Non-volatile memristors as prospective radio frequency switches for 6G communications - A review

Summary

As communication technologies continue to improve to the 6G standard, highly performing radio frequency (RF) switches that have short switching times, high energy efficiency, high cutoff frequencies, capabilities of operating beyond 100 GHz while maintaining reasonable RF performance, and high data transmission rates of 100 Gbit s⁻¹ are needed. Existing RF switches based on solid-state and microelectromechanical systems (MEMS) technologies fall short of meeting these requirements because of their volatility and parasitic inductive effects that lead to undesirably high insertion losses at high frequencies. Non-volatile memristors are emerging as likely candidates to replace traditional switches. They offer numerous benefits, including virtually zero static energy consumption, fast switching speeds, and excellent scalability. This review will provide an overview of the progress and challenges associated with implementing four types of non-volatile memristive technologies as RF switches: Resistive random access memory (ReRAM), conductive bridge random access memory (CBRAM), phase change memory (PCM), and memory based on two-dimensional (2D) materials, specifically molybdenum disulfide (MoS₂) and hexagonal boron nitride (hBN). The performance of these prospective switches is compared against one another and subsequently benchmarked against existing switch technologies. The review will provide an objective critique of the future readiness for 6G for the four main types of memristive switch technologies based on factors such as RF performance, CMOS compatibility, and fabrication maturity. Special attention is paid to 2D material-based memristors involving MoS₂ and hBN, which have attracted significant research interest for multiple benefits that are claimed to be distinctive. Finally, it will also propose research directions to further the progress of future memristive RF switches.

Graphical abstract

Condensed matter physics; Nanoscience; Telecommunication engineering; Electronic engineering; Materials science.

Introduction

In a radio frequency (RF) front-end module, RF switches are essential components that control signal flow between different blocks, such as antennas, low noise amplifiers (LNAs), power amplifiers (PAs), mixers, filters,¹ and hence help in realizing reconfigurable front-ends in wireless and satellite communication systems.^1,2 Due to parasitic effects, in the ON-state, the attenuation of the transmitted signal is known as insertion loss, whereas the ability of the switch to block signal transmission between different blocks in the OFF-state is termed isolation.

The rapid growth of the Internet of Things (IoT) has led to the widespread availability of services offered by large networks of interconnected objects, which is now brought closer to end users through the Internet of Everything (IoE).³ In order to enable super-IoT and IoE functions, 6G will demand a 100× to 1000× increase in KPI relative to 5G as recorded in Table 1. Some key enabling technologies needed to drive 6G include edge intelligence, blockchain, and terahertz communication. To facilitate these requirements, highly performing RF switches capable of handling high frequencies from 100 to 500 GHz with realistic insertion loss and isolation are needed.⁴ The switches must be able to handle data transmission rates up to 100 Gbit s⁻¹ in order to support existing mobile networks, as well as the implementation of novel features such as augmented/virtual reality (AR/VR), artificial intelligence with enhanced security and privacy, and IoT.^4,5

Table 1.

Summary of KPIs from 5G to 6G

	5G	B5Ga	6G
Data rateb	1 Gbps	100 Gbps	1 Tbps
E2Ec delay	5 ms	1 ms	below 1 ms
Radio delay	100 ns	100 ns	10 ns
Processing delay	100 ns	50 ns	10 ns
E2E reliability	99.999%	99.9999%	99.99999%

Table adapted with permission from ref.³ Copyright 2023, IEEE.

^aBeyond-5G.

^bAverage per user.

^cEnd-to-end.

Conventional RF switches are based on mature technologies that are mass-fabricated in foundries and are CMOS-compatible, and are broadly classified as semiconductor-based solid-state and microelectromechanical systems (MEMS) devices.¹ In solid-state switches, switching occurs via the controlled flow of electrons through the semiconductor.¹ Some examples of solid-state devices include PIN diodes, hybrid diodes, and field-effect transistors (FETs). Although they are reliable and compact, solid-state switches suffer from poor power handling and exhibit poor isolation at mm-Wave frequencies.^6,7 MEMS switches make use of mechanical movements triggered by a voltage or current to connect or disconnect an ohmic or capacitive contact.^8,9 They outperform solid-state switches in RF performance and power handling but at the expense of reliability issues such as dielectric charging, contact welding, stiction, and beam warping. Furthermore, MEMS switches have slow switching speeds and require high voltages for switching, as well as specialized packaging, which limits their widespread usage.¹⁰ A feature that both types of switches have in common is that they are volatile, meaning that they consume static energy from the standby DC bias used to maintain the ON- and OFF-states.¹¹ Consequently, the energy efficiency of the switch is reduced, rendering it unfavorable for high-frequency operations.⁴ Also, due to parasitic inductive effects, the insertion loss of these switches deteriorates substantially at high frequencies.¹²

The theoretical concept of memory resistors (otherwise known as memristors) was first proposed by Chua in 1971,¹³ while practical metal-oxide resistive switching devices demonstrated decades later established the foundation for modern resistive memory technologies.¹⁴ Owing to their attractive features such as non-volatility, low switching energy consumption (∼ pJ per switching¹⁵), fast switching speed (<ns¹⁶), high endurance (>10¹² cycles¹⁷), excellent scalability (4F²¹⁸), and compatibility with CMOS^19,20 in non-volatile memory and non-conventional computing, memory-based resistors (also known as memristors) have gained significant attention as prospective RF switches for 6G applications.⁴ Namely, resistive random access memory (ReRAM), conductive bridge random access memory (CBRAM), phase change memory (PCM), and memristors based on two-dimensional (2D) materials, in particular molybdenum disulfide (MoS₂) and hexagonal boron nitride (hBN), are being explored and studied.

This review discusses the progress and challenges of the four main classes of non-volatile memristive technologies being considered for potential RF switches. A comparison is then made between these four technologies based on their performance metrics, CMOS compatibility, and fabrication maturity. Subsequently, the memristive switch technologies are compared with their conventional counterparts. The review concludes with an overview of the current state of the field and the future directions needed to advance it further.

Memristive RF switch fundamentals

Figure 1 depicts the basic equivalent circuit models of an RF switch in both the ON- and OFF-states. In the ON-state, the switch is modeled as a parasitic resistor, where its resistance (R_ON) governs the switch’s insertion loss owing to the impedance that it poses to the signal path according to the following relation: IL (dB) ≈ 20 lg (1 + $\frac{{\mathrm{R}}_{\text{ON}}}{2{\mathrm{Z}}_0}$),²¹ where Z₀ refers to the load impedance. Conversely, the switch’s OFF-state capacitance (C_OFF) results in signal leakage through the switch and is thus responsible for its isolation, defined as Iso (dB) ≈ 10 lg $\left(1+{\left(\frac{1}{2{\mathrm{z}}_0\mathrm{\omega }{\mathrm{C}}_{\text{OFF}}}\right)}^2\right)$,²¹ where ω = 2πf is the angular frequency. A key figure of merit (FoM) used to quantify the switch’s performance is the cutoff frequency (F_CO), at which the impedances in both the ON- and OFF-states are equivalent. As F_CO = $\frac{1}{2\mathrm{\pi }{\mathrm{R}}_{\text{ON}}{\mathrm{C}}_{\text{OFF}}}$, it is indicative of the combined insertion loss and isolation performance of the switch.¹ To fulfill the demanding requirements for 6G communications, low R_ON (<10 Ω) and low C_OFF (<10 fF) are needed for the switch to have low insertion loss and high isolation.⁴ This would then translate to a high F_CO in the THz range, which is used for benchmarking potential switches.⁴ Other FoMs include actuation voltage, operating frequency bandwidth, switching time and energy, and cycle endurance.^21,22

Figure 1.

Basic equivalent circuit models of an RF switch

(A) ON-state, modeled as a resistor.

(B) OFF-state, modeled as a capacitor.

Power handling is another important FoM of RF switches, defined as the maximum power that the switch can transmit in the ON-state while maintaining excellent isolation in the OFF-state.²² As passive components, RF switches exhibit a linear relationship between their input and output powers over a range of input powers. High linearity is essential as it demonstrates the ability of an RF switch to preserve the amplitude and spectral integrity of an RF signal over its operating power range without experiencing significant gain or loss, compression, or intermodulation distortion (IMD).²³ Two key indicators of linearity are the 1 dB compression point (P1dB) and the input third-order intercept point (IIP3). P1dB refers to the point at which the switch’s insertion loss increases by 1 dB from its corresponding input power, while IIP3 is a measure of IMD, where a higher IIP3 indicates less intermodulation distortion. The parameters are measured by injecting two nearly equal tones, f₁ and f₂ (e.g., 1.000 GHz and 1.001 GHz) into a switch to obtain two outputs: one containing the desired signals (f₁ and f₂), and the other the third-order distortion products (2f₁ - f₂ and 2f₂ - f₁). As the input power increases, the desired output power increases proportionally, whereas the third-order distortion product output power increases by approximately three times. By extrapolating these two straight lines, IIP3 is determined at their intersection point.²⁴ Schematic diagrams for power handling, P1dB and IIP3 are presented in Figure 2.

