2 min readEnergy

The 2023 Microsoft and TempoQuest wind forecast benchmark, explained

In April 2023 Microsoft and TempoQuest published a benchmark of WRF on Azure CPUs against AceCAST on Azure GPUs. Here is what was measured, on which releases, and how to read it now.

Wind turbines on a green field under a blue sky with scattered white clouds.

First published May 1, 2023 · Updated November 16, 2026

A wind operator or a utility that wants its own high-resolution forecast has to answer a plain question first: what will each run cost, and will it finish in time? In April 2023, Microsoft and TempoQuest published a benchmark that put numbers on that question for one setup on Microsoft Azure. Energy Theory, an energy news site, covered the work on May 1, 2023. Here is TempoQuest's own account: what was measured, on which software, and how to read the figures today.

Why wind forecasting is a computing problem

Wind power is sensitive to small forecast errors. Below a turbine's rated output, power rises with roughly the cube of wind speed, so a small miss in speed becomes a large miss in megawatts. Grid operators plan generation a day ahead, and every megawatt the wind forecast gets wrong has to be covered some other way.

A model grid has to be fine enough to see the hills and coastlines that steer the wind at a site, and finer grids need more computing. That is the gap the 2023 work addressed: running WRF, the Weather Research and Forecasting model maintained by NCAR, on GPUs instead of CPUs.

A historical Azure study

The NVIDIA Technical Blog published the study with Microsoft on April 28, 2023. It measured earlier AceCAST releases, 2.1 and 3.0.1, in separate comparisons on Azure. See the benchmark and cost example.

The worked example distinguishes the node-to-node result from the sixteen-CPU-node comparison and shows what the recorded November 2026 rates imply. Neither establishes a performance promise for your own forecast.

What this means for an energy team today

The benchmark is history, but the question behind it is current. Here is what TempoQuest offers for it today.

  • AceCAST. It runs WRF on NVIDIA GPUs with WRF-compatible inputs, namelists and outputs, on Linux systems on premises or in the cloud. Wind and solar planning is one of the uses it is built for.
  • Aerava. If you want a forecast without running a model, Aerava runs a wind forecast at 1 km for a US site: a 400 km run, 6 to 24 hours ahead, with wind at up to 16 heights you choose, including hub height, and a turbine cut-out check. The price is shown before you run.
  • See AceCAST 5, available now, for its physics, outputs and upgrade guidance.

Whichever route you take, score the forecast against your own measurements before you rely on it. A met mast, a lidar or the turbines' own anemometers will tell you more about your site than any published benchmark. Our article on forecasting for wind and solar operators explains why small wind errors are so costly.

What to do next

Price a forecast, or see wind energy uses.

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.