Guide to Spintronic Crossbar Arrays: From Binary Switches to Analog Computing

1. Introduction: Why Crossbars?

In a conventional computer, the processor and memory are separated. To perform a calculation, data must travel from memory to the processor, be processed, and then travel back. This back-and-forth, known as the von Neumann bottleneck, limits speed and consumes energy.

To quantify this, the von Neumann bottleneck is essentially a bandwidth limitation. If a processor operates at frequency $f$ and the data bus transfers $B$ bits per cycle, the processor often sits idle waiting for data. In-memory computing effectively makes the memory its own processor, eliminating this delay.

A crossbar array offers an alternative. Imagine a grid of wires: horizontal rows and vertical columns. At each intersection, a tiny electronic device connects the row to the column. When voltages are applied to the rows, currents flow through these devices and are collected at the columns. The output currents are the result of a matrix-vector multiplication, performed in a single, parallel step. This is the essence of in-memory computing.

–${\mathrm{V}}_4$ are applied to the horizontal rows; output currents ${\mathrm{I}}_1$–${\mathrm{I}}_4$ are collected from the vertical columns. The conductance at each intersection ${\mathrm{G}}_{\mathrm{i}\mathrm{j}}$ determines how much current flows from a given row to a given column. The column currents are the weighted sums of the row voltages, performing a matrix-vector multiplication in a single analog step.

Figure 1. Conceptual crossbar array. Input voltages ${\mathrm{V}}_1$

Figure 1. Conceptual crossbar array. Input voltages ${\mathrm{V}}_1$

2. What is a Spintronic Device?

A spintronic device, short for “spin-based electronic device”, uses the quantum spin of electrons, not just their charge, to store and process information. An electron can have two spin states: “spin up” or “spin down.” In a magnetic material, these spin states correspond to different magnetic orientations.

The most common spintronic device for crossbars is the magnetic tunnel junction (MTJ). An MTJ consists of two ferromagnetic layers separated by a thin insulating barrier:

• Reference layer: Fixed magnetization.
• Free layer: Magnetization can be switched.

When the magnetizations of the two layers are aligned (parallel), the resistance is low. When they are opposite (anti-parallel), the resistance is high. This difference, called tunnel magnetoresistance (TMR), allows the device to store a binary state (0 or 1). We measure the efficiency of this state change using the TMR ratio, which defines the “read window” or the contrast between our 0s and 1s:

$\mathrm{TMR}={\displaystyle \frac{R_{\mathrm{AP}}-R_{\mathrm{P}}}{R_{\mathrm{P}}}}\times 100\mathrm{\%}$Plain-text equation: TMR = (R_AP − R_P) / R_P

where $R_{\mathrm{P}}$ is the low-resistance parallel state and $R_{\mathrm{AP}}$ is the high-resistance anti-parallel state. A higher TMR makes the device more resilient to electronic noise.

This binary state is what gives the crossbar its programmable conductance.

Figure 2. illustrates an MTJ structure. Magnetic tunnel junction (MTJ). (a) Parallel magnetization—low resistance. (b) Anti-parallel magnetization—high resistance. The tunnel barrier separates the two ferromagnetic layers.

Figure 2. illustrates an MTJ structure. Magnetic tunnel junction (MTJ). (a) Parallel magnetization—low resistance. (b) Anti-parallel magnetization—high resistance. The tunnel barrier separates the two ferromagnetic layers.

3. The Crossbar Architecture: Inputs and Outputs

A crossbar array is built by placing spintronic devices at the intersections of perpendicular nanowires. Each intersection is called a crossbar junction. The array has:

• Rows (word lines): Carry input voltages. In our context, these are the signals we want to process, for example, sensor voltages.
• Columns (bit lines): Carry output currents. These are the results of the computation.
• Device at (i,j): Has a programmable conductance $G_{ij}$

. The conductance is set by the state of the spintronic device (binary or multi-level).

The output current at column $i$ is the physical result of two fundamental laws of physics working together:

• Ohm’s Law: Multiplication is performed by the conductance at the junction ($I=V\cdot G$

).

2.

Kirchhoff’s Current Law: Addition is performed as currents from all rows merge into the column wire.

