195 Review scientific paper  MIDEM Society Spintronics as a Non-Volatile Complement to Modern Microelectronics Viktor Sverdlov1, Josef Weinbub2, and Siegfried Selberherr1 1Institute for Microelectronics, TU Wien, Wien, Austria 2Christian Doppler Laboratory for High Performance TCAD, Institute for Microelectronics, TU Wien, Wien, Austria Abstract: Continuous miniaturization of semiconductor devices has been the main driver behind the outstanding increase of speed and performance of integrated circuits. In addition to a harmful active power penalty, small device dimensions result in rapidly rising leakages and fast growing stand-by power. The critical high power consumption becomes incompatible with the global demands to sustain and accelerate the vital industrial growth, and an introduction of new solutions for energy efficient computations becomes paramount. A highly attractive option to reduce power consumption is to introduce non-volatility in integrated circuits. Preserving the data with- out power eliminates the need for refreshment cycles and related leakages as well as the necessity to initialize the data in temporarily unused parts of the circuit. Spin transistors are promising devices, with the charge-based functionality complemented by the electron spin. The non-volatility is introduced by making the source and drain ferromagnetic. Recent advances in resolving several fundamental problems including spin injection from a metal ferromagnet to a semiconductor, spin propagation and relaxation, as well as spin ma- nipulation by the electric field, resulted in successful demonstrations of such devices. However, the small relative current ratio between parallel/anti-parallel source and drain alignment at room temperature remains a substantial challenge preventing these devices from entering the market in the near future. In contrast, a magnetic tunnel junction is an excellent candidate for realizing power-reducing approaches, as it possesses a simple structure, long retention time, high endurance, fast operation speed, and yields high integration density. Magnetic tunnel junctions with large magnetoresistance ratio are perfectly suited as key elements of non-volatile magnetoresistive memory compatible with the complementary metal-oxide-semiconductor technology and capable to replace dynamic and potentially static random access memo- ries. We review the present status of the technology, remaining challenges, as well as approaches to resolve the remaining problems. Regarding active power reduction, delegating data processing capabilities into the non-volatile segment and combining non-volatile elements with CMOS allows for efficient power gating. It also paves the way for a new low-power and high-performance in-memory processing paradigm-based on an intrinsic logic-in-memory architecture, where the same non-volatile elements are used to store and to process the information. Spintronics kot trajno dopolnilo moderne mikroelektronike Keywords: Spintronics; non-volatility; spin field-effect transistors; spin relaxation; magnetic tunnel junctions; magnetic random access memory (MRAM), spin-transfer torque MRAM, spin-orbit torque MRAM; voltage-controlled MRAM; logic-in-memory Sestavljivi 2D Vernier TDC na osnovi obročnih oscilatorjev Izvleček: Nenehno zmanjševanje polprevodniških naprav je bila glavna gonilna sila izjemnega povečanja hitrosti in zmogljivosti integriranih vezij. Poleg škodljivega povečevanja aktivne moči majhne dimenzije naprav vplivajo tudi na višje uhajalne tokove in večjo porabo v stanju mirovanja. Visoka poraba energije je postala nekompatibilna z zahtevami vzdržnosti rastoče industrije, zato so potrebne nove rešitve učinkovite rabe energije. Zelo privlačna možnost zmanjšanja porabe energije je uvedba trajnosti v integriranih vezjih. Ohranjanje podatkov brez porabe energije odpravlja potrebo po osveževalnih ciklih in uhajanjih ter nujnosti inicializiranja podatkov v začasno neuporabljenih delih vezja. Spin tranzistorji so obetavni elementi, ki temeljijo na osnovi naboja z dopolnitvijo vrtilne količine elektrona (spin). Trajnost se ustvari tako, da sta vir in ponor feromagnetna. Nedavni napredek pri reševanju številnih temeljnih problemov, vključno s injektiranjem vrtilne količine iz kovinskega feromagnetika v polprevodnik, vzpostavitev in relaksacija vrtilne količine ter upravljanje vrtilne količine z električnim poljem, Journal of Microelectronics, Electronic Components and Materials Vol. 47, No. 4(2017), 195 – 210 196 1 Introduction The breathtaking increase in performance and speed of integrated circuits has been enabled by continu- ous miniaturization of complementary metal-oxide semiconductor (CMOS) devices. On this exciting path numerous outstanding technological challenges have been resolved. Among the most crucial technological changes recently adopted by the semiconductor in- dustry to boost CMOS performance while maintaining gate control over the semiconductor channel are the introduction of strain [1], high-k gate dielectrics and metal gates [2], and a three-dimensional (3D) tri-gate transistor architecture [3-5]. The successes and innova- tive solutions developed for the microelectronics tech- nology have been always supported by sophisticated simulation tools, which allow reducing the research and development costs by 35-40% [6]. Although transistor sizes are scaled down, the on-cur- rents cannot be further decreased due to the need to charge/discharge the load capacitances and to main- tain the clock, which is saturated at approximately 4.0 GHz. Increasing the clock frequency results in an active power penalty, while continuous transistor scaling re- sults in growing leakages and stand-by power. Novel revolutionary approaches are desperately needed in the long run to sustain the vital societal and industrial progress in computing performance whilst simultane- ously reducing power consumption. The ultimate solution to one of the primary issues – the power reduction – is to introduce non-volatility into the circuits. Non-volatility is the ability to preserve data, when the supply power is turned off. It enables stand- by power-free integrated circuits as no information is je bil ključen za uspešno demonstracijo takšnih naprav. Kljub temu ostaja relativno majhno razmerje tokov med vzporednim/nasprotnim položajem izvora in ponora pri sobni temperaturi ključen izziv, ki preprečuje vstop teh naprav na trg v bližnji prihodnosti. V nasprotju s tem je magnetni tunelski spoj odličen kandidat za uresničevanje pristopov zmanjšanja moči, saj ima preprosto struk- turo, dolg čas zadrževanja, visoko vzdržljivost, hitro obratovalno hitrost in visoko gostoto integracije. Magnetni tunelski spoji z visokim razmerjem magnetoresonance so idealni za uporabo v trajnih megnetorezistivnih spominih, ki so združljivi s komplementarno kovina- oksid-polprevodnik tehnologijo in sposobni zamenjati dinamičen in statičen spomin z naključnim dostopom. V članku je podan pregled tehnologije, izzivi in postopki, kako rešiti obstoječe probleme. Zmanjševanje porabe energije se lahko doseže s prenosom obdelave podatkov v trajne segmente in kombinacijo trajnih elementov s CMOS. Prav tako se utira pot novi paradigmi procesiranja v pomnilniku z nizko porabo energije in visoko zmogljivostjo, ki temelji na arhitekturi logike v pomnilniku, kjer se za shranjevanje in obdelavo podatkov uporabljajo isti trajni elementi. Ključne besede: Spintronika; trajnost; tranzistorji z vrtilnim poljem; relaksacija vrtilne količine; magnetni tunelski spoj; magnetni spomin z naključnim dostopom (MRAM); navor prenosa vrtilnosti MRAM; navor orbite vrtilnosti MRAM; napetostno krmiljen MRAM; spominska logika * Corresponding Author’s e-mail: sverdlov@iue.tuwien.ac.at lost and there is no need to recover the data, when the power is turned on. Non-volatility is crucial for elimi- nating the leakage power dissipation and data refresh- ment cycles. Apart from stand-alone applications, e.g., critical program and data storage devices in extreme environments employed in the air and space industry, it is particularly promising to use non-volatility in the main computer memory as a replacement of conven- tional volatile CMOS-based dynamic random-access memory (DRAM) [7], which will drastically reduce en- ergy consumption. In modern multicore processors, much of the energy consumption appears in the hier- archical multi-level cache memory structure. To reduce this energy consumption, a viable approach is to re- place the caches with a non-volatile memory technol- ogy which also offers a reduced memory cell size com- pared to static random-access memory (SRAM) [7]. This will help bridging the speed gap between the last-level cashes and main memory, since CMOS SRAM is much faster compared to CMOS DRAM. To be competitive with the traditional volatile memory technologies and also with non-volatile flash memo- ry, emerging non-volatile memories must offer a fast switching time, a high integration density supported with good scalability, a long retention time, a high en- durance, and a low power consumption. At the same time, they must possess a simple structure to reduce fabrication costs and the new non-volatile circuit ele- ments must be compatible with CMOS technology to benefit from advantages provided by the well-devel- oped CMOS fabrication technology. A spin field-effect transistor (SpinFET) is a promising fu- ture semiconductor device with a performance poten- V. Sverdlov et al; Informacije Midem, Vol. 47, No. 4(2017), 195 – 210 197 tially superior to that achieved in the present transis- tor technology. The non-volatility in SpinFETs is added by replacing the non-magnetic source and drain in a FET