SNI-GNN uses NVIDIA BlueField-3 SmartNICs to speed GNN training 3.6x

Full-graph GNN training offers high accuracy but struggles to scale across multi-server clusters because of heavy, irregular inter-node embedding exchanges. SNI-GNN, from researchers Guofan Yu, Sitian Chen, Zhenheng Tang, Xiaowen Chu, and Amelie Chi Zhou (accepted at ICDE 2026), tackles this bottleneck by offloading prediction to SmartNICs. The system deploys a lightweight linear-trend predictor directly on NVIDIA BlueField-3 DPUs to refine cached historical embeddings, avoiding the need to fetch fresh embeddings from remote GPUs for every layer. An importance-based boundary-node sampling policy further reduces communication, while an asynchronous DPU-GPU data pipeline reuses intermediate results to keep the hardware busy.

The results are significant: SNI-GNN cuts communication by 21-45% and delivers 1.3-3.6x end-to-end speedups over BNS-GCN and up to 1.29x over the SANCUS baseline, all while keeping accuracy loss at or below 0.01. The authors also provide formal error and convergence bounds, showing that predictor bias remains controlled under bounded second-order dynamics and that the system converges in standard non-convex settings with inexact gradients. Already integrated with state-of-the-art full-graph systems on BlueField-3, SNI-GNN demonstrates that in-network prediction is a practical complement to partitioning and compression for scaling GNN training to 16 GPUs and graphs with tens of millions of edges.