Accelerating GNN Inference for High-Speed Particle Track Reconstruction

Graph Neural Networks (GNNs) process large, sparse, and irregular graph, making them useful for particle track reconstruction. This project uses GNNs with TrackML which is a particle tracking challenge with the goal of reconstructing the paths of particles that traveled through detectors, and doing so as fast as possible without sacrificing accuracy. The GNN input graphs are processed using float32, but it is possible to also use Automated Mixed Precision (AMP) or manually convert to different precisions.