SmartNICs Accelerate Full-Graph GNN Training with In-Network Prediction

Training Graph Neural Networks (GNNs) on full graphs, while offering high accuracy, faces significant scalability challenges in multi-server clusters. The primary bottleneck is the heavy and irregular exchange of embeddings between nodes, which consumes substantial network bandwidth and processing time. This communication overhead limits the efficiency of large-scale GNN training. Researchers have developed SNI-GNN, a novel system that leverages SmartNICs (Smart Network Interface Cards) to address this issue. SNI-GNN reduces inter-node communication by offloading a lightweight linear-trend predictor onto the SmartNICs. This predictor refines cached historical embeddings directly in the network, minimizing the need for frequent data transfers. The system also incorporates an importance-based boundary-node sampling policy and an asynchronous DPU-GPU data pipeline, further optimizing data flow and resource utilization. The theoretical analysis provides error and convergence bounds, demonstrating that the predictor's bias remains controlled and yields standard non-convex convergence. Implemented on NVIDIA BlueField-3 SmartNICs, SNI-GNN integrates seamlessly with existing full-graph GNN systems. It achieves a 21-45% reduction in communication, leading to 1.3-3.6 times end-to-end speedups over BNS-GCN and up to 1.29 times over baseline SANCUS, all while maintaining an accuracy loss of less than 0.01. This approach scales efficiently to 16 GPUs on graphs with millions of edges, proving that SmartNIC-based in-network prediction is a practical and powerful complement to other communication-reduction techniques for large-scale GNN training.