The total current for any column $i$ is:

$I_i=\sum _{j=1}^n{V}_j\cdot G_{ij}$Plain-text equation: I_i = Σ_j (V_j · G_ij) (junction currents summed by KCL along a column)

In matrix form this is the classic Matrix-Vector Multiplication (MVM):

Plain-text matrix form: I = G · V (currents = conductance matrix × voltage vector)

Mini example (2×2 crossbar):

Let the input voltages be V = [V1, V2]ᵀ = [0.20 V, 0.10 V]ᵀ.

Let the conductance matrix be G = [[1 mS, 2 mS], [0.5 mS, 1 mS]].

Then I = G·V gives:

I1 = (1 mS)(0.20 V) + (2 mS)(0.10 V) = 0.20 mA + 0.20 mA = 0.40 mA

I2 = (0.5 mS)(0.20 V) + (1 mS)(0.10 V) = 0.10 mA + 0.10 mA = 0.20 mA

Interpretation: the crossbar produces two analog output currents that already contain the multiply–accumulate result.

Important: even if each cell stores only 0/1 (two conductance values), the column output is still an analog current sum. It becomes ‘digital’ only after sensing and thresholding/ADC.

$\mathbf{I}=\mathbf{V}\cdot \mathbf{G}$where $\mathbf{V}$ is the row voltage vector and $\mathbf{G}$ is the conductance matrix.

. Output currents (I1–I4) are the sum of contributions from all rows.

Figure 3. shows a 4×4 crossbar array with devices at each junction. A 4×4 crossbar array. Input voltages (V1–V4) are applied to rows. Each junction contains a spintronic device with conductance $G_{ij}$

Figure 3. shows a 4×4 crossbar array with devices at each junction. A 4×4 crossbar array. Input voltages (V1–V4) are applied to rows. Each junction contains a spintronic device with conductance $G_{ij}$

4. Binary (Single-Level) Operation

In its simplest form, each spintronic device is binary: it can be either in a low-resistance state (closed latch) or a high-resistance state (open latch). These states represent logic values “1” (low resistance) and “0” (high resistance), respectively.

In this mode, the conductance $G_{ij}$ is a discrete variable restricted to a set of two values:

$G_{ij}\in \{G_L,G_H\}$where $G_H$ (high conductance) represents a logic 1 and $G_L$ (low conductance) represents a logic 0. This makes the array act like a massive parallel “pattern matcher”.

In this mode, the conductance $G_{ij}$is restricted to two values:

$G_{ij}=\left\{\begin{array}{cc}{G}_{\mathrm{high}}& \left(\mathrm{logic}1\right)\\ G_{\mathrm{low}}& \left(\mathrm{logic}0\right)\end{array}\right.$Switching: The state can be changed by applying a current pulse through the device. If the current exceeds a certain threshold, the magnetization of the free layer flips, changing the resistance. This is done during a programming phase before normal operation.

Advantages of binary devices:

• Non-volatile: state is retained without power.
• Fast: switching in nanoseconds.
• Endurance: virtually unlimited write cycles.

Output: The output current at each column is a weighted sum of the inputs, where the weights are either $G_H$ (representing 1) or $G_L$ (representing 0). This produces a binary-weighted analog output. For example, if we set $G_H=100\mu \mathrm{S}$ and $G_L=1\mu \mathrm{S}$, then the output is a sum of the input voltages multiplied by either 100 or 1, depending on the stored bits.

Note: the crossbar itself is still performing analog physics. ‘Binary’ here refers to the stored weight levels (two conductances). To use the result in digital logic, we sense the column current and ota s sa using a comparator or ADC.

Figure 4. shows the binary switching characteristic. Binary switching of an MTJ. A current pulse above threshold switches the device from high resistance (R_high) to low resistance (R_low). The inset shows the hysteresis loop of magnetization.

Figure 4. shows the binary switching characteristic. Binary switching of an MTJ. A current pulse above threshold switches the device from high resistance (R_high) to low resistance (R_low). The inset shows the hysteresis loop of magnetization.

5. Multi-Level (Analog) Operation

Binary devices store only one bit per cell. For applications that require continuous weight values, such as neural networks or advanced control algorithms, multi-level devices are needed. In a multi-level spintronic crossbar, each device can have more than two distinct conductance levels, allowing it to represent a continuous analog value.

