Atomistic GPU Batching

gpu computingparallel computingmolecular dynamicsmolecular simulationsperformance optimizationcetz

Comparison of three strategies for running multiple atomistic simulations on GPUs:

1. Unbatched - Sequential execution with minimal GPU utilization (0-5%)
2. Binned Auto-Batcher - Fixed-size batches improve GPU utilization (40-60%) but waste resources as structures complete
3. In-Flight Auto-Batcher - Dynamic reallocation eliminates GPU idle time by immediately swapping in new structures when others complete, maintaining high GPU utilization (80-90%)

Key points:

• GPU resources are wasted when not fully utilized by a single simulation
• Fixed batching improves utilization but suffers as structures complete at different rates
• Dynamic in-flight batching maximizes GPU utilization by maintaining a full batch at all times
• The overall time to process all structures is significantly reduced with in-flight batching
• Effective GPU utilization leads to higher throughput and energy efficiency