Research Papers in Progress

We present a novel partitioned crossbar architecture that mitigates IR drop effects in large-scale memristor arrays. Our approach divides a 256×256 crossbar into 128×128 sub-arrays, achieving 45% reduction in maximum voltage drop and 3.2× improvement in computation accuracy. Experimental results on fabricated HfO₂-based devices demonstrate 20 TOPS/W energy efficiency for neural network inference.

This work presents a multi-scale modeling framework for HfO₂-based memristors, bridging atomistic simulations with system-level performance. We develop a compact model that captures switching dynamics, variability, and reliability metrics. The model is validated against experimental data from over 10,000 devices, showing excellent agreement across six orders of magnitude in timescale.

We introduce a novel training methodology that accounts for device-to-device and cycle-to-cycle variations in memristor crossbars. Our approach uses stochastic gradient descent with hardware-aware noise injection, achieving 96.2% accuracy on CIFAR-10 despite 15% device variation. The method reduces accuracy degradation by 8.5× compared to conventional training.

We demonstrate real-time object detection running entirely on memristor crossbar arrays, achieving 1000+ FPS on 640×480 images with 82.3% mAP on COCO dataset. The implementation uses a custom YOLO-variant optimized for analog computing, consuming only 2.5W total power.

This paper presents the first demonstration of 3D-stacked memristor crossbar arrays with 8 active layers. We develop novel via technology for inter-layer connectivity and thermal management solutions. The 3D architecture achieves 12.5 Gb/cm² density while maintaining switching characteristics comparable to 2D arrays.

We demonstrate reliable 4-bit (16-level) programming in HfO₂-based memristors using optimized pulse sequences. Our approach achieves 2% cell-to-cell variation and >10⁵ endurance cycles. The multi-level capability increases effective storage density by 4× without area overhead.