CiMC: A Computing-in-Memory Controller for Memristive Crossbar Array

Abstract

Listen

Translate article

The significant challenges for the exascale system are the relatively limited computational resources and data movement between processors and memories. Memristor has the potential to build a universal digital nonvolatile computing system to break the energy-efficiency bottleneck caused by the separation of storage and processing in the traditional computing framework. Computing-in-Memory(CiM) based on memristor is a potential approach to addressing the data transfer bottleneck between processor and memory in a computing system. CiM could reduce or even remove the data movement between processors and memory. The unique properties of Memtistior allow both logic and memory function within an element, opening up the possibility of directly computing on the data stored in memory. We propose a Compute-in-Memory Controller, a design for computing and storing in a memristive crossbar array. Both logic operation and arithmetic functions are implemented with the controller. We also propose a two-layer instruction set to perform CiM functions, which addresses the question of how a CiM memory could be integrated into a general-purpose computing system.