The Circuit That Remembers: How Memristors Could Slash Power Converter Energy by 92%

Introduction

The server rooms that power the internet consume electricity at a staggering rate—enough to rival the energy use of some small nations. At the heart of every power converter that steps voltage up or down, every smartphone charger, every electric vehicle inverter, sits a component that flickers millions of times per second, dictating how efficiently energy flows. That component is the pulse width modulation controller, and it is where a surprising amount of power gets wasted.

A team of electrical engineers at Penn State has just demonstrated a circuit that could slash the energy consumed by these controllers by 92%—using a device that, until recently, most engineers had never even encountered outside of research laboratories. The device is called a memristor, and it记忆—memories—electricity. It remembers how much current has passed through it and adjusts its resistance accordingly, like a faucet that learns the pressure of the water flowing through it over time.

The finding isn't merely incremental. Memristors have been championed for their potential in artificial intelligence and neuromorphic computing, where they've shown energy reductions of 99.9% compared to conventional computing architectures. But their integration into power converter control—where the rubber meets the road for anything that plugs into a wall—has remained incomplete. The Penn State team has now built the missing bridge: a programmable pulse width modulation circuit that operates at around 200 kilohertz and consumes less than one-tenth the power of the digital signal processors currently dominating the field.

The Science

The Problem With How We Control Power Today

Every time you charge your phone, every time an electric motor spins in a factory robot, every time a solar panel feeds energy into the grid, a power converter is at work. These devices take electricity at one voltage and transform it into electricity at another—stepping down a 400-volt EV battery to the 5 volts your USB-C cable delivers, for instance, or converting the variable output of solar panels into the steady alternating current that your home appliances expect.

Controlling these converters requires something that turns power on and off with extraordinary precision. This is the job of pulse width modulation, or PWM. Imagine blinking a light switch at 200,000 times per second—each flash lasting just the right fraction of that microsecond interval to regulate how much energy flows. The "duty cycle"—the percentage of time the switch is on versus off—determines the output voltage. A 50% duty cycle produces half the maximum voltage; an 80% duty cycle produces four-fifths.

Two fundamentally different approaches currently generate these PWM signals. The first uses digital signal processors, chips that perform the calculation mathematically: a counter increments, a comparison happens, and a digital signal emerges (Wu et al., 2026). The second uses analog electronics: a triangle-shaped waveform oscillates against a reference voltage, and a comparator produces the PWM output. Each approach carries a hidden cost.

DSPs are programmable—you can change frequencies, adjust duty cycles, update parameters on the fly—but they're computationally wasteful for this task. Before a memristor-based AI accelerator can control a power converter through a DSP, the memristor's analog result must be converted to a digital number, stored in memory, retrieved, processed, and then converted back into an analog PWM signal. Each conversion step burns energy. "Although DSP-based PWM modules provide high programmability, they are not directly compatible with fully memristive computing platforms," the researchers note. "This signal-conversion and memory-access process introduces additional interface complexity, latency, and power consumption."

Analog PWM generators, meanwhile, are energy-efficient but inflexible. "Changes in switching frequency or operating range typically require hardware modification," the paper explains. Modern power converters increasingly need adaptive behavior—SiC and GaN semiconductors, which are replacing traditional silicon in high-efficiency applications, operate best at switching frequencies ranging from tens to hundreds of kilohertz, and different conditions may demand different frequencies. Analog circuits can't readily accommodate these requirements.

The Memristor: A Device That Remembers

Memristors occupy a peculiar corner of electrical engineering's vocabulary—they were theoretically predicted in 1971 by Leon Chua but not practically demonstrated until 2008, when a team at HP Labs showed that a particular arrangement of titanium dioxide could exhibit resistance that depended on its electrical history. The word combines "memory" with "resistor."

Physically, a memristor is deceptively simple. Two electrodes—typically a top electrode and a bottom electrode—sandwich a material layer whose resistance changes when current flows through it. Apply voltage in one direction, and the device's resistance decreases; reverse the direction, and it increases. Unlike a regular resistor, which maintains a fixed value, a memristor "remembers" what has passed through it. When you remove the voltage, the resistance state persists.

