Conference paper Open Access
Giannopoulos, Iason; Le Gallo, Manuel; Jonnalagadda, Vara Prasad; Eleftheriou, Evangelos; Sebastian, Abu
The explosive growth in data-centric artificial intelligence related applications necessitates exploration of non-von Neumann computing paradigms such as in-memory computing. The ability to perform certain computational tasks within the memory unit will reduce dramatically the time and energy that is spent into shuttling the data from the memory to the processing unit. However, the nanoscale resistive memory devices that are useful for these technologies suffer from non-ideal characteristics. In this work we deal with the computational precision loss due to the strong and inhomogeneous temperature dependence of resistive devices and in particular phase-change memory. We describe a temperature compensation method that applies to resistive crossbar arrays and its realization as a peripheral circuit. We derive array-level temperature compensation functions that are remarkably effective for projected phase-change memory devices. We simulate the system and experimentally validate its efficacy in the task of matrix-vector multiplications. The computational precision is found to be equivalent to an 8-bit multiplier at elevated temperatures.