Exploring Large‑Scale Graph Neural Network Applications and System Optimizations at Tencent

The article shares Tencent's exploration of large‑scale graph neural network (GNN) applications and recent system paradigms, beginning with an overview of three classic approaches—layer‑wise, node‑wise layer‑wise, and graph‑wise sampling—to reduce memory and computation costs.

While sub‑graph sampling has enabled successful industrial deployments (e.g., recommendation), it merely sidesteps the core challenge of performing full‑graph GNN updates, leading to inefficiencies, accuracy loss, and high communication overhead in domains such as drug discovery.

To address these issues, Tencent extends the classic GAS (Gather‑Apply‑Scatter) model to a SAGA paradigm, adding four steps—Scatter, Apply Edge, Gather, Apply Vertex—where GPU‑intensive and CPU‑intensive phases alternate, enabling more flexible graph processing.

Several system‑level optimizations are discussed: (1) graph partitioning, including locality‑aware, model‑based, and vertical feature partitioning; (2) node‑feature transmission improvements such as static caching, DistGNN blocking, and hybrid parameter mechanisms; and (3) pipeline and communication strategies that employ asynchronous or half‑synchronous updates to reduce inter‑machine traffic.

The future direction highlights that SAGA does not suit all GNN architectures, especially emerging Graph Transformers and models that incorporate geometric information (e.g., protein or catalyst simulations). Scaling such models requires new system designs that handle large feature dimensions and spatial data efficiently.

The Q&A section emphasizes the market gap for end‑to‑end graph computing platforms that combine system engineering with AI, noting strong demand in drug discovery, finance, and other scientific applications.

Overall, the talk underscores rapid progress in large‑scale GNN research and the ongoing need for integrated hardware‑software solutions.