Figure 2.

Schematic illustration diagrams. Diagrams were created with the aid of ChatGPT 5.5, and Gemini 3.5 was used to further edit the originally created diagrams

(A) Power handling.

(B) P1dB.

(C) IIP3.

Overview of four different memristive technologies

Conventionally used for memory and neuromorphic computing applications,¹ ReRAM devices consist of an amorphous transition metal oxide layer, such as TiO_x,²⁵ and HfO_x,²⁶ sandwiched between two metal electrodes, otherwise known as a metal-insulator-metal (MIM) configuration. Oxygen vacancy sub-stoichiometric conductive filaments are formed via the electroforming process²⁷ when a positive DC bias voltage is applied to the transition metal oxide layer to move oxygen ions within it toward an oxygen reservoir at the anode interface. When this occurs, the device is said to be in the SET or low resistance state (LRS). Reversing the polarity of the DC bias voltage partially restores the stoichiometry of the transition metal oxide layer, thus rupturing the conductive filament and switching the device to the RESET or high resistance state (HRS). Re-applying a positive DC bias voltage reforms the sub-stoichiometric conductive filaments, returning the device to the SET state. Advances in HfO₂-based interfacial, trilayer, and forming-free devices have improved analog switching linearity, energy efficiency, and variability, making ReRAM increasingly attractive for neuromorphic and in-memory computing.^28,29

CBRAM devices, otherwise known as electrochemical metallization cells, have MIM structures reminiscent of those of ReRAM devices, except for an ion exchange membrane or air gap instead of a transition metal oxide layer as the active region and electrochemically asymmetric electrodes made of two different metals.⁴ Ion exchange membranes can be inorganic (e.g., GeSe₂ chalcogenide glass,³⁰ SiO₂³¹) or organic (e.g., PMMA,³² Nafion^{11,24,33,34,35}) in nature. One electrode is said to be active, as the metal atoms at the electrode oxidize readily compared to the other electrode, which is termed the inert electrode. Ag and Cu are common materials selected for the active electrode, whereas W, Pt, and Au have been used for the inert electrode.⁴ Under the influence of a positive DC bias voltage, CBRAM devices undergo switching due to the formation of conductive filaments. Instead of oxygen vacancies, the conductive filaments originate from metal cations that diffuse through the ion exchange membrane to the inert electrode, where they are reduced back to atoms. This then forms a low-resistance path for electrons to flow, causing the switch to be in the SET state. Applying a negative DC bias voltage to the active electrode causes the metal ions to return to the active electrode, thus rupturing the conductive filaments and reverting the switch to the RESET state.³¹ In CBRAM devices, a limiting current, otherwise known as the compliance current, is supplied to the device to prevent the active layer from being damaged.³¹ Recent studies on CBRAM devices have focused on controlling electrochemical filament evolution, improving analog/binary stability, and extending electrochemical metallization concepts into 2D-confined systems.^36,37

PCMs utilize chalcogenide materials as their switching layer, which ranges from elemental Sb in their simplest form to ternary or quaternary alloys such as Ge₂Sb₂Te₅ and Ge₂Sn₂Sb₂Te₇.³⁸ Due to their ability to form interatomic resonance bonds, chalcogenides are poor glass formers, which makes them suitable as phase change materials because of their rapid crystallization property.³⁹ Chalcogenides exist in one of two states: amorphous (RESET) or crystalline (SET). In the amorphous state, the atoms are arranged in a disordered fashion, which lowers the mean free path of conduction for electrons because of electron scattering, resulting in the amorphous state having high resistance. By contrast, the atoms are ordered in the crystalline state, causing the resistance of the state to be low.⁴⁰ The resistance ratio between the amorphous and crystalline states can exceed 10⁴.⁴¹ To trigger a phase transition in the chalcogenide from the amorphous to crystalline state, a low-amplitude, long-duration (usually >1 μs) electrical pulse heats the PCM such that the atoms have sufficient time to arrange themselves in an orderly manner. Conversely, a high-amplitude, short-duration (usually <100 ns) electrical pulse melts the PCM. The orientation of the atoms becomes disorganized, and a rapid quench locks the position of the atoms, thus changing the state of the chalcogenide from crystalline to amorphous.⁴⁰ The main progress for PCM lies in analog in-memory computing, rapid learning, drift-aware architectures, and edge-AI acceleration rather than conventional storage alone.^42,43

The switching mechanism of memristors consisting of either 2D MoS₂ or hBN mono- or multilayers stacked between metal electrodes has been elucidated using ab initio simulations. Taking Au as an example for the metal electrodes, Au⁺ cations oxidized from the Au anode under a SET voltage fill vacancies (S in MoS₂ and B in hBN) in the 2D layer, where they are subsequently reduced back to Au atoms by captured electrons. As the percentage of Au atoms that occupy vacancies increases, a conductive pathway is formed, and the density of states surrounding the Fermi level increases, triggering the transition from HRS to LRS. Applying a RESET voltage displaces the Au atoms from the vacancies and causes them to return to the electrodes, thus restoring the switch to HRS.⁴⁴ Unlike the linear filamentary structures observed in CBRAM devices, 2D MoS₂ and hBN memristive devices that use Au as their electrodes have conductive pathways consisting of segmented clusters. This is because Au has a high redox potential, which causes the conductive pathways to grow at non-uniform rates. The low mobility of Au also restricts long-range diffusion, which consequently causes the reduction of Au⁺ cations to be concentrated around initial nucleation sites that gradually develop into disjointed nanoclusters of varying sizes.⁴⁵ The latest research on MoS₂ and hBN-based systems has produced significant advances in device scalability, switching speed, reliability, and heterogeneous integration. Recent studies have demonstrated improved control of defect- and filament-mediated switching in atomically thin MoS₂ and hBN layers, enabling forming-free operation, enhanced device uniformity, ultra-fast switching, and ultra-low switching energy. Concurrently, these material systems have been integrated into large-scale memristor arrays and silicon-compatible architectures, highlighting their growing potential for neuromorphic computing, in-memory computing, and future high-frequency electronic applications.^46,47

Unified switching diagrams for these four memristive technologies are illustrated in Figure 3.

Figure 3.

Unified switching diagrams for ReRAM, CBRAM, PCM, and 2D MoS₂ and hBN memristors

Diagrams were created with the aid of ChatGPT 5.5, and Gemini 3.5 was used to further edit the originally created diagrams.

ReRAM RF switches

The first experimental demonstration of a ReRAM-based RF switch was performed by Badulescu using a heavily reduced 30 nm thick TiO_2-x transition metal oxide layer with Pt electrodes.⁴⁸ The switch was fabricated with a coplanar waveguide (CPW) design and tested up to 10 GHz, achieving insertion loss and isolation values of 2.1 dB and 32 dB, respectively. The R_ON and C_OFF of the switch were 28 Ω and 2.5 fF, which yielded an F_CO value of 2.27 THz.

Chen et al. published their work on a ReRAM-based RF switch, which featured 5 nm thick HfO₂ as the transition metal oxide layer stacked between two Ni electrodes, each with a thickness of 50 nm⁴⁹ Figure 4 shows a schematic of the switch structure. This switch was designed for flexible electronic applications involving 6G communications, and RF performance characterizations were conducted up to 110 GHz. The switch’s insertion loss remained roughly constant at 2 dB, whereas its isolation fell from 60 to 10 dB over the entire frequency range starting from 1 GHz. The lowest R_ON of the switch was 32 Ω, and its C_OFF had an average value of 5.8 fF, resulting in an F_CO value of 0.857 THz. The switch demonstrated good endurance, reliably switching even after 10⁶ cycles. It also maintained its insertion loss and isolation performance at 110 GHz after withstanding 2500 bending cycles. In addition, it had a switching time of 10 ns?

Figure 4.

