Full Paper Int. J. on Recent Trends in Engineering and Technology, Vol. 8, No.

2, Jan 2013

MTJ-Based Nonvolatile 9T SRAM Cell

S. Kushwaha, D. Kumar, M. Saw, A. Islam
Department of Electronics and Communication Engineering Birla Institute of Technology, Mesra, Ranchi, Jharkhand, India dhruva1157.09@bitmesra.ac.in, monisha1216.09@bitmesra.ac.in
Abstract This paper presents a spin-transfer torque- magnetic tunnel junction (STT-MTJ) based non-volatile 9-transistor (9T) SRAM cell. The cell achieves low power dissipation due to its series connected MTJ elements and read buffer which offer stacking effect. The paper studies the impact of PVT (process, voltage, and temperature) variations on the design metric of the SRAM cell such as write delay and compares the results with non-volatile 8T SRAM cell (NV8T). The proposed design consumes lower leakage power and exhibits narrower spread in write delay compared with NV8T. Index Terms variability, SRAM, RSNM, WSNM, MTJ, hold power

I. INTRODUCTION Spintronics is the field of electronics that exploit the spin property of electrons giving rise to spin transistors; contrary to charge of electrons in traditional electronic devices [1-7]. It gives a new direction for research where these spin transistors can be combined with ordinary transistors to extract the maximum potential of both. The promising features of spin transistors are non-volatility and re-configurability. These features are very useful and suitable for future onchip, high-speed, high-density, low-power and nonvolatile memory circuits. The nonvolatile property of spin transistor is best suited for power-gating systems and field programmable gate arrays [8]. There are two kinds of magneto-resistive devices spin valve (SV) and magnetic tunnel junction (MTJ). Spin valve (SV) devices consist of at least two ferromagnetic metal (FM) layers separated by an ultra-thin nonmagnetic metal (NM) layer. In an MTJ, two ferromagnetic layers one free layer and the other fixed layer are separated by an oxide layer. The insulating oxide layer is of either AlOx or MgO. The MgO is now preferred to AlOx due to its higher tunnel magnetoresistance (TMR) ratio defined as (RAP RP)/RP, where RAP and RP are the resistances of MTJ in anti-parallel and parallel magnetic orientation [9]. Spin-transfer torque (STT) phenomenon is used to change the magnetic orientation of the free layer relative to the fixed layer to give a parallel and an antiparallel direction. The parallel orientation of MTJ offers low resistance and conversely, an antiparallel orientation offers high resistance. These two resistance states can be used to denote a logic state of 0 and 1 respectively. These resistance states persist even after the removal of switching current passing through the MTJs. This property is utilized in nonvolatile memory design. The MTJ is also CMOS-compatible with high stability and reliability [10]. Moreover, the STT- MTJ can be fabricated on top of CMOS 2013 ACEEE DOI: 01.IJRTET.8.2.50 26

devices to reduce the area overhead [11]. All these features raise hope for building a nonvolatile memory and logic circuit that do not consume OFF-state leakage current and support ultra-low power operation. This work utilizes MTJ in conjunction with conventional MOSFET to mitigate variation in design metrics of memory circuit such as 9-transistor SRAM cell [12]. The memory has evolved into a multi-leveled structure with SRAM making up the fastest access memory. In modern microprocessors, this trend has continued, such that SRAM cache itself has multiple levels. For instance, in recent Intel and AMD microprocessors, L1 (level one) cache contains a small amount of the fastest memory cells. At the next level, L2 cache contains a large amount of cells that are slightly slower, and so on. 6T-SRAMs are widely used in L1 and L2 cache arrays while DRAMs are used as large off-chip memory arrays or as embedded DRAMs (eDRAM) as higher level, highdensity caches. Conversely, the design of traditional chargebased SRAMs, DRAMs and eDRAMs pose challenges such as higher standby power and lower noise margin in sub-45 nanometer nodes and beyond. Alternative to 6T-SRAM bitcells, subthreshold 10T [13], low-power 10T [14], and lowleakage 11T [15] SRAM cells have been proposed for lowvoltage and low-power operations. However, the problem of standby power in semiconductor memory remains to be solved. CMOS technology scaling driven by the benefit of integration density, higher speed of operation and lower power dissipation, has overcome many barriers over the past few decades. Presently, it is facing even more complications, which are more severe than earlier. One of them is variability. Variability is becoming a metric of equal importance as power, delay, and area. TG-based fully differential 8T SRAM bitcell proposed in [16] could mitigate variability in design metrics of SRAM cell, but failed to improve read static noise margin (RSNM). This work attempts to mitigate variability in design metrics of MTJ-STT-based nonvolatile memory cell with enhanced read static noise margin (RSNM) in highly scaled technology node such as 22-nm. The non-volatility of the proposed design solves the problem of leakage power in memory. Extensive simulations were performed in HSPICE using 22-nm Predictive Technology Model (PTM) [17] to support the concluding remarks. The rest of the paper is organized as follows. Section II presents the proposed design. Simulation results are discussed and compared in Section III. Finally, the concluding remarks are provided in Section IV.

