Perpendicular STT-MRAM: Fundamentals and Advances

• p-STT-MRAM is a magnetic memory type that utilizes perpendicular MTJs with out-of-plane magnetization and spin-transfer torque for efficient data switching.
• PSA variants leverage thick storage layers where shape anisotropy complements interfacial perpendicular magnetic anisotropy to maintain high thermal stability at nanoscale dimensions.
• Advanced reversal dynamics, including coherent rotation, domain-wall propagation, and assisted-writing methods, optimize write speeds, reduce error rates, and enable robust low-temperature operation.

Perpendicular spin-transfer-torque magnetic random-access memory (p-STT-MRAM) exploits a magnetic tunnel junction (MTJ) whose free layer is magnetized out-of-plane. In the canonical two-terminal realization, a spin-polarized current from a reference layer traverses an MgO barrier and exerts a Slonczewski torque on the free layer; in perpendicular-shape-anisotropy (PSA) variants, the free layer is made sufficiently thick that shape anisotropy adds a positive contribution to the effective perpendicular anisotropy. Across recent work, p-STT-MRAM encompasses ultrathin CoFeB/MgO nanopillars, thick PSA storage layers, thermoelectric magnonic writing, cryogenic operation, RF-assisted switching, and detailed micromagnetic treatment of synthetic-antiferromagnet (SAF) alignment, stray fields, and nonuniform reversal (Zhang et al., 2021, Perrissin et al., 2018).

1. Governing physics and canonical device model

In a two-terminal perpendicular MTJ, the free-layer magnetization unit vector $\mathbf m(t)$ is commonly described by the Landau–Lifshitz–Gilbert equation with a spin-torque term,

$$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$

with

$$\tau_{\rm STT} = \frac{\hbar}{2e}\frac{I}{M_s V P}\eta(\theta)\,\mathbf m\times(\mathbf m_p\times\mathbf m),$$

where $\gamma$ is the gyromagnetic ratio, $\alpha$ the Gilbert damping, $V$ the free-layer volume, $P$ the spin-polarization, $\eta(\theta)$ a geometry-dependent efficiency, and $\mathbf m_p$ the fixed-layer magnetization direction (Rehm et al., 2019). In equivalent current-density notation used for perpendicular geometry,

$$\frac{d\mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d\mathbf m}{dt} + \gamma a_J\,\mathbf m\times(\mathbf m\times\mathbf p),$$

with $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$0 along $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$1 in the standard perpendicular configuration (Song et al., 2015).

Two scalar figures organize most p-STT-MRAM analysis. The first is the thermal stability factor,

$$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$2

with $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$3 the barrier between the two stable perpendicular states. The second is the macrospin critical current density,

$$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$4

which shows the dependence on damping, free-layer moment, thickness, spin-polarization efficiency, and effective anisotropy field (Mihajlovic et al., 2019). Long-pulse threshold voltages and fast-switching characteristic times are often summarized by $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$5 and $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$6, with the ballistic-regime relation $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$7 used to fit switching-probability phase diagrams (Rehm et al., 2019).

The macrospin description remains useful but is not exhaustive. In PSA-STT-MRAM, three-dimensional geometry, spatially nonuniform spin accumulation, and edge-localized modes become central. Self-consistent micromagnetic and transport simulations for a 20 nm diameter, 20 nm thick pillar found “flower” states at the upper and lower surfaces and a field-like torque comparable to the damping-like torque, indicating that standard thin-film intuition is incomplete for thick perpendicular pillars (Meneguolo et al., 19 Mar 2025).

2. Stack architectures and anisotropy engineering

The reference pMTJ stack used for cryogenic switching studies contains a SAF and a CoFeB-based perpendicular MTJ: SAF2 is $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$8, SAF1 is $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$9, the reference layer is CoFeB(0.9 nm), the tunnel barrier is MgO(1 nm), and the free layer is a CoFeB(1.5 nm)/W(0.3 nm)/CoFeB(0.8 nm) composite, with nominal diameters of 40 nm, 50 nm, and 60 nm (Rehm et al., 2019). In these devices, perpendicular anisotropy arises from both Co/Pt multilayers in the SAF and interfacial anisotropy at the CoFeB/MgO interfaces.

