Regional forecasts on one NVIDIA GPU

Your guide to installing wrf_gpu and running a WRF-compatible forecast.

wrf_gpu rewrites WRF v4's ARW dynamics and physics for JAX/XLA/Pallas. Start with prepared WRF inputs and a supported namelist, run time integration on your GPU, and use familiar NetCDF tools to read the history files.

Get your first forecast running¶

1. Install

Check Python, CUDA JAX, your NVIDIA GPU and WRF runtime tables.

2. Run your first forecast

Try the bundled 3 km Alpine case, then inspect its output.

3. Forecast your region

Set up a case from open GFS data with gpuwrf quickstart (experimental).

4. Read your results

WRF history, full output, cadence and checkpoint files.

Prefer help from an AI coding assistant? Follow the AI quickstart. The repository's operator runbook uses the same public commands.

Already have a WRF case?¶

Bring namelist.input, wrfinput_d0* and wrfbdy_d01. Keep your preprocessing and downstream analysis. For fixed nests, use --domains-from-namelist. The input guide explains preparation and running the model shows the commands. Products, station verification, ensembles and export are on tools and commands.

Choose your configuration¶

Check the physics menu, namelist compatibility and nesting guide for your case. Options have explicit support status and unsupported selections stop with a named reason. What is not covered yet is on the limitations page.

See the evidence and plan your workload¶

Original Fortran CPU-WRF comparisons cover selected regional forecasts through 24 and 72 hours. Validation explains the method, the fixed limits and the results, with every hourly error plot. These comparisons test agreement with WRF, not forecast skill against observations.

Measured regional workloads show substantial speed gains on one RTX 5090. See performance and energy for the exact hardware, cases, clock definitions and measured numbers. Independent forecasts can share the card.

Read how the rewrite works or the release history when you want more detail. This independent implementation is not affiliated with UCAR/NCAR; see credits and licensing.

What's new in v0.4.0¶

Released 17 October 2026. More than 80 WRF 4.7.1 physics options now run on the GPU and passed a comparison with the original CPU-WRF, together with nudging (FDDA), digital-filter initialization, two-way nesting, stochastic physics, ensembles and forecast tools such as products, station verification and export. New and experimental: gpuwrf quickstart sets up a forecast for your own region from open GFS data (see the region guide). The comparisons ran 66 configurations over 24 hours (Goddard radiation: six hours), all within the fixed acceptance limits. Read the release notes, the physics and namelist pages for the status of each option and the validation details. The large microphysics lookup tables for v0.4.0 are not yet attached to the GitHub release; until they are, runs that need them stop with an “asset unavailable” message.