Virgil Watkins1,Diana Kim1,Laszlo Cline1,Yiyang Li1
University of Michigan1
Virgil Watkins1,Diana Kim1,Laszlo Cline1,Yiyang Li1
University of Michigan1
Analog neuromorphic computing can decrease the energy consumption of data-intensive tasks like machine learning by orders of magnitude through conducting matrix vector multiplication using Ohm’s and Kirchoff’s Laws. Electrochemical random access memory (ECRAM) stores and switches analog resistance states by electrochemically modulating the concentration of oxygen vacancies in a transition metal oxide. Unfortunately, most ECRAM devices are volatile and revert to equilibrium, with retention times orders of magnitude lower than the typical 10 year, 85C requirement. In our work we develop a nonvolatile ECRAM cell using tantalum oxide. Our core innovation is the use of phase separating materials in the miscibility gap. In this configuration, all resistance states have the same chemical potential, thereby eliminating the driving force for volatility. This result not only exceeds the expected 10 year lifetime, but can provide a memory cell with potentially indefinite retention at elevated temperatures.