by its ferromagnetic counterparts. The two ferro- magnetic contacts (source and drain) are linked by a non-magnetic semiconductor channel region. Metallic ferromagnetic contacts serve not only as an injector/ detector of the spin-polarized electron charge current in the channel, but, because of their magnetization, the source and drain electrodes provide an additional current modulation due to their capabilities to inject/ detect spins [8]. Indeed, the electron current gets en- hanced in the case of parallel alignment between the source/drain electrodes as electrons are injected with spins parallel to the drain magnetization and can easily escape from the channel to the drain, while the current is suppressed for anti-parallel magnetization align- ment [8]. As the magnetization of the source/drain can be manipulated by means of an external magnetic field and/or current (by means of the spin-transfer torque), the two on-current states for parallel/anti-parallel mag- netization alignment potentially enable reprogramma- ble logic [9]. Importantly, the relative magnetization orientation between source and drain is preserved without external power, which makes reprogramma- ble logic partly non-volatile. Below we discuss recent advances and remaining challenges to realize SpinFET- based logic in detail. We only stress the most impor- tant, in our opinion, shortcoming to overcome, namely, a small relative difference between the on-currents in parallel and anti-parallel source/drain alignment. Magnetoresistive random access memory (MRAM) and in particular spin transfer torque (STT) MRAM possesses many, if not all, of these advantages and is considered as a perfect candidate for future universal memory ap- plications. MRAM is CMOS-integrable, which increases its potential to replace the typical processor-embed- ded SRAM and DRAM. With STT MRAM currently emerging as a commercial product for stand-alone applications, it will be critically important to introduce STT MRAM in the main com- puter memory, i.e., to replace conventional DRAM and SRAM. This will create a new innovative multi-billion dollar industry and will sustain the breathtaking path of electronics by delivering cheaper, faster, and envi- ronmentally friendlier compact and mobile devices. Bringing STT MRAM into the vast computer memory market as embedded and stand-alone applications for traditional high-performance and low-power mobile platforms will result in an exponential growth of the non-volatile memory market share in the near future. Regardless of the first commercial STT MRAM-based products for stand-alone applications being available, one critical aspect of the currently used STT MRAM technology is a relatively high switching current, which again is in sharp contrast to the overall demand for re- duced energy consumption. This obviously prevents the MRAM from successfully entering the vast com- puter memory market. The problem of high switching current and large active writing energy jeopardizes the advantages provided by non-volatility, such as zero stand-by power, no data refreshment, and no data re- covery. Several plausible approaches to address these issues are conceivable including the replacement of the in-plane magnetization orientation in magnetic tunnel junctions (MTJs) with perpendicular magneti- zation, the use of composite free recording layers, de- coupling the write and read current paths, controlling magnetization by voltage, and employing new materi- als with improved properties and characteristics. 2 Spin transistor In a SpinFET [8] schematically shown in Fig.1 the elec- trons with spin aligned to the drain magnetization di- rection can easily leave the channel to the drain thus contributing to the current. The total current through the device depends on the relative angle between the magnetization direction of the drain and the electron spin polarization at the end of the semiconductor channel. The electron spin orientation at the end of the channel is determined by the source magnetization and can be additionally manipulated by the modula- tion which is achieved by tuning the strength of the effective spin-orbit interaction in the channel induced by the gate voltage. As a non-equilibrium quantity, the injected spin relaxes to its equilibrium zero value while propagating through the channel. Spin relaxation is an important detrimental ingredient as it reduces the cur- rent modulation and affects the SpinFET functionality. Figure 1: Datta-Das SpinFET [8]. Spin-polarized elec- trons are injected from a ferromagnetic source and ab- sorbed by in a ferromagnetic drain. The electron spins in the channel are manipulated by means of the gate voltage-dependent spin-orbit interaction. 2.1 Spin-Orbit Interaction The spin relaxation is governed by the spin-orbit inter- action (SOI) and scattering, both spin-dependent and V. Sverdlov et al; Informacije Midem, Vol. 47, No. 4(2017), 195 – 210 198 spin-independent, and manifests itself differently in semiconductors with and without the inversion sym- metry [10,11]. In crystals obeying the inversion symmetry (silicon, germanium) the spin relaxation is governed be the Elli- ott-Yafet mechanism [10,11]. The wave function with a fixed spin projection (defining the quantization axis) is not an eigenstate of the Hamiltonian due to the elec- tron momentum-dependent SOI. In other words, the SOI forces the eigenstate wave function to possess a small but finite contribution with an opposite spin pro- jection in the fixed basis. Therefore, the small but finite amplitude to flip the electron spin appears at every spin-independent scattering event – the Elliott process [12]. This is complemented by the Yafet spin-flip events due to SOI-dependent electron-phonon scattering. In silicon the electron spin relaxation is determined by the inter-valley transitions [12] and can be efficiently controlled by stress [13]. In silicon channels, uniaxial stress generating shear strain is particularly efficient to suppress the spin relaxation [14] as it lifts the degen- eracy between the two unprimed subband ladders [15]. In addition, choosing the spin injection direction also boosts the spin lifetime by a factor of two [16], as shown in Fig.2. Figure 2: Spin lifetime in (001) thin Si film as a function of shear strain. The spin lifetime is a factor of two longer for spins injected in-plane as compared to the time for spins injected perpendicular to the film. The factor of two is preserved for both inter- and intra-valley scat- tering and is independent of the scattering mechanism (electron-phonon and surface roughness scattering). In III-V materials without inversion symmetry the de- generacy between the up and down spin states with the same electron momentum is lifted, and the spin relaxation is governed by the Dyakonov-Perel mecha- nism [10,11]. However, the SOI does not always play a detrimental role. In semiconductor channels the SOI may also be used efficiently to manipulate the electron spins in the channel [8]. The inversion symmetry in the channel can additionally be violated by applying the gate voltage. In this case the strength of the effective SOI depends on the effective electric field perpendicu- lar to the channel [17]. The strength of the gate voltage- dependent SOI can be used to modulate the current between the ferromagnetic source and drain by means of an additional spin modulation in the channel result- ing in a different spin orientation relative to the drain as compared to the case without SOI [8]. Importantly, as the strength of the SOI in the channel depends on the effective electric field, the suggested method provides a purely electrical mean of manipulating the electron spins and thus the current in the channel. The voltage-induced SOI in III-V materials can be used for an efficient spin injection in the channel from the point contacts [18]. Additional gates are used to cre- ate the point contacts to the two-dimensional (2D) electron gas by confining the 2D gas in the III-V chan- nel under the gates. Application of different voltages to these gates generates the spin-orbit Rashba field perpendicular to the point contact. By properly tuning the chemical potential one can achieve that, due to this SOI, all electrons moving to the right are spin-polarized (while electrons moving to the left are polarized in the opposite direction). Thereby an efficient and purely electrical spin injection/detection is achieved. Using this injection scheme, the ever first reliable demonstra- tion [18] of a working SpinFET [8] suggested in 1990 was achieved. Not long ago, new 2D materials (graphene, transition metal dichalcogenides), become attractive for future microelectronics applications. Recently, a new concept for realizing a spin switch with a graphene channel was demonstrated [19]. Spin-polarized electrons are in- jected into the graphene, a good spin conductor due to low SOI, and reach the drain electrode, if the elec- trochemical potential in MoS2 is tuned into the energy gap. In this case the electrons do not enter MoS2. The situation is completely changed, if the electrochemical potential is tuned by the gate into the MoS2 conduction band. In this case a parallel path for electrons through a material with high SOI is open, which results in strong spin relaxation, so that the current reaching the drain is not spin-polarized. Regardless of the successful demonstration of the SpinFET, the conductance modulations were only re- solved at temperatures far below 300K. The spin switch [19] discussed above was demonstrated to work at 0 0.6 1.2 1.8 2.4 εxy (%) 10-4 10-3 10 -2 10-1 100 101 102 10 3 104 10 5 106 Sp in li fe tim e [n s] in-plane out-of-plane