Temporal-encoded 6T-RRAM with Bidirectional Control for Future Neuromorphic Systems

Neuromorphic systems are of high importance in artificial intelligence (AI) to eschew challenges inherent to deep learning acceleration in conventional systems. Throughout the development history of neuromorphic systems, in-memory computing with emerging memory technologies, such as resistive random-access memory (RRAM), offer advantages by executing operations in-situ, exactly where the data are located, providing significant improvement in data throughput and energy efficiency. Herein, we introduce a temporal-encoded memristive crossbar that offers faster yet more efficient in-memory operations to accelerate AI algorithms. Specifically, our memristive crossbar is built based upon the hafnium-oxide RRAM in a 6-transistor-1-RRAM (6T1R) structure and controlled by a series of temporal-encoded voltage pulses, where individual 6T1R cell can be programmed to encode either positive or negative weight values through a bidirectional switching mechanism. In some cases, individual 6T1R cell can also be programmed directly to the high-impedance state, either encoding zero weight value or quarantining faulty RRAM from the network. In practice, proof-of-concept validations were conducted using a custom 65nm CMOS/RRAM technology node, yielding an average classification accuracy of 94%, 95%, and 99%, respectively, when tested with the banknote authentication, balance scale, and iris plant datasets. Better yet, the parallel inter-tile communications of our prototype achieves up to 1.6 trillion operations per second.