AIFS 2.0 runs without Ampere — but any hardware does not mean fast

The main obstacle to running an open weather model turned out to be not the weights themselves but the dependency on flash-attn. AIFS Single 2.0 originally expects an NVIDIA GPU of Ampere architecture or newer — an A100, H100 or RTX 30xx, for instance. Without such a card the "open" model remained for many more of a demo than a working tool.

A compatibility patch changes that: flash-attn calls are redirected to SDPA from PyTorch. So a forecast can be built on any GPU through CUDA, on Apple MPS or on CPU. But an important clarification: being able to run and running fast are different things. On CPU the model technically works, yet inference will be slow.

What exactly AIFS opens up

AIFS is not an ordinary weather widget but ECMWF's global medium-range forecasting model. It is trained on meteorological data and produces a forecast instead of solving the physical equations of the atmosphere step by step. By ECMWF's estimate a single forecast requires roughly 1,000 times less energy than a traditional system based on physical models.

That said, the system does not get by entirely "without physics". For the initial state AIFS uses the same assimilation data as IFS, which is based on physical laws. In other words, machine learning computes the atmosphere's evolution more cheaply, but the starting conditions still rest on the infrastructure of classical numerical forecasting.

The AIFS Single 2.0 weights are published under a CC BY 4.0 licence. That gives researchers and developers not merely the chance to look at the model but to run their own experiments: change the forecast horizon, compare results with observations, study errors and build applied prototypes for energy, maritime tasks or teaching.

The simplest route is the cloud

If you have no powerful graphics card of your own, a forecast can be run through Hugging Face Jobs. The guide uses an NVIDIA T4 with 16 GB of memory: the job downloads the data and the model, builds a 48-hour forecast and saves the result to a Hugging Face dataset.

Execution takes about 10 minutes, and the configuration itself costs roughly $0.40 an hour with per-minute billing. You pay for the time to launch and run the job; building the container is not billed. This is no longer "running on your own hardware", but the barrier to entry turns from buying a GPU into paying for a small cloud experiment.

The local scenario is more useful for regular work and for learning. Initial conditions are downloaded from ECMWF Open Data free of charge and without an account, then stored in a local cache. After that you can choose the forecast horizon, fields such as temperature and pressure, and build maps of the results.

Limitations that are easy to forget

AIFS does not replace the whole meteorological system. It requires two consecutive analyses six hours apart, and the first data download takes several minutes. On weak hardware a forecast will run noticeably longer. In the cloud you have to account separately for compute cost and result storage.

Besides, the claim of "up to 20% better" quality relates to forecasting tropical cyclone tracks and is given with reference to ECMWF data — it is not a universal guarantee of superiority over physical models for any location, variable and weather situation. And the AIFS ENS ensemble configuration with 51 members does not turn a local run of Single into a full assessment of uncertainty either.

And here is the practical conclusion: opening the weights solves almost nothing by itself if running them requires a rare stack and an expensive GPU. Real usefulness appears only together with compatibility, a ready environment and an intelligible way to obtain the initial data. In the case of AIFS that last metre has finally been made short enough — though not free and not instant.

If you needed to build your own forecasts regularly, would you choose a T4 in the cloud or a slow local run on affordable hardware? Source: All articles / Machine Learning / Habr