notesum.ai

Published at November 6

Efficient Message Passing Architecture for GCN Training on HBM-based FPGAs with Orthogonal Topology On-Chip Networks

cs.AR

cs.LG

Released Date: November 6, 2024

Authors: Qizhe Wu¹, Letian Zhao¹, Yuchen Gui¹, Huawen Liang Xiaotian Wang

Aff.: ¹University of Science and Technology of China, Hefei, Anhui, China

Arxiv: http://arxiv.org/abs/2411.03857v1

	On-chip-Mem.	40MB	54MB	43MB
		GPU	HP-GNN(Lin et al., 2022)	Ours
Platform	Device	Nvidia A100	Alveo U250	VCU128
	Peak Perf.	19.5TFLOPS	1.8TFLOPS	2TFLOPS
NS-GCN	Flickr	0.21( $0.47\times$ )	0.16( $1\times$ )	0.09( $1.6\times$ )
	Reddit	6.59( $0.16\times$ )	1.09( $1\times$ )	1.05( $1.03\times$ )
	Yelp	2.90( $0.38\times$ )	1.35( $1\times$ )	1.11( $1.22\times$ )
	AmazonP.	5.06( $0.38\times$ )	3.49( $1\times$ )	1.92( $1.81\times$ )
NS-SAGE	Flickr	0.29( $0.75\times$ )	0.22( $1\times$ )	0.12( $1.54\times$ )
	Reddit	3.05( $0.51\times$ )	1.56( $1\times$ )	1.37( $1.13\times$ )
	Yelp	3.51( $0.52\times$ )	1.85( $1\times$ )	1.64( $1.12\times$ )
	AmazonP.	6.83( $0.70\times$ )	4.83( $1\times$ )	3.65( $1.32\times$ )