Design of a GPU register file using a monolithic 3D technique

Chang Hyun Kim, Hun Jae Lee, Sung Woo Chung, Seon Wook Kim

Research output: Contribution to journalArticlepeer-review

Abstract

While graphics processing units (GPUs) have been widely used to handle large amounts of data in recent years, their significant power consumption causes many usage constraints. Among the many components of a GPU, register files are critical to performance, and the crossbar network to configure them is the most significant factor in power consumption. Currently studied as a way to increase chip aggregation is 3D integration technology, like monolithic 3D (M3D) with finegrain characteristics. We suggest in this study how M3D technology can be applied to the crossbar of GPU register files to reduce power consumption. Our performance analysis shows that we obtain 40.2% and 18.9% savings in dynamic and leakage power, respectively, along with a 64.7% delay reduction.

Original languageEnglish
Pages (from-to)499-505
Number of pages7
JournalIEIE Transactions on Smart Processing and Computing
Volume8
Issue number6
DOIs
Publication statusPublished - 2019 Dec 30

Keywords

  • Crossbar network
  • Energy
  • GPU register file
  • Monolithic 3D integration
  • Power

ASJC Scopus subject areas

  • Signal Processing
  • Electrical and Electronic Engineering

Fingerprint

Dive into the research topics of 'Design of a GPU register file using a monolithic 3D technique'. Together they form a unique fingerprint.

Cite this