AceCAST 5: what's new
AceCAST 5 is available now. Here is what changed: WRF 4.8, four physics updates and four new model outputs.
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 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.
A forecast run has four stages, and all of them are on the clock.
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.
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.
If the model run takes less time, a forecast team can spend the saving in different ways.
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.
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.
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.
Request a trial and time your own configuration, or price a forecast first.

AceCAST 5 is available now. Here is what changed: WRF 4.8, four physics updates and four new model outputs.

Aerava runs a weather model around the site you pin and checks the result against your own storm, wind and turbine cut-out limits. Here is what you can run today, what it costs and what it says about accuracy.

One published 2023 benchmark and two Azure list prices, worked line by line: what a GPU VM costs against one CPU VM and against sixteen, and what the sums leave out.
Your forecast
Tell us the decision it has to serve. One short form reaches the team, and we review every request.