3 min readAceCAST

Time to solution: how to get a high-resolution forecast out sooner

A forecast is only useful if it arrives before the decision it informs. What drives time to solution for a high-resolution WRF run, and what running WRF on GPUs changes.

A red and white striped windsock held out by a steady wind on a lattice mast, against a blue sky with thin white cloud.

First published June 1, 2022 · Updated November 16, 2026

A forecast is only useful if it arrives before the decision it is meant to inform. A power trader needs tomorrow's wind before the market closes. An airport needs the afternoon's storms before the morning planning call. Time to solution is the name for that clock: the wall-clock time from the moment a model run starts to the moment its output is ready to use. This article explains what drives it for a high-resolution WRF forecast, and what running WRF on GPUs with AceCAST changes.

What time to solution includes

A forecast run has four stages, and all of them are on the clock.

  1. Getting the starting data. The run waits for a global or national model to supply its initial and boundary conditions.
  2. Pre-processing. That data is interpolated onto your own grid.
  3. The model run. The model steps the atmosphere forward in short time steps until it reaches the end of the forecast.
  4. Output and post-processing. Files are written, and maps, tables or data feeds are made from them.

For a large, fine grid, the model run is usually the longest stage. It is also the one that grows fastest when you ask for more detail.

Why resolution sets the clock

A weather model divides the atmosphere into grid cells. Grid spacing limits the detail a forecast can represent. To resolve the storm, or the ridge line a wind farm sits on, the cells have to shrink.

Finer grids need much more computing.

On a fixed CPU cluster, the extra computing forces a trade. You can have the fine grid, or the large area, or the early delivery, but not all three.

Three ways to spend a faster model

If the model run takes less time, a forecast team can spend the saving in different ways.

  • Deliver sooner. Keep the same forecast and hand it over earlier.
  • Add detail. Keep the same delivery time and run a finer grid or a larger area.
  • Run more often. Start a new run each time new data arrives, or run several versions of the forecast to see how far they disagree.

Which of these is worth most depends on the decision. A trading desk usually wants the early delivery. A team forecasting for complex terrain usually wants the finer grid. A team that has to state its confidence usually wants the extra runs.

What has been measured

AceCAST runs WRF, the Weather Research and Forecasting model maintained by NCAR, on NVIDIA GPUs. It keeps the WRF workflow, with WRF-compatible inputs, namelists and outputs, so the model run is the stage that changes.

Two published sources give measured results. Both are for earlier AceCAST releases.

Read these as measurements, not promises. Speed depends on the domain, the physics options and the hardware. That is one reason the two sources above report different figures. The only number that matters for your operation is the one from your own configuration, and a trial is how you get it.

Cost is a separate question from speed. A GPU machine usually costs more per hour than a CPU machine, so what counts is the cost of each finished forecast. Our article on the 2023 Azure benchmark works through one example using list rates recorded in November 2026.

AceCAST today

For what each release supports, see the AceCAST documentation.

See AceCAST 5, available now, for its physics, outputs and upgrade guidance.

If you would rather not run a model at all, Aerava runs a forecast for a site you pin and shows the price before you run, and Custom Weather Data delivers datasets to your specification.

What to do next

Request a trial and time your own configuration, or price a forecast first.

Your forecast

Have a forecast question of your own?

Tell us the decision it has to serve. One short form reaches the team, and we review every request.