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.
Weather data is arriving faster than most forecast operations can use it. Why more observations need more computing before they improve a forecast, and where GPUs fit.

Weather data is arriving faster than most forecast operations can use it. New satellites, radars and ground sensors report more often and in finer detail every year. But an observation improves a forecast only if a model can take it in and finish running before the forecast is needed. For an operations lead, that turns a science question into a budget question: how much computing does it take to turn more data into better decisions?
Every numerical forecast starts from an estimate of the atmosphere as it is now. Errors in that starting picture grow as the model runs forward, so a forecast can be no better than the observations behind it allow.
The coverage is uneven. Upper-air stations measure temperature, humidity and wind through the depth of the atmosphere, and across much of the world they are far apart. The World Meteorological Organization map below shows the spacing of upper-air reports by country. Dark red means the widest gaps.

Closing those gaps pays. A World Bank report on the value of surface-based meteorological observation data estimated that every dollar invested in new observational data could return at least twenty-six dollars in socio-economic benefit.
Satellites have changed the supply side. The chart below, from Christopherson, Chandra and Quanbeck (2019), counts satellites by launch year and type of user. Commercial launches rise sharply from 2014.

More data does not help by itself. It has to pass through data assimilation, the step that blends new observations with a recent short forecast to produce the starting picture for the next run. Assimilation is among the most expensive parts of a forecast system, because the leading methods run the model many times over to work out how much weight each observation deserves.
The cost does not stop there. Once the starting picture holds more detail, the forecast model needs a finer grid to keep it. Finer grids need much more computing. Data, assimilation and resolution all pull on the same budget.
The clearest public statement of the problem came from NOAA's Science Advisory Board. As HPCwire reported in January 2022, the board's report found that improvements in weather forecasts were directly limited by the computing available to develop, test and run new models and data assimilation. It called for on the order of 100 times NOAA's operational computing capacity of the time over the following decade.
That was a statement about a national weather service in 2022, but the logic applies to any organization that runs its own forecasts: buying more data without more computing leaves the data unused.
There are two ways to get more computing. One is more of the same processors. The other is processors that do more of this kind of work each. Running the forecast model on GPUs is the second route.
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, on Linux systems on premises or in the cloud. AceCAST's part is the forecast run itself. The steps that prepare observations and starting conditions stay with the tools your team already uses, and AceCAST reads the WRF-compatible input they produce.
See the benchmark and cost example.
More observations and a finer grid are investments, and they should be checked like investments. Forecast verification scores each forecast against stations, lidar or radar, using measures such as bias, mean absolute error and root-mean-square error. Run it before and after a change, over enough days to be sure. Separately, TempoQuest is developing a Forecast Verification Tool, designed to improve the verification of weather forecasts.
Two options avoid the hardware question. Aerava runs a forecast for a site you pin and shows the price before you run. Custom Weather Data delivers project-specific model outputs, with typical requests at around 1 km or 3 km resolution and at 15-minute, 30-minute, hourly or daily cadence.
See AceCAST 5, available now, for its physics, outputs and upgrade guidance.

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.