Schematic illustration of the device structure of the flexible HfO₂ ReRAM RF switch by Chen et al. Figure adapted from ref.⁴⁹ under the CC BY-NC-ND 4.0 license

ReRAM devices benefit from mature fabrication technology, excellent CMOS compatibility, and fast switching speeds.^1,4,50 However, as seen in both switches above, they typically have R_ON values above 10 Ω, resulting in significant insertion losses that restrict their use as RF switches at frequencies in the THz range.¹ Experimental studies on implementing ReRAM technology for prospective RF switches are scarce, much less demonstrations of RF characterization above 100 GHz, which is required for 6G communications. Hence, further studies on ReRAM-type switches should focus on RF characterization tests in the sub-THz/THz regimes and refining their performance at these frequencies through the use of optimized CPW device designs. Excluding Chen et al.’s HfO₂-based switch, a known problem of ReRAM switches is their poor endurance, limited to only tens of cycles. Through the use of device engineering methods, such as optimizing oxide sputtering conditions, improving the properties of metal electrodes, and varying the oxide area,⁴⁸ the endurance of several TaO_x-based ReRAM memory devices has been shown to drastically improve, reaching over 10¹¹ cycles.^11,50

CBRAM RF switches

The implementation of CBRAM technology for RF switches began with Nessel et al. in 2008,⁴³ featuring an active Ag anode and an inert Ni cathode separated by a 10 μm gap filled with Ag⁺ ion-doped GeSe2 chalcogenide glass. Later, Pi et al. showcased their work on a CBRAM switch that featured a 35 nm air gap as its switching medium.⁵¹ The performance of both of these switches has already been discussed in previous review articles.^1,4 A recent advancement in CBRAM RF switches for a reconfigurable X-band bandpass filter utilized VO_x as the dielectric layer.⁵² An SEM image of this switch is available in Figure 5. The switch SETs at an average voltage of 2.55 V and RESETs at ∼ −0.9 V. Additionally, it was capable of reaching an endurance of ∼ 10³ switching cycles and a rapid switching speed of 100 ns? Furthermore, the R_ON and C_OFF of the switch obtained were 5 Ω and 7 fF, and thus the switch had an F_CO of 4.5 THz. High-frequency performance of the switch was characterized up to 67 GHz, revealing insertion loss and isolation to be 0.74 dB and 13 dB, respectively. The P1dB of the switch was analyzed through power-dependent loss measurements. In the ON-state, no noticeable compression (>0.01 dB) in the output power was observed up to an input power of 10 dBm, which was the measurement limit of the VNA source used. The same was true for the OFF-state, whereby the switch was able to maintain its high isolation performance, producing losses of >35 dB in the output power.

Figure 5.

An SEM image of the VO_x CBRAM RF switch by Seo et al. Figure adapted from ref.⁵² under the CC BY-NC-ND 4.0 license

Scale bar: 20 μm.

Another contemporary CBRAM RF switch consists of an Ag/WO_3-x/WSe₂/Au vertical layer stack in the overlap region that forms the switch area,⁵³ as shown in Figure 6. The switch had a SET voltage between 2 and 3 V and a RESET voltage between −1.5 and −2 V. It also displayed a rapid switching speed of 3.2 ns and a low switching energy of 3.7 nJ. However, its endurance was poor, with only 32 switching cycles. The R_ON and C_OFF of the switch were 8.2 Ω and 0.77 fF, causing the F_CO of the switch to be 25.2 THz. This corresponds to an insertion loss and isolation of 0.7 dB and 30 dB, respectively, at 67 GHz. The linearity metrics for this switch were reported to be 16 dBm for P1dB, for which the power handling was also considered, and 38.7 dBm for IIP3.

Figure 6.

The WO_3-x/WSe₂ heterojunction CBRAM RF switch by Cao et al. Figure adapted with permission from ref.⁵³ Copyright 2025, Wiley-VCH

(A) Schematic illustration of ground-signal-ground (GSG) CPW.

(B) Enlarged view of the Ag/WO_3-x/WSe₂/Au vertical MIM configuration.

In addition to the ongoing work on CBRAM RF switches, Zhang et al. introduced a HfO₂/MoS₂ heterojunction in their switch architecture.⁵⁴ SEM images of the fabricated switch are shown in Figures 7A and 7B. In amorphous oxides, the formation and rupture of conductive filaments show stochastic behavior because of the uniform distribution of the electric field and disordered orientation of atoms within the film. This causes the conductive filaments to branch out, and thus rupture can occur at random points along the filament, as visualized in Figures 7C and 7D. Placing a thin MoS₂ layer at the interface between the bottom Pt electrode and the HfO₂ dielectric creates a van der Waals heterojunction, whereby Ag⁺ ions in HfO₂ diffuse rapidly through the MoS₂ layer owing to its higher diffusion coefficient, forming conical conductive filaments.⁵⁵ During RESET, filament rupture is bound to a specific region in the conductive filament, and the filament is reconstructed at that same region during subsequent SET operations, as shown in Figures 7E and 7F. This then stabilizes the formation and rupture of the conductive filaments, which allows for more consistent switching, leading to better performance than switches without the heterojunction. With the inclusion of the MoS₂ layer, the SET and RESET voltages of the switch were reduced from ∼0.9 V and ∼ −0.7 V–0.7 V and −0.6 V, respectively. Furthermore, for the measured frequency range from DC–43.5 GHz, the insertion loss becomes more uniform, varying from 2.4 to 3.3 dB without the MoS₂ layer, to 1.3–1.7 dB with the MoS₂ layer. However, this is also accompanied by a dip in the isolation, going from 44.6 to 18.8 dB without the MoS₂ layer, to 32.6–9.2 dB with the MoS₂ layer. In addition, the switch demonstrated low-latency switching at 125 ns, consuming ∼1 nJ of switching energy in the process.

Figure 7.

A HfO₂/MoS₂ CBRAM RF switch by Zhang et al. Figure adapted from ref.⁵⁴ under the CC BY-NC-ND 4.0 license

(A) SEM image. Scale bar: 300 μm.

(B) Zoomed-in view of the device overlap region. Scale bar: 20 μm.

(C) Conductive filament formation without MoS₂ layer during SET.

(D) Conductive filament rupture without MoS₂ layer during RESET.

(E) Conductive filament formation with MoS₂ layer during SET.

(F) Conductive filament rupture with MoS₂ layer during RESET.

Zhou et al. also utilized HfO₂ as the switching material for their CBRAM switch, along with an Au bottom electrode coupled with Ag for the top electrode and GSG CPW pads, as shown in Figure 8.⁵⁶ HfO₂ has a strong affinity toward oxygen vacancies, which then serve to create accessible routes for Ag⁺ cation migration to the bottom electrode, thus effectively lowering the LRS resistance of the device to a great extent. At ∼2.2 V, the device SETS and enters LRS, where its current is capped at 1 mA, the compliance current limit imposed on the device. Subsequently, RESET occurs at −1.1 V, restoring the device to its original state. At room temperature, a retention time of 10⁵ s and an ON/OFF ratio of ∼ 10⁹ were obtained. Under a SET voltage of 4 V and a RESET voltage of −3.5 V, the device had a maximum endurance of 2300 cycles. RF characterization was performed up to 43.5 GHz, at which the insertion loss and isolation were 0.53 dB and 26.8 dB, respectively. Upon varying the compliance current from 0.1 to 10 mA, the R_ON of the switch was also reduced proportionally from 197.5 to 5.8 Ω, which the authors used to validate the dependence of R_ON on compliance current. By taking advantage of this property of CBRAM, the authors then implemented their switches into a π-type tunable attenuator, which consisted of one switch connected in series, and two switches in a parallel arrangement. Adjusting the compliance current allows the R_ON of each switch to be toggled between various programmable states to achieve the desired attenuation outcome. For this π-type tunable attenuator consisting solely of a memristor network, 12 attenuation states ranging from 2 to 24 dB could be attained across a frequency range of 10 MHz–43.5 GHz. Since Ag has high ionic mobility, which could promote the formation of conductive filaments more readily, the authors conducted the same experiments on switches that had both Ag top and bottom electrodes. As compared to the Au bottom electrode switches, the Ag/HfO₂/Ag switches had significantly lower SET (∼0.38 V) and RESET (∼−0.34 V) voltages. The retention time was 10⁴ s with an ON/OFF switching ratio of 10¹⁰, but only 400 endurance cycles were reached. The Ag/HfO₂/Ag switches had R_ON and C_OFF values of 4.1 Ω and 5.9 fF, giving a slightly lower F_CO value of 6.6 THz. The insertion loss of these switches was 0.36 dB, lower than that of switches with Au bottom electrodes, which could be attributed to the lower R_ON value. The isolation, however, was lower by 1 dB at 25.8 dB as a result of the larger C_OFF value. The range of R_ON values that could be varied for the Ag/HfO₂/Ag switches was narrower (4.1–22 Ω), and the attenuation states produced in the π-type tunable attenuator were identical to those for the Au bottom electrode switches.

Figure 8.

A HfO₂ CBRAM RF switch by Zhou et al. Figure adapted from ref.⁵⁶ under the CC BY-NC-ND 4.0 license

(A) Schematic illustration of GSG CPW.

(B) Enlarged view of the Ag/HfO₂/Au vertical stack.

