PipeGNN: A Bandwidth-Efficient GNN Accelerator with Node-Level Pipelined Push Execution

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Graph Neural Networks (GNNs) have become a fundamental tool for learning over graph-structured data. Under the message-passing framework, mainstream GNN models alternate between feature transformation and neighborhood aggregation. Fusing these two phases into a node-level pipelined push dataflow, in which each node’s transformed feature streams directly into aggregation, establishes a single-pass I/O target that reads each input tensor exactly once per layer. No existing hardware sustains this dataflow at full concurrency. Prior fusion accelerators fall short of this target in three ways: designs that gather the raw input features for aggregation or write transformed features off-chip incur redundant traffic, designs that stream transformed features on-chip still keep the two phases as separate matrix kernels without a unified execution granularity, and the designs that aggregate at message granularity serialize per-destination aggregation or allocate state that does not scale to large graphs. We propose PipeGNN, a bandwidth-efficient GNN accelerator whose streaming microarchitecture sustains this dataflow as a fully concurrent pipeline within a fixed on-chip buffer budget. PipeGNN decouples execution from topological boundaries by aligning execution granularity with both memory burst size and aggregation compute throughput, and resolves four coupled microarchitectural barriers through coordinated mechanisms: compute-aligned edge batching, contention-aware message reduction, adaptive cross-batch scheduling, and tiered state residency. Evaluated across diverse GNN workloads, PipeGNN achieves geometric mean speedups of 10.1 × over AWB-GCN and 12.2 × over FlowGNN, and 2.4 × to 46.9 × over a GPU baseline across graph scales, reduces off-chip traffic by 2.69 × over AWB-GCN and 5.93 × over FlowGNN, and improves energy efficiency by 56.81 × over AWB-GCN and 1.53 × over FlowGNN, with a 154 × traffic reduction and a 116.81 × energy reduction against HyGCN.