Skip to main navigation Skip to search Skip to main content

A Dynamic Virtual Memory Management System for LLMs on AI Chips

Research output: Chapter in book / Conference proceedingConference article published in proceeding or bookAcademic researchpeer-review

Abstract

With the widespread application of large-scale DNN models, efficient training and inference with limited resources has become a popular research area. However, memory fragmentation is a significant barrier to efficient training and inference. In this article, we propose an efficient memory management strategy for execution on the Ascend platform based on the latest NPU virtual memory management APIs and develop efficient memory management logic for these. Indeed, the modification is mainly within the caching allocator of the NPU PyTorch extension, designing related garbage collection, allocation, splitting, reallocation, and fusing small fragmented blocks. In a four-card study on the Ascend 910B platform with training and inference accelerating architecture, MindSpeed, the greatest performance improvement was achieved by reducing the average fragmentation rate from 10% to 6%.

Original languageEnglish
Title of host publicationProceedings - 2025 IEEE 43rd International Conference on Computer Design, ICCD 2025
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages389-392
Number of pages4
ISBN (Electronic)9798331503468
DOIs
Publication statusPublished - Dec 2025
Event43rd International Conference on Computer Design, ICCD 2025 - Richardson, United States
Duration: 10 Nov 202512 Nov 2025

Publication series

NameProceedings - IEEE International Conference on Computer Design: VLSI in Computers and Processors
ISSN (Print)1063-6404

Conference

Conference43rd International Conference on Computer Design, ICCD 2025
Country/TerritoryUnited States
CityRichardson
Period10/11/2512/11/25

Keywords

  • AI Chip
  • LLM Training and Fine-Tuning
  • Virtual Memory Management System

ASJC Scopus subject areas

  • Hardware and Architecture
  • Electrical and Electronic Engineering

Fingerprint

Dive into the research topics of 'A Dynamic Virtual Memory Management System for LLMs on AI Chips'. Together they form a unique fingerprint.

Cite this