Generally, CBRAM-based RF switches can achieve smaller R_ON values than ReRAM switches because of the low resistance of metal ions. The R_ON of CBRAM switches can be further reduced by increasing the compliance current owing to the inverse relationship between current and resistance.⁵⁷ Furthermore, because of their filamentary nature, the R_ON of the switches is independent of their device areas,⁵⁸ which allows the switches to be scaled to smaller areas to reduce C_OFF at higher operating frequencies without any consequence on R_ON, which then also enhances their F_CO. CBRAM switches also offer switching voltages that can go as low as < 1 V, along with fast switching times and energies in the nanosecond and nanojoule regimes, respectively.^53,54,56 However, a serious disadvantage that CBRAM switches face is their severe limitation in endurance, which currently has a maximum of 10³ cycles only.^11,51 As recommended for improving the endurance of ReRAM devices, the same can be done for CBRAM devices by employing materials engineering strategies. For example, in the case of Pi et al.’s air gap switch, which only has a maximum endurance of 716 cycles when tested in deionized water, it has been demonstrated that scaling the air gap of memristors which share similar structures to a much shorter length can boost the endurance to as high as 10¹¹ cycles.⁵⁹ The stochastic nature of the conductive filament growth results in cycle-to-cycle and device-to-device variability.⁶⁰ However, as proven by Zhang et al., the use of heterojunctions can aid in localizing filament formation and breaking, leading to better switching uniformity.⁵⁴ The power handling for CBRAM switches is also hindered by the stability of the filament when subjected to high RF powers.³⁵ As a means to overcome this limitation, the filament geometry and electrode design ought to be fine-tuned in order to raise the device’s current carrying capacity.³⁵

PCM RF switches

PCM RF switches exist as two-, three-, or four-terminal devices. Two-terminal PCM switches are directly heated by passing a DC current through the PCM with two heater electrodes that are connected to the PCM,⁶¹ or via direct contact to the RF paths.⁶² Three- and four-terminal PCM switches, on the other hand, can be directly or indirectly heated, where the latter involves an electrical signal heating an intermediate highly resistive material. The material then functions as a heater and transfers the heat energy to the PCM via the Joule effect.^61,63

Early works on the use of PCMs for RF switch applications began with Chua et al. in 2010,⁶⁴ which featured a GeTe-based reconfigurable switch with an R_ON value of 180 Ω. Over the years, PCM RF switches have evolved to provide better RF performance. In 2014, Moon et al. presented a two-terminal GeTe switch.⁶⁵ This switch attained low R_ON and C_OFF values of 0.5 Ω and 4 fF, giving a large F_CO of 80 THz. The switch also exhibited excellent RF performance, with an insertion loss of 0.2 dB and 30 dB isolation at a maximum operating frequency of 67 GHz, accompanied by a remarkably high power handling capability of 40 dBm without any self-actuation limit. Singh et al. proposed a four-terminal GeTe switch design in 2021, as illustrated in Figure 9.⁶⁶ This switch had R_ON and C_OFF values of 1.6 Ω and 7 fF, and an F_CO value of 14.2 THz. At the maximum operating frequency of 67 GHz, the insertion loss and isolation values measured were 0.5 dB and 14 dB, respectively. For a 30 μm² channel area, the switch had a high maximum power handling capability of 39 dBm in the ON-state but self-actuated at 26 dBm in the OFF-state. The endurance of this switch design was reported to be as high as 10⁶ cycles, with switching times ranging from 200 ns to 1.2 μs＀ The IIP3 of the switch was 41 dBm.

Figure 9.

A four-terminal GeTe PCM RF switch by Singh et al. Figure adapted with permission from ref.⁶⁶ Copyright 2021, IEEE

(A) Optical micrograph. Scale bar: 100 μm.

(B) 3D-rendered view.

(C) SEM micrograph of switch junction shows channel width (w_s) and microheater length (l_h). Scale bar: 2 μm.

(D) Zoomed-in false colored SEM micrograph of the PCM channel showing the channel length (L_s) and microheater width (w_h). Scale bar: 0.5 μm.

In 2020, El-Hinnawy announced the results of a commercial PCM RF switch manufactured by Tower Semiconductor. The switch had an R_ON and C_OFF of 2.3 Ω and 2.7 fF, and therefore, an F_CO value of 25.2 THz. The insertion loss and isolation were 0.25 dB and 22 dB at 40 GHz, respectively. However, what distinguishes this switch is its large endurance of 1 billion cycles.⁶⁷ The endurance was then enhanced to 10 billion cycles in a work published three years later, along with excellent power handling beyond 30 dBm and a switching speed of 700 ns.⁶⁸ In 2025, the company reported the design of a single-pole double-throw (SPDT) switch that was integrated with CMOS drivers.⁶⁹ The switch was tested up to 110 GHz, yielding an insertion loss of 2 dB and an isolation of 20 dB. The power handling of the switch was 29 dBm.

In a recent publication in 2025, Chen et al. used a monolayer of elemental Sb as the chalcogenide material in a PCM.⁷⁰ Figure 10 shows an optical microscope image of the switch. Unlike GeTe, Sb does not experience the same problem of component segregation and structural decomposition (since it consists of a single element), which eventually leads to crystallization failure of the phase change material.^71,72 The Sb PCM switch had a low total energy consumption of 1.1 μJ and showed an ultrafast response of only 300 ns because of rapid recrystallization, which can be explained by glass homogeneity along with topological rearrangement of the three-to 4-fold rings. Unfortunately, this switch had a low endurance of only 82 cycles. The R_ON and C_OFF of the switch were 9.2 Ω and 2.7 fF, corresponding to a F_CO of 6.4 THz. At the maximum operating frequency of 67 GHz, the insertion loss and isolation of the switch were 1.2 dB and 18 dB, respectively.

Figure 10.

An optical microscope image of the Sb PCM RF switch by Chen et al. Figure adapted with permission from ref.⁷⁰ Copyright 2025, American Physical Society

Scale bar: 5 μm.

Another paper that was published in 2025 utilized 2D MoTe₂ as the phase change material.⁷³ The device structure of the switch is presented in Figure 11. MoTe₂ is heated directly, and switching occurs through reversible phase transformations between its semiconducting 2H phase (HRS) and metallic 1 T′ phase (LRS). The R_ON of the switch had an average value of 10.3 Ω, while the C_OFF had a very small value of only 0.157 fF. As such, the switch obtained a F_CO of 72 THz. The switch was tested up to 110 GHz, yielding a very low insertion loss of 0.3 dB and an isolation of 19 dB. The switch showed high power-handling performance of 29 dBm. Linearity measurements were also conducted on the device. At a maximum input power of 16 dBm, the output power compression did not exceed 0.25 dB, revealing that the P1dB lay beyond the maximum input power up to which the device was measured. Moreover, the device had an IIP3 of 42 dBm.

Figure 11.

A MoTe₂ PCM RF switch structure by Wu et al. Figure adapted with permission from ref.⁷³ Copyright 2025, ACS

(A) Schematic illustration.

(B) Magnified top view SEM image of the device region. Scale bar: 1 μm. The inset shows an SEM image of the switch in GSG CPW configuration. Scale bar: 40 μm.

In the case of GeTe, PCM RF switches exhibit high power handling capabilities, high endurance, and low insertion loss while maintaining moderate to high isolation values. They have been integrated into a 200 mm back-end-of-line (BEOL) wafer foundry manufacturing process^67,74 and have been implemented in various mm-Wave devices such as multiport RF switches,⁷⁵ switch matrices,⁷⁶ and variable attenuators.⁷⁷ However, the heterogeneous integration of PCM switches with other devices is constrained by the choice of substrates and conductors available that can withstand the high heating temperatures required to activate phase transitions.²⁴ Heating can be confined within the active volume instead of over the entire device region through the use of nanoscale microheaters, thermal confinement layers, and plasmonic heating. This would then reduce the thermal budget and also the heating temperature needed to induce phase transitions.⁷⁸ The need for an integrated microheater in a lateral configuration switch complicates fabrication and BEOL integration, contributing to switching speeds ranging from nanoseconds to microseconds,⁷⁹ which are considerably slow by future communication standards. Materials with rapid crystallization properties, such as elemental Sb that was discussed in this section, should be investigated as alternative phase change materials to reduce the switching time to trigger phase transitions. Regardless of whether the PCM is heated directly or indirectly, a significant amount of thermal energy is required to change the state of the switch, which reduces the energy efficiency of PCM switches.⁷⁹ By doping a SiO_x layer with Ag, Hou et al. were able to make use of an Ag conductive filament formed by a pulse SET voltage within GeTe to directly heat the material, inducing phase transitions.⁸⁰ A schematic of the switch device is presented in Figure 12. Due to the high current density generated by the conductive filaments, the GeTe layer undergoes complete crystallization, thus reducing the switch’s LRS resistance. The energy required for the SET process is only 19.2 nJ, marginally lower than those of both directly^62,63,81 and indirectly heated PCM switches.^67,82,83 Unlike CBRAM switches, PCM switches encounter a tradeoff between R_ON and C_OFF upon area scaling, which causes the isolation to be increased at the expense of insertion loss when the area is reduced.⁶⁷ Zhang et al. were able to mitigate this effect by modifying the device design of conventional GeTe PCM switches.⁸⁴ The difference in designs is depicted in Figures 13A and 13B. The ground plane at one end of the CPW is joined to a bias port in order to improve the isolation by reducing the effect of parasitics contributed by the bias signal. However, doing so also increases the ON-state energy loss of the switch, which subsequently worsens the insertion loss as well. As such, the GeTe layer was segmented into an array of multiple parallel blocks in the channel to reduce the resistance of crystalline GeTe and lower the insertion loss. From Figure 13C, it can clearly be seen that there is a 20 dB improvement in isolation for the bias-decoupled model when compared to the original model without substantially compromising the insertion loss.