intra-valley inter-valley V. Sverdlov et al; Informacije Midem, Vol. 47, No. 4(2017), 195 – 210 199 temperatures below 200K and requires additional cool- ing, which modern microelectronics working at room temperature is striving to avoid. Recently, the first suc- cessful demonstration of a silicon spin metal-oxide- semiconductor field-effect transistor (SpinMOSFET) at room temperature [20] was presented. In silicon the strength of the Rashba SOI is much smaller compared to that in III-V semiconductors. Therefore, the SOI can- not be used for spin manipulation. Thus, the current modulation in the SpinMOSFET is achieved by altering the relative magnetization between the ferromagnetic source and drain. This way, a high difference between the on-currents in parallel and anti-parallel source/ drain configuration was demonstrated. However, the relative ratio of the currents, a characteristic similar to the tunnel magnetoresistance (TMR) ratio, is still sever- al orders of magnitude lower [20] than the TMR in MTJs. Fig.3 shows the TMR in a silicon SpinFET as a function of the SOI strength, for several channel lengths for an ideal case when the spin relaxation is neglected. In or- der to facilitate the spin injection and detection, delta function-like Schottky barriers between the source, drain, and the channel of the strength 2 2 F F EhU mπ = were h is the Plank constant, EF and mF are the Fermi energy and the electron mass in the ferromagnetic contacts, are assumed. Even in this ideal case the TMR is about 10%, much inferior to that in MTJs. A TMR less than 1% was experimentally observed at room tem- perature [20]. Figure 3: The ratio of the resistance difference in an- tiparallel and parallel configurations to the resistance in parallel configuration of the source and drain in Si- based SpinFETs as a function of the spin-orbit interac- tion strength, for several channel lengths. 2.2 Spin Injection and Spin-Dependent Tunneling The device described above can function only if the electron spins are efficiently injected/extracted in/ from the channel. As there are no semiconductor fer- romagnets at room temperature, to achieve the ef- ficient injection/detection from metal ferromagnets in the semiconductor channel and vice versa, a thin tunneling barrier must be placed between the elec- trodes and the channels [21] to mitigate the spin im- pedance mismatch. However, the signal attributed to the spin injection [22] appears to be much weaker as compared to the large effect [23] currently attributed to the spin-dependent resonant tunneling [24-26], and the development of efficient ways to electrically inject spins from a ferromagnetic metal in a semiconductor has become an area of active research. To summarize, although many fundamental challenges have been resolved and both a SpinFET and a SpinMOS- FET have been successfully demonstrated, an enhance- ment of the on-current ratio between the parallel and anti-parallel source/drain magnetization alignment at room temperature remains one of the main challenges. In addition, both SpinFET and SpinMOSFET still rely on the charge current to transfer the spin, which may set some limitations for the applicability of such devices in main-stream microelectronics in the future. Non-vola- tile devices based on MTJs possess the TMR suitable for practical applications and are reviewed below. 3 Magnetoresistive random access memory Applications driven by magnetic moments and in- duced magnetic fields have a large share in typical information technology products. Coupling between magnetic fields and currents in coils was employed in the first electronic devices. However, the coupling is relatively weak, resulting in low efficiency and high energy supply costs. An efficient coupling between the electrical and the magnetic degree of freedom is possi- ble on a quantum mechanical level and was discovered in 1986 as a phenomenon called the giant magneto- resistance (GMR) effect. This facilitated a reliable, purely electrical read operation of the information encoded in the magnetization orientation. Based on this principle hard drive storage devices with extremely high density appeared on the market. The enormous impact of this discovery on the development of information technol- ogy was recognized by awarding the inventors the No- bel Prize in 2007 [27,28]. 0 10 20 30 40 50 β [µeVnm] -0.1 -0.05 0 0.05 0.1 TM R L=2µm L=4µm L=8µm V. Sverdlov et al; Informacije Midem, Vol. 47, No. 4(2017), 195 – 210 200 The next generation of storage devices with higher density is based on the unique properties of the MTJ. It was discovered that, if the non-magnetic metal layer in a GMR memory element is substituted by a thin di- electric, the tunneling current through the structure strongly depends on the relative polarization of the fer- romagnetic contacts (Fig.4). The difference in the MTJ resistivity can reach several hundred percent at room temperature [29]. Thanks to this technology a new gen- eration of hard drives with even higher storage densi- ties has been developed. Figure 4: A magnetic tunnel junction possesses low (high) resistance for parallel (antiparallel) relative ori- entation between the ferromagnetic electrodes. In order to be used in memories, MTJs must be com- plemented with the ability to efficiently convert charge information into magnetic moment orientation. Writ- ing the state by the magnetic field is currently used in toggle switching commercial MRAM. This method, however, is not scalable as the magnetic field is gener- ated by the current, which leads to a current increase with scaling [30]. 3.1 Spin Transfer Torque Magnitoresistive Random Access Memory The STT effect [31,32] has been proven to be a perfect alternative to the magnetic field for magnetization switching. The STT is used for purely electrical data writing by passing the current through the MTJ. The memory technology based on MTJs and the STT effect has resulted in the development of STT MRAM. STT MRAM is characterized by lower power-consumption and better scalability than conventional MRAM, where the switching is performed by the magnetic field [33] generated by an electric current passing through the write lines next to the cell. Several cells are arranged in a matrix connected with bit and world lines. A cell in this cross-point architecture is written by simultane- ously selecting the cell with current pulses applied to the corresponding world and bit lines. The problem of half-selected cells [30] is solved by the application of a certain pulse sequence to the lines supplemented with a special design of the free layer arranged as a synthetic anti-ferromagnet [34]. This results in a deterministic, toggle-like fast switching of the free layer. The magnetic field employed for switching prevents MRAM from scaling down beyond 90nm [35] as the current needed for the field generation rapidly in- creases with scaling. STT [31,32] opened a new way of manipulating magnetization dynamics by using spin- polarized currents instead of magnetic fields. The spin- polarized current allows writing the information into the memory cell by purely electrical means. When elec- trons pass through a fixed ferromagnetic layer, their spins become aligned with the magnetization. When these spin-polarized electrons enter the free layer, they become aligned with the magnetization of the free layer within a transition layer of a few angstroms. The electron spins change results in a torque exerted on the magnetization of the free layer. This torque causes magnetization switching, if it is large enough to over- come the damping. By altering the current polarity the magnetization of the free layer can be switched from the anti-parallel to the parallel state and back with re- spect to the reference layer. The interest in STT MRAM has increased significantly after the observation of spin torque induced switch- ing in AlOx-based [36] and MgO-based [37] STT MRAM cells. Depending on the orientation of the layer mag- netizations the magnetic pillars can be divided into two categories: (1) perpendicular with out-of-plane magnetization direction and (2) in-plane with the mag- netization lying in the plane of the magnetic layer. The introduction of STT MRAM with in-plane magnetiza- tion orientation to the market has already begun with the first demonstration by Everspin Technologies [38] of a 64Mb chip. The size of the MTJ bits is 80–90nm with an aspect ratio of 2 and 3. An MgO barrier was used and MTJs with a TMR ratio above 110% were exploited. The chip is able to operate within a broad temperature range. Switching the magnetization can occur spontaneously due to thermal fluctuations. This is an undesired event, which leads to the loss of the stored information. An important parameter of MRAM is the thermal stability V. Sverdlov et al; Informacije Midem, Vol. 47, No. 4(2017), 195 – 210 201 factor which is defined as the ratio of the thermal sta- bility barrier to the operating temperature. For gigabit applications the thermal stability barrier should be at least 80kT to guarantee the required retention time of 10 years. Achieving large thermal stability and a low switching current for fast switching simultaneously represents one of the main challenges to engineer a good MRAM cell. Perpendicular MTJs (p-MTJs) with the thermal barrier equal to the switching barrier are pre- ferred for applications, because they allow to reduce the switching current. In addition, p-MTJs are better suited for high-density memory [39]. Both field-induced and STT switching can be comple- mented with heat assisted switching [30]. This tech- nology was routinely used in hard drives in order to facilitate writing. Presently