In the case of the ReRAM crossbar, it is gaining attention as a next-generation deep learning accelerator because it can calculate MVM very quickly from analog due to its structure. However, the results of MVM cannot be displayed normally by the case of many non-idealities in the ReRAM, and for this reason, practical use in deep learning is not yet possible. Many methods have been studied to solve non-idealities, but in the case of those currently being studied, problem is that a lot of preprocessing or post-processing is required for each crossbar. In this paper, we introduce a method that statistically processes the non-idealities of ReRAM based on batch normalization so that there is no hardware overhead and shows good resilience.
Publisher
Ulsan National Institute of Science and Technology (UNIST)