3 min readCompany news

AceCAST and AMD Instinct GPUs: what was announced in March 2022

In March 2022 TempoQuest announced work with AMD to port AceCAST to AMD Instinct MI200 series accelerators. What that does and does not tell you today, and how to check hardware support.

An AMD Instinct MI210 accelerator card, long and black with a gold edge connector, shown at an angle on a white background.

First published March 22, 2022 · Updated November 16, 2026

A team buying GPUs for weather modeling is making a decision it will live with for years. The first question is which hardware the forecast software runs on. In March 2022, TempoQuest announced work with AMD to port AceCAST to AMD Instinct MI200 series accelerators. Here is what that announcement does and does not tell you today, and how to check hardware support before you spend money.

What was announced in March 2022

In March 2022, TempoQuest announced that it was working with AMD to port AceCAST, its GPU version of the WRF weather model, to AMD Instinct MI200 series accelerators.

A 2022 announcement is not a statement of what AceCAST supports today. If AMD hardware matters to your purchase, ask us directly before you commit.

What AceCAST runs on today

AceCAST runs WRF, the Weather Research and Forecasting model maintained by NCAR, on NVIDIA GPUs: Linux systems, on premises or in the cloud. It keeps the WRF workflow, with WRF-compatible inputs, namelists and outputs. The AceCAST documentation is the reference for requirements.

See the benchmark and cost example.

Why the GPU vendor matters for a weather model

A CPU program usually moves between processor brands with little effort. GPU software is different. Each GPU vendor has its own programming tools and libraries, and code written for one does not automatically run on another.

A weather model is a very large body of scientific code. It has a core that moves air around the grid and many optional physics schemes for clouds, rain, radiation, turbulence and the land surface. Supporting a second GPU vendor means engineering work across all of it, followed by careful checking that the forecasts still agree with the reference model.

For a buyer, three things follow.

  • Check support for the exact hardware. "Runs on GPUs" is not enough. Ask which vendor, which models and which operating system.
  • Check your own configuration, not a demonstration. Support for the hardware and support for the physics options in your namelist are separate questions. You need a yes to both.
  • Compare cost per forecast, not cost per card. What matters is how long your forecast takes on a machine and what that machine costs per hour.

How to check before you buy

Hardware is hard to return. Three steps keep the risk low.

  1. Read the requirements. Start with the AceCAST documentation, then ask us about anything it does not cover.
  2. Run your own case. An AceCAST trial runs for 14 days on one machine. Renting GPUs in the cloud for the test lets you measure before you buy.
  3. Price it. Cloud list prices give a reference point. On November 11, 2026, the Azure list price of a Standard_ND96amsr_A100_v4 virtual machine, which has 8 NVIDIA A100 GPUs, was $32.77 per hour for Linux, pay-as-you-go, in the East US region. Prices change, so check the current figure. The estimator on our pricing page turns hourly rates into a cost per run.

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

If you would rather not own hardware at all, Aerava runs forecasts for a site you pin and shows the price before you run, and Custom Weather Data delivers datasets to your specification. In both cases the hardware question is ours to answer, not yours.

What to do next

Ask about hardware, or request a trial.

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.