This behavior arises from ion migration within the device. Oxygen vacancies—missing oxygen atoms carrying a positive charge—drift through the material under the influence of the applied electric field, effectively doping the semiconductor and changing its conductivity. The process is nonlinear and exhibits a hysteresis loop in the current-voltage relationship, a signature that distinguishes memristive behavior from ordinary resistance.

The commercial memristors used in this study come from Knowm Inc., a company that manufactures devices based on alternating current-driven switching mechanisms. Two types appear in the research: tungsten-based (W-type) and chromium-based (Cr-type) devices. Both exhibit the pinched hysteresis loop characteristic of memristors, though their exact switching dynamics differ.

The critical parameter for this application is the threshold voltage—the applied voltage above which the device begins to change its resistance state. The Knowm Cr-type memristors have a typical threshold voltage of approximately 0.33 volts. Below this threshold, the device can be "read"—its resistance measured—without disturbing the programmed state. Above it, the resistance changes.

Programming Memory Into Matter

To use a memristor for frequency control, you must first program it to a precise resistance value. But here's the catch: the switching behavior is stochastic. Apply the same sequence of voltage pulses twice, and you'll rarely get exactly the same final resistance. The process involves probabilistic ion migration, and cycle-to-cycle variation is inherent to currently available devices.

The Penn State team solved this with closed-loop programming. After each write pulse, they measure the resistance. If it's too low, they adjust and try again; if it's too high, they do the same. The iterative procedure continues until the device lands within an acceptable tolerance window—±10% of the target resistance, which the researchers deemed sufficient for the PWM application.

The programming platform uses a Knowm Discovery board, which interfaces with the memristor through a series resistor. A pulse generator supplies both write and read pulses; a voltage divider forms between the memristor and the series resistor, allowing the resistance to be calculated from the measured voltage.

The mathematics of the voltage divider sets practical limits on the resistance range that can be accurately monitored. With a series resistor of 10 kilo-ohms and a read pulse of 0.1 volts, the detectable range becomes roughly 1 kilo-ohm to 100 kilo-ohms—below 1 kilo-ohm, the voltage division ratio becomes too extreme and noise dominates; above 100 kilo-ohms, the same problem occurs in the opposite direction. For the Cr-type memristors used in the main experiments, this range proved workable.

The statistical results validate the approach. Programming nine different target resistance values, each tested 50 times, yielded median errors close to zero across all targets—the closed-loop method doesn't introduce systematic bias. The standard deviation for the 21 kilo-ohm target, for instance, was 946.55 ohms, about 4.5% of the target value, comfortably within the ±10% tolerance. The probability of reaching the target window across all trials was 100%.

But the stochastic nature doesn't disappear entirely. The number of write pulses required to program the memristor from a high-resistance state (>100 kilo-ohms) to 17 kilo-ohms varied from 2 to 11 pulses across trials, with most (approximately 82%) converging within 3 to 9 pulses. Fixed-pulse-count programming would be unreliable; closed-loop verification after each pulse is essential.

The Circuit Design

Replacing a Resistor to Enable Programmability

The proposed circuit comprises two main stages: a triangular waveform generator and a hysteresis comparator that produces the final PWM output. In a conventional analog PWM generator, a fixed resistor determines the carrier frequency. The Penn State innovation replaces this resistor with a memristor—and then rethinks the circuit topology to make that replacement useful.

The triangular carrier generation relies on a Schmitt trigger and an integrator working in concert. The Schmitt trigger produces a square wave, which the integrator integrates into a linearly ramping voltage; when that voltage hits a threshold, the Schmitt trigger flips, and the integrator ramps in the opposite direction. The result is a clean triangle wave.

The oscillation frequency of this carrier depends on several component values, but crucially, it depends on the resistance in the feedback path. The formula governing frequency is:

$$f=\frac{R_2}{4R_1{C}_1(R_5+R_{M1})}$$

where the variables represent component values in the circuit, and importantly, $R_{M1}$ is the memristor resistance. Change the memristor resistance, change the frequency. Program the memristor to a new value, and you can tune to a new frequency—with no hardware changes.

