Scalable Deep Learning for Greenland Ice Bed Topography Prediction

Abstract

The rapid evolution of deep learning and GPU-accelerated computing has revolutionized large-scale geospatial modeling, yet the application of these advances to high-resolution ice-sheet bed topography prediction remains challenging due to the sheer volume of input data and the computational demands of deep convolutional neural networks. In this thesis, we investigate the performance tradeoffs of increasing spatial extents on the Greenland Upernavik dataset. The experiments are conducted on distributed multi-node multi-GPU training pipelines on the BedTopoCNN architecture. Three spatial extents—600×600, 1200×1200, and 2400×2400 pixels—are benchmarked on both the Frontera and CHIP highperformance computing clusters via PyTorch’s DistributedDataParallel framework over up to four nodes. We measure epoch time, strong scaling efficiency, and analyze how larger input patches impact convergence speed and resource utilization. Overall, our benchmarks show that while smaller 600×600 patches deliver the best throughput, enabling full 20000-epoch runs in under 48 hours with excellent scaling and stable accuracy. After expanding to 1200×1200, it hits a practical limit on single GPUs but gains efficiency from multi-GPU parallelism. Finally, pushing to 2400×2400 overwhelms even 16-GPU configurations. But adding a bottleneck layer helps substantially to meet the 48-hour training time at the largest spatial extent.