- https://doi.org/10.1109/sci68648.2025.11333847
HPC-Accelerated Graph Multi-Scale Spatio-Temporal Transformer for Traffic Forecasting: A Multi-GPU Implementation on METR-LA
- Dec 9, 2025
- M S Deepika +3 more
Traffic prediction is a critical component of Intelligent Transportation Systems (ITS), requiring advanced models to capture complex spatio-temporal dependencies across sensor networks while meeting real-time constraints. This paper presents an HPC-optimized implementation of Gated Multi-Scale Spatio-Temporal Graph Transformer (GMSTGT) on the METR-LA dataset, utilizing distributed training across 4 GPUs on Param Utkarsh HPC system for simultaneous multi-horizon prediction. The proposed GMSTGT architecture integrates dynamic graph construction with hierarchical attention mechanisms, reducing memory complexity from $\mathbf{O}\left(\mathbf{N}^{2}\right)$ to $\mathrm{O}(\mathrm{N} \log \mathrm{N})$ for 207 -sensor networks. Experimental validation demonstrates Mean Absolute Error (MAE) of 2.68 (15-min), 3.24 $(30-\mathrm{min})$, $3.89(45-\mathrm{min})$, and $4.45(60-\mathrm{min})$ with 15.8 ms inference latency at batch size 64. The distributed implementation achieves nearly linear scaling ($\approx 4$) speedup over sequential approaches while maintaining prediction accuracy within 0.1% of single-GPU baselines.
