Low-Memory Graph Neural Network Training through Kernel-Level Optimization

Abstract

Graph neural networks have become a central computational method for learning over relational data, but their training remains constrained by severe memory pressure, irregular data access, and expensive message-passing operations. This paper examines low-memory graph neural network training through the lens of kernel-level optimization, arguing that meaningful memory reduction cannot be achieved solely through model compression or graph sampling. Instead, it requires coordinated redesign of execution kernels, memory hierarchies, scheduling policies, and data movement across heterogeneous hardware. The analysis situates kernel-level optimization within broader system architecture, infrastructure, governance, sustainability, robustness, and fairness debates. It considers how fused kernels, sparse memory layouts, recomputation, quantization, unified memory, and distributed partitioning alter the trade-offs among throughput, accuracy, energy use, and deployability. The paper also explores the institutional consequences of memory-efficient graph learning, including the possibility of more inclusive access to graph artificial intelligence, the risk of new forms of technical lock-in, and the need for governance frameworks that treat efficiency as a socio-technical property rather than a purely computational metric. It concludes that kernel-level memory optimization is not merely an implementation detail but a foundational condition for scalable, equitable, and environmentally responsible graph learning systems. The discussion is intended for researchers and practitioners who design graph learning platforms, hardware accelerators, distributed training infrastructures, and policy frameworks for responsible artificial intelligence.