Device-physics co-design of hybrid RRAM-MRAM crossbars: Overcoming stochastic non-idealities for energy efficient edge speech recognition

Abstract

Memristive in-memory computing leverages intrinsic device physics to overcome von Neumann bottlenecks in edge artificial intelligence (AI), yet non-idealities in resistive switching materials limit deployment robustness. Here, we demonstrate a physics-guided hybrid memristor architecture integrating resistive random-access memory (RRAM) and magnetoresistive random-access memory (MRAM) for energy-efficient keyword spotting. By co-designing a binarized depthwise separable convolutional neural network (790-bit model) with device-aware algorithms, we exploit the complementary physical properties of RRAM (high density, filamentary switching) and MRAM (high endurance, spintronic switching). Convolutional layers are mapped to RRAM crossbars for parallel analog vector-matrix multiplication via Ohm's law, while fully connected layers utilize MRAM's endurance for in situ training. A statistic-aware training strategy incorporates device stochasticity (Gaussian resistance variation, shorts/opens) during backpropagation, enhancing fault tolerance. Implemented on a 1 kb TiN/HfOx/TaOx RRAM and CoFeB/MgO-based MRAM array, the system achieves 3.72 ms latency and 14.95 μW power at >85% accuracy under 20% defect rates. This work establishes a device-physics-driven framework for reliable edge AI accelerators.