ReGNN: A ReRAM-based Heterogeneous Architecture for General Graph Neural Networks
TimeWednesday, July 13th11:37am - 12pm PDT
Location3002, Level 3
Event Type
Research Manuscript
In-memory and Near-memory Computing
DescriptionIn this work, we propose a ReRAM-based processing-in-memory (PIM) architecture
called ReGNN for GNN acceleration. ReGNN is composed of an analog PIM (APIM) module and an digital PIM (DPIM) module. Particularly, we design an new algorithm in DPIM to break the limitation that ReRAM can only accelerate MVM operations. Moreover, ReGNN maps data according to the degree of vertices in the aggregation engine to
improve data parallelism. Experimental results shows that ReGNN can speed up GNN inference by 53.6× and 2.3×, and reduce energy consumption by 84.8× and 3.55×, compared with GPU (baseline) and the state-of-the-art ReGraphX, respectively.