Universal Approach for Calibrating Large‐Scale Electronic and Photonic Crossbar Arrays

Abstract

Analog electronic and photonic crossbar arrays have been emerging as energy‐efficient hardware implementations to accelerate computationally intensive general matrix–vector and matrix–matrix multiplications in machine learning (ML) algorithms. However, the inevitable nonuniformity in large‐scale electronic and optoelectronic devices and systems prevents scalable deployment. Herein, a calibration approach is reported that enables accurate calculations in crossbar arrays despite hardware imperfections. This approach is experimentally validated in a small‐scale free‐space photonic crossbar array based on cascaded spatial light modulators and demonstrated the scalability and universality of this approach in various large‐scale electronic and photonic crossbar arrays. The improved performance of calibrated crossbar arrays in an ML model inference is further demonstrated to classify handwritten digital images.

Two-dimensional material-based memristive devices for alternative computing

“…The reported result also mentions cycle-to-cycle variability of 2.42%, while also reporting accuracy on MNIST datasets as 93.34%. Another recent result presented the synthesis of 2H-MoTe 2 with vertically aligned grain boundaries in order to create a device with more uniformity across its fabrication [ 165 ]. The MoTe 2 was fabricated using chemical vapour deposition, and the results obtained show cycle-to-cycle variation of 8.3%, a yield of 83.7% with device variability between 8.3% and 14.2%.…”

“…The result shows an improvement in processing time, voltage drops, latency, and footprint due to the densely packed layers, with a further reduction in these metrics as the number of stacked layers was increased progressively. Another approach to improving integration in large arrays utilizes calibration in order to mitigate hardware imperfections causing errors [ 165 ]. The approach aims to find the difference between the ideal behavior of the device and the actual behavior device by performing current and voltage sweeps over the operating region of the device and using the result as a correction metric.…”