notesum.ai

Published at November 25

A Data-Driven Approach to Dataflow-Aware Online Scheduling for Graph Neural Network Inference

cs.AR

Released Date: November 25, 2024

Authors: Pol Puigdemont¹, Enrico Russo², Axel Wassington¹, Abhijit Das¹, Sergi Abadal¹, Maurizio Palesi

Aff.: ¹Universitat Politècnica de Catalunya, Spain; ²University of Catania, Italy

Arxiv: http://arxiv.org/abs/2411.16342v1

Refer to caption

Tiling	Aggregation			Combination
$\omega$	T_Va	T_Fa	T_N	T_Vc	T_Gc	T_Fc
a	$*$	$\min\left(\text{Num. PEs},F\right)$	$1$	$*$	$1$	$\min\left(\text{Num. PEs},F\right)$
b	$*$	$\min\left(2,F\right)$	$\lfloor F/2\rfloor$	$*$	$1$	$\min\left(2,F\right)$
c	$*$	$\min\left(8,F\right)$	$\lfloor F/2\rfloor$	$*$	$1$	$\min\left(8,F\right)$
d	$*$	$1$	$1$	$*$	$1$	$1$
e	$*$	$\min\left(18,F\right)$	$\lfloor F/2\rfloor$	$*$	$1$	$\min\left(18,F\right)$
f	$*$	$1$	$\min\left(18,V\right)$	$*$	$1$	$1$
g	$*$	$\min\left(18,F\right)$	$1$	$*$	$1$	$\min\left(85,F\right)$
h	$*$	$1$	$\min\left(18,V\right)$	$*$	$1$	$\min\left(85,F\right)$