Close

Presentation

PIMGCN: A ReRAM-Based PIM Design for Graph Convolutional Network Acceleration
Time
Location
Event Type
Research Manuscript
Virtual Programs
Hosted in Virtual Platform
Keywords
AI/ML System Design
Topics
Design
DescriptionGCNs mainly include two phases with distinct execution patterns: Aggregation phase with a dynamic and irregular execution pattern, and Combination phase with a static and regular execution pattern.
In this paper, we propose PIMGCN, a Processing-In-Memory (PIM) architecture for GCNs on ReRAMs. PIMGCN implements a pipelined search-execute model to leverage the high intra-vertex parallelism, and leverage the inter-vertex parallelism using our proposed optimization algorithm. Comparison shows that PIMGCN demonstrates considerable performance speedup and energy reduction over HyGCN~(219x and 95.3x), and demonstrate up to 1.6x speedup and 238x energy efficiency improvement over the SOTA GCN accelerator AWB-GCN.