How it works: Instead of having a uniform free layer that switches all at once, a multi-level MTJ uses a granular nanostructure. The free layer consists of many small magnetic grains (8–10 nm) that switch at slightly different currents. As the programming current increases, more grains switch, and the overall resistance changes in small steps.

The “analog” behavior in granular MTJs is actually the statistical sum of many microscopic magnetic flips. If a device contains $N$ grains, the total conductance is approximately the sum of all grains that have been switched to the parallel state:

$G_{\mathrm{total}}\approx \sum _{k=1}^N{\mathrm{state}}_k\cdot g_k$where $g_k$ is the conductance contribution of a single grain and ${\mathrm{state}}_k\in \left\{\mathrm{0,1}\right\}$ indicates whether that grain is in the parallel state. This allows us to map digital “weights” (like those in a neural network) to specific physical conductance levels.

By carefully controlling the current pulse amplitude, any intermediate resistance can be set.

State of the art: Research has demonstrated up to 2 bits per cell (4 levels) reliably, and development is ongoing toward 4 bits per cell (16 levels). These multi-level devices are still in the research phase, but they hold great promise for true analog computing.

Output: In a multi-level crossbar, each conductance $G_{ij}$ can take one of many values (e.g., 16 levels). The output current at column $i$ is still $I_i={\sum }_j{V}_j\cdot G_{ij}$, but now the weights are continuous, allowing precise analog computations.

Figure 5. illustrates the granular nanostructure. Granular nanostructure for multi-level MTJ. Bottom granular layer (CoPt alloy) with magnetic grains (8–10 nm). Top continuous layer (CoFeB) coupled via exchange. The magnetization of the top layer follows the average of the bottom grains, enabling intermediate states.

Figure 5. illustrates the granular nanostructure. Granular nanostructure for multi-level MTJ. Bottom granular layer (CoPt alloy) with magnetic grains (8–10 nm). Top continuous layer (CoFeB) coupled via exchange. The magnetization of the top layer follows the average of the bottom grains, enabling intermediate states.

6. Programming the Crossbar

To use a crossbar for a specific computation, we must first set the conductances $G_{ij}$ to the desired values. This is done during a programming phase, separate from the normal operation. For binary crossbars, programming is simply writing bits (0 or 1) into each cell using current pulses. For multi-level crossbars, we apply precisely controlled current pulses to achieve the desired intermediate conductance.

Practical note (multi-level weights): programming is usually iterative (“program-and-verify”). A controller applies a short pulse, reads the resulting conductance, and repeats until the target level is reached within tolerance. This compensates for device variability and drift.

Once programmed, the crossbar acts as a fixed analog computing engine. The input voltages are applied to the rows, and the output currents appear at the columns in nanoseconds. No further programming is needed until we want to change the function.

In this report, we do not go into the details of how the programming circuits are designed or the specific materials used. Instead, we focus on the fundamental operation: inputs (voltages on rows), outputs (currents on columns), and the role of the programmed conductances.

7. What the Crossbar Reads and Outputs

7.1. Inputs

• Rows: Analog voltages (continuous values) or digital voltages (binary). These are the signals we want to process, for example, sensor readings, cell voltages, or control signals.
• Programming: During programming, current pulses are applied to set the conductances. This is done once before operation.

7.2. Outputs

Plain-text equation: I_i = Σ_j (V_j · G_ij) (junction currents summed by KCL along a column)

• Columns: Analog currents, given by $I_i={\sum }_j{V}_j\cdot G_{ij}$

• Binary case: Output is a weighted sum with binary weights (either $G_H$

or $G_L$).

• Multi-level case: Output is a true analog weighted sum with continuous weights.

• Directly used as control signals (e.g., to trigger comparators).
• Converted to digital via ADCs for further processing. The resolution of the Analog-to-Digital Converter (ADC) must match the precision of the crossbar. For example:

• Fed into another crossbar for multi-layer computations.