methods using thermally assisted switching in MRAM have already been devel- oped. In addition to assisting switching, heat can fa- cilitate new unique functionalities, for instance, using an MRAM cell with a soft reference layer as a magnetic logic unit [40]. This extends the MRAM research and development area towards logic-in-memory architec- tures and non-volatile computing. Regardless of the undisputable success of the first MRAM products on the market, several important chal- lenges remain. The general requirements for any mem- ory type including MRAM are: - ability to write the data with low energy without damaging the device; - data retention within a given long time interval; - ability to read the recorded data without destroy- ing the data. Improving one or two aspects of its functionality usu- ally leads to a degradation of the remaining function- ality [30]. Therefore, a careful parameter optimization specific to a particular technology is the main subject which must be addressed in order to facilitate produc- tion of high density memory arrays suitable for replac- ing SRAM caches and DRAM-embedded main comput- er memory. One of the main problems of STT MRAM is a relatively high critical current required for STT-induced mag- netization switching. This fact has several implications. Firstly, due to the relatively high energy required for writing, the current generation of STT MRAM cannot be used in high-level processor caches due to the high activity factor in these elements and the high level of generated heat. The necessity to switch memory fre- quently negates the benefits of non-volatility provided by MRAM. Secondly, large switching currents are sup- plied via an access transistor. This potentially puts scal- ing limitations on the transistor dimensions of a one- transistor (1T)-1MTJ memory cell. However, a careful and innovative design yielded already a successful im- plementation of 8Mb 1T-1MTJ STT MRAM embedded in a 28nm CMOS logic platform [41]. Finally, a large switching current density can result in serious reliabil- ity issues like MTJ’s resistance drift and eventually its dielectric breakdown. The critical current density de- pends on the switching pulse duration, with a substan- tial current increase for faster, sub 10ns switching. A plausible way to reduce the switching current density is to work with p-MTJs. The problem of data retention is related to thermally agitated magnetization fluctuations. During these fluc- tuations the magnetization can switch spontaneously via a potential barrier separating the two states with opposite magnetization directions. As already noted, for about 10 years data retention the thermal stabil- ity barrier must be at least 80kT for gigabit MRAM ar- rays. However, increasing the barrier also results in an increase of the switching current density, which is proportional to the thermal barrier for p-MTJs. In order to reduce the switching current density and preserve the large thermal barrier at the same time one has to reduce the Gilbert damping and increase the spin cur- rent polarization. An interface-induced p-MTJ structure with a composite free layer CoFeB/Ta/CoFeB with two MgO interfaces [42] allows simultaneously boosting the thermal barriers and reducing damping. For in-plane MTJs, the faster switching can also be achieved, when the composite free layer is made of two half-ellipses separated by a narrow gap. The pecu- liarities of the magnetization dynamics of the two parts of the composite free layer [43,44], which occur in op- posite senses to each other, lead to the magnetization switching in-plane. This way the large demagnetiza- tion penalty of the magnetization getting out of plane is avoided, and the switching barrier becomes equal to the thermal barrier. Because the thermal barrier depends on the free layer volume, the required large thermal stability factors of ~80kT are easily achieved in this structure. A large TMR ratio is needed for reliably reading the information in MRAM. Indeed, the middle reference resistance to which the low and high resistance MTJ states are compared must be well separated from ei- ther of them. However, since a bit-to-bit resistance vari- ation within a memory array is increasingly difficult to control with device sizes scaling down, the dispersion increases and so must the TMR. Obtaining a large TMR is more difficult in interface-induced p-MTJs, because the layer width must be reduced in order to boost the magnetic anisotropy; however, a TMR ratio as large as 350% has been demonstrated [45]. V. Sverdlov et al; Informacije Midem, Vol. 47, No. 4(2017), 195 – 210 202 With growing data services such as Big Data analysis the need for additional memory capacity and speed as well as in-memory computing has increased dramati- cally. The last-level cache memory must be increased [46-48] to bridge the memory-bandwidth gap between central processing units (CPUs) and the main memory. The CPU performance can be significantly boosted by using fast non-volatile memories in cache for data stor- ing without the need to address the main memory. In particular, the use of STT MRAM as the last-level cache memory helps bridging the memory-bandwidth gap between multi-core CPUs and the main memory. Ultra-large volatile DRAM devices are available; how- ever, due to the high refresh rate and thus high power consumption their use as last-level caches has been limited. The introduction of non-volatility to reduce en- ergy consumption in last-level cache memory can in- crease the CPU performance significantly by using this cache for data storing and processing without the need to address the main memory. Advanced STT MRAM is characterized by high-speed access with less than 10ns. It is thus suitable for last- level cashes where it guarantees about ten times pow- er reduction [49-51], while other types of non-volatile memories are much slower and cannot provide such a high speed access. 4Gbit density STT MRAM arrays with p-MTJs and compact memory cell were recently report- ed [39]. On May 26th, 2017, Samsung [52] reaffirmed the beginning of production of embedded STT MRAM based on the 28nm silicon-on-insulator technology node [41] in 2018. On September 15th, 2017, Global- foundries announced the beginning of embedded STT MRAM production based on the 22nm fully-depleted silicon-on-insulator technology [53]. We are therefore witnessing the beginning of non-volatile STT MRAM entering the DRAM and potentially SRAM markets, tra- ditionally dominated by CMOS-based volatile devices. If successful, it will result in an exponential expansion of the STT MRAM market with a momentous impact on information storage and processing in the near future. 3.2 Advanced MRAM Although STT MRAM is competitive with DRAM for em- bedded memory applications and can also be used in level three caches in CPUs, increasing write currents for faster switching prevents it from being used in level one caches, where very fast switching is required. An ultimate swap to p-MTJs and Gilbert damping reduc- tion are two common paths to reduce the switching current; however, these efforts are counteracted by the necessity to maintain high thermal stability which requires high perpendicular magnetic anisotropy [30]. There are indications that by downscaling the p-MTJ di- ameter the switching current decreases faster than the thermal stability factor, which has been shown to be as high as 120 in p-MTJs with a diameter of 30nm [54]. Nevertheless, it is preferred to have an alternative way to switch the free layer. Interface-induced perpendicular magnetic anisotropy materials provide a sufficiently large thermal stability factor for free layers with diameters down to 12-14nm. Since the anisotropy is determined by the interface properties, it can be altered by applying an electric field. The electric field polarizes the charge densities of the interfacial atoms, thereby modifying overlap integrals and exchange interactions. This may soften the perpendicular magnetic anisotropy thus reducing the switching energy barrier and even changing it to in-plane. The magnetization can easily be pushed over the barrier by a small current and stabilized in the state with an opposite magnetization after the voltage is re- moved. 3.2.1 Voltage-controlled MRAM An MRAM controlled by voltage [55-58] is a viable op- tion for last-level cache applications. The voltage-con- trolled MRAM switching principle is based on voltage- mediated removal of the potential barrier separating the two stable magnetization orientation states. With- out the barrier the magnetization precesses around the effective magnetic field and can be put into the alter- native magnetization state, when the potential barrier is re-introduced at the end of the voltage pulse [59]. Because the voltage-induced switching is unipolar, the voltage controlled MRAM is free from the read disturb which is characteristic to STT MRAM. Although voltage controlled MRAM is a two-terminal device, the separa- tion between read and write is performed by alternat- ing the polarities during these two operations. Voltage controlled MRAM has a few unsolved issues so far preventing it from being broadly used in applica- tions. One of the problems originates in the preces- sion at switching and thus depends on the initial state determined by the fluctuating thermal and unwanted variability. This variability results in write errors and must be suppressed. The second problem is a larger resistance of the memo- ry cell compared to STT MRAM, which results in smaller currents. Small currents lead to a longer delay while reading the state by a sense amplifier. As it was shown recently [60], both problems can be solved by carefully tailoring and optimizing the entire circuit. V. Sverdlov et al; Informacije Midem, Vol. 47, No. 4(2017), 195 – 210 203 Extending the ideas of voltage-controlled magnetic anisotropy, the voltage pulse can be applied not only to lower the potential barrier between the two mag- netization states but also to boost it to make the switching harder [60]. The switching is mediated by a spin-orbit and/or spin Hall torque generated by the current flowing through a conductive line made of a heavy metal underneath the magnetic MTJ, ensuring the write operation without an external magnetic field [61]. If reading is performed by applying the voltage pulse with its polarity opposite to that used for writing, the potential barrier is increased, hardening the cell im- munity against read disturb errors. 