GPU vs CPU: choose a path

Use this page to decide whether you need a GPU build and how to run one successful GPU-capable binary without mixing toolchains.

Decision tree

        flowchart TD
  A[Start] --> B{Do you need CUDA or ROCm FFT / device-resident solves at scale?}
  B -->|No| C[CPU tree: simpler, fewer moving parts]
  B -->|Yes| D{Can you build HeFFTe with CUDA or HIP matching your toolkit?}
  D -->|No| E[Fix HeFFTe + toolkit first — see INSTALL]
  D -->|Yes| F[Separate GPU build directory + matching OpenPFC flags]
  C --> G[mpirun CPU tungsten / examples]
  F --> H[JSON backend + tungsten_cuda or tungsten_hip]

  • Stick to CPU if you are learning the framework, debugging physics, or your cluster does not expose GPUs consistently. The same spectral APIs exist; only the FFT backend and runtime differ.

  • Move to GPU when you have a HeFFTe build with CUDA or ROCm enabled that matches your nvcc / ROCm stack, and you use shipped GPU binaries (tungsten_cuda, tungsten_hip) or your own code paths that pull in device FFT.

Preconditions (all must align)

Piece

Rule

MPI

Same implementation for OpenPFC, HeFFTe, and mpirun (INSTALL.md).

HeFFTe

GPU-enabled build (e.g. 2.4.1-cuda / ROCm variant) on CMAKE_PREFIX_PATH.

OpenPFC

-DOpenPFC_ENABLE_CUDA=ON or -DOpenPFC_ENABLE_HIP=ON in a fresh GPU build dir (build_cpu_gpu.md).

Golden path (after CPU works)

  1. Keep your working CPU tree (build-cpu).

  2. Create build-gpu (or build-hip) and configure with GPU flags + correct CMAKE_PREFIX_PATH to the GPU HeFFTe prefix (build_cpu_gpu.md).

  3. Run the GPU quickstart: tutorials/gpu_app_quickstart.md (backend in JSON, tungsten_cuda / tungsten_hip).

  4. Compare log shapes to CPU: example_run_output.md.

See also