Hybrid Memristor-based Computing Technology for Edge Application
±è°æ¹Î
KAIST

Analog memristors enable compact neuromorphic computing with low power consumption. One of the issues in the technology is slow speed for precise analog data programming. In this presentation, a novel analog data programming method is proposed utilizing an in-memory logic technique. The method can transfer any resistance values from reference resistors to the target memristor accurately inside a crossbar array by performing an appropriate voltage clocking. We propose an ideal memristor model based on the method and evaluated a Ti-doped NbOx charge trap memristor as a promising candidate for application. The characteristic error of the Ti-doped NbOx memristor device was about 5 % on average, compared to the ideal memristor, and configuring optimum parallel resistors in the circuit further improved this to 2.95 %. We then applied the method to program a memristive neural network and confirmed this error was negligible, and thus, the proposed method is viable. Such hybirid computing technology allows low power and low cost device suitable for edge applications in the IoT era.