Floating Point Multiplication Mapping on ReRAM Based In-memory Computing Architecture

Low leakage power, high endurance and non-volatile storage capabilities have made memristive devices, such as Resistive RAM (ReRAM) popular. ReRAMs also offer in-memory computing capabilities by means of stateful logic operations. However, there are no standard libraries for floating point operations on ReRAM-based in-memory computing platforms. In this paper, we propose a mapping for one such mathematical function, namely multiplication for floating point numbers. We undertook a detailed study to derive a mapping with low memory footprint on ReVAMP architecture, which is an in-memory computing platform based on ReRAM crossbar arrays. For multiplication of two IEEE-754 compliant single (or double) precision numbers, the proposed mapping requires 1.944 Kb (or 8.904 Kb) of computation memory, 374.4 Kb (or 3457 Kb) of instruction memory while consuming an energy equivalent to 5.6 pJ (or 471 pJ).