Workshop: IA^3 2026 — 16th Workshop on Irregular Applications: Architectures and Algorithms
Authors: Marco Siracusa, Adrià Armejach, and Miquel Moretó (Barcelona Supercomputing Center, Universitat Politècnica de Catalunya)
Abstract: Graph neural networks address critical problems in content recommendation, fraud detection, and drug discovery. Unlike dense machine-learning models, graph models are often memory bound rather than compute bound, making HBM-based processors a promising alternative to GPUs. This paper first characterizes how graph input locality shapes the architectural requirements of these workloads. It then evaluates locality-aware architectural optimizations, including decoupled memory access and near-memory processing, that enable HBM-based processors to outperform GPUs by up to 7x.
Back to IA^3 2026 — 16th Workshop on Irregular Applications: Architectures and Algorithms Archive Listing
Back to Full Workshop Archive Listing