PSA-STT-MRAM changes the free-layer geometry. Instead of an ultrathin storage layer, the storage layer thickness is made comparable to or larger than the pillar diameter so that shape anisotropy reinforces rather than opposes perpendicularity. In the formulation used for PSA modeling,

$$\tau_{\rm STT} = \frac{\hbar}{2e}\frac{I}{M_s V P}\eta(\theta)\,\mathbf m\times(\mathbf m_p\times\mathbf m),$$0

with

$$\tau_{\rm STT} = \frac{\hbar}{2e}\frac{I}{M_s V P}\eta(\theta)\,\mathbf m\times(\mathbf m_p\times\mathbf m),$$1

For $$\tau_{\rm STT} = \frac{\hbar}{2e}\frac{I}{M_s V P}\eta(\theta)\,\mathbf m\times(\mathbf m_p\times\mathbf m),$$2, $$\tau_{\rm STT} = \frac{\hbar}{2e}\frac{I}{M_s V P}\eta(\theta)\,\mathbf m\times(\mathbf m_p\times\mathbf m),$$3 becomes negative so that the shape term favors out-of-plane magnetization (Zhang et al., 2021). The closely related expression

$$\tau_{\rm STT} = \frac{\hbar}{2e}\frac{I}{M_s V P}\eta(\theta)\,\mathbf m\times(\mathbf m_p\times\mathbf m),$$4

is used in device-scaling analyses of cylindrical PSA pillars (Perrissin et al., 2018).

Representative architectures reported across the literature are summarized below (Rehm et al., 2019, Perrissin et al., 2018, Mojumder et al., 2011, Caçoilo et al., 2023).

Architecture	Defining feature	Representative reported values
Conventional pMTJ	SAF/RL/MgO/CoFeB-based free layer	40–60 nm diameters; MgO 1 nm; $$\tau_{\rm STT} = \frac{\hbar}{2e}\frac{I}{M_s V P}\eta(\theta)\,\mathbf m\times(\mathbf m_p\times\mathbf m),$$5 at 300 K
PSA-STT-MRAM	Thick storage layer with $$\tau_{\rm STT} = \frac{\hbar}{2e}\frac{I}{M_s V P}\eta(\theta)\,\mathbf m\times(\mathbf m_p\times\mathbf m),$$6	$$\tau_{\rm STT} = \frac{\hbar}{2e}\frac{I}{M_s V P}\eta(\theta)\,\mathbf m\times(\mathbf m_p\times\mathbf m),$$7 at 8 nm diameter; $$\tau_{\rm STT} = \frac{\hbar}{2e}\frac{I}{M_s V P}\eta(\theta)\,\mathbf m\times(\mathbf m_p\times\mathbf m),$$8 down to 4 nm diameter
Thermoelectric p-STT-MRAM	Four-terminal ferrite/metal/MTJ/Peltier stack	$$\tau_{\rm STT} = \frac{\hbar}{2e}\frac{I}{M_s V P}\eta(\theta)\,\mathbf m\times(\mathbf m_p\times\mathbf m),$$9–$\gamma$0 ns; write energy $\gamma$1–$\gamma$2 fJ
Core-shell PSA-MTJ	Dipolar-coupled core and shell storage layer	$\gamma$3 at $\gamma$4 nm; $\gamma$5 mT at 20 nm pitch

The PSA concept was introduced precisely to extend downsize scalability. Analytical and micromagnetic work predicts that, once the storage-layer thickness exceeds a few nanometers, the PSA and interfacial PMA contributions add constructively, allowing $\gamma$6 even at very small diameters; for FeCoB/MgO-based PSA devices, $\gamma$7 was reported for MTJs as small as 8 nm in diameter, with the possibility to maintain $\gamma$8 down to 4 nm diameter (Perrissin et al., 2018).

Direct magnetic imaging has confirmed the physical reality of PSA in tall nanopillars. Off-axis electron holography on FeCoB/NiFe pillars with a NiFe storage layer 60 nm high and $\gamma$9 nm in diameter showed a clear dipole along the cylinder’s long axis, no vortex or multi-domain pattern, and maintenance of PSA up to at least $\alpha$0; the reconstructed induction yielded $\alpha$1 T at $\alpha$2 (Almeida et al., 2022).

A further extension is the dipolar-coupled core-shell PSA-MTJ. There, a cylindrical FeCoB core is surrounded by a coaxial Co or Ni shell, and antiparallel core-shell coupling raises the barrier while reducing array stray fields. For $\alpha$3 nm, $\alpha$4 nm, $\alpha$5 nm, and $\alpha$6 nm, the macrospin barrier estimate gives $\alpha$7 versus $\alpha$8 for the isolated core (Caçoilo et al., 2023).