But there's a constraint. The memristor must remain below its threshold voltage during normal operation; if it exceeds 0.33 volts, its resistance will drift unintentionally, destabilizing the oscillator. This requirement limits the maximum allowable resistance:

$$R_{M1,\max }=\frac{2V_{th}}{V_{dd}}{R}_2$$

where $V_{th}$ is the threshold voltage and $V_{dd}$ is the supply voltage. The key insight is that increasing the feedback resistor $R_2$ allows a larger memristor resistance range while keeping the memristor voltage below threshold—and larger $R_2$ simultaneously increases the oscillation frequency. In the previous topology the team had developed, by contrast, increasing the resistance to extend the allowable range actually reduced the operating frequency. "Extending the programmable resistance range in the previous topology comes at the cost of a lower operating frequency," the paper notes, "whereas the proposed topology can improve high-frequency operation by increasing $R_2$."

The previous design also required both positive and negative voltage rails; the new design operates from a single positive rail, simplifying implementation.

Noise Immunity Through Hysteresis

The triangular carrier signal feeds into a hysteresis comparator, which compares it against a reference voltage to generate the PWM output. Without hysteresis—the intentional offset between the rising-edge and falling-edge switching thresholds—noise on either signal could cause spurious rapid switching, creating jitter and instability in the output.

The hysteresis is implemented with two resistors that feed part of the output voltage back into the reference input. The result is two distinct threshold voltages: $V_{th+}$ for the low-to-high transition and $V_{th-}$ for the high-to-low transition. Noise must exceed this deadband to trigger a transition, suppressing false switching and ensuring clean edges.

The Carrier Amplitude Challenge

There's a subtlety: the carrier signal must have sufficient amplitude to dominate noise. With no series resistor, the maximum peak-to-peak carrier amplitude would be:

$$V_{c,pp,\max }=2V_{th}$$

For the 0.33-volt threshold of the Knowm memristors, that's a maximum of 660 millivolts, and practical operation with safety margins would likely keep it below 200 millivolts. That's dangerously close to the noise floor for many environments. The series resistor $R_5$ addresses this by allowing the carrier amplitude to scale with the supply voltage:

$$V_{c,pp,\max }=\frac{R_5}{R_2}{V}_{dd}+2V_{th}$$

With appropriate resistor choices, the carrier amplitude can be increased to provide robust noise immunity while maintaining the memristor below its threshold.

What They Found

Frequency Programmability at Practical Levels

The experimental validation produced PWM signals at four different programmed switching frequencies: 144.7 kilohertz, 159.6 kilohertz, 181.7 kilohertz, and 204.2 kilohertz. These weren't theoretical calculations—they were measurements taken from a prototype circuit built with commercially available Knowm M+SDC Cr-type memristors.

The relationship between frequency and resistance followed expectations: lower resistance produced higher frequency. At 14.73 kilo-ohms, the switching frequency was 144.7 kilohertz. At 13.73 kilo-ohms, it rose to 159.6 kilohertz. At 10.79 kilo-ohms, it reached 181.7 kilohertz. At 8.35 kilo-ohms, it peaked at 204.2 kilohertz.

A discrepancy appeared between calculated and measured frequencies, which the researchers attributed to nonideal effects in the prototype: parasitic capacitances, board-level parasitics, and component tolerances. The memristor model assumes ideal behavior; real hardware introduces second-order effects. But the measured values clustered in the expected range, demonstrating that the principle works at the intended operating frequencies.

PWM Controller Power Consumption Comparison

Label	Value
DSP-based (100%)	100
Conventional Analog	55
Memristor-based	8

Power consumption relative to DSP baseline. The memristor-based circuit achieves 92% reduction compared to a TI F28379D DSP while maintaining full programmability.

The tunable range—from 144.7 to 204.2 kilohertz—represents a span of approximately 34% above and below the center frequency. For comparison, SiC and GaN power converters, which represent the cutting edge of power electronics efficiency, typically operate in the 50 to 500 kilohertz range. A 60-kilohertz window within this band is practically useful.

The Power Comparison That Matters

The more striking finding concerns power consumption. The researchers compared their memristor-based PWM circuit against a Texas Instruments LAUNCHXL-F28379D development board—a widely used DSP platform in power converter applications—running a buck converter at 100 kilohertz. Both setups generated a single PWM signal to control the same power stage.