Full Paper Int. J. on Recent Trends in Engineering and Technology, Vol. 8, No. 2, Jan 2013 II. PROPOSED DESIGN AND DEVICE SIZING This paper proposes a STT-MTJ-based nonvolatile differential memory cell (see Fig. 1) (hereafter called NV9T). Its design metrics are assessed and compared with its differential counterpart namely STT-MTJ-based nonvolatile 8T SRAM cell (see Fig. 2) (hereafter called NV8T). The transistors in NV9T cell was sized to maintain pull-up (PU) transistors < pass-gate (PG) transistors < pull-down (PD) transistors. Transistor widths for NV9T cell WPU/WPG/WPD were chosen as 22nm/33nm/44nm respectively. For a fair comparison, transistors of basic 6T portion of both the cells had the same dimensions. Widths of the extra transistors in NV9T (MN5/6/7), were equal to 33nm/33nm/44nm respectively. For extra transistors MN5/6 in NV8T a width of 44 nm was selected. Channel length of all transistors of both the cells was taken to be 22 nm. ratio and ratio of 6T portion of both the cells were taken to be 1.33 and 0.67 respectively, where ratio = driver/access and ratio = pull-up/access. This is because typical values of ratio ranges from 1.2 to 3 to avoid read upset in conventional 6T SRAM cell [18]. As write-ability of the SRAM cell is determined by the ratio or pull-up ratio, the ratio was decided to be d 1.8 to maintain good write-ability [19]. Therefore, to maintain an appreciable read stability, writeability and feedback pull-up strength ratio = 1.33 and ratio = 0.67 were chosen for the proposed design [13][16]. The critical design strategy of the proposed cell was the series connection of an MTJ with each driver. These MTJs help in storing and restoring the cell content. A read buffer (MN5/6/7) was used for the cell to isolate storage nodes while reading. This helped in achieving higher read stability. The proposed cell would exhibit very small RSNM without this read buffer since the drivers (MN1/2) were very small and MTJs were in series with them. Thus, this read buffer improved the read stability of the proposed cell.

Fig. 2. Nonvolatile 8T SRAM cell (NV8T) [20]

variations occur due to process and environmental parameters such as operating voltage and temperature. The basic reasons of cropping up variations are lateral and vertical dimensions scaling of devices. The problem of variability (defined as standard deviation () to mean () ratio of a design metric) becomes more severe with further miniaturization and hence, it is imperative that this problem be addressed. The solution of the variability problem is to explore new design techniques. This work attempted to mitigate variation in design metrics of SRAM cell. A. Simulation Setup This section presents comparison of various design metrics, which were estimated during MC (Monte Carlo) simulation using 22-nm PTM. In order to analyze variability with MC simulations, channel length (L), channel-doping concentration (NDEP), oxide thickness (tox) and threshold voltage (Vt) of all the devices in the cells were considered as independent random variables and assumed to have Gaussian distribution with 3 variation of 10% [21]. Anticipated fluctuation in VDD is 10% in future technology generations such as 22-nm [22]. Therefore, most of the design metrics were estimated by varying the supply voltage by 10% around the nominal VDD of 0.8 V. Authors in [23] showed that a sample size of 2000 ensures a lower than 4% inaccuracy in the estimation of standard deviation. Hence, a sample size of 5000 [24] was used in this work to achieve even higher accuracy. All parameters were estimated using the above simulation setup. B. Store and Restore Operation In normal SRAM operation of NV8T SR and CTRL signals remain grounded. Storage operation in NV8T is initiated by raising SR and CTRL lines to VVDD (virtual VDD). At the end of storage operation, SR and CTRL pulses are turned low. Then global VDD and GND are disconnected from the VVDD and VGND (virtual GND) by sleep transistors (see Fig. 3) and MRAM array enters into inactive mode in which no read / write operations are performed. Values of the storage nodes are stored in MTJ in the form of RAP/RP (anti-parallel resistance/parallel resistance) in the MTJ. During restore operation, SR and CTRL lines remain grounded and VV DD is raised high by switching sleep transistors ON. Store and restore operation of proposed NV9T SRAM cell is illustrated in Fig. 4. During the store operation,