Figure 12.

A schematic illustration of the GeTe PCM RF switch with an Ag conductive filament as a heater. Figure adapted with permission from ref.⁸⁰ Copyright 2024, IEEE

Figure 13.

A modified GeTe PCM RF switch design by Zhang et al. Figure adapted with permission from ref.⁸⁴ Copyright 2025, IEEE

(A) Bias-decoupled design.

(B) Conventional design.

(C) Simulation results of the RF performance for the original and bias-decoupled models.

2D MoS₂ and hBN memristive RF switches

Non-volatile resistive switching has been reported in a plethora of 2D materials, such as graphene oxide mixtures, black phosphorus, transition metal dichalcogenides (TMDs), and hexagonal boron nitride (hBN).^85,86 Conventional vertical MIM metal oxide-based devices face challenges in nanometer scaling due to excessive leakage currents. In their work, Zhao et al. reported that the use of monolayer TMDs, such as MoS₂, MoSe₂, WS₂, and WSe₂, can resolve this issue due to their high degree of crystallinity and large bandgaps, which suppress large leakage currents.⁸⁷ However, based on analysis of data extracted from published I-V characteristics of selected monolayer RF switches, the HRS leakage currents read at 0.1 V showed substantial variability, ranging from as low as 0.04 pA up to 50 μA. The leakage currents for all of these switches in ascending order are tabulated in Table 2.

Table 2.

Leakage currents of selected monolayer 2D material RF switches

Monolayer material	Leakage current (nA)	References
hBN	0.00004	Yang et al.88
hBN	0.081	Kim et al.89
MoS2	10	Kim et al.90
MoS2	15	Kim et al.91
MoS2	15	Ge et al.92
hBN	44	Kim et al.22
MoS2	286	Kim et al.93
MoS2	11800	Kim et al.94
MoS2	50000	Ge et al.44

As these atomically thin 2D monolayers can exhibit memristive properties at atomic thicknesses, they are collectively termed “atomristors”.⁹² The concept of using atomristors as RF switches was shown by Kim et al. in their work on a MoS₂-based RF switch using Au electrodes in 2018.⁹⁰ Since then, much interest has been sparked in this field of research. The configuration of this MoS₂ switch is illustrated in Figure 14. The MoS₂ monolayer was grown using chemical vapor deposition (CVD). Low R_ON and C_OFF values of 3 Ω and 4 fF were obtained, giving this switch an F_CO of 13.2 THz. As one can expect from the low R_ON and C_OFF values, the switch had low insertion loss and high isolation values of 0.3 dB and 30 dB, respectively, at an operating frequency of 50 GHz.

Figure 14.

A crossbar monolayer MoS₂ RF switch by Kim et al. Figure adapted with permission from ref.⁹³ Copyright 2018, IEEE

(A) Schematic illustration.

(B) Top view optical image of the MoS₂ crossbar switch. Scale bar: 20 μm.

Solution processing has been performed to electrochemically exfoliate nanosheets of MoS₂ to produce dense monolayers with narrow thickness distributions.^95,96 As the uniformity of MoS₂ monolayers created by solution processing is high, this technique is capable of supporting large-volume fabrication and minimizing device-to-device variability. Mingates et al. reported endurance of 100 cycles, short switching time of 40 ns, and switching energy of merely 330 pJ for their solution-processed MoS₂ switch.⁹⁶ R_ON, C_OFF, and F_CO for the switch were 6.8 Ω, 4.3 fF, and 5.4 THz. The insertion loss and isolation of the switch at 110 GHz were 0.9 dB and 18 dB, respectively.

Aside from solution processing, another method that has been used in the fabrication of MoS₂ switches is inkjet printing, which is economically feasible, environmentally friendly, and viable for integrating devices onto flexible substrates. Xiao et al. have implemented this technique to print MoS₂ switches with Ag electrodes onto a graphene metasurface for non-volatile reconfigurability.⁹⁷ The switch’s endurance was tested up to 300 cycles and retention up to 1.8 × 10⁷ s under ambient conditions. The R_ON and C_OFF of the switch were 8.9 Ω and 31.92 fF, equivalent to an F_CO of 0.56 THz. RF characterization at 12 GHz yielded an insertion loss of 0.7 dB and isolation of 10 dB.

MoS₂ switches have been tested at high frequencies beyond the sub-terahertz regime (>300 GHz), showing potential for 6G communication applications. In a 2022 publication, Kim et al. developed a monolayer MoS₂ switch capable of operating up to 480 GHz.⁹¹ It also featured a high switching speed of 500 ps with a correspondingly low switching energy of 50 pJ. This switch had an R_ON of 8 Ω and C_OFF of 0.9 fF, and thus an F_CO of 22 THz. Its insertion loss was relatively flat at approximately 1 dB, along with moderate isolation of 15 dB at 480 GHz. The IIP3 of the switch was measured to be 31 dBm. Skrzypczak et al. went further by extending the measurement frequency to 750 GHz, which returned reasonable insertion loss and isolation results of 6 dB and 12 dB, respectively, at such a high frequency.⁹⁸ R_ON and C_OFF values extracted from the S-parameter results were 12.5 Ω and 0.31 fF, and the F_CO of the switch was 41.1 THz.

Aside from MoS₂, another 2D material that has been commonly used in atomristors is hBN. Kim et al. were the first to report on an RF switch based on monolayer hBN.⁸⁹ An optical microscope view of the GSG CPW and an SEM image of the switch device are presented in Figures 15A and 15B, respectively. The switch had R_ON and C_OFF values of 1.6 Ω and 2.3 fF, resulting in an F_CO value of 43 THz. At the maximum operational frequency of 110 GHz, the hBN switch had low insertion loss of 0.2 dB and moderate isolation of 15 dB. The monolayer hBN switch has a maximum power handling capability of 30 dBm in the ON-state and self-actuates at 20 dBm in the OFF-state, which is a 10-fold improvement to that of Ge et al.’s crossbar monolayer MoS₂ switch⁴⁴ at 10 dBm because of its larger bandgap and thermomechanical stability.

Figure 15.

A monolayer hBN RF switch by Kim et al. Figure adapted with permission from ref.⁸⁹ Copyright 2019, IEEE

(A) Optical microscope image. The dashed circle indicates the area containing the switch. Scale bar: 100 μm.

(B) Enlarged top view SEM (false color) image. Scale bar: 1 μm.

Subsequently, Kim et al. developed their monolayer hBN switch, whose maximum operating frequency was extended from 110 to 220 GHz.²² At 67 GHz, the R_ON and C_OFF were 2.8 Ω and 0.44 fF, thus yielding the highest-ever recorded F_CO value of 129 THz. At the maximum operating frequency of 220 GHz, the switch had low insertion loss but poor isolation, measuring 0.5 dB and 10 dB, respectively.

Instead of Au, Yang et al. used Ag electrodes for their monolayer hBN switch, which attained lower SET voltages and leakage currents compared to the former.⁸⁸ A schematic of the switch is presented in Figure 16. The switch was actuated at 0.34 V and consumed low switching energy (1.2 nJ) and power (0.23 W). However, the switch had limited endurance of merely 20 cycles. R_ON and C_OFF of this switch were 6.1 Ω and 0.68 fF, giving an F_CO of 38 THz. RF measurements were taken across the D-band (110–170 GHz), returning insertion loss of 0.5 dB and isolation of 23 dB. In the ON-state, the switch’s power handling was 20 dBm, but it displayed self-actuation at 4 dBm in the OFF-state as a consequence of its low SET voltage of 0.34 V.

Figure 16.

A monolayer hBN RF switch by Yang et al. Figure adapted with permission from ref.⁸⁸ Copyright 2023, ACS

(A) Schematic illustration of GSG CPW.

(B) Enlarged view of the MIM overlap area.

In addition to monolayer hBN, the utilization of multiple layers in RF switches has been reported. In 2024, Pazos et al. announced the use of 12–18 layers of hBN for RF switches in series, shunt, and series-shunt configurations, whose designs are presented in Figure 17.⁹⁹ The F_CO of this multilayer switch in series configuration was 6.7 THz with R_ON and C_OFF values of 4.5 Ω and 5.1 fF. The insertion loss and isolation values were 0.9 dB and 10 dB, respectively, at 120 GHz. In shunt configuration, the multilayer hBN switch was able to attain a higher isolation of 27.5 dB at an operational frequency of 260 GHz, and an even higher isolation of 35 dB at 120 GHz in series-shunt configuration.

Figure 17.