Programmable PWM Frequency at Different Memristor Resistances

Label	Value
14.73	144.7 kΩ
13.73	159.6 kΩ
10.79	181.7 kΩ
8.35	204.2 kΩ

Measured switching frequencies at four programmed memristor resistance values. Frequency increases as resistance decreases, demonstrating tunable control from 144.7 kHz to 204.2 kHz.

The memristor-based circuit achieved a 92% reduction in power consumption compared to the DSP. The researchers also tested a conventional analog PWM implementation for comparison, which showed intermediate power consumption between the two extremes. The analog circuit is more efficient than the DSP but less efficient than the memristor approach—and crucially, lacks programmability.

The exact power numbers weren't fully detailed in the available paper excerpt, but the relative reduction is stark: for the same functional outcome—controlling a buck converter through PWM—the memristor circuit consumes roughly one-eighth the energy of the DSP equivalent.

Duty Cycle Control

Frequency control is one dimension of PWM; duty cycle control is another. The buck converter validation tested duty cycles of 20%, 40%, 60%, and 80% at the 181.7 kilohertz operating point. The measured duty cycles—20.04%, 40.06%, 59.44%, and 80.95%—tracked closely to the targets, confirming that the hysteresis comparator approach produces accurate duty cycle resolution in addition to programmable frequency.

The output voltages across these duty cycles validated correct converter operation. At a 15-volt input, the expected output for a buck converter is the input multiplied by the duty cycle. The measured values aligned with theory, confirming that the PWM signal faithfully drives the power stage.

Programmable Resistance States

Beyond the main experiments, the researchers explored a wider resistance programming space. Five distinguishable resistance states were programmed within a range from approximately 1 to 8 kilo-ohms, each showing reproducible behavior across repeated trials. Red markers in the documentation showed individual measurements; black horizontal lines showed the mean of each programmed state.

The ability to establish multiple stable resistance states is essential for the frequency-programmability claim. If the memristor could only be set to a single value, there would be no programmability—only a fixed frequency. The demonstration of five distinct states, reliably reproducible, validates the approach.

Why This Changes Things

The Efficiency Frontier in Power Electronics

Power converter efficiency matters enormously—and not only for the obvious reason of reducing electricity bills. In high-power applications, even small percentage improvements translate to megawatts of saved generation capacity, reduced cooling requirements, and lower infrastructure costs. A server data center consuming 100 megawatts that improves its power conversion efficiency by 1% saves a megawatt continuously; at typical electricity prices, that's approximately a million dollars per year in operating costs.

The 92% figure deserves context. It refers to the PWM generation circuit specifically, not the entire converter. The power stage—the actual transistors and magnetics that switch and transform energy—dominates the losses in many converters. But control circuitry power consumption grows more significant as the main stage becomes more efficient, and it becomes critical in battery-powered and edge-computing applications where every milliwatt matters.

In electric vehicles, where the inverter switching at tens of kilohertz determines how efficiently battery energy reaches the motor, low-power control electronics mean more range per charge. In solar microinverters that attach directly to panel backs, where the electronics must operate reliably for 25 years with no maintenance, lower power consumption reduces thermal stress and extends component lifetimes.

Closing the Neuromorphic Loop

Memristors have been celebrated for their potential in neuromorphic computing—systems that mimic the architecture of biological brains. Neural networks running on memristor crossbars perform matrix multiplications with extraordinary efficiency because the computation happens directly within the memory array, eliminating the constant data movement that burns energy in conventional processors. Researchers have reported energy savings of approximately 99.9% for in-memory computing applications compared with conventional architectures, and up to 308 times improvement over SRAM-based AI accelerators.

But practical deployment has lagged the theoretical promise. A memristor-based AI accelerator may process sensor data with extraordinary efficiency—but if it then feeds that result into a conventional DSP to generate the PWM signal that actually controls the motor, the energy savings are partially negated by the conversion overhead.

"Since PWM generation is the final essential stage of power converter control," the paper states, "a programmable memristor-based PWM circuit is needed to fully realize the energy-efficiency benefits of memristive computing." The circuit demonstrated at Penn State provides exactly this bridge: the first stage where memristive computation can flow directly to physical actuation without an intervening digital conversion step.

