Technical Papers Archive

TempGNN: An Efficient Real-Time TGNN Inference Accelerator via Target-Centric Dependency Packets


Authors: Hui Yu and Wei Zhang (The Hong Kong University of Science and Technology), Ligang He (University of Warwick), Jin Zhao and Yu Zhang (Huazhong University of Science and Technology (HUST)), Carol Jingyi Li (The Hong Kong University of Science and Technology), Zixiao Wang (Huazhong University of Science and Technology (HUST)), and Jiang Xu (The Hong Kong University of Science and Technology)

Abstract: Temporal graph neural networks (TGNNs) effectively model dynamic graphs, but inference suffers from limited useful parallelism and irregular memory accesses because strict temporal dependencies require timestamp-ordered vertex updates. We present TempGNN, a target-centric accelerator for high-performance real-time TGNN inference. TempGNN reorganizes sampled temporal computation into target dependency packets (TDPs), each capturing the minimal temporally valid dependency states required to update a target vertex. By decoupling dependency preparation from ordered target-state commitment, TempGNN preserves temporal correctness while exposing branch-level parallelism across partially overlapping targets. To realize this execution model, TempGNN integrates a dependency-driven TDP construction and data-loading scheme together with an overlap-aware synchronization strategy that eliminates redundant dependency processing and supports efficient on-chip execution. Implemented on a Xilinx Alveo U280 FPGA, TempGNN achieves 28.2× speedup and 33.5× energy savings over an A100 GPU, and further delivers 3.8×–7.6× speedups with 6.5×–10.2× energy savings over prior TGNN accelerators.


Presentation: file

Back to Technical Papers Archive Listing