LGNNIC's SmartNIC offloads GNN preprocessing for 62x faster training

Graph Neural Networks (GNNs) are essential for social networks, chip design, and recommendation systems, but as graphs scale past single-node memory, distributing them across remote nodes causes severe inter-node network congestion. LGNNIC, proposed by Liad Gerstman, Aditya Dhakal, Dejan Milojicic, and Avi Mendelson, tackles this by exploiting SmartNICs already co-located with remote memory nodes in modern systems. Instead of streaming raw graph data to training nodes, LGNNIC offloads key preprocessing tasks onto the SmartNICs, dramatically reducing the volume of data that crosses the network. The two complementary techniques—Neighbor Sampling for mini-batch selection and Quantization of the sampled batches—run entirely on the SmartNIC, shrinking both transaction size and frequency during distributed training.

The researchers built a proof-of-concept with one NVIDIA BlueField-2 SmartNIC and one compute node equipped with an A100 GPU, evaluating both a low-overhead DOCA-DMA synchronization mechanism and a high-overhead socket-based baseline. Across standard GNN workloads, Neighbor Sampling achieved up to 62.4x speedup with Sockets and 17.5x with DOCA-DMA, driven primarily by reduced data transaction time. Quantization added further gains of up to 3.6x (Sockets) and 1.3x (DOCA-DMA). These results show that co-locating SmartNICs with remote memory nodes can turn network bottlenecks into a major acceleration opportunity. While the PoC uses just one remote node, the architecture is designed to scale across multi-node clusters, potentially reshaping how large-scale GNN training is deployed in data centers.