Evaluation of 2T0C DRAM-Based Processing-in-Memory Systems for Accelerating Deep Neural Network Models

Jihoon Jang, Inseong Hwang, Hyun Kim

Abstract: This paper presents the simulator design and evaluation of a 2TOC DRAM-based Processing-in-Memory (PIM) system designed to optimize deep learning applications. By customizing DRAMsim3 to simulate the unique characteristics of 2TOC DRAM, experimental results with deep learning models such as VGG-8 and AlexNet demonstrate that the 2TOC DRAM-based PIM system achieves computational speeds up to 30 x and energy efficiency up to 23 x compared to conventional CPU systems.