8. Types of Spintronic Crossbars: State of the Art

Type	Device	Levels	Output	Status	Reference
Binary STT-MRAM	Standard MTJ	2	Binary-weighted analog	Commercial	[4] (overview of spintronic memristors) or product datasheet
Multi-level STT-MRAM	Granular MTJ	4+	Continuous analog	Research	[1] (Gupte et al., npj Unconventional Computing, 2026)
SOT-MRAM	MTJ with heavy metal	2	Binary-weighted analog	Emerging	[3] (Zvezdin, SpinEdge seminar, 2025) or [5] (Deng et al., Advanced Functional Materials, 2024)
Domain-Wall	MTJ with extended track	Multiple	Analog	Research	[4] (general spintronics review) or classic reference (e.g., Parkin et al., Science, 2008)

8.1. Binary STT-MRAM

• Mature technology already in production.
• High density, good endurance, CMOS-compatible.
• Used for embedded memory and in-memory computing.

8.2. Multi-Level Granular MTJ

• Novel structure using granular nanostructures.
• Near-continuous states without special lithography.
• Enables analog synaptic weights for neuromorphic computing.

8.3. Spin-Orbit Torque (SOT) Devices

3.

Three-terminal configuration: separate read and write paths.

4.

Faster writing, lower power, but larger area.

5.

Can also be adapted for multi-level operation.

8.4. Domain-Wall Devices

• Use motion of magnetic domain walls along a track.
• Multiple states by positioning the wall.
• Require special geometry (e.g., notches, curved tracks).

8.5. Commercial Status: What Exists Today (Binary vs. Multi-Level)

Binary (commercially available): spintronic memory (MRAM, especially STT-MRAM) is a mature product category. Several vendors sell MRAM memory chips that use binary MTJs (two stable resistance states), mainly as non-volatile embedded or stand-alone memory. These parts are ‘spintronic’ and are commercially available, but they are sold as memory, not as analog matrix-multiply accelerators. [15,16]

Compute-in-memory (prototype demos): major semiconductor groups have demonstrated MRAM-based in-memory computing macros (using MRAM arrays plus sensing circuits) to accelerate multiply–accumulate operations. This shows a technology path from commercial MTJ memory toward computing arrays, but these are still closer to advanced prototypes than mass-market products. [17]

Multi-level (mostly research): robust multi-bit MTJ ‘synapses’ and true analog crossbar computing are active research topics. Multi-level operation requires tighter control of device states, programming/verify loops, and higher-precision sensing. For this reason, most multi-level spintronic crossbar results are reported in research papers rather than product datasheets. (See, for example, [1].)

9. Summary: What a Spintronic Crossbar Can Do

Capability	Description
In-memory computing	Compute directly where data is stored, eliminating von Neumann bottleneck. [2,8,9]
Parallel operation	All rows and columns active simultaneously.
Matrix-vector multiplication	$I=U\times G$in nanoseconds.
Binary or analog	Single-level (binary) or multi-level (continuous) weights.
Non-volatile	States retained without power.
Reconfigurable	Program $G$matrix for different functions.

10. Current Limitations and Recent Advances

While spintronic crossbars promise exceptional speed and energy efficiency, they are not yet “perfect components.” Real designs must account for temperature dependence, manufacturing variability, stability (especially for multi-level states), and array-level non-idealities such as IR drop and sneak-path currents. The good news is that many of these limitations are engineering challenges rather than fundamental blockers, and the last few years have delivered major improvements in stability and manufacturability (particularly for perpendicular MTJ stacks and mature MRAM processes).

10.1. Temperature Sensitivity

Spintronic devices rely on magnetic ordering, so key parameters drift with temperature. As temperature rises, the tunnel magnetoresistance (TMR) ratio and the resistance contrast between states typically decrease, reducing read margin. Switching thresholds (critical current/voltage) and retention margins also vary with temperature because they depend on the thermal stability factor. Industrial MRAM products therefore specify retention targets and validate them through high-temperature qualification, showing that stable operation is achievable at scale. For sensing applications, the same MTJ technology is widely used in TMR sensors, where temperature-dependent sensitivity and offset drift are mitigated through calibration and compensation methods (hardware or software). These sensor techniques are directly relevant to crossbar readout and long-term stability management. [18,21,22,23,24]