3.2.2 Spin-orbit torque MRAM Among the newly discovered physical phenomena suitable for next-generation MRAM are the spin Hall ef- fect and the spin-orbit torque (SOT) switching [62-66]. Current passing in a material with a high spin Hall an- gle/SOI results in spin-orbit torques capable to switch the free layer of an MTJ. This way the read and write cur- rents are decoupled, which prevents the tunnel barrier from damage and improves device reliability. The spin Hall effect and/or SOT alone do not provide switching in devices with perpendicular magnetiza- tion. To provide switching, it is required to apply an external magnetic field. In addition, innovative materi- als are required to increase the torques and to boost the switching efficiency. New materials with a strong SOI, e.g., topological insulators, allow the current to flow only at their interface states [67]. Due to the spin- momentum locking characteristic to these states the passing current results in a large spin accumulation at the interface [68,69] and the SOT aids the magnetiza- tion to switch. A potential disadvantage of the write and read cur- rent paths’ separation is that these devices appear in a three-terminal cell configuration [70]. Therefore, they can be used only for applications in which the density is not the top priority, but for high-speed reliable op- eration competing with SRAM. There exists a different design of a three-terminal MRAM cell, where the switching is done by the current induced fast domain wall motion within a ferromag- netic material between the two ferromagnetic elec- trodes, while reading is done by means of an additional ferromagnetic contact grown on top of the ferromag- netic layer [70]. The domain wall is pushed by both STT and spin SOT, with the relative strength of each con- tribution tuned by proper engineering the magnetic layer structure. The domain walls can be moved very fast [71,72], which is attractive for high speed applica- tions. However, the critical current densities obtained experimentally are still high. A reduction of the critical currents by minimizing the domain wall pinning in do- main wall MRAM may compromise the data retention. Similar to SOT-based MRAM, the reduction of the cur- rent density, while maintaining the domain wall speed, remains a critical challenge for domain wall motion based MRAM. The need to lower the critical current in advanced MRAM accelerates the search for new materials with large SOI. A promising candidate for such a material is a topological insulator, for which a large spin Hall angle has been demonstrated [67]. A general form of the relevant torque terms in the presence of spin-orbit interaction can be determined by symmetry considera- tions [73,74]. As SOTs appear at the interface between a material with high SOI and a ferromagnet, a description of these torques by means of boundary conditions was recently suggested [75,76]. The corresponding bound- ary conditions allow to relate the non-equilibrium spin accumulation at both sides of the interface in presence of the in-plane current and couple them to the mag- netization dynamics. 4 Non-volatile logic MRAM is CMOS compatible and attractive to use with CMOS-based logic applications. Fast non-volatile mem- ory combined with non-volatile processing elements is a fertile ground for realizing the first microprocessors with reduced power consumption working on an en- tirely new principle. In addition, MRAM arrays are em- bedded directly on top of CMOS logic [77]. This allows reducing the length of interconnects and the corre- sponding delay time. 4.1 CMOS-MRAM hybrid logic The computer architecture where non-volatile ele- ments are located on a chip with CMOS devices is traditionally called logic-in-memory, although as of yet no information is processed in non-volatile ele- ments. Power-efficient MRAM-based logic-in-memory concepts have already been demonstrated [78]. They include field-programmable gate arrays and ternary content addressable memory as well as other variants. These CMOS/spintronic hybrid solutions are already competitive in comparison to the conventional CMOS technology with respect to power consumption and speed. The power consumption problem in modern integrat- ed circuits with ultra-scaled CMOS devices is becoming critical, which prompts various power reduction tech- nologies to be used for keeping the heat dissipation V. Sverdlov et al; Informacije Midem, Vol. 47, No. 4(2017), 195 – 210 204 under control. The techniques based on reduced volt- age operation, clock gating, and power gating modes allow to address the problem to a certain extent, how- ever, they also result in an increase of the time delay to get into or out from these modes. The use of non-vola- tile MRAM-based devices [79-81] with fast access to the stored data allows cutting out the penalty of stand-by power and eliminates the delays when using energy saving modes. The first microcontroller unit with zero standby power featuring non-volatile elements is op- erating at 8MHz [82]. In order to boost the operating frequency, spin-based non-volatile flip-flops were re- cently used to demonstrate a power-gating microcon- troller unit [78] fabricated with standard 90nm CMOS technology with an additional MTJ process. The chip features a very short delay in entering/exiting power- on/power-off with the potential to be further reduced by optimizing parasitic capacitances. Another new circuit example is a field programmable gate array built with non-volatile devices. Here, tempo- ral data is quickly saved in magnetic tunnel junctions before the power is turned off. This has a great poten- tial to reduce the power consumption, which becomes a critical issue in conventional SRAM-based gate arrays [83-85]. By using a logic-in-memory structure [86,87], replacing SRAM cells with non-volatile flip-flops [88] and smartly connected redundant MTJs to avoid resist- ance variations [89], the area of a six-input look-up ta- ble is shown to be reduced by about 50% [78]. Ternary content-addressable memory (TCAM) is able to perform a very high-speed search to match an input [90]. CMOS-based TCAM suffers from standby power losses and relatively high costs due to its complex struc- ture [90]. Employing a 2T-2MTJ structure for the equality search logic part reduces the TCAM cell area [91,92]. A 1Mb non-volatile TCAM chip with a 6T-2MTJ cell struc- ture fabricated in 90nm CMOS and perpendicular MTJ technologies has been demonstrated [93], with 9T- 2MTJ [94], 7T-2MTJ [95], 4T-2MTJ [96], and 5T-4MTJ [97] modifications for high-speed accessibility and reduced variation effects have been also reported. Currently, the TCAM cell structure design as well as the word segmen- tation algorithm optimization is under intense investiga- tion [78] in order to increase speed and reduce the area. With the continued rapid development of smartphones and mobile video applications it becomes necessary to introduce non-volatile elements into important cir- cuits responsible for performance acceleration. A mo- tion-vector prediction circuit is critical for performing mobile video compression by finding motion vectors between two adjacent frames. It has been demon- strated that the introduction of non-volatile elements to implement a full adder helps making the circuit compact, fast, and stable [98-102]. The introduction of non-volatility and a logic-in-memory architecture helps reduce power consumption by 45% [103]. With the activation ratio of embedded clusters decreased a reduction of 97% is possible [78]. Another example of an application specific circuit currently under thorough investigation is a brain-inspired computing network with non-volatile elements, which also demonstrates a large, i.e., more than 90%, power reduction on average when compared to its CMOS based counterparts [78]. 