Skip to content

GPU workflows

Broader topicCPU and GPU execution

Choose an execution path with explicit device, transfer, and lifecycle requirements.

Follow the science

GPU workflows

Prepare resident simulations, manage transfers and checkpoints, or use direct kernels.

Accelerated simulations benefit from retaining state on the device, but preparation, workspace capacity, checkpoints, and stochastic state become part of the workflow. Prepared examples expose these choices alongside scientific diagnostics. A faster run is useful only if its supported model and evidence answer the same question.

A question to investigate

For a representative workload, how much time is spent advancing physics versus transferring or inspecting state, and what numerical agreement is required?

Start here

Prepared GPU simulations

Helpful background

Concepts to have in mind before starting.

Connect the ideas

Explore a neighboring question.

Read, calculate, investigate

Also called: CUDA workflows.