Access Official Technical PDF

Share Summary:LinkedInEmail

A passive memristor crossbar could help bring larger neural-network workloads onto specialized chips by storing synaptic weights directly in a dense array. But density alone is not enough: analog computing depends on devices that can be programmed to consistent conductance values, even when cells share lines and can affect one another. A 64×64 metal-oxide crossbar reported in 2019 offers a useful case study in addressing that challenge, with measured conductance tuning and a hardware demonstration of vector-matrix multiplication.

Why passive memristor crossbars are challenging

Memristors, also known as resistive RAM (ReRAM) devices, change electrical resistance in response to applied voltage. In an analog crossbar, each device can represent a weight through its conductance. Applying input voltages to one set of lines and measuring currents on the other performs vector-matrix multiplication in the array—the core operation used throughout neural-network inference.

Unlike a one-transistor, one-resistor (1T1R) memory cell, a passive crosspoint has no dedicated select transistor. That arrangement can support high device density and avoid the cell-level transistor, but it also makes programming more demanding. A write pulse intended for one device can partially affect other devices sharing its rows or columns. This half-select disturbance, combined with variation in device switching behavior, can make it difficult to set and preserve accurate analog weights.

For industrial automation and edge-computing designers, the distinction matters: a dense array is useful only if its cells can be programmed precisely enough for the required computation. The reported work assessed both the fabrication of a larger passive array and its ability to represent analog values across many devices.

Fabrication choices for a larger memristive crossbar

The circuit contained 64×64 crosspoints using titanium, aluminum, and titanium nitride electrodes, with an aluminum oxide and titanium oxide switching layer. Its process used etch-based patterning and chemical-mechanical polishing (CMP) to produce planar, isolated electrodes. The fabrication flow was designed with a low temperature budget and compatibility with foundry processes in mind, including potential back-end-of-line integration.

Several process decisions addressed issues that can become more consequential as arrays grow. Aluminum was selected for electrode conductivity, with titanium nitride used to protect it from oxidation. Etching, rather than lift-off, enabled the fabrication of electrodes with a larger aspect ratio. CMP helped improve electrode surfaces and limit sidewall residue that could create unwanted connections or interfere with line isolation.

The switching layer and thermal annealing also played a role in device behavior. The authors used annealing to adjust the oxygen-vacancy profile in the titanium oxide, helping reduce as-fabricated conductance and improve device uniformity. An aluminum oxide layer suppressed current at small voltages and helped limit leakage. These steps illustrate how cell materials, interfaces, and patterning must be considered together when building a passive array.

Measured yield and conductance tuning

The 64×64 array contained 4,096 crosspoints. The reported device yield was approximately 99%, with 45 devices identified as unswitchable. Average normalized variation in switching threshold voltages was about 26%. That variation did not prevent accurate array-level programming in the reported tests, but it underlines why a programming method must accommodate device-to-device differences.

To set conductance values, the researchers used a write-verify approach: apply a programming pulse, read the device conductance, and adjust subsequent pulses based on the measured result. A half-biasing scheme reduced disturbance to unselected devices during writes. Programming the full array in multiple cycles allowed devices affected by earlier writes to be tuned again.

In one demonstration, the array represented a grayscale image across its 4,096 devices. Excluding the unswitchable cells, about 98% of devices were tuned within 5% of their target, and the average tuning error was approximately 3.76%. The work also demonstrated 1% tuning error for a specific device across selected conductance levels. These results are measurements from the reported experiments, rather than a general guarantee for other devices or fabrication processes.

Testing vector-matrix multiplication with MNIST

The team also tested whether the crossbar could support a simple classification task. A software-trained, single-layer perceptron used 64 inputs and 10 outputs. Images from the MNIST handwritten-digit dataset were down-sampled from 28×28 pixels to 8×8, allowing each image to be represented by a 64-value input vector. The network’s weights were mapped to conductances in a 64×10 portion of the array.

Input voltages represented pixel values, and currents measured at the output lines represented the vector-matrix multiplication results. The output with the largest current indicated the predicted digit. Neuron functions, including bias, were handled in software; the hardware test therefore focused on the crossbar’s multiply operation rather than implementing a complete neural network on the chip.

Measured classification performance was close to the software-modeled result for the same network, with the reported experimental result 1.87% behind the ideal software model at the most accurate weight import. Errors included effects from unswitchable devices and cases where the output currents for competing classes were close. The demonstration shows why array-level performance depends not just on individual device precision, but also on how programming errors and faulty cells interact with the computation.

Practical implications for automation hardware

For engineers evaluating analog in-memory computing, the study offers evidence that passive crossbars can combine substantial device counts with useful analog tuning accuracy. It also clarifies the integration work involved: electrode resistance, line isolation, process uniformity, switching variation, and half-select effects all influence whether a crossbar can reliably perform its intended computation.

The results do not establish a complete production-ready neuromorphic system. The classifier was small, used ex-situ software training, and emulated neuron functions outside the array. The work also identifies further engineering needs, including integration with conventional semiconductor circuits and reducing device operating and write currents.

Conclusion

A 64×64 passive memristor crossbar demonstrated approximately 99% device yield, average conductance-tuning error below 4% for a grayscale pattern, and MNIST classification close to the software-model limit for the tested network. For industrial automation decision-makers, the key takeaway is that dense analog memory depends on coordinated advances in fabrication, device uniformity, and programming methods. Those factors determine whether crossbar density can translate into practical, reliable vector-matrix computation.

Access Official Technical PDF

About The Author