Bulk switching RRAM technologies have been developed to address nonidealities of filamentary RRAM for embedded compute-in-memory applications. However, high density 3D integration and scalability to nano regime have yet to be experimentally demonstrated. Here, we present a scalable, filament-free 3D 8-layer vertical bulk RRAM (b-RRAM) technology optimized for embedded compute-in-memory (CIM) applications. This forming-free device features reliable cycling, multi-level switching, and enhanced speed via hydrogen doping. Guided by multiscale device simulations to optimize the switching stack, we demonstrate 40×40 nm2 b-RRAM cells with MΩ-level resistance and current nonlinearity, enabling accurate, energy-efficient matrix-vector multiplications (MVM) in selector-less crossbars. A hyperdimensional computing-based continual learning algorithm is implemented on 3D b-RRAM for edge AI tasks, achieving ~90% accuracy—comparable to high-precision floating-point (FP) baselines—while delivering substantial energy savings.
Zhou et al. (Sat,) studied this question.