst I, 64, 122(2017).

[18] Y L Zhou, H Cai, L Xie et al. A self-timed voltage-mode sensing scheme with successive sensing and checking for STT-MRAM. IEEE Trans Circuits Syst I, 67, 1602(2020).