4.2 Intrinsic logic-in-memory The current age of Big Data requires an unprecedented level of data storage capacity complemented with ef- ficient processing capabilities. The data processing is typically confined in large data centers, which appears to customers as a cloud computing environment to en- able resource flexibility. The scale of data centers (and their power consumption) is increasing exponentially. One of the limitations of current computing systems is the overhead of transferring data between memory and processors. As already mentioned, the problem can be solved by placing the main memory closer to processors. Another efficient solution will be to per- form at least part of the data processing already in the storage by designing a memory architecture with en- hanced functionalities capable to directly perform a set of Big Data-oriented, memory-centric operations. This methodology promises a dramatic reduction in the need for data transfers between memory and proces- sor, eliminating the interconnection bottleneck, by cre- ating a new high performance and low power efficient computing paradigm, where the data is not only stored but also analyzed by non-volatile stand-by-power free processing units. Placing the actual computation into the magnetic domain reduces the need of converting magnetically stored information into currents and voltages for pro- cessing and helps not only to simplify the circuit layout but also increases the integration density. The idea is to use MTJs as elementary blocks for non-conventional logic-in-memory architectures. Our invention, which shows that on an MRAM array any two of the coupled 1T-1MTJ cells can serve simultaneously as non-volatile memory and computing units by performing a logical implication operation, has been granted a patent [104]. These structures inherently realize non-volatile logic- in-memory circuits with zero-standby power, where the same elements are used for storing and also pro- cessing information. They have a great potential for Big Data storing and computing, as they are also opening a path for developing computing architectures concep- tually different from the still standard Von Neumann architecture. A new design of an implication-based full V. Sverdlov et al; Informacije Midem, Vol. 47, No. 4(2017), 195 – 210 205 adder involves six 1T-1MTJ cells with 27 subsequent FALSE and material implication operations [105]. The paradigm of employing memory for information processing is perfectly suited for the Big Data revolu- tion we are experiencing now by providing computa- tion capabilities within memory itself, thus eliminating the need for data communication between memory and the processor. However, precisely because of the absence of a clear division between memory and the computing unit this intrinsic logic-in-memory archi- tecture is completely different from the Von Neumann architecture currently employed, and a development of a conceptually new calculation paradigm using this architecture is needed. An alternative option is to follow a more conventional path with memory and computing units separated, where, however, both elements are non-volatile and implemented in a magnetic domain. Placing the actual computation into the magnetic domain reduces the need of converting magnetically stored information into the currents and voltages for processing. It also simplifies the circuit layout and boosts the integra- tion density. The idea of combining MTJs with a com- mon free layer enables the realization of an efficient nanooscillator [106] and a non-volatile magnetic flip- flop [107]. The computation unit is represented by the STT based non-volatile majority gate, with non-volatile magnetic flip-flops [107] used as memory registers. Be- cause all the data processing elements are performed in the magnetization domain, the flip-flops could be put at the legs of the majority gates thus removing the need of data transfer and interconnects between the processing unit and memory. By using the non-volatile computation unit the realization of a 1-bit full adder in magnetic domain is demonstrated [108]. Finally, we mention a completely different neural net- work based approach to calculations. Non-volatile MTJs fit for neural network realizations as they can be considered as a current-driven programmable resistor – memristor – and they significantly advance program- ming and storage functions. MTJ based neural net- works have been demonstrated featuring non-volatile synapses [109] for high-speed pattern recognition with about 70% reduction of gate count and 99% improve- ment in speed. Neuromorphic computing is becoming a reality, with the first self-learning chip revealed on September 25th, 2017 by Intel [110]. 5 Conclusions Spin transistors have been recently successfully dem- onstrated, however, an enhancement of the on-current ratio between the parallel and anti-parallel source/ drain magnetization configuration at room tempera- ture remains one of the main challenges. As both Spin- FET and SpinMOSFET still rely on the charge current to transfer the spin, it sets limitations for the applicability of such devices in main-stream microelectronics, and new ideas are needed for the future. Non-volatile devices based on MTJs possess a TMR suit- able for practical applications. Several companies an- nounced embedded STT MRAM production in 2018. Although STT MRAM is positioned as a successor not only for flash, but also for CMOS-based main computer memory, the relatively high switching current and pow- er may confine STT-MRAM to replacing flash memory in data-intensive and low-power mobile, automotive, or Internet of Things applications. Because of the large switching currents and insufficient speed, STT-MRAM is unlikely to replace SRAM in high-level core caches. Novel innovative non-volatile devices with improved switching characteristics and low power consumption are required for processor-embedded memories. Finally, the successful adoption of non-volatility in mi- croelectronic systems by developing various logic-in- memory architectures and in-memory processing will inevitably result in increasing disseminations of this technology for other applications such as ultra-low- power electronics, high-performance computing, the Internet of Things, and Big Data analysis. 6 Acknowledgments Fruitful discussions with Dr. T. Windbacher and the fi- nancial support by the Austrian Federal Ministry of Sci- ence, Research and Economy and the National Foun- dation for Research, Technology and Development are gratefully acknowledged. 7 References 1. S.-E. Thompson, M. Armstrong, C. Auth et al., A 90-nm Logic Technology Featuring Strained- Silicon, IEEE Trans. Electron Dev., vol. 51, pp. 1790– 1797, 2004. DOI: 10.1109/TED.2004. 836648 2. K. Mistry, C. Allen, C. Auth et al., A 45nm Logic Technology with High-k+Metal Gate Transistors, Strained Silicon, 9 Cu Interconnect Layers, 193nm Dry Patterning, and 100% Pb-Free Packaging, IEDM Techn. Digest, pp. 247–250, 2007. DOI: 10.1109/IEDM.2007.4418914 V. Sverdlov et al; Informacije Midem, Vol. 47, No. 4(2017), 195 – 210 206 3. S. Natarajan, M. Armstrong, M. Bost et al., A 32nm Logic Technology Featuring 2nd-Generation High- k + Metal-Gate Transistors, Enhanced Channel Strain and 0.171μm2 SRAM Cell Size in a 291Mb Array, IEDM Techn. Digest, pp. 941–943, 2008. DOI: 10.1109/IEDM.2009.5424253 4. R. Xie, P. Montanini, K. Akarvardar et al., A 7nm FinFET Technology Featuring EUV Pattern- ing and Dual Strained High Mobility Chan- nels, IEDM Techn. Digest, pp. 47–50, 2016. DOI: 10.1109/IEDM.2016.7838334 5. S.-Y. Wu, C. Y. Lin, M. C. Chiang et al., 7nm CMOS Platform Technology Featuring 4th Generation FinFET Transistors with a 0.027um2 High Density 6-T SRAM cell for Mobile SoC Applications, IEDM Techn. Digest, pp. 43–46, 2016. DOI: 10.1109/IEDM.2016.7838333 6. International Technology Roadmap for Semicon- ductors (2016). Available: http://www.itrs2.net/ 7. H. Ohno, M. Stiles, B. Dieny, Spintronics, Proc. of the IEEE, vol.104, pp. 1782–1786, 2016. DOI: 109/JPROC.2016.260116301163 8. S. Datta, B. Das, Electronic Analog of the Electro- Optic Modulator, Appl. Phys. Lett., vol. 56 (7), pp. 665–667, 1990. DOI: 10.1063/1. 102730 9. S. Sugahara, J. Nitta, Spin-Transistor Electronics: An Overview and Outlook, Proc. of the IEEE, vol. 98, pp. 2124–2154, 2010. DOI: 10.1109/JPROC.2010.2064272 10. I. Zutic, J. Fabian, S. Das Sarma, Spintronics: Fun- damentals and Applications, Rev. Mod. Phys., vol. 76, pp. 323–410, 2004. DOI: 10.1103/RevModPhys.76.323 11. J. Fabian, A. Matos-Abiaguea, C. Ertler, P. Stano, I. Zutic, Semiconductor Spintronics, Acta Phys. Slo- vaca, vol. 5, pp. 565–907, 2007. http://www.phys- ics.sk/aps/pubs/2007/aps-07-04/aps-07-04.pdf 12. P. Li, H. Dery, Spin-Orbit Symmetries of Conduc- tion Electrons in Silicon, Phys. Rev. Lett. vol. 107, 107203, 2011. DOI: 10.1103/PhysRevLett.107.107203 13. O. Chalaev, Y. Song, H. Dery, Suppressing the Spin Relaxation of Electrons in Silicon, Phys. Rev. B, vol. 95, 035204, 2017. DOI: 10.1103/PhysRevB.95.035204 14. V. Sverdlov, S. Selberherr, Silicon Spintronics: Progress and Challenges, Phys. Rep., vol. 585, pp. 1–40, 2015. DOI:10.1016/j.physrep.2015.05.002. 15. V. Sverdlov, Strain-induced Effects in Advanced MOSFETs, Springer, 2011. 16. V. Sverdlov, J. Ghosh, S. Selberherr, Universal De- pendence of the Spin Lifetime in Silicon Films on the Spin Injection Direction, in Abstracts Work- shop on Innovative Devices and Systems (WINDS), p. 7, 2016. 17. Y. Bychkov, E. Rashba, Properties of a 2D Electron Gas with Lifted Spectral Degeneracy, JETP Lett. vol. 39, pp. 78–81, 1984. Available: http://www. jetpletters.ac.ru/ps/1264/article_19121.shtml 18. P. Chuang, S.-C. Ho, L.W. Smith et al., All-electric All-semiconductor Spin Field-effect transistors, Nature Nanotechnol., vol. 10, pp. 35–39, 2015. DOI: 10.1038/nnano.2014.296 19. W. Yan, O. Txoperena, R. Llopis et al., A Two-dimen- sional Spin Field-effect Switch, Nature Communi- cations, vol. 7, 13372, 2016. DOI: 10.1038/ncomms13372 20. T. Tahara, H. Koike, M. Kameno, et al., Room-tem- perature Operation of Si Spin MOSFET with High on/off Spin Signal Ratio, Appl. Phys. Express, vol. 8, 11304 2015. DOI: 10.7567/APEX.8.113004 21. E. I. Rashba, Theory of Electrical Spin Injection: Tunnel Contacts as a Solution of the Conductiv- ity Mismatch Problem, Phys. Rev. B, vol. 62, pp. R16267–R16270, 2000. DOI: 10.1103/PhysRevB.62.R16267 22. T. Tahara, Y. Ando, M. Kameno et al., Observation of Large Spin Accumulation Voltages in Non- degenerate Si Spin Devices due to Spin Drift Ef- fect: Experiments and Theory, Phys. Rev. B, vol. 93, 214406, 2016. DOI: 10.1103/PhysRevB.93.214406 23. R. Jansen, Silicon Spintronics, Nature Materials, vol. 11, pp. 400–408, 2012. DOI: 10.1038/nmat3293 24. Y. Song and H. Dery, Magnetic-Field-Modulated Reso- nant Tunneling in Ferromagnetic-Insulator-Nonmag- netic Junctions, Phys. Rev. Let. vol. 113, 047205, 2014. DOI: 10.1103/PhysRevLett.113.047205 25. Z. Yue, M. C. Prestgard, A. Tiwari, M. E. Raikh, Resonant Magnetotunneling between Normal and Ferromagnetic Electrodes in Relation to the Three-terminal Spin Transport, Phys. Rev. B, vol. 91, 195316 2015. DOI: 10.1103/PhysRevB.91.195316 26. V. Sverdlov, J. Weinbub, S. Selberherr, Spin-De- pendent Trap-Assisted Tunneling in Magnetic Tunnel Junctions: A Monte Carlo Study, in Ab- stract Book Intl. Workshop on Computational Na- notechnology (IWCN), pp. 88 – 90, 2017. 