ON- and OFF-state circuit diagrams and optical microscope images of the multilayered hBN RF switches by Pazos et al. Figure adapted with permission from ref.⁹⁹ Copyright 2024, Nature

(A) Series configuration.

(B) Shunt configuration.

(C) Series-shunt configuration.

In summary, 2D memristive RF switches based on MoS₂ and hBN typically offer high FoMs in terms of F_CO, maximum operating frequency, and switching time and energy. Additionally, they can achieve relatively low insertion losses and, in some cases, high isolation values, even at high frequencies exceeding 100 GHz. For atomristors, the insertion loss decreases as a function of frequency because of C_ON, which is parallel to the quasi-static R_ON. Conversely, the insertion loss increases with frequency for other switch technologies due to parasitic inductive effects.²² The different effects on the insertion loss as a function of frequency for 2D monolayer atomristor switches versus PCM, FET and MEMS switches, along with their associated equivalent circuit diagrams, are shown in Figure 18. Similar to CBRAM devices, the R_ON and C_OFF are unaffected by each other upon area scaling due to the localized nature of the filaments.^58,59,90 2D hBN and MoS₂ RF switches actuate at SET voltages <1 V, but this also poses a restriction to their power handling capacity.⁸⁸ The highest power handling capability recorded is 20 dBm for hBN switches before self-actuation happens in the OFF-state, which is inadequate for 6G communication applications requiring high power. The power handling of future 2D MoS₂ and hBN memristive switches could be improved by stabilizing the conductive filament, optimizing electrode geometry, and improving heat dissipation through thermally conductive substrates and encapsulation layers.^22,100,101

Figure 18.

Comparison of insertion loss behavior for atomristor versus PCM, MEMS and FET switches

Equivalent circuit models explain why insertion loss decreases with frequency for atomristors but increases for conventional technologies because of parasitic inductance. Figure adapted with permission from ref.⁸⁹ Copyright 2019, IEEE.

Currently, 2D memristor-based RF switches face poor integration on silicon wafers, and wafer-scale fabrication is still in its infancy, resulting in large numbers of defects that lower the yield to <75% and increase the variability across devices.⁸⁶ CVD and solution processing methods for monolayers need to be developed further to mature wafer-scale fabrication. Subsequently, the yield can be enhanced via better material quality, defect engineering, and stricter process control.^96,102 Significant breakthroughs in the current state of BEOL-compatible integration processes and heterogeneous CMOS-2D strategies are necessary to drive the progress of CMOS compatibility of 2D MoS₂ and hBN memristive switches.^96,102

Besides poor integration, 2D MoS₂ and hBN switches also face constraints in their endurance. Monolayer MoS₂ and hBN switches typically exhibit endurance of only hundreds of cycles,^44,88,89 while that of the multilayer hBN switch is limited to 2000 cycles at best.⁹⁹ Possible approaches to improve the endurance of 2D MoS₂ and hBN memristor-based switches include defect (vacancy) engineering, optimization of electrode interfaces, and tailored switching protocols.^103,104

In recent years, 2D MoS₂ and hBN have attracted great attention as contemporary RF switches due to numerous acclaimed benefits. However, not many comparisons have been made with other types of switches to support the advantages that these 2D memristor-based switches have over them. To benchmark their performance, the insertion loss and isolation results of the 2D hBN and MoS₂ switches were compared with other switch technologies at 40 GHz. Figures 19A and 19B show plots of insertion loss against R_ON and isolation against C_OFF, respectively, for the switches being compared. From both plots, no consistent trend is observed in either insertion loss or isolation performance to substantiate that in all cases, 2D hBN and MoS₂ switches outperform their counterparts, at least at a frequency of 40 GHz. For example, from Figure 19A, it can be observed that one of the monolayer MoS₂ atomristor switches has an insertion loss of 1.06 dB, which is higher than the other types of switches being compared except for the ReRAM switch.

Figure 19.

Comparison of RF performance between different RF switch technologies at 40 GHz. Data compiled from ref.^{24,44,49,51,52,53,56,65,67,70,73,89,90,91,92,93,94,98,105,106,107,108,109,110}

(A) Insertion loss against R_ON.

(B) Isolation against C_OFF.

Comparative assessment of memristive RF switch technologies

Table 3 compiles the performance of the four memristive RF switch technologies that have been discussed across their various metrics. Direct comparison of RF switch performance reported in the literature should be interpreted with caution because the reported metrics were obtained under different experimental conditions. These include variations in operating frequency, device geometry, switching area, CPW dimensions, substrate material, calibration and de-embedding procedures, and RF measurement setups. Furthermore, metrics such as insertion loss, isolation, power handling, and FCO are frequency-dependent, while extracted parameters including R_ON and C_OFF may vary depending on the equivalent circuit model and extraction methodology employed. Consequently, the comparisons presented in this review should be regarded as representative indicators of the current state of the art rather than absolute rankings between technologies. Insertion losses and isolation were taken at the maximum operating frequency for each device. Unless already stated in their original works, F_CO was derived from R_ON and C_OFF values using the expression stated in the section: “Memristive RF switch fundamentals”. Finally, only RF switch demonstrations were considered in the quantitative comparison. Memory-device-only results were completely excluded from the performance ranges of the RF switch metrics.

Table 3.

Comparative performance of reported memristive RF switch

	ReRAM	CBRAM	PCM	2D MoS2 and hBN Memristors
RON(Ω)	28–3248,49	3.6–8.230,51,52,53,54,56	0.5–10.365,66,67,68,69,70,73	1.6–13.522,44,88,89,90,91,92,93,94,96,98,99
COFF(fF)	2.5–5.848,49	0.77–930,51,52,53,54,56	0.157–765,66,67,68,69,70,73	0.31–7.722,44,88,89,90,91,92,93,94,96,98,99
FCO(THz)	0.857–2.2748,49	2.95–3530,51,52,53,54,56	6.4–8065,66,67,68,69,70,73	5.87–12922,44,88,89,90,91,92,93,94,96,98,99
Maximum operating frequency (GHz)	10–11048,49	6–6730,51,52,53,54,56	40–11065,66,67,68,69,70,73	50–75022,44,88,89,90,91,92,93,94,96,98,99
Insertion loss at maximum operating frequency (dB)	2–2.148,49	0.33–1.730,51,52,53,54,56	0.2–265,66,67,68,69,70,73	0.2–622,44,88,89,90,91,92,93,94,96,98,99
Isolation at maximum operating frequency (dB)	10–3248,49	9–3430,51,52,53,54,56	14–3065,66,67,68,69,70,73	10–3022,44,88,89,90,91,92,93,94,96,98,99
Endurance (cycles)	10–10648,49	32–10351,52,53,54,56	82–101067,68,70,73	20–10444,88,89,90,96,99
Switching Speed	10 ns49	3.2–125 ns51,52,53,54	200 ns – 1.2 μs66,68,70	500 ps – 40 ns22,44,89,91,93
Switching Energy	∼ pJ48,49	1–3.7 nJ51,52,53,54	19.2 nJ–1.1 μJ70,80	50 pJ–1.2 nJ22,88,91,96
Power handling (dBm)	Limited (Not reported)48,49	10–2730,51,52,53,54	26–4065,66,68,73	4–2022,44,88,93
IIP3 (dBm)	Limited (Not reported)48,49	38.7–4451,52,53	41–4566,73	3191
CMOS compatibility	Excellent49	Excellent52,53,54	Good63,64,65,66	Moderate90,91,97
Fabrication Maturity	High	High	High	Low

Values correspond to RF switch demonstrations reported under different measurement frequencies, device dimensions, substrates, calibration techniques, and de-embedding procedures. Performance ranges should therefore be interpreted as indicative of technology trends rather than direct one-to-one comparisons.

With the exception of ReRAM switches, all the memristive switch technologies possess low R_ON values and are generally capable of meeting the target specification of <10 Ω for 6G requirements. Among them, PCM and 2D MoS₂ and hBN memristors stand out for having the lowest R_ON values. All four memristive switch technologies also satisfy the C_OFF target requirement of <10 fF, with CBRAM, PCM, and 2D MoS₂ and hBN switches even achieving ultra-low C_OFF values of <1 fF. The F_CO values of the switches span a wide range, starting from sub-THz to as high as > 100 THz. As a result of having both the lowest R_ON and C_OFF values, both PCM and 2D MoS₂ and hBN memristors stand above ReRAM and CBRAM switches in F_CO. To date, the maximum operating frequency of CBRAM switches is only 67 GHz, making it the only memristive switch technology that has not been characterized past 100 GHz. Several RF measurements beyond 100 GHz have been reported for 2D MoS₂ and hBN switches, such as Yang et al.'s monolayer hBN switch (from 110 to 170 GHz),⁸⁸ Kim et al.'s monolayer hBN switch (up to 220 GHz),²² Pazos et al.'s multilayer hBN switch (up to 260 GHz),⁹⁹ Kim et al.'s monolayer MoS₂ switch (up to 480 GHz),⁹¹ and Skrzypczak et al.‘s multilayer MoS₂ switch (up to 750 GHz),⁹⁸ to name a few. As a consequence of their large R_ON, it should be expected that the ReRAM switches have higher insertion losses of ∼2 dB as compared to their counterparts. Skrzypczak et al.‘s multilayer MoS₂ switch is an anomaly, having a high insertion loss of 6 dB corresponding to a very high maximum operating frequency of 750 GHz. For isolation, similar ranges are observed across the four memristive switch technologies despite there being some differences in the C_OFF variation for each memristive switch technology.

