Skip to content

GPU execution

Start with the prepared simulation workflow: construct a native CPU aerosol, upload once, evolve supported processes on CUDA, and download a detached snapshot for analysis. Warp is a declared Particula dependency; CUDA requires an available NVIDIA GPU and compatible driver. Device selection is explicit. Direct kernels remain available for manual resource control.

Task Guide
Run native aerosol processes on CUDA Prepared simulations
Choose the supported execution path Execution, support matrix, backend selection
Move state between CPU and Warp Transfers
Manage fixed-capacity particle storage Slots, capacity exhaustion
Configure direct condensation Condensation, configuration
Configure direct coagulation Coagulation, configuration
Apply direct dilution or wall loss Dilution and wall loss
Inspect, finalize, or restart a resident session Checkpoints
Capture a stable resident loop on native CUDA Graph capture
Diagnose rejected calls Troubleshooting

Use the GPU examples for runnable entry points. CUDA capture has no CPU fallback. Qualification gates remain separate from implementation support and do not provide measured performance recommendations.

Explore this topic: GPU workflows