TrackGNN: A Highly Parallelized and Self-Adaptive GNN Accelerator for Track Reconstruction on FPGAs

Track reconstruction is essential in high-energy physics (HEP) to identify charged particles and understand their properties. Graph neural networks (GNNs) show great promise as an alternative. However, the stringent real-time and low-latency requirements in trigger decision pose significant challenges for deploying GNN inference in modern particle tracking systems. In this work, we achieve real-time track reconstruction on FPGAs by proposing TrackGNN, a high-performance FPGA-based dataflow accelerator. TrackGNN leverages a streaming architecture with multiple levels of parallelism. In particular, we introduce a novel self-adaptive renaming mechanism to improve data locality in the graph structure, reducing initiation delays and pipeline stalls. We also explore a fixed-point quantization strategy to improve energy efficiency and maintain model accuracy. Experimental results show that TrackGNN achieves 27.6 × and 4.9 - 101.1 × speedup over CPUs and GPUs, respectively, along with 55.36 × and 515.46 × better energy efficiency. Compared to the state-of-the-art FPGA-based GNN accelerator, TrackGNN delivers a 16.67% performance boost with FlowGNN [1],a gain directly attributable to the proposed renaming process, and 5.7 × speedup over overlay architectures. TrackGNN is fully open-sourced and available at https://github.com/silvenachen/TrackGNN.