Fig. 1. Proposed nonvolatile 9T SRAM cell (NV9T)

III. SIMULATION RESULTS AND DISCUSSION This work focuses on the challenges faced in designing static random access memory in nanometer regime, where 27 2013 ACEEE DOI: 01.IJRTET.8.2.50

Full Paper Int. J. on Recent Trends in Engineering and Technology, Vol. 8, No. 2, Jan 2013

Fig. 3. Power gating with sleep transistors Fig. 5. Read delay comparison Table I. COMPARISON OF WRITE DELAY AND ITS VARIABILITY @ 27C.

Fig. 4. Timing diagram of storage nodes during store and restore operation of proposed nonvolatile 9T SRAM cell (NV9T)

VH drops due to discharging of storage node H through the leaky devices. When the restore operation is initiated by switching the sleep transistor ON the storage node H is restored to its previous high level. C. Read Access Time or Read Delay Read delay is estimated as the time required by BL (bit line)/BLB (bit line bar) to discharge by 50 mV [24]. The 50 mV differential between BL and BLB is good enough to be detected by sense amplifier, thereby avoiding misread [24], [25]. The proposed cell exhibits 38% penalty in read delay at nominal VDD compared with NV8T (see Fig. 5). This penalty occurs due to larger body effect offered by the read buffer and larger bit line capacitance. D. Write Access Time and Its Variability Write delay is estimated as the time required by the node storing 0 to rise up to 90% of VDD or the node storing 1 to fall by 10% of VDD. As pull-up and access devices are mainly responsible for the write delay and they are of same sizes in both the cells, therefore, only marginal differences in the write delay is observed due to tail MTJ elements in NV9T compared with NV8T (see Table I and Fig. 6). However, the proposed design proves its robustness by exhibiting narrower spread (6% @ nominal VDD of 0.8) in write delay variability. E. Write-ability The write-ability of a cell is measured by write static noise margin (WSNM), which is estimated graphically using read and write VTCs (voltage transfer curves). The write VTC, 2013 ACEEE DOI: 01.IJRTET.8.2.50 28

Fig. 6. Write delay and its variability comparison

while writing 1 to L, is gauged by sweeping VL (Fig. 7, yaxis) with BLB & WWL high and BL low and monitoring VH (Fig. 7, x-axis). This write VTC is used in combination with the read VTC, which is measured by sweeping VH (Fig. 7, xaxis) and monitoring VL (Fig. 7, y-axis). The side length of the smallest square, which can be embedded between the read and write VTCs of the same SRAM cell at the lower half of the curves, represents WSNM [26]. When WSNM falls below zero, write VTC intersects read VTC, thereby indicating positive write margin and signifying write failure. As can be observed from Fig. 7, the proposed design NV9T exhibits equal write-ability to NV8T . The proposed bitcell exhibits negative write margin, indicating successful write operation. As the proposed cell exhibits higher WSNM (=340 mV) than RSNM (= 330 mV), the cell is RSNM limited (see Fig. 8). F. Read Stability The standard 6T SRAM cell is found to be unstable at

Full Paper Int. J. on Recent Trends in Engineering and Technology, Vol. 8, No. 2, Jan 2013
Table II. C OMPARISON OF OPERATING FREQUENCY

be completely overlapped with the decoder operation. The estimated results are presented in Table II. As observed from the table, the proposed design exhibits lower operating frequency compared with NV8T because of larger body effect and bit line capacitance. H. Leakage Power Dissipation In nanoscale technology, thinner oxide and lower Vt result in significant rise in gate leakage and sub-threshold leakage currents. Therefore, leakage power is now a significant contributor to the total chip power. It is imperative to reduce leakage power to prolong battery life. The comparison of hold power (HPWR) of NV9T and NV8T is shown in Fig. 9 and 10. As can be observed, the proposed design consumes lower leakage power at all temperatures and supply voltages. This is due to the larger body effect offered by read buffer and the use of MTJ elements at the source end of the pull-down drivers.

