Implementing Spatio-Temporal Graph Convolutional Networks on Graphcore IPUs

• Published in: 2022 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW) pp. 45 - 54
• Main Authors: Moe, Johannes, Pogorelov, Konstantin, Schroeder, Daniel Thilo, Langguth, Johannes
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.05.2022
• Subjects: AI accelerator; Computer architecture; Graph Neural Networks; Graphics processing units; IPU; Recurrent neural networks; Software; Software algorithms; Speech recognition; STGCN; Tensors
• ISBN: 9781665497480
• DOI: 10.1109/IPDPSW55747.2022.00016

Artificial neural networks have been used for a multitude of regression tasks, and their descendants have expanded the domain to many applications such as image and speech recognition, filtering of social networks, and machine translation. While conventional and recurrent neural networks work well on data represented in Euclidean space, they struggle with data in non-Euclidean space. Graph Neural Networks (GNN) expand recurrent neural networks to directly process sparse representations of graphs, but they are computationally expensive, which invites the use of powerful hardware accelerators. In this paper, we investigate the viability of the Graphcore Intelligence Processing Unit (IPU) for efficient implementation of Spatio-Temporal Graph Convolutional Networks. The results show that IPUs are well suited for this task.