3. Reversal dynamics and switching modes

The nominal reversal pathway in a perpendicular STT cell is not universal. In an undamaged circular cell, simulations show a sequence of coherent in-phase precession, then domain-wall nucleation at the rim, then domain-wall propagation toward the center (Song et al., 2015). This picture is already more complex than a strict macrospin, and it becomes still less uniform under structural modification or increased aspect ratio.

Edge-damaged cells illustrate how strongly the reversal pathway can change. When the anisotropy at a 5 nm peripheral rim is reduced, the rim magnetization tilts off axis at equilibrium, the local spin torque becomes immediately large at the edge, and reversal proceeds by rapid edge switching without a well-formed domain wall. The edge then acts as an “exchange-mediated seed” that accelerates reversal of the center (Song et al., 2015). This altered mode lowers the current much more strongly than it lowers the barrier.

PSA pillars introduce a separate aspect-ratio-controlled transition. For 20 nm lateral size, micromagnetic simulations found macrospin-like coherent rotation for aspect ratio $\alpha$9, while for $V$0 the reversal becomes non-coherent. Around $V$1–40 nm a buckling-like mode appears, and for $V$2 nm a transverse domain wall nucleates at the MgO interface and propagates along the vertical axis. The inverse switching time follows a strict linear relationship with applied voltage,

$V$3

and the slope decreases as the reversal moves from coherent to domain-wall-dominated dynamics (Caçoilo et al., 2020).

Three-dimensional PSA modeling adds another layer. In a 20 nm by 20 nm pillar, self-consistent transport simulations found edge canting of about $V$4 in the surface “flower” states, an effective ratio $V$5, and high-order three-dimensional ferromagnetic-resonance edge modes with $V$6–45 GHz above a threshold current density $V$7. These modes shorten or eliminate the incubation stage typical of macrospin reversal (Meneguolo et al., 19 Mar 2025). A plausible implication is that in thick pillars the write process is governed as much by localized mode excitation as by the average anisotropy barrier.

Thermoelectric p-STT-MRAM uses a different excitation mechanism. A Peltier-controlled heat flux produces an interfacial temperature differential $V$8 of up to $V$9 K, generating a magnon flux in an adjacent ferrite and a spin current at the ferrite-metal interface. The resulting torque enters the stochastic Landau–Lifshitz–Gilbert equation in the same Slonczewski form,

$P$0

but the spin current is now set by $P$1 rather than tunnel-current flow. Under a short pulse with peak $P$2 K and duration 150 ps, and with a tilt $P$3 between ferrite and free-layer easy axes, micromagnetic simulation shows complete P$P$4AP or AP$P$5P reversal in $P$6–0.8 ns (Mojumder et al., 2011).

4. Performance metrics and operating envelopes

At room temperature and below, p-STT-MRAM performance is commonly reported through switching phase diagrams, characteristic times, write energy, write-error rate (WER), and tunnel magnetoresistance (TMR). Cryogenic measurements on 40–60 nm pMTJ nanopillars show that the fitted characteristic time $P$7 decreases with temperature at fixed pulse-voltage overdrive, contrary to naïve macrospin expectations: for AP$P$8P switching, $P$9 drops from 1.48 ns at 295 K to 0.94 ns at 4 K, and for P$\eta(\theta)$0AP from 1.38 ns to 1.03 ns. At 4 K, 40 nm devices exhibit measured optimal switching energies of about 103 fJ for AP$\eta(\theta)$1P and 286 fJ for P$\eta(\theta)$2AP, while 4 ns pulses reach $\eta(\theta)$3 (Rehm et al., 2019).

Cooling also modifies the read margin. In the same nanopillars, TMR increases from about $\eta(\theta)$4 at 295 K to about $\eta(\theta)$5 at 4 K (Rehm et al., 2019). This has direct significance for cryogenic memory, because lower-temperature operation simultaneously changes switching speed, energy, and readout contrast.

