Self-Selective Multi-Terminal Memtransistor Crossbar Array for In-Memory Computing.

Two-terminal resistive switching devices are commonly plagued with longstanding scientific issues including interdevice variability and sneak current that lead to computational errors and high-power consumption. This necessitates the integration of a separate selector in a one-transistor-one-RRAM (1T-1R) configuration to mitigate crosstalk issue, which compromises circuit footprint. Here, we demonstrate a multi-terminal memtransistor crossbar array with increased parallelism in programming via independent gate control, which allows in situ computation at a dense cell size of 3-4.5 F2 and a minimal sneak current of 0.1 nA. Moreover, a low switching energy of 20 fJ/bit is achieved at a voltage of merely 0.42 V. The architecture is capable of performing multiply-and-accumulate operation, a core computing task for pattern classification. A high MNIST recognition accuracy of 96.87% is simulated owing to the linear synaptic plasticity. Such computing paradigm is deemed revolutionary toward enabling data-centric applications in artificial intelligence and Internet-of-things.

MoS2 /Polymer Heterostructures Enabling Stable Resistive Switching and Multistate Randomness.

Resistive random-access memories (ReRAMs) based on transition metal dichalcogenide layers are promising physical sources for random number generation (RNG). However, most ReRAM devices undergo performance degradation from cycle to cycle, which makes preserving a normal probability distribution during operation a challenging task. Here, ReRAM devices with excellent stability are reported by using a MoS2 /polymer heterostructure as active layer. The stability enhancement manifests in outstanding cumulative probabilities for both high- and low-resistivity states of the memory cells. Moreover, the intrinsic values of the high-resistivity state are found to be an excellent source of randomness as suggested by a Chi-square test. It is demonstrated that one of these cells alone can generate ten distinct random states, in contrast to the four conventional binary cells that would be required for an equivalent number of states. This work unravels a scalable interface engineering process for the production of high-performance ReRAM devices, and sheds light on their promising application as reliable RNGs for enhanced cybersecurity in the big data era.