notesum.ai

Published at October 31

DynaSplit: A Hardware-Software Co-Design Framework for Energy-Aware Inference on Edge

cs.DC

cs.LG

cs.SE

Released Date: October 31, 2024

Authors: Daniel May¹, Alessandro Tundo¹, Shashikant Ilager², Ivona Brandic

Aff.: ¹Institute of Information Systems Engineering, TU Wien, Vienna, Austria; ²Informatics Institute, University of Amsterdam, Amsterdam, Netherlands

Arxiv: http://arxiv.org/abs/2410.23881v1

Parameter

Parameter Type

Domain

CPU Frequency (

CPU_{f}

)

Numerical

(low: 0.6, high: 1.8, step: 0.2)

{0.6, 0.8, 1.0, 1.2, 1.4, 1.6, 1.8}

Edge TPU Frequency (

TPU_{f}

)

Categorical

{off, std, max}

Use Cloud GPU (GPU)

Categorical

{Yes, No}

VGG16 (Simonyan and Zisserman, 2015) Split Layer (

L_{VGG}

)

Numerical

(low: 0, high: 22, step: 1)

{0, 1, 2, …, 22}

\hdashlineVision Transformer (Dosovitskiy et al., 2021) Split Layer (

L_{ViT}

)

Numerical

(low: 0, high: 19, step: 1)

{0, 1, 2, …, 19}