The analogy to biological motor control is instructive. In the human nervous system, the brain's "neural computation" flows through spinal cord circuits directly to muscle fibers, with no conversion to digital representation and back. The memristor-based PWM circuit enables a similar direct path in engineered systems. A memristor crossbar could process sensor inputs, generate control signals, and produce PWM outputs—all in the analog domain, with no ADC or DSP in the critical path.

Prior Work and the Frequency Gap

The Penn State work builds on earlier attempts, but substantial improvements separate it from prior demonstrations. The first memristor-based PWM generator, by Hassanein and colleagues, was limited to 10 hertz—far below the audio range, let alone the hundreds-of-kilohertz frequencies required for power electronics. A preliminary circuit by the same Penn State team improved this to 20 kilohertz, still below the practical range for modern converters.

The key advancement was topological. Previous designs placed the memristor directly in the feedback loop or in a voltage divider, where increasing the programmable resistance range to accommodate real-world memristor characteristics came at the cost of lower operating frequency. The new design decouples these parameters—allowing both wide resistance range and high frequency simultaneously.

The researchers explicitly note the target application: "Such frequencies are necessary for modern SiC/GaN converters, where switching frequencies typically range from tens to hundreds of kilohertz." Silicon carbide and gallium nitride semiconductors are rapidly displacing traditional silicon in high-efficiency applications because they can switch faster with lower losses. A PWM generator that works at 200 kilohertz can control these semiconductors; one that works at 20 hertz cannot.

The Programmability Advantage Over Pure Analog

Conventional analog PWM generators achieve low power consumption—but at the cost of flexibility. Change the desired switching frequency, and you change the hardware: swap a resistor, recalculate capacitor values, verify the timing tolerances. In a research laboratory, this is inconvenient. In a commercial product, it's prohibitive.

The memristor-based approach preserves the low-power advantage of analog circuits while adding the programmability of digital control. The frequency is set by writing a resistance value to the memristor; to change frequencies, write a different resistance. No hardware changes. No redesign. The same circuit board serves multiple applications.

This programmability enables adaptive control strategies that would be impractical with fixed analog circuits. A power converter could adjust its switching frequency based on load conditions—operating at lower frequency (lower switching losses) when efficiency matters most, and higher frequency when physical size matters more. It could implement frequency dithering to spread electromagnetic interference across a wider band. It could respond to thermal conditions, reducing frequency when overheating to reduce losses.

Practical Implementation Considerations

The memristor programming workflow requires a brief calibration step before operation. The closed-loop programming procedure typically converges within a few pulses, and the researchers demonstrated 100% success rate in reaching the ±10% tolerance window. But this is still a step that doesn't exist in conventional PWM generators, which power up with predefined frequencies.

For production systems, the programming could happen during manufacturing calibration, with the resistance state locked in. Alternatively, in systems requiring frequent reconfiguration, the programming time represents a small overhead relative to the operating period—microseconds of programming versus milliseconds or seconds of continuous operation.

The stochastic nature of memristor programming is addressed by closed-loop verification, but it does mean that the exact resistance achieved varies slightly around the target. The ±10% tolerance represents an engineering compromise: tight enough for stable frequency control, loose enough to be achievable with current device technology. Future memristors with more deterministic switching could enable tighter tolerances and more precise frequency control.

What's Next

The Path to Practical Deployment

The demonstrated prototype operates at 200 kilohertz with a 60-kilohertz tuning range. For SiC/GaN converters, this puts the technology in a practically useful band—between the tens-of-kilohertz operation common for large industrial drives and the hundreds-of-kilohertz operation targeted for compact high-frequency designs.

Extending the frequency range in both directions would broaden applicability. Lower frequencies matter for large-scale power conversion where magnetic components are sized for efficiency at lower switching rates; higher frequencies matter for aerospace applications where weight is paramount and the lower losses of wide-bandgap semiconductors justify aggressive miniaturization.