Filamentary-type switches (ReRAM, CBRAM, and 2D MoS₂ and hBN-based memristors) display nanosecond switching speeds and very low switching energies of only up to a few nanojoules. However, they share common problems of poor endurance and limited power handling. Conversely, PCM switches have significantly higher power handling capabilities and endurance because of thermally induced phase transitions rather than metallic filament formation, but tend to also have slower switching speeds and higher switching energies owing to the need for thermal activation. Currently, few studies have reported IIP3 for these four memristive switch technologies, of which Wu et al.‘s 2D MoTe₂ PCM switch displays the best performance of 42 dBm.

ReRAM and CBRAM rank highest in CMOS compatibility. As previously mentioned, GeTe PCM switches have already been integrated into BEOL wafer manufacturing processes in commercial fabs. However, the need for a microheater and the high heating temperatures also pose challenges in integration. Lastly, because of the novelty of 2D materials, the development of CMOS-compatible processes to enhance their integration onto silicon wafers is still underway. The fabrication processes for ReRAM, CBRAM, and PCM are well-understood and established, whereas those for 2D MoS₂ and hBN switches remain immature at this stage because of issues in large-area growth, yield, uniformity, and transfer processes.

Although absolute quantitative comparisons remain challenging because of differences in measurement conditions, several consistent trends emerge across the reported literature. PCM switches consistently display superior power handling and endurance, ReRAM devices benefit from mature CMOS-compatible fabrication, CBRAM switches offer an attractive compromise between RF performance and switching energy, while 2D MoS₂ and hBN memristors exhibit the highest F_CO owing to their atomically thin switching layers and extremely low C_OFF. These trends remain consistent despite variations in device architectures and measurement methodologies.

Comparison with conventional RF switch technologies

In this section, the following parameters (area, F_CO, maximum operating frequency, insertion loss, isolation, power handling, and switching time) of non-volatile memristive switch technologies, as compiled in Table 4 later in discussion, will be compared with those of conventional volatile switch technologies. To facilitate a fair comparison with conventional switch technologies, insertion loss and isolation values were normalized to 40 GHz because it lies within the overlapping operating range of all the switches being assessed in this comparison, thereby providing a common reference point for benchmarking insertion loss and isolation. Other metrics, including switching speed, endurance, and power handling, were taken directly from the respective literature. The benchmarking methodology and associated limitations have already been discussed in the previous section: “Comparative assessment of memristive RF switch technologies”. Differences in device geometry, fabrication processes, and measurement methodology remain and should be considered when interpreting the results.

Table 4.

Benchmark comparison between non-volatile RF switches and conventional RF switch technologies

Switch technology	Volatile	Area (μm2)	Switching voltage (V)	FCO (THz)	Maximum operating Frequency (GHz)	Insertion loss at 40 GHz (dB)	Isolation at 40 GHz (dB)	Power handling (dBm)	Switching time (ns)	Reference
22 nm FDSOI CMOS	Yes	26000	0.8	1.46	220	0.65	35	N/A	N/A	Wu et al.109
0.35 μm SiGe HBT	Yes	4–8	N/A	0.95	70	2.09	27.2	1	0.075	Thian et al.111
70 nm GaAS m-HEMT	Yes	610000	N/A	0.82	77	0.67	6.9	19.5	N/A	Kim et al.106
GaAs p-HEMT	Yes	N/A	N/A	0.54	80	3	20	N/A	1.4–1.8	Wang et al.112
PIN diodes	Yes	50	10	2.47	50	0.3	2.81	26	N/A	Boles et al.107
MEMS	Yes	62500	80–90	12.4	40	0.56	14	40	5000	Patel et al.105
MIT	Yes	4	N/A	24.1	290	0.23	22	37	10000	Hillman et al.108,110
ReRAM	No	25	−1	0.86	110	2	16	10	10	Chen et al.49
CBRAM	No	0.6	3.7	35	40	0.33	30	17	1–100	Pi et al.51
CBRAM	No	2.15	2.55	4.54	67	0.81	16.3	10	100	Seo et al.52
CBRAM	No	4	2	25.2	67	0.57	38.4	16	3.2	Cao et al.53
PCM	No	103	N/A	80	67	0.2	36.3	40	N/A	Moon et al.65
PCM	No	30	7.8	14.2	67	0.71	18.5	26	200–1200	Singh et al.66
PCM	No	0.25	0.7	72	110	0.44	23.9	29	N/A	Wu et al.73
Monolayer MoS2	No	0.03	1	70.4	67	0.23	33.7	10	30	Ge et al.44
Monolayer MoS2	No	0.04	1.5	22	480	0.54	29.3	N/A	0.5	Kim et al.91
Monolayer hBN	No	0.03	0.9	129	220	0.5	10	20	15	Kim et al.22
Monolayer hBN	No	0.05	1.5	43	110	0.12	25.6	20	15	Kim et al.89
Multilayer hBN (Series)	No	0.81	3	6.7	260	0.33	18.7	N/A	2.5	Pazos et al.99
Multilayer MoS2	No	0.0625	1.75	41.1	750	1.25	27.2	N/A	N/A	Skrzypczak et al.98

Noticeably, all the non-volatile switches have much smaller footprints compared to the 22 nm FDSOI CMOS, 70 nm GaAs m-HEMT, and MEMS switches. Two of the PCM switches^65,66 and the ReRAM switch are the only non-volatile switches that are larger than the 0.35 μm SiGe HBT and MIT switches. In terms of F_CO, two of the CBRAM switches,^51,53 two of the PCM switches,^65,73 and the following 2D MoS₂ and hBN switches^22,44,89,98 have values exceeding that of the MIT switch, which is the highest among the volatile switches. Among these, two MoS₂ and hBN atomristor switches^22,44 have F_CO values that far surpass that of the MIT switch.

All of the non-volatile switches have maximum operating frequencies that are at least comparable with the MEMS switch at 40 GHz, which is the lowest among the volatile switches listed. The highest frequencies that have been measured are up to 750 GHz for the multilayer MoS₂ switch, followed by a monolayer MoS₂ switch⁹¹ at 480 GHz. At 40 GHz, most of the non-volatile switches except for the multilayer MoS₂ switch and ReRAM switch achieve low insertion losses up to 0.81 dB, which is substantially lower than those of the 0.35 μm SiGe HBT and GaAs p-HEMT switches. For the isolation, both the volatile and non-volatile switches produce similar values. The 22 nm FDSOI CMOS switch has the best isolation value of 35 dB among the volatile switches, which is similar to those of a CBRAM switch⁵³ and a monolayer MoS₂ switch.⁴⁴

In terms of power handling, many of the non-volatile switches do not show a significant improvement over their volatile counterparts except for one PCM switch.⁶⁵ PCM switches show the best power handling among the non-volatile switches. Both a PCM switch⁶⁵ and a MEMS switch¹⁰⁵ are tied with the highest power handling capacity of 40 dBm among all the tabulated switches, followed by the MIT switch at 37 dBm.

Among the non-volatile switches, short switching times of at most 100 ns have been achieved by the ReRAM, CBRAM, and 2D MoS₂ and hBN switches. A monolayer MoS₂ switch⁹¹ and the multilayer hBN switch have very short switching times ranging from subnanosecond values to a few nanoseconds, which are close to those of the GaAs p-HEMT switches, but still not as fast as that of the 0.35 μm SiGe HBT switch. Due to the need for an integrated microheater, as mentioned in the section: “PCM RF switches”, the switching time reported for a PCM switch⁶⁶ is relatively long as compared to the other non-volatile switches. Nonetheless, the longest switching time for this PCM switch is still much shorter than those of the MEMS and MIT switches.

Conclusion and outlook

RF switches form the backbone of reconfigurable systems and RF front-end modules by routing signals among components such as antennas, LNAs, mixers, and filters, while supporting multiple frequency bands. Hence, high-performance RF switches are needed to drive the continuous evolution and development of wireless communication networks and devices. 6G communication standards mandate that RF switches need to be able to operate at high frequencies beyond 100 GHz while maintaining acceptable insertion loss and isolation performance, as well as handle high data transmission rates up to 100 Gbit s⁻¹ to enable advanced functions such as AR/VR, artificial intelligence with tighter security and privacy measures, and IoT. This has proven to be a challenge for existing RF switch technologies (solid-state and MEMS) for two main reasons: poor energy efficiency due to their volatile nature and degradation of insertion loss because of parasitic inductance.