10.2. Manufacturing Variability and Materials Control

In a crossbar array, each MTJ is fabricated in a dense matrix; even with state-of-the-art lithography there are small variations in film thickness, grain structure, and etch profile across a wafer. Because an MTJ is a nanoscale multilayer (often CoFeB/MgO/CoFeB), the barrier/interface quality strongly affects TMR, resistance-area (RA) product, and switching distributions. These variations broaden the resistance (or conductance) distributions of the nominal ‘same’ state and can reduce the number of reliably distinguishable levels, especially for multi-level operation. Practical multi-level systems therefore rely on iterative program-and-verify and periodic recalibration to compensate variability and drift. [1,18,19,20]

10.3. Endurance, Retention, and Read Disturb

MTJs can offer very high write endurance, which is excellent for repeated reprogramming. However, repeated reads can cause read-disturb if the read current is too high; modern sensing schemes keep read currents well below switching thresholds. Retention is the complementary requirement: the stored state (binary or intermediate) must remain stable over time and temperature. Binary MRAM is already qualified for long retention, whereas multi-level analog states require tighter control and often depend on verify/re-write loops to maintain precision. [1,21]

10.4. Sneak Paths and Selector/Access Devices

In a passive crossbar, every junction shares row and column wires. During read or compute operations, unintended current can flow through neighbouring devices (sneak paths), corrupting the measured output. To mitigate this, many practical arrays use access devices (e.g., 1T-1MTJ in MRAM macros) or selector elements that enforce non-linear I–V behaviour. Selectors add area and complexity but become increasingly important as arrays scale. [9,12,21]

10.5. IR Drop, Parasitics, Noise, and Crossbar Scaling

When voltages are applied to long rows, currents through the metal lines create voltage gradients (IR drop), so the atrice sint voltage is not atrice sinto the array. Parasitic capacitances set a finite settling time, limiting how quickly outputs can be sampled. Finally, the readout chain (TIA/ADC or comparators) introduces noise and quantization error, which can dominate when conductance differences are small (multi-level operation). Common mitigation strategies include tiling large atrice sinto smaller sub-arrays, adding calibration, and designing sensing circuits matched to the required precision. [9,12]

10.6. Recent Advances: Why Stability Has Improved

Despite these challenges, the last few years have seen rapid progress that improves stability and practical deployability:

• Higher TMR and better thermal/anneal stability: perpendicular MTJ stacks have demonstrated strong thermal stability at small dimensions and high magnetoresistance, improving sensing margin and robustness. [18,19,20,21]
• Multi-level repeatability: granular MTJ approaches combined with program-and-verify can achieve tighter level distributions (multi-bit ‘synapses’), improving practical analog operation. [1]
• Manufacturing maturity: commercial MRAM products and embedded MRAM process options demonstrate that MTJ fabrication can meet industrial reliability targets. [15,16,17,21]
• Array-level mitigation: crossbar-aware circuit techniques (tiling, calibration, access devices/selectors) reduce sneak paths and IR-drop effects in larger arrays. [9,12]
• Spintronics beyond crossbars (TMR sensors): TMR sensors are widely used in industry and automotive; recent work shows effective temperature compensation (hardware/software), validating long-term stability and manufacturability of MTJ technology. [22,23,24]

Crossbar matrices still face non-idealities, but stability and manufacturability have improved significantly. Binary MRAM is already commercial, and research progress is steadily pushing multi-level crossbars toward practical engineering reality.

11. Conclusions

Spintronic crossbar arrays represent a fundamental shift in how computation can be performed. By combining storage and processing in a single array, they overcome the von Neumann bottleneck that limits conventional digital systems. Binary devices are already commercial; multi-level devices based on granular nanostructures are opening the door to true analog computing, where weights are continuous and matrix operations happen in a single parallel step. The challenges of temperature sensitivity, variability, and sneak paths have been addressed by advanced materials, selector devices, and iterative programming schemes, leading to rapid adoption of spintronic technology in embedded memory and magnetic sensing. This momentum makes the prospect of large-scale analog crossbar computing increasingly realistic.

This foundation is essential for understanding advanced applications, from ultra-fast bypass and MPPT in photovoltaic systems to neuromorphic computing and energy-harvesting IoT nodes. The same crossbar architecture serves as a versatile platform for intelligent energy management across all scales.