Training Graph Neural Networks (GNNs) on large-scale data is essential in domains such as transportation, recommendation systems, and molecular biology. As graph sizes increasingly surpass main memory capacities, the out-of-core (OOC) systems have been proposed, storing graphs on external storage, sequentially loading and processing smaller partitions. However, existing OOC solutions face key challenges, including excessive data migration between storage and memory, as well as reduced model accuracy. In this paper, we theoretically and empirically analyze the limitations of state-of-the-art (SOTA) OOC-based GNN systems and identify opportunities for optimization. Guided by our theoretical insights, we design a novel system, named VeloxGNN, to improve data migration efficiency while maintaining high model accuracy. First, we introduce a novel algorithm called Delayed Gradient Propagation (DGP), specifically designed for OOC-based GNN training. DGP leverages both historical node embeddings and gradients to achieve two key objectives: minimizing data migration (by ensuring that the graph is read at most once) while maintaining model accuracy. To support DGP, we then propose system-level optimizations, including dynamic memory management, a DGP-aware loading order, and a new graph partitioning scheme that separates labeled and unlabeled data. The experimental results in real systems show that VeloxGNN achieves in-memory model accuracy while reducing training time by 17.7% to 73.3% across various datasets and GNN models, outperforming state-of-the-art methods. This result highlights VeloxGNN’s potential for efficient and scalable GNN training on large-scale graph data.