27. A. Fert, Nobel Lecture: Origin, Development, and Future of Spintronics, Rev. Modern Phys., vol. 80, pp. 1517–1530, 2008. DOI: 10.1103/RevModPhys.80.1517 28. P. A. Grunberg, Nobel Lecture: From Spin Waves to Giant Magnetoresistance and Beyond, Rev. Mod- ern Phys., vol. 80, pp. 1531–1540, 2008. DOI: 10.1103/RevModPhys.80.1531 29. S. Ikeda, J. Hayakawa, Y. Ashizawa et al., Tunnel Magnetoresistance of 604% at 300 K by Suppres- sion of Ta Diffusion in CoFeB/MgO/CoFeB Pseudo- spin-valves Annealed at High Temperature, Appl. Phys. Lett., vol. 93, 082508, 2008. DOI: 10.1063/1.2976435 V. Sverdlov et al; Informacije Midem, Vol. 47, No. 4(2017), 195 – 210 207 30. D. Apalkov, B. Dieny, and J. M. Slaughter, Magne- toresistive Random Access Memory, Proc. of the IEEE, vol. 104, pp. 1796–1830, 2016. DOI: 10.1109/JPROC.2016.2590142 31. J. Slonczewski, Current-driven Excitation of Mag- netic Multilayers, J. Magn. Mater., vol. 159, pp. L1– L7, 1996. DOI: 10.1016/0304-8853(96)00062-5 32. L. Berger, Emission of Spin Waves by a Magnetic Multilayer Traversed by a Current, Phys. Rev. B, vol. 54, pp. 9353–9358, 1996. DOI: 10.1103/PhysRevB.54.9353 33. Electronicdesign.com, 4-Mbit Device is First Com- mercially Available MRAM, Electronic Design, 2006. Available: http://electronicdesign.com/ dsps/4-mbit-device-first-commerciallyavailable- mram 34. L. Savtchenko, B. Engel, N. Rizzo, M. Deherrera, J. Janesky, Method of Writing to Scalable Magne- toresistance Random Access Memory Element, US Patent 6545906 B1, 2003. https://www.goog- le.com/patents/US6545906 35. R. Sbiaa, H. Meng, and S. N. Piramanayagam, Ma- terials with Perpendicular Magnetic Anisotropy for Magnetic Random Access Memory, Phys. Stat. Solidi (RRL) – Rapid Research Letters, vol. 5, pp. 413–419, 2011. DOI: 10.1002/pssr.201105420 36. Y. Huai, F. Albert, P. Nguyen et al., Observation of Spin-transfer Switching in Deep Submicron-sized and Low-resistance Magnetic Tunnel Junctions, Appl. Phys. Lett., vol. 84, pp. 3118–3120, 2004. DOI: 10.1063/1.1707228 37. Z. Diao, D. Apalkov, M. Pakala et al., Spin Transfer Switching and Spin Polarization in Magnetic Tun- nel Junctions with MgO and AlOx Barriers, Appl. Phys. Lett., vol. 87, 232502, 2005. DOI: 10.1063/1.2139849 38. N. D. Rizzo, D. Houssameddine, J. Janesky et al., A Fully Functional 64 Mb DDR3 ST-MRAM Built on 90 nm CMOS Technology, IEEE Trans. Magn., vol. 49, pp. 4441–4446, 2013. DOI: 10.1109/TMAG.2013.2243133 39. S.-W. Chung, T. Kishi, J. W. Park et al., 4Gbit Density STT-MRAM Using Perpendicular MTJ Realized with Compact Cell Structure, IEDM Techn. Digest, pp. 659–662, 2016. DOI: 10.1109/IEDM.2016.7838490 40. B. Dieny, R. Sousa, S. Bandiera et al., Extended Scalability and Functionalities of MRAM Based on Thermally Assisted Writing, IEDM Techn. Digest, pp. 1.3.1–1.3.4, 2011. DOI: 10.1109/IEDM.2011.6131471 41. Y. J. Song, J. H. Lee, H. C. Shin et al., Highly Func- tional and Reliable 8Mb STT-MRAM Embedded in 28nm Logic, IEDM Techn. Digest, pp. 663–666, 2016. DOI: 10.1109/IEDM.2016.7838491 42. H. Sato, M. Yamanouchi, S. Ikeda et al., MgO/ CoFeB/Ta/CoFeB/MgO Recording Structure in Magnetic Tunnel Junctions with Perpendicular Easy Axis, IEEE Trans. Magn., vol. 49, pp. 4437– 4440, 2013. DOI: 10.1109/TMAG.2013.2251326 43. A. Makarov, V. Sverdlov, D. Osintsev, S. Selberherr, Reduction of Switching Time in Pentalayer Mag- netic Tunnel Junctions with a Composite-Free Layer, Phys. Stat. Solidi (RRL) – Rapid Research Let- ters, vol. 5, pp. 420–422, 2011. DOI: 10.1002/pssr.201105376 44. A. Makarov, T. Windbacher, V. Sverdlov, S. Selber- herr, CMOS-Compatible Spintronic Devices: A Review, Semicond. Sci. and Tech., vol. 31, 113006, 2016. DOI: 10.1088/0268-1242/31/11/113006 45. M. Krounbi, V. Nikitin, D. Apalkov et al., Status and Challenges in Spin-Transfer Torque MRAM Tech- nology, in Proc. of ECS Meeting, 2015. Available: http://ecst.ecsdl.org/content/69/3/119.abstract 46. M. T. Chang, P. Rosenfeld, S. L. Lu B. Jacob, Tech- nology Comparison for Large Last level Caches (L3 Cs): Low leakage SRAM, Low Write-energy STT-RAM, and Refresh-optimized eDRAM, in Proc. 19th Intl. Symp. High Performance Computer Architecture (HPCA), pp. 143–154, 2013. DOI: 10.1109/HPCA.2013.6522314 47. T. Endoh, T. Ohsawa, H. Koike et al., Restructuring of Memory Hierarchy in Computing System with Spintronics-based Technologies, in Symp. VLSI Technol., pp. 89–90, 2012. DOI: 10.1109/VLSIT.2012.6242475 48. T. Endoh, Nonvolatile Logic and Memory Devices Based on Spintronics, in Proc. of IEEE Intl. Symp. Circ. Syst., pp. 13–16, 2015. DOI: 10.1109/ISCAS.2015.7168558 49. H. Noguchi, K. Ikegami, N. Shimomura et al., High- ly Reliable and Low-power Nonvolatile Cache Memory with Advanced Perpendicular STT- MRAM for High-performance CPU, in Symp. VLSI Circuits, pp. 97–98, 2014. DOI: 10.1109/VLSIC.2014.6858403 50. G. Jan, L. Thomas, S. Le et al., Achieving Sub‐ns Switching of STT‐MRAM for Future Embedded LLC Applications through Improvement of Nu- cleation and Propagation Switching Mechanisms, in Symp. VLSI Technol., pp. 18–19, 2016. DOI: 10.1109/VLSIT.2016.7573362 51 H. Noguchi, K. Ikegami, S. Takaya et al., 7.2 4Mb STT-MRAM-based Cache with Memory-access- aware Power Optimization and Write-verify-write / Read-modify-write scheme, in ISSCC Dig. Tech. Papers, pp. 132–133, 2016. DOI: 10.1109/ISSCC.2016.7417942 52. https://www.mram-info.com/tags/companies/ samsung V. Sverdlov et al; Informacije Midem, Vol. 47, No. 4(2017), 195 – 210 208 53. https://www.globalfoundries.com/news-events/ press-releases/globalfoundries-launches-embed- ded-mram-22fdxr-platform 54. L. Thomas, G. Jan, J. Zhu et al., Perpendicular Spin Transfer Torque Magnetic Random Access Memo- ries with High Spin Torque Efficiency and Thermal Stability for Embedded Applications, J. Appl. Phys., vol. 115, 172615, 2014. DOI: 10.1063/1.4870917 55. T. Nozaki, Y. Shiota, M. Shiraishi et al., Voltage- induced Perpendicular Magnetic Anisotropy Change in Magnetic Tunnel Junctions, Appl. Phys. Lett., vol. 96, 022506, 2010. DOI: 10.1063/1.3279157 56. Y. Shiota, T. Nozaki, F. Bonell et al., Induction of Co- herent Magnetization Switching in a Few Atomic Layers of FeCo Using Voltage Pulses, Nature Mate- rials, vol. 11, pp. 39-43, 2012. DOI: 10.1038/nmat3172 57. Y. Shiota, T. Nozaki, S. Tamaru et al., Evaluation of Write Error Rate for Voltage-driven Dynamic Mag- netization Switching in Magnetic Tunnel Junc- tions with Perpendicular Magnetization, Appl. Phys. Express, vol. 9, 013001, 2015. DOI: 10.7567/APEX.9.013001 58. K. L. Wang, X. Kou, P. Upadhyaya et al., Electric-Field Control of Spin-Orbit Interaction for Low-Power Spintronics, Proc. of the IEEE, vol. 104, pp. 1974- 2008, 2016. DOI: 10.1109/JPROC.2016.2573836 59. M. K. Niranjan, CG. Duan, S. S. Jaswal, E. Tsymbal, Electric Field Eeffect on Magnetization at the Fe/MgO(001) Interface, Appl. Phys. Lett., vol. 96, 222504, 2010. DOI: 10.1063/1.3443658 60. H. Noguchi, K. Ikegami, K. Abe et al., Novel Voltage Controlled MRAM (VCM) with Fast Read/Write Circuits for Ultra Large Last Level Cache, in IEDM Techn. Digest, pp.675–678, 2016. DOI: 10.1109/IEDM.2016.7838494 61. H. Yoda, N. Shimomura, Y. Ohsawa et al., Voltage- Control Spintronics Memory (VoCSM) Having Po- tentials of Ultra-Low Energy-Consumption and High-Density, in IEDM Techn. Digest, pp.679–682, 2016. DOI: 10.1109/IEDM.2016.7838495 62. I. M. Miron, G. Gaudin, S. Auffret et al., Current- driven Spin Torque Induced by the Rashba Effect in a Ferromagnetic Metal Layer, Nature Materials, vol. 9, pp. 230–234, 2010. DOI: 10.1038/nmat2613 63. I. M. Miron, K. Garello, G. Gaudin et al., Perpen- dicular Switching of a Single Ferromagnetic Layer Induced by In-plane Current Injection, Nature, vol. 476, pp. 189–193, 2011. DOI: 10.1038/nature10309 64. L. Liu, O. J. Lee, T. J. Gudmundsen et al., Current-in- duced Switching of Perpendicularly Magnetized Magnetic Layers Using Spin Torque from the Spin Hall Effect, Phys. Rev. Lett., vol. 109, 096602, 2012. DOI: 10.1103/PhysRevLett.109.096602 65. A. Brataas and K. M. D. Hals, Spin-orbit Torques in Action,” Nature Nanotechnol., vol. 9, pp. 86–88, 2014. DOI: 10.1038/nnano.2014.8 66. M. Cubukcu, O. Boulle, M. Drouard et al., Spin- orbit Torque Magnetization Switching of a Three- terminal Perpendicular Magnetic Tunnel Junc- tion, Appl. Phys. Lett., vol. 104, 042406, 2014. DOI: 10.1063/1.4863407 67. Y. Fan, P. Upadhyaya, X. Kou et al., Magnetization Switching through Giant Spin–orbit Torque in a Magnetically Doped Topological Insulator Heter- ostructure, Nature Materials, vol.13, pp. 699–704, 2014. DOI: 10.1038/nmat3973 68. C. H. Li, O. M. J. Van‘t Erve, J. T. Robinson et al., Elec- trical Detection of Charge-current-induced Spin Polarization due to Spin-momentum Locking in Bi2Se3, Nature Nanotechnol., vol. 9, pp. 218–224, 2014. DOI: 10.1038/nnano.2014.16 69. C. H. Li, O. M. J. Van‘t Erve, S. Rajput et al., Direct Comparison of Current-induced Spin Polarization in Topological Insulator Bi2Se3 and InAs Rashba States, Nature Communications, vol. 7, 13518, 2016. DOI: 10.1038/ncomms13518 70. S.