The power consumption advantage over DSPs—92% reduction—represents an enormous headroom for optimization. As the researchers refine the circuit implementation, further reductions may be achievable. Every milliwatt saved in control electronics is a milliwatt that doesn't need to be dissipated as heat, a milliwatt that extends battery life, a milliwatt that reduces cooling infrastructure.

Integration Challenges

The current prototype uses discrete components mounted on a printed circuit board. Commercial deployment would require integration—either onto a mixed-signal integrated circuit that co-locates the memristor with its programming and control circuitry, or onto a module that integrates the memristor with conventional CMOS driving and protection circuitry.

Memristor integration with standard CMOS processes has been demonstrated by several research groups, though manufacturing yield and consistency remain challenges. The Penn State work provides strong motivation for pursuing this integration: the efficiency gains are substantial enough to justify significant engineering investment.

Temperature stability is another consideration. Memristor resistance can vary with temperature, which could shift the programmed frequency during operation. The hysteresis comparator provides some tolerance for carrier amplitude variations, but frequency stability over temperature would need characterization and possibly compensation for demanding applications.

The Broader Memristive Computing Landscape

The PWM circuit is one piece of a larger vision. A complete memristive computing system might comprise memristor crossbars for matrix-vector multiplication, analog circuits for signal conditioning and activation functions, and the PWM generation stage for physical control output. All of these stages would operate in the analog domain, with no digital conversion overhead except at system boundaries.

Such a system would be particularly suited to edge computing applications, where sensor data must be processed locally under tight power budgets—smart sensors in industrial monitoring, battery-powered medical devices, autonomous vehicle controllers. The combination of AI capability and efficient power conversion control in a single low-power platform could enable previously impractical distributed intelligence.

"This work closes the missing interface between memristive computing and practical power converter control," the researchers state, "enabling a fully programmable low-power control architecture." The interface—long identified as a missing piece—is now, in prototype form, present.

Memristor Technology Trajectory

Current commercial memristors like the Knowm devices used in this work represent early-generation technology. They exhibit stochastic switching behavior, limited resistance windows, and threshold voltages that constrain circuit design. The history of semiconductor technology suggests that as manufacturing processes mature, these characteristics will improve.

Higher threshold voltages would allow larger read pulses with improved noise immunity and broader resistance ranges. More deterministic switching would reduce programming times and improve consistency. Lower variability would enable tighter tolerance bands and more precise frequency control.

The memristor is no longer a laboratory curiosity. Companies like Knowm, Weebit Nano, and others are working to commercialize the technology for memory, computing, and signal processing applications. The Penn State work provides a concrete application-pull: a clear market need (efficient PWM generation) that memristors can uniquely address.

Open Questions

Several questions remain open. How does the memristor's long-term reliability compare with conventional components? Memristors can undergo billions of write cycles, but the resistance drift over time and with temperature cycling must be characterized for safety-critical applications like automotive power conversion. How does the circuit perform across the full temperature range required for automotive or industrial environments? Parasitic effects that are second-order at room temperature may become dominant at extremes. Can the frequency range be extended further while maintaining programmability? The current design achieves approximately 60 kilohertz of tuning range; wider range would enable more adaptive control strategies.

The researchers acknowledge that the discrepancy between calculated and measured frequencies points to nonideal effects—parasitic capacitances, component tolerances—that a production design would need to address. The prototype demonstrates feasibility; productization requires engineering.

A Niche That Becomes Mainstream

Every transformative technology follows a similar arc: early demonstrations show possibility, engineering refines the implementation, manufacturing reduces costs, and applications expand beyond the original niche. Memristors are at the early stages of this arc, with the Penn State PWM circuit representing a demonstration of possibility.

The 92% power reduction isn't a laboratory curiosity—it's a number that changes the economics of power converter design. At that efficiency level, the trade-offs that once favored DSPs (flexibility at the cost of power) no longer clearly win. A programmable low-power solution suddenly becomes competitive on both dimensions.

Silicon carbide and gallium nitride semiconductors are creating new demands for high-frequency control electronics. The market for efficient wide-bandgap power converters is growing rapidly as these materials displace silicon in automotive, industrial, and renewable energy applications. A PWM controller that matches the frequency capabilities of these semiconductors while consuming 92% less power arrives at precisely the right moment.

The memristor remembered. Now it's ready to control.