In-Memory Annealing Unit (IMAU): Energy-Efficient (2000 TOPS/W) Combinatorial Optimizer for Solving Travelling Salesman Problem

Ming Chun Hong, Le Chih Cho, Chih Sheng Lin, Yu Hui Lin, Po An Chen, I. Ting Wang, Pei Jer Tzeng, Shyh Shyuan Sheu, Wei Chung Lo, Chih I. Wu, Tuo Hung Hou*

*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

5 Scopus citations

Abstract

An in-memory annealing unit (IMAU) as an energy-efficient combinatorial optimizer for solving the travelling salesman problem (TSP) has been demonstrated for the first time. A hardware-algorithm co-optimization approach is adopted to overcome the challenges of solving TSP using IMAU, such as large problem size, insufficient weight precision, and inaccurate analog computing. The high-capacity (1152x1024) binary RRAM-based IMAU with an embedded simulated annealing (SA) function achieves an extremely high throughput of 90 TOPS and energy efficiency of 2000 TOPS/W. A new multi-step SA algorithm is proposed to solve the otherwise floating-point TSP using merely 5-level (2.3 bit) weights and achieves the floating point-equivalent shortest route for the 10-city TSP in IMAU.

Original languageEnglish
Title of host publication2021 IEEE International Electron Devices Meeting, IEDM 2021
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages21.3.1-21.3.4
ISBN (Electronic)9781665425728
DOIs
StatePublished - 2021
Event2021 IEEE International Electron Devices Meeting, IEDM 2021 - San Francisco, United States
Duration: 11 Dec 202116 Dec 2021

Publication series

NameTechnical Digest - International Electron Devices Meeting, IEDM
Volume2021-December
ISSN (Print)0163-1918

Conference

Conference2021 IEEE International Electron Devices Meeting, IEDM 2021
Country/TerritoryUnited States
CitySan Francisco
Period11/12/2116/12/21

Fingerprint

Dive into the research topics of 'In-Memory Annealing Unit (IMAU): Energy-Efficient (2000 TOPS/W) Combinatorial Optimizer for Solving Travelling Salesman Problem'. Together they form a unique fingerprint.

Cite this