Thermoelectric p-STT-MRAM targets a different operating envelope. Relative to conventional electrically written p-STT-MRAM with PMA, the reported comparison gives a maximum $\eta(\theta)$6 pulse of $\eta(\theta)$7 K, an equivalent critical spin-current density of about $\eta(\theta)$8 versus 14 MA/cm$\eta(\theta)$9 electric current density, a switching time of 0.6–0.8 ns versus at least 1 ns, and a write energy of about 3–4 fJ versus about 6–8 fJ. The reported WER remains below $\mathbf m_p$0, thermal stability is about $\mathbf m_p$1, retention at $\mathbf m_p$2 exceeds 10 years, and TMR is about $\mathbf m_p$3 rather than about $\mathbf m_p$4 because the MgO barrier can be thickened to 1.5 nm when no write current passes through it (Mojumder et al., 2011).

PSA devices emphasize stability at reduced diameter rather than the minimum absolute switching energy. For fixed lateral size $\mathbf m_p$5 nm, simulations show $\mathbf m_p$6 growing from about 60 at $\mathbf m_p$7 nm to much greater than 200 at $\mathbf m_p$8 nm, but the higher aspect ratio also pushes the threshold voltage from about 0.6–0.8 V in the macrospin-like regime to about 1.5 V once domain-wall propagation dominates (Caçoilo et al., 2020). This is the central PSA trade-off: large volume stabilizes the bit, but the same volume can slow or complicate the write trajectory.

Core-shell PSA aims to moderate that trade-off. For $\mathbf m_p$9 nm, $$\frac{d\mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d\mathbf m}{dt} + \gamma a_J\,\mathbf m\times(\mathbf m\times\mathbf p),$$0 nm, and $$\frac{d\mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d\mathbf m}{dt} + \gamma a_J\,\mathbf m\times(\mathbf m\times\mathbf p),$$1 V, the isolated core switches in about 2 ns, the Ni-shell design remains close to 2 ns, while the Co-shell design gives about 3–4 ns depending on shell damping. In dense arrays, the same architecture reduces the estimated average stray field: at 20 nm pitch, a conventional p-MTJ gives about 40 mT, a uniform PSA pillar about 75 mT, and the core-shell design about 10 mT (Caçoilo et al., 2023).

5. Reliability, non-idealities, and parasitic couplings

A recurrent misconception is that the apparent switching efficiency of p-STT-MRAM is controlled only by the intrinsic STT term. Experiments on p-MRAM cells with variable resistance-area product show otherwise. As RA increases from $$\frac{d\mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d\mathbf m}{dt} + \gamma a_J\,\mathbf m\times(\mathbf m\times\mathbf p),$$2 to $$\frac{d\mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d\mathbf m}{dt} + \gamma a_J\,\mathbf m\times(\mathbf m\times\mathbf p),$$3, the P$$\frac{d\mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d\mathbf m}{dt} + \gamma a_J\,\mathbf m\times(\mathbf m\times\mathbf p),$$4AP switching current density falls by about $$\frac{d\mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d\mathbf m}{dt} + \gamma a_J\,\mathbf m\times(\mathbf m\times\mathbf p),$$5 and the AP$$\frac{d\mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d\mathbf m}{dt} + \gamma a_J\,\mathbf m\times(\mathbf m\times\mathbf p),$$6P current density by about $$\frac{d\mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d\mathbf m}{dt} + \gamma a_J\,\mathbf m\times(\mathbf m\times\mathbf p),$$7, even though $$\frac{d\mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d\mathbf m}{dt} + \gamma a_J\,\mathbf m\times(\mathbf m\times\mathbf p),$$8, $$\frac{d\mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d\mathbf m}{dt} + \gamma a_J\,\mathbf m\times(\mathbf m\times\mathbf p),$$9, and $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$00 remain essentially constant and TMR changes only weakly. The reported explanation combines STT with self-heating and voltage-controlled magnetic anisotropy (VCMA), using

$$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$01

with $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$02, and a VCMA field coefficient $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$03. When these terms are included, the full dataset is reproduced with RA-independent $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$04, $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$05, $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$06, and $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$07 (Mihajlovic et al., 2019).

Interfacial Dzyaloshinskii–Moriya interaction (DMI) acts in the opposite direction. Micromagnetic calculations for free layers 10–40 nm in diameter and 1 nm thick show that increasing the DMI constant from 0 to $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$08 lowers the thermal barrier and raises the switching current density. For a 30 nm cell, the reported barrier drops from 0.82 eV to 0.55 eV, the stability factor from 31.8 to 21.3, and the switching current density from $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$09 to $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$10. The mechanism is twofold: DMI lowers domain-wall energy and promotes nonuniform, frustrated reversal (Jang et al., 2015).

