Characterizing the Communication Requirements of GNN Accelerators: A Model-Based Approach

TL;DR

This paper develops analytical models to evaluate data movement and communication needs in GNN accelerators, aiding in understanding their scalability and enabling comparison across different designs.

Contribution

It introduces a model-based approach to characterize data movement and communication requirements in GNN accelerators, addressing a gap in understanding their hardware efficiency.

Findings

Models accurately predict data movement in GNN accelerators

Scalability characteristics vary with hardware and graph parameters

Enables comparison of different GNN accelerator designs

Abstract

Relational data present in real world graph representations demands for tools capable to study it accurately. In this regard Graph Neural Network (GNN) is a powerful tool, wherein various models for it have also been developed over the past decade. Recently, there has been a significant push towards creating accelerators that speed up the inference and training process of GNNs. These accelerators, however, do not delve into the impact of their dataflows on the overall data movement and, hence, on the communication requirements. In this paper, we formulate analytical models that capture the amount of data movement in the most recent GNN accelerator frameworks. Specifically, the proposed models capture the dataflows and hardware setup of these accelerator designs and expose their scalability characteristics for a set of hardware, GNN model and input graph parameters. Additionally, the proposed approach provides means for the comparative analysis of the vastly different GNN accelerators.