Synergistic Development of Perovskite Memristors and Algorithms for Robust Analog Computing

Nanyang Ye, Qiao Sun, Yifei Wang, Liujia Yang, Jundong Zhou, Lei Wang, Guang-Zhong Yang, Xinbing Wang, Chenghu Zhou, Wei Ren, Leilei Gu, Huaqiang Wu, Qinying GuarXiv 2024 · arXiv · OpenReview

Abstract

Analog computing using non-volatile memristors has emerged as a promising solution for energy-efficient deep learning. New materials, like perovskites-based memristors are recently attractive due to their cost-effectiveness, energy efficiency and flexibility. Yet, challenges in material diversity and immature fabrications require extensive experimentation for device development. Moreover, significant non-idealities in these memristors often impede them for computing. Here, we propose a synergistic methodology to concurrently optimize perovskite memristor fabrication and develop robust analog DNNs that effectively address the inherent non-idealities of these memristors. Employing Bayesian optimization (BO) with a focus on usability, we efficiently identify optimal materials and fabrication conditions for perovskite memristors. Meanwhile, we developed "BayesMulti", a DNN training strategy utilizing BO-guided noise injection to improve the resistance of analog DNNs to memristor imperfections. Our approach theoretically ensures that within a certain range of parameter perturbations due to memristor non-idealities, the prediction outcomes remain consistent. Our integrated approach enables use of analog computing in much deeper and wider networks, which significantly outperforms existing methods in diverse tasks like image classification, autonomous driving, species identification, and large vision-language models, achieving up to 100-fold improvements. We further validate our methodology on a 10$\times$10 optimized perovskite memristor crossbar, demonstrating high accuracy in a classification task and low energy consumption. This study offers a versatile solution for efficient optimization of various analog computing systems, encompassing both devices and algorithms.

Introduction

Analog in-memory computing (IMC) using memristors is a key technology for overcoming the "memory wall" in energy-efficient AI. By performing computations directly within the storage medium, IMC avoids the energy-intensive data movement typical of von Neumann architectures. Perovskite materials have recently gained attention for these devices due to their low cost and flexibility. However, the immaturity of perovskite fabrication leads to high device variability, and inherent analog non-idealities—such as random conductance drift and non-monotonic behavior—often degrade the performance of deep neural networks (DNNs) deployed on such hardware.

Methodology

The researchers developed a synergistic hardware-software co-design framework to address these challenges simultaneously.

Overview of the synergistic hardware-software co-design framework, including perovskite memristor fabrication, Bayesian optimization of experimental parameters, and the BayesMulti robust training algorithm.

Bayesian Fabrication Optimization

To navigate the complex search space of 8,400 possible fabrication configurations, the team employed Bayesian Optimization (BO). They optimized five critical factors: perovskite type, nanowire dimensions, lead deposition time, and electrode thickness. A novel "Usability" metric was introduced to quantitatively evaluate devices based on their monotonicity and stochastic stability, providing a clear target for the optimization algorithm.

Iterative Bayesian optimization process for memristor fabrication, highlighting common non-idealities like incomplete filling and overfilling that were mitigated through the search for optimal parameters.

BayesMulti Algorithm

On the software side, the "BayesMulti" strategy was developed to enhance DNN resilience. This method injects multinomial noise into the network during training. The noise parameters are themselves optimized via BO to match the specific non-ideality profiles of the perovskite hardware, effectively "immunizing" the model against hardware imperfections.

Results and Discussion

The synergistic approach yielded significant improvements across both hardware and software domains:

  • Fabrication Efficiency: The BO-guided search improved the memristor usability score from 0.36 to 0.93 in just 12 iterations, successfully mitigating issues like incomplete filling or overfilling of nanowires.
  • Algorithmic Robustness: BayesMulti demonstrated superior performance compared to standard Empirical Risk Minimization (ERM). In 3D object detection tasks (KITTI dataset), the BayesMulti-trained models maintained high accuracy under noise levels where standard models failed completely.
  • System Performance: A physical 10×10 perovskite crossbar achieved 270 times higher energy efficiency than an NVIDIA Tesla V100 GPU. The framework also successfully scaled to complex tasks, including Large Vision-Language Models (LVLMs) like MiniGPT-4, showing up to 100-fold improvements in prediction consistency.

Conclusion

This study demonstrates that a hardware-software co-design approach is essential for realizing the potential of analog computing with emerging materials. By combining Bayesian optimization for fabrication with noise-aware training algorithms, the researchers created a robust and highly efficient system capable of handling modern, large-scale AI tasks. Future work will focus on optimizing peripheral circuitry and scaling the crossbar arrays for even larger models.

Summary updated

Abstract

Analog computing using non-volatile memristors has emerged as a promising solution for energy-efficient deep learning. New materials, like perovskites-based memristors are recently attractive due to their cost-effectiveness, energy efficiency and flexibility. Yet, challenges in material diversity and immature fabrications require extensive experimentation for device development. Moreover, significant non-idealities in these memristors often impede them for computing. Here, we propose a synergistic methodology to concurrently optimize perovskite memristor fabrication and develop robust analog DNNs that effectively address the inherent non-idealities of these memristors. Employing Bayesian optimization (BO) with a focus on usability, we efficiently identify optimal materials and fabrication conditions for perovskite memristors. Meanwhile, we developed "BayesMulti", a DNN training strategy utilizing BO-guided noise injection to improve the resistance of analog DNNs to memristor imperfections. Our approach theoretically ensures that within a certain range of parameter perturbations due to memristor non-idealities, the prediction outcomes remain consistent. Our integrated approach enables use of analog computing in much deeper and wider networks, which significantly outperforms existing methods in diverse tasks like image classification, autonomous driving, species identification, and large vision-language models, achieving up to 100-fold improvements. We further validate our methodology on a 10$\times$10 optimized perovskite memristor crossbar, demonstrating high accuracy in a classification task and low energy consumption. This study offers a versatile solution for efficient optimization of various analog computing systems, encompassing both devices and algorithms.
BibTeX
@misc{ye2024synergistic,
  author = {Nanyang Ye and Qiao Sun and Yifei Wang and Liujia Yang and Jundong Zhou and Lei Wang and Guang-Zhong Yang and Xinbing Wang and Chenghu Zhou and Wei Ren and Leilei Gu and Huaqiang Wu and Qinying Gu},
  title = {{Synergistic Development of Perovskite Memristors and Algorithms for Robust Analog Computing}},
  year = {2024},
  eprint = {2412.02779},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2412.02779},
}
Figure