Lego: Dynamic Tensor-Splitting Multi-Tenant DNN Models on Multi-Chip-Module Architecture

Zhou Yu Xuan, Ching Jui Lee, Tsung Tai Yeh

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

Abstract

Modern deep neural network (DNN) accelerators target the acceleration of a single DNN model and limit the throughput for multi-tenant DNN data center applications. The multi-chip-module (MCM) architecture breaks a monolithic accelerator into multiple small chiplets. The MCM is a promising approach that dispatches DNN models across chiplets with equal PEs. However, it is challenging to distribute data of DNN model layers with different parameters across chiplets while maximizing the chiplet utilization. This work proposes Lego MCM architecture that dynamically adapts to the size of DNN model layers and improves the throughput of multi-tenant DNN applications by increasing the chiplet utilization. Lego's dynamic scheduler achieves the geometric average 1.51× speedup over a monolithic DNN accelerator.

Original languageEnglish
Title of host publicationProceedings - International SoC Design Conference 2022, ISOCC 2022
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages173-174
Number of pages2
ISBN (Electronic)9781665459716
DOIs
StatePublished - 2022
Event19th International System-on-Chip Design Conference, ISOCC 2022 - Gangneung-si, Korea, Republic of
Duration: 19 Oct 202222 Oct 2022

Publication series

NameProceedings - International SoC Design Conference 2022, ISOCC 2022

Conference

Conference19th International System-on-Chip Design Conference, ISOCC 2022
Country/TerritoryKorea, Republic of
CityGangneung-si
Period19/10/2222/10/22

Fingerprint

Dive into the research topics of 'Lego: Dynamic Tensor-Splitting Multi-Tenant DNN Models on Multi-Chip-Module Architecture'. Together they form a unique fingerprint.

Cite this