Over the last decade, non-volatile memristors (ReRAM, CBRAM, PCM, and 2D MoS₂ and hBN) have been sought out as plausible candidates for the next generation of RF switches. While traditional CMOS switches expend nearly zero energy for switching, their volatile nature requires a continuous DC bias to supply static standby power to maintain their state. In contrast, although the switching energy for memristors is higher, the standby power needed after programming is essentially zero. Therefore, the total energy consumed over long idle periods is much lower in the case of memristors, giving them higher energy efficiency over their volatile technology counterparts. Other attractive features that make these memristors viable options for high-performance RF switches include their fast switching speeds, excellent scalability, and compatibility with CMOS. For each of the four classes of memristors listed above, their current state of research, challenges as well as proposed research directions have been outlined in their dedicated sections. A comparative assessment of key RF switch parameters was then made between the four classes of memristive switch technologies before extending the comparison to include volatile switch technologies.

Table 5 provides a summary of the primary strengths, major bottlenecks toward 6G, possible research directions that can be embarked on, and future readiness for 6G. PCM currently represents the most mature candidate for practical high-frequency RF switching because of its superior endurance, power handling, and commercial foundry demonstrations. ReRAM offers a CMOS-compatible alternative but still requires substantial improvements in insertion loss. CBRAM provides an attractive balance between RF performance and scalability, although endurance remains a concern. Lastly, 2D MoS₂ and hBN memristors remain the least mature even though they exhibit the highest F_CO. Despite the recent discovery of 2D materials, it is surprising that MoS₂ and hBN are already being heavily considered for their prospects as RF switches at frequencies extending into the mm-Wave and THz regimes, which more mature memristive switch technologies have struggled to reach. As a result of their novelty, 2D materials are still in their early stages of development. Hence, challenges still persist, such as minimizing the number of defects and parasitics in large-scale fabrication, increasing their endurance to over 10⁹ cycles to meet industry standards, and improving their integration on silicon wafers to enhance yield. Unlike PCM switches, which have been extensively studied for RF switch applications, filamentary-type memristors are still relatively unexplored in this field, with few works reported in the literature on their development toward 6G communications standards. Due to their filamentary nature, they offer the advantage of having their R_ON and C_OFF decoupled from each other, thus allowing them to maintain low insertion losses while increasing their isolation at higher frequencies upon area scaling.

Table 5.

Summary of the current state of memristive switch technologies toward 6G

	ReRAM	CBRAM	PCM	2D MoS2 and hBN Memristors
Primary Advantages	•mature fabrication technology •excellent CMOS compatibility •fast switching speed	•low RON •excellent scalability (RON independent of device area •low switching voltage and energy	•outstanding endurance exceeding industry standards •high power handling capability •commercial foundry demonstrations	•best FCO •excellent scalability (RON independent of device area •insertion loss decreases as a function of frequency because of CON in parallel to quasi-static RON •lowest switching speed and energy
Major Bottlenecks Toward 6G	1.high RON leading to higher insertion loss 2.limited RF demonstrations above 100 GHz 3.endurance remains far below commercial requirements	1.poor endurance (Limited to 103 cycles) 2.stochastic conductive filament causes variability 3.limited power handling	1.high switching energy 2.relatively slow switching speed 3.heterogeneous integration challenges caused by high heating temperatures 4.area scaling introduces tradeoff between RON and COFF, and consequently insertion loss and isolation	1.wafer-scale fabrication remains immature 2.poor device yield and reproducibility 3.limited endurance 4.limited power handling 5.CMOS integration remains challenging
Possible Research Directions	1.optimize oxide sputtering conditions 2.extend RF characterization into sub-THz/THz regime using optimized CPW device designs 3.improve conductive filament stability through oxide engineering, optimized electrodes, and device scaling	1.improve electrolyte composition and interface engineering to enhance endurance 2.introduce heterojunctions (e.g., HfO2/MoS2) to localize filament growth and rupture 3.optimize filament geometry and electrode design to improve current-carrying capability	1.develop localized Joule-heating and filament-assisted heating architectures to reduce switching energy 2.explore fast-crystallizing materials such as elemental Sb and 2D MoTe2 3.optimize thermal confinement to improve heater efficiency and reduce heating temperatures 4.employ segmented PCM structures and bias-decoupled RF layouts to reduce insertion loss while maintaining isolation	1.develop scalable CVD growth and solution-processing techniques for large area monolayers 2.improve material quality, defect engineering, and process control to improve fabrication yield 3.engineer vacancy sites, electrode interfaces, and switching pathways to improve endurance 4.improve power handling by stabilizing conductive filaments, optimizing electrode geometry, and improving heat dissipation 5.develop BEOL-compatible integration processes and heterogeneous CMOS-2D integration strategies
Future readiness toward 6G	medium term	medium term	near term	long term

Other than concentrating efforts solely on 2D materials, further investigations should also be conducted on other materials for ReRAM and CBRAM RF switches suitable for 6G communications. For example, chalcogenide glasses such as germanium sulfide (GeS) have shown exceptional performance when used as the switching medium in CBRAM devices for memory and neuromorphic computing applications, achieving endurance as high as 10⁹ cycles, far higher than that achieved by currently reported CBRAM RF switches.^113,114 The last report on the use of chalcogenide glasses in CBRAM RF switches was published by Nessel et al. in 2008,³⁰ who used GeSe₂ doped with Ag⁺ ions as the switching layer. Despite the age of their work, the power handling of 27 dBm for their switch has yet to be surpassed by any CBRAM RF switches today. Thus, with high endurance and power handling properties, chalcogenide glasses demonstrate promising potential to overcome the present limitations encountered by CBRAM RF switches. Until more studies are carried out on other materials for filamentary-type switches, an objective assessment cannot be made to justify the superiority of 2D MoS₂ and hBN memristor switches over other switch technologies.

Overall, the prospects of non-volatile memristors outperforming and eventually replacing conventional solid-state and MEMS switches are bright, especially with major advances being made recently. This is evident from Figure 20, which contains plots of the performance trends of the non-volatile memristive switches from the past decade. At present, not many reports are available on linearity studies conducted on memristive RF switch technologies. This aspect of memristive RF switches should not be overlooked since linearity is an indicator of how well a switch is able to maintain its linear relationship between its input and output powers under the influence of large signals before significant distortion takes place, and future communication systems will employ many high-performance transmission applications such as high-order modulation, carrier aggregation, beamforming, and simultaneous multi-band operation, all of which are highly sensitive to nonlinear distortion. Thus, it is paramount to ensure that future switches not only have low insertion loss and high isolation performance, but also high P1dB (>30 dBm) and high IIP3 (>50 dBm) to satisfy the linearity requirements for 6G communication systems.¹¹⁵ Several research directions can be pursued over short-, medium-, and long-term time frames to bring future memristive switch technologies closer to practical realization. These future research directions are outlined according to their respective time frames in the technology roadmap as depicted in Figure 21.

Figure 20.

Performance trends of memristive switch technologies in the past decade

(A) F_CO against year. Data compiled from ref.^{22,49,52,53,56,66,67,70,73,88,91,96,98,99}

(B) Switching speed against year. Data compiled from ref.^{22,52,53,54,66,67,70,73,91,96,98,99}

(C) Maximum operating frequency against year. Data compiled from ref.^{22,52,53,54,56,66,67,70,73,88,91,96,98,99}

(D) Insertion loss against maximum operating frequency. Data compiled from ref.^{22,51,52,53,54,56,66,67,70,73,88,91,96,98,99}

(E) Isolation against maximum operating frequency. Data compiled from ref.^{22,51,52,53,54,56,66,67,70,73,88,91,96,98,99}

Figure 21.

A technology roadmap outlining future research directions in the short-, medium-, and long-term time frames to further advance the development of memristive switch technologies for 6G deployment

The original version of the figure was created by ChatGPT 5.5, and the design was then fully redrawn for submission of the finalized manuscript.

Acknowledgments

This work was jointly supported by Singapore Economic Development Board and GlobalFoundries, Singapore, under the Industrial Postgraduate Program.

Author contributions

E.W.K.S. conducted the literature investigation and synthesis, prepared the figures, schematics, comparative analyses, and tables, and wrote the original manuscript. R.T.T. and D.S.A. supervised the development and technical direction of the work. All authors critically reviewed and edited the manuscript and approved the final version.

Declaration of interests

The authors declare no competing interests.

Declaration of generative AI and AI-assisted technologies in the writing process

During the preparation of the manuscript revision, the authors used ChatGPT 5.5 to aid in the creation of the following figures: schematic illustration diagrams for power handling and linearity (Figure 2), unified switching diagram (Figure 3), technology roadmap outlining future research directions for RF switches (Figure 21). Gemini 3.5 was subsequently used to make further edits to Figures 2 and 3. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.