Fig. 7. Static voltage characteristics of SRAM cells during write opera tion

Fig. 8. Static voltage characteristics of SRAM cells during read opera tion

nanoscale technology generation. It and its variants fail to meet operational requirements due to a low RSNM. The RSNM is estimated graphically as the length of a side of the largest possible square that can be inscribed inside the smaller wing of the butterfly curve. Fig. 8 plots the butterfly curves for RSNM of NV9T and NV8T. As can be observed from Fig. 8, NV9T offers higher RSNM compared with NV8T (1.94) due to isolation of storage nodes from bit lines. Cells with RSNM of at least 25% of VDD are generally considered to have excellent read stability [27]. The proposed cell meets this requirement. G.. Performance Analysis For the estimation of the operating frequency, the precharge periods was not taken into account because it can 2013 ACEEE DOI: 01.IJRTET.8.2.50 29

Fig. 9. Leakage or hold power versus temperature

Fig. 10. Leakage or hold power versus VDD

Full Paper Int. J. on Recent Trends in Engineering and Technology, Vol. 8, No. 2, Jan 2013 CONCLUSIONS This work proposes a low-power and variability-aware nonvolatile 9T SRAM cell. It analyzes the impact of PVT variations on write delay. It consumes lower standby power at all considered temperatures and supply voltages. It also achieves higher read stability (RSNM) without reduction of its write-ability. The proposed design is, therefore, an attractive alternative for low power applications in scaled technology node such as 22-nm and beyond. REFERENCES
[1] S. A. Wolf, D. D. Awschalon, R. A. Buhrman, J.M. Daughton, S. vonMolnar,M. L. Roukes, A. Y. Chtchelkanova, and D. M. Treger, Spintronics: A spin-based electronics vision for the future, Science, vol. 294,no. 5546, pp. 14881495, Nov. 2001. [2] I. Zutic, J. Fabian, and S. Das Sarma, Spintronics: Fundamentals and applications, Rev. Mod. Phys., vol. 76, no. 2, pp. 323410, Apr. 2004. [3] J. Fabian, A. Matos-Abiaguea, C. Ertlera, P. Stano, and I. Zutic, Semiconductor spintronics, Acta Physica Slovaca, vol. 57, no. 45, pp. 565907, Aug.Oct. 2007. [4] S. Bandyopadhyay and M. Cahay, Introduction to Spintronics. Boca Raton, FL: CRC Press, 2008. [5] J.-M. George, M. Elsen, V. Garcia, H. Jaffres, and R. Mattana, Spintonic with semiconductors, C. R. Physique, vol. 6, no. 9, pp. 966976, Nov. 2005. [6] C. Chappert, A. Fert, and F. N. Van Dau, The emergence of spin electronics in data storage, Nature Mater., vol. 6, no. 11, pp. 813823, Nov. 2007. [7] D. D. Awschalom and M. E. Flatte, Challenges for semiconductor spintronics, Nature Phys., vol. 3, no. 3, pp. 153159, Mar. 2007. [8] S. Sugahara and J. Nitta, Spin-trasistor electronics: an overview and outlook, Proc. IEEE, vol.98. no. 12, pp. 2124 2154, Dec. 2010. [9] W. H. Butler, X.