Design of the switchable mode TaO x -based device for highly stable performance in neuromorphic computing

Abstract Resistive Random Access Memory (RRAM) has the advantages of good read/write performance, low power consumption, high density storage and scalability. However, the efficiency and reliability of RRAM based storage-computing systems are limited by non-ideal factors, such as sneak current and voltage drop problems. To address these problems, a TaO x -based device which can be switched between the bipolar memristive mode and Complementary Resistive Switch(CRS) mode was designed. The distinct threshold voltages observed for the two operational modes permits mode-optimized read/write signaling in crossbar arrays. Consequently, the device achieves non-destructive reading in bipolar memristive mode, and writing/programming in CRS mode to reduce the non-ideal factors effects, which also provides powerful device data for the subsequent array-level simulations. The back propagation algorithm is then used to simulate the device at the array level, and the performance of the storage-computation fully parallel image recognition task are tested. The results show that a series of non-ideal factors such as leakage current and complex read/write methods in the array can be effectively avoided, and excellent accuracy results can be obtained in a fully parallel image recognition task, which shows great potential in neuromorphic computing applications.