-W. Lee, K.-J. Lee, Emerging Three-Terminal Mag- netic Memory Devices, Proc. of the IEEE, vol. 104, pp. 1831–1843, 2016. DOI: 10.1109/JPROC.2016.2543782 71. L. Thomas, K.-S.Ryu, S.-H. Yang, and S. S. P. Parkin, Chiral Spin Torque at Magnetic Domain Walls, Na- ture Nanotechnol., vol. 8, pp. 527–533, 2013. DOI: 10.1038/nnano.2013.102 72. S. Emori, U. Bauer, S.-M. Ahn et al., Current-driven Dynamics of Chiral Ferromagnetic Domain Walls, Nature Materials, vol. 12, pp. 611–616, 2013. DOI: 10.1038/nmat3675 73. R. G. Elías, N. Vidal-Silva, and A. Manchon, Steady Motion of Skyrmions and Domains Walls un- der Diffusive Spin Torques, Phys. Rev. B, vol. 95, 104406, 2017. DOI: 10.1103/PhysRevB.95.104406 74. E. van der Bijl and R. A. Duine, Current-induced Torques in Textured Rashba Ferromagnets, Phys. Rev. B, vol. 86, 094406, 2012. DOI: 10.1103/PhysRevB.86.094406 75. V. P. Amin and M. D. Stiles, Spin Transport at Inter- faces with Spin-orbit Coupling: Formalism, Phys. Rev. B, vol. 94, 104419, 2016. DOI: 10.1103/PhysRevB.94.104419 76. V. P. Amin and M. D. Stiles, Spin Transport at Inter- faces with Spin-orbit Coupling: Phenomenology, Phys. Rev. B, vol. 94, 104420, 2016. DOI: 10.1103/PhysRevB.94.104420 77. J.-G. Zhu, Magnetoresistive Random Access Mem- ory: The Path to Competitiveness and Scalability, Proc. of the IEEE, vol. 96, pp. 1786–1798, 2008. DOI: 10.1109/JPROC.2008.2004313 V. Sverdlov et al; Informacije Midem, Vol. 47, No. 4(2017), 195 – 210 209 78. T. Hany, T. Endoh, D. Suzuki et al., Standby-Pow- er-Free Integrated Circuits Using MTJ-Based VLSI Computing, Proc. of the IEEE, vol. 104, pp. 1844- 1863, 2016. DOI: 10.1109/JPROC.2016.2574939 79. T. Endoh, S. Togashi, F. Iga et al., A 600 MHz MTJ- based Nonvolatile Latch Making Use of Incubation Time in MTJ Switching, IEDM Techn. Digest, pp.75– 78, 2011. DOI: 10.1109/IEDM.2011.6131487 80. N. Sakimura, Y. Tsuji, R. Nebashi et al., A 90 nm 20 MHz Fully Nonvolatile Microcontroller for Stand- by-Power-Critical Applications, in Proc. of IEEE Intl. Solid-State Circuits Conf., p. p. 184–185, 2014. DOI: 10.1109/ISSCC.2014.6757392 81. H. Koike, T. Ohsawa, S. Ikeda et al., A Power-gated MPU with 3-microsecond Entry/Exit Delay Using MTJ-based Nonvolatile Flip-Flop, in Proc. of IEEE A- SSCC., pp. 317–320, 2013. DOI: 10.1109/ASSCC.2013.6691046 82. S. C. Bartling, S. Khanna, M. P. Clinto et al., An 8 MHz 75 μA/MHz Zero-leakage Non-volatile Log- ic-based Cortex-M0 MCU SoC Exhibiting 100% Digital State Retention at VDD = 0 V with < 400 ns Wakeup and Sleep Transition, in Proc. IEEE Intl. Solid-State Circuits Conf., pp. 432–433, 2013. DOI: 10.1109/ISSCC.2013.6487802 83. Y. Guillemenet, L. Torres, G. Sassatelli et al., A Non- volatile Run-time FPGA Using Thermally Assisted Switching MRAMs, in Proc. Intl. Conf. Field-Pro- grammable Logic, pp. 421–426, 2008. DOI: 10.1109/FPL.2008.4629974 84. Y. Y. Liauw, Z. Zhang, W. Kim et al., Nonvolatile 3D- FPGA with Monolithically Stacked RRAM-based Configuration Memory, in Proc. IEEE Intl. Solid- State Circuits Conf., pp. 406–408, 2012. DOI: 10.1109/ISSCC.2012.6177067 85. Z. Zhang, Y. Y. Liauw, C. Chen, S. S. Wong, Monolithic 3-D FPGAs, Proc. of the IEEE, vol. 103, pp. 1197–1210, 2015. DOI: 10.1109/JPROC.2015.24339543954 86. D. Suzuki, M. Natsui, S. Ikeda et al., Fabrication of a Nonvolatile Lookup-table Circuit Chip Us- ing Magneto/Semiconductor-hybrid Structure, for an Immediate-power-up Field Programma- ble Gate Array, in Symp. VLSI Circuits, pp. 80–81, 2009. Available: http://ieeexplore.ieee.org/docu- ment/5205282/ 87. D. Suzuki, M. Natsui, T. Endoh et al., Six-input Lookup Table Circuit with 62% Fewer Transistors Using Nonvolatile Logic-in-memory Architecture with Series/Parallel-connected Magnetic Tunnel Junctions, J. Appl. Phys., vol. 111, 07E318, 2012. DOI: 10.1063/1.3672411 88. S. Yamamoto, Y. Shuto, and S. Sugahara, Nonvola- tile Power-gating Field-programmable Gate Array Using Nonvolatile Static Random Access Memory and Nonvolatile Flip-flops Based on Pseudo-spin- transistor Architecture with Spin-transfer-torque Magnetic Tunnel Junctions, Jpn. J. Appl. Phys., vol. 51, pp. 11PB021–11PB025, 2012. DOI: 10.1143/JJAP.51.11PB02 89. D. Suzuki, M. Natsui, A. Mochizuki et al., Fabrica- tion of a 3000-6-input-LUTs Embedded and Block- level Power-gated Nonvolatile FPGA Chip Using p-MTJ-based Logic-in-memory Structure, in Symp. VLSI Circuits, pp. 172–173, 2015. DOI: 10.1109/VLSIT.2015.7223644 90. K. Pagiamtzis and A. Sheikholeslami, Content- addressable Memory (CAM) Circuits and Archi- tectures: A Tutorial and Survey, IEEE J. Solid-State Circuits, vol. 41, pp. 712–727, 2006. DOI: 10.1109/JSSC.2005.864128 91. S. Matsunaga, K. Hiyama, A. Matsumoto et al., Standby-power-free Compact Ternary Content- addressable Memory Cell Chip Using Magnetic Tunnel Junction Devices, Appl. Phys. Express, vol. 2, pp. 0230041–0230043, 2009. DOI: 10.1143/APEX.2.023004 92. S. Matsunaga, M. Natsui, S. Ikeda et al., Implemen- tation of a Perpendicular MTJ-based Read-disturb- Tolerant 2T-2R Nonvolatile TCAM Based on a Re- versed Current Reading Scheme, in Proc. Asia South Pacific Design Autom. Conf., pp. 475–476, 2012. DOI: 10.1109/ASPDAC.2012.6164998 93. S. Matsunaga, A. Katsumata, M. Natsui et al., Fab- rication of a 99%-energy-less Nonvolatile Multi- functional CAM Chip Using Hierarchical Power Gating for a Massively-parallel Full-text-search Engine, in Symp. VLSI Circuits, pp. 106–107, 2013. Available: http://ieeexplore.ieee.org/document/ 6576611/ 94. S. Matsunaga, A. Katsumata, M. Natsui et al., De- sign of a Nine Transistor/Two-magnetic-tunnel Junction Cell- based Low-energy Nonvolatile Ter- nary Content-addressable Memory, Jpn. J. Appl. Phys., vol. 51, pp. 02BM061–02BM066, 2012. DOI: 10.1143/JJAP.51.02BM06 95. S. Matsunaga, A. Katsumata, M. Natsui et al., De- sign of a 270 ps-Access 7T-2MTJ-cell Nonvolatile Ternary Content-addressable Memory, J. Appl. Phys., vol. 111, pp. 07E3361–07E3363, 2012. DOI: 10.1063/1.3677875 96. S. Matsunaga, S. Miura, H. Honjou et al., A 3.14 µm2 4T-2MTJ-cell Fully Parallel TCAM Based on Non- volatile Logic-in-memory Architecture, in Symp. VLSI Circuits, pp. 44–45, 2012. DOI: 10.1109/VLSIC.2012.6243781 V. Sverdlov et al; Informacije Midem, Vol. 47, No. 4(2017), 195 – 210 210 97. S. Matsunaga, A. Mochizuki, N. Sakimura et al., Complementary 5T-4MTJ Nonvolatile TCAM Cell Circuit with Phase-selective Parallel Writ- ing Scheme, IEICE Electron. Express, vol. 11, pp. 201402971–201402977, 2014. DOI: 10.1587/elex.11.20140297 98. S. Matsunaga, J. Hayakawa, S. Ikeda et al., Fabrica- tion of a Nonvolatile Full Adder Based on Logic- in-memory Architecture Using Magnetic Tunnel Junctions, Appl. Phys. Express, vol. 1, pp. 0913011– 0913013, 2008. DOI: 10.1143/APEX.1.091301 99. H.-P. Trinh, W. Zhao, J.-O. Klein et al., Magnetic Ad- der Based on Racetrack Memory, IEEE Trans. Cir- cuits Systems, vol. 60, pp. 1469–1477, 2013. DOI: 10.1109/TCSI.2012.2220507 100. E. Deng, Y. Zhang, W. Kang et al., Synchronous 8-bit Non-volatile Full-adder Based on Spin Trans- fer Torque Magnetic Tunnel Junction, IEEE Trans. Circuits Systems, vol. 62, pp. 1757–1765, 2015. DOI: 10.1109/TCSI.2015.2423751 101. M. Natsui, D. Suzuki, N. Sakimura et al., Nonvola- tile Logic-in-memory Array processor in 90 nm MTJ/MOS Achieving 75% Leakage Reduction Us- ing Cycle-based Power Gating, in Proc. of IEEE Intl. Solid-State Circuits Conf., pp. 194–195, 2013. DOI: 10.1109/ISSCC.2013.6487696 102. M. Natsui, D. Suzuki, N. Sakimura et al., Nonvola- tile Logic-in-memory LSI Using Cycle-based Pow- er Gating and its Application to Motion-vector Prediction, IEEE J. Solid-State Circuits, vol. 50, pp. 476–489, 2015. DOI: 10.1109/JSSC.2014.2362853 103. Y. Ma, T. Shibata, T.Endoh, An MTJ-Based Nonvola- tile Associative Memory Architecture with Intelli- gent Power-saving Scheme for High-speed Low- power Recognition Applications, in Proc. of IEEE Intl. Symp. Circ. Syst., pp. 1248–1251, 2013. DOI: 10.1109/ISCAS.2013.6572079 104. H. Mahmoudi, T. Windbacher, V. Sverdlov, S. Selberherr, RRAM Implication Logic Gates, Pat- ent: Intl., No. Wo 2014/079747 A1; EP 12193826.0. http://www.google.ch/patents/EP2736044A1 105. H. Mahmoudi, V. Sverdlov, S. Selberherr, MTJ- based Implication Logic Gates and Circuit Archi- tecture for Large-Scale Spintronic Stateful Logic Systems, in Proc. of ESSDERC, pp. 254 – 257, 2012. DOI: 10.1109/ESSDERC.2012.6343381 106. A. Makarov, V. Sverdlov, S. Selberherr, Geometry Optimization of Spin-Torque Oscillators Com- posed of Two MgO-MTJs with a Shared Free Layer, in Proc. of the Intl. Conf. on Nanoscale Magnetism (ICNM), p.69, 2013. Available: www.iue.tuwien. ac.at/pdf/ib_2013/CP2013_Makarov_2.pdf 107. T. Windbacher, H. Mahmoudi, V. Sverdlov, S.  Selberherr, Spin Torque Magnetic Integrated Circuit, Patent: Intl., No. Wo 2014/154497 A1; Pat- ent priority number EP 13161375.4. https://www. google.com/patents/EP2784020B1 108. T. Windbacher, A. Makarov, V. Sverdlov, S. Selber- herr, A Universal Nonvolatile Processing Envi- ronment, in “Future Trends in Microelectron- ics - Journey into the Unknown”, S.Luryi, J.Xu, A.Zaslavsky (ed); J.Wiley&Sons, pp. 83–91, 2016, DOI: 10.1002/9781119069225.ch1-6 109. Y. Ma and T. Endoh, A Novel Neuron Circuit with Nonvolatile Synapses Based on Magnetic-tunnel- junction for High-speed Pattern Learning and Recognition, in Proc. Asia-Pacific Workshop Fun- dam. Appl. Adv. Semicond. Devices (AWAD), vol. 4B- 1, pp. 273–278, 2015. 110. https://newsroom.intel.com/editorials/intels- new-self-learning-chip-promises-accelerate-arti- ficial-intelligence/ Arrived: 31. 08. 2017 Accepted: 20. 12. 2017 V. Sverdlov et al; Informacije Midem, Vol. 47, No. 4(2017), 195 – 210