-G. Zhang, T. C. Schulthess, and J. M. MacLaren, Spin dependent tunneling conductance of Fe/MgO/ Fe sandwiches, Phys. Rev. B, Condens. Matter, vol. 63, no. 5, pp. 054 416, Jan. 2001. [10] S. Ikeda, J. Hayakawa, M. L. Young, F. Matsukura, Y. Ohno, T. Hanyu, and H. Ohno, Magnetic tunnel junctions for spintronic memories and beyond, IEEE Trans. Electron Devices, vol. 54, no. 5, pp. 9911002, May 2007. [11] S. Matsunaga, J. Hayakawa, S. Ikeda, K. Miura, H. Hasegawa, T. Endoh, H. Ohno, and T. Hanyu, Fabrication of a nonvolatile full adder based on logic-in-memory architecture using magnetic tunnel junctions, Appl. Phys. Express, vol. 1, no. 9, pp. 091 301-1091 301-3, 2008. [12] Z. Liu, and V Kursun, Characterizatuon of a novel ninetransistor SRAM cell, IEEE Trans. Very Large Scale Integr. (VLSI) Syst., vol. 16, no. 4, pp. 488-492, Apr. 2008. [13] A. Islam , Mohd. Hasan, and Tughrul Arslan, Variation Resilient Subthreshold SRAM Cell Design Technique, International Journal of Electronics, vol. 99, no. 9, pp. 12231227, September 2012. [14] A. Islam and Mohd. Hasan, Leakage characterization of 10T SRAM cell, IEEE Trans. Electron Devices, vol. 59, no. 3, pp. 631 638, March 2012. [15] A. Islam and Mohd. Hasan, Variability Aware Low Leakage Reliable SRAM Cell Design Technique, Microelectronics reliability, vol. 52, no. 6, pp. 1247 1252, June 2012. [16] A. Islam and Mohd. Hasan, A Technique to Mitigate Impact of Process, Voltage and Temperature Variations on Design Metrics of SRAM Cell, Microelectronics Reliability, vol. 52, no. 2, pp. 405 411, February 2012. [17] Nanoscale Integration and Modeling (NIMO) Group, Arizona State University (ASU). [Online]. Available: http:// ptm.asu.edu/. [18] N. Varma, J. Kong, and A. P. Chandrakasan, Nanometer MOSFET variation in minimum energy subthreshold circuits, IEEE Trans. Electron Devices, vol. 55, no. 1, pp. 163-174, Jan 2008. [19] Jan M. Rabaey, Anantha Chandrakasan, and Borivoje Nikolic, Digital Integrated Circuits: A Design Perspective, 2 ed. New Delhi, India: Prentice-Hall of India, 2005. [20] Y. Shuto, et al., Nonvolatile SRAM architecture using MOSFET-based spin-transistors, arxiv.org/pdf/0901.2980. [21] R. Vaddi, S. dasgupta, and R.P . Agarwal, Device and circuit co-design robustness studies in the subthreshold logic for ultralow-power applications for 32 nm CMOS, IEEE Trans. Electron Devices, vol. 57, no. 3, pp. 654-664, Mar. 2010. [22] Semiconductor Industry Association (SIA), International Technology Roadmap for Semiconductors 2009 Edition, http:/ /www.itrs.net/Links/2009ITRS/Home2009.htm. [23] M. Alioto, G. Palumbo, and M. Pennisi, Understanding the effect of process variations on the delay static and domino logic, IEEE Trans. Very Large Scale Integr. (VLSI) Syst., vol. 18, no. 5, pp. 697710, May. 2010. [24] J. P. Kulkarni, K. Kim, and K. Roy, A 160mV robust schmitt trigger based subthreshold SRAM, IEEE J. Solid-State Circuits, vol. 42, no. 10, pp. 23032313, Oct. 2007. [25] H. Noguchi, S. Okumura, Y. Iguchi, H. Fujiwara, Y. Morita, K. Nii, H. Kawaguchi, and M. Yoshimoto, Which is the best dual-port SRAM in 45-nm process technology?8T, 10T single end, and 10T differential, in Proc. IEEE ICICDT, Jun. 2008, pp. 5558. [26] Z. Guo, A. Carlson, L-T. Pang, K. T. Doung, T-J. K. Liu, and B. Nikolic, Large-scale SRAM variability characterization in 45nm CMOS, IEEE J. Solid-State Circuits, vol. 44, no. 11, pp. 3174-3192, Nov. 2009. [27] A. E. Carlson, Device and circuit techniques for reducing variation in nanoscale SRAM, Ph.D. dissertation, Univ. California Berkeley, Berkeley, CA, May 2008.

2013 ACEEE DOI: 01.IJRTET.8.2.50

30