By contrast, not all damage is detrimental. Numerical studies of edge-damaged perpendicular MRAM cells show that a 5 nm rim with degraded anisotropy can increase $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$11 by factors of 2–3. For $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$12 nm and $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$13 nm, the undamaged cell has $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$14 and $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$15, while the all-damaged case gives $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$16 and $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$17. The reduction in barrier is moderate because the damaged volume is small, whereas the switching mode is strongly altered by rim tilting (Song et al., 2015).

Reference-layer alignment and stray fields add another level of device-specific complexity. Systematic micromagnetic phase diagrams of 30 nm-diameter three-layer p-STT-MRAM nanopillars show four equilibrium groups—APc, APnc, Pc, and Pnc—as functions of bilinear and biquadratic interlayer exchange coupling. Across 4374 parameter sets, SAF asymmetry in saturation magnetization, anisotropy, or thickness reduces the coupling needed to stabilize antiparallel states, but in noncollinear antiparallel states it can raise SAF reversal barriers while lowering the free-layer barrier. SAF stray fields shift the free-layer barrier by up to $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$18, and increasing free-layer thickness or $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$19 suppresses the APc and APnc regions in $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$20 space (Terko et al., 9 May 2026).

6. Temperature dependence, scaling limits, and assisted writing

Temperature enters p-STT-MRAM not merely through the denominator of $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$21 but through the temperature dependence of the anisotropy itself. For PSA-STT-MRAM, the model

$$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$22

leads to a corrected barrier

$$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$23

Because $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$24, elevated temperature can collapse the PSA contribution faster than a purely interfacial barrier. The reported design implication is that PSA allows stable operation down to $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$25–10 nm with $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$26 at 300 K, whereas at 400 K one must increase $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$27 or $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$28 (Zhang et al., 2021).

Direct thermal imaging of PSA pillars supports the basic stability claim. For a FeCoB/NiFe storage layer with aspect ratio about 3:1, the induction decreases only from about $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$29 T at $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$30 to about $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$31 T at $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$32, with the magnetization remaining aligned along the pillar axis throughout the in-situ heating series. Using the stated approximation $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$33, the inferred $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$34 changes from about 120 at 293 K to about 57 at 523 K (Almeida et al., 2022).

Assisted-write schemes attempt to relax the conventional energy-speed-endurance trade-off. Thermoelectric p-STT-MRAM decouples the read and write paths by using a Peltier element and magnonic spin current rather than tunnel-current injection; the reported challenges are integration of Peltier elements in back-end-of-line CMOS, materials optimization for high magnonic thermal conductance and low phononic leakage, engineering of ferrite/metal interfaces for maximal $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$35 and minimal thermal resistance, cell-level and array-level thermal crosstalk control, and system-architecture co-design (Mojumder et al., 2011).

RF-assisted switching provides a different assist mechanism. In perpendicular MTJs with diameters of 85, 65, 45, or 25 nm, a 30 ns RF pulse applied before the DC write pulse increases the switching probability. For a 45 nm device at $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$36 GHz, $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$37 V, and $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$38, the baseline $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$39 at $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$40 V and $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$41 ns rises to $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$42 under RF+DC, and the maximum reported assist reaches $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$43 for partial overlap. The effect is larger at lower RF frequency, and the reported write benefit is that $$\frac{d \mathbf m}{dt} = -\gamma\,\mathbf m\times\mathbf H_{\rm eff} + \alpha\,\mathbf m\times\frac{d \mathbf m}{dt} + \tau_{\rm STT},$$44 can be shortened by about 10–20% for the same target probability, reducing both write energy and MgO stress (Hayward et al., 13 Dec 2025).

Taken together, these results define p-STT-MRAM as a family of perpendicular MTJ technologies rather than a single device template. Conventional ultrathin pMTJs remain the reference for compact two-terminal operation; PSA extends stability to sub-20 nm and even 4 nm design points but introduces three-dimensional reversal physics; thermoelectric writing removes tunnel-barrier write stress at the cost of thermal-integration complexity; cryogenic operation improves TMR and characteristic switching time; and auxiliary control knobs such as RF pre-excitation, SAF asymmetry, and deliberate edge engineering reshape the practical design space (Rehm et al., 2019, Perrissin et al., 2018).