부산대학교 도서관

In-memory computing is a promising candidate for overcoming the von Neumann memory wall and accelerating data processing in neural network applications. We propose a ITIR-based approach to carry out an efficient in-memory computation of the minimum logic function to be executed in the form of a series of 1-bit search operations similar to modern associative processors. We study the proposed design in the context of Kohonen Neural Networks that are also known as self-organizing maps. They are based on competitive learning where the minimum-search operation is a crucial part in the learning process. We study the trade off of different low-to-high resistance state windows associated with different endurance levels in the presence of process variations. We demonstrate ~0-4% reduction in the MNIST test dataset accuracy for the different scenarios. We report a maximum of 6fJ for the 1-bit search operation.