Limitations on Pitch Design due to Thermal Crosstalk in Pr1-xCaxMnO3 RRAM Crossbar Arrays

Networks designed with memristor crossbar arrays have been proposed for Artificial Intelligence (AI) and Machine Learning (ML)-based applications for enhanced network efficiency. The area efficiency can be achieved by increasing the density of the crossbar structures (reducing the pitch) while parallel reading multiple memory units can improve the computational performance of networks. When using emerging memories such as Resistive Random-Access Memories (RRAMs) in crossbar architectures, the variability in the memristive weights will impact the network accuracy. One of the sources of weight perturbation is heat generation and dissipation while programming the devices. The phenomenon is famously known as thermal crosstalk. In this work, we have proposed limitations on array pitch design considering thermal crosstalk using an experimentally calibrated TCAD model for Pr
1−x
Ca
x
MnO
3
(PCMO) based RRAM arrays. 21% and 48% error for 25nm pitch in computing Multiply and Accumulate (MAC) output is observed for a five-device network for the same and different conductance states respectively.