Nvidia's Earth digital twin AI model, once trained, can predict weather 3,000 times faster than a supercomputer using 1,000 times less energy.

Judge as of March 2024. True if NVIDIA's published materials or the related research (e.g. FourCastNet / Earth-2 / CorrDiff) report that the trained AI weather model makes forecasts about 3,000 times faster and with about 1,000 times less energy than a traditional supercomputer numerical model. Approximate figures of the same order (e.g. 1,000x faster, or energy savings in the thousands) that NVIDIA has published make it mostly true. Mixed if one of the two figures (speed or energy) is supported and the other is not. Mostly false if the figures come from a narrow benchmark (e.g. one downscaling task) stated as general weather prediction, or are much larger than published results. False if published results contradict the figures outright (e.g. gains of only tens of times).

Set before the research began.

1. E1 NVIDIA Announces Earth Climate Digital Twin NVIDIA Corporation (GlobeNewswire release, investor relations site) · 18 March 2024 Own record

   “Earth-2’s APIs offer AI models and employ a new NVIDIA generative AI model called CorrDiff, using state-of-the-art diffusion modeling, that generates 12.5x higher resolution images than current numerical models 1,000x faster and 3,000x more energy efficiently.”

2. E2 Nvidia creates digital twin of whole Earth — we might finally get an accurate weather forecast Tom's Guide · 21 March 2024

   “Nvidia says the images are 12.5 times higher resolution than previous models, can run models 1,000 times faster and is 3,000 times more energy efficient due to running on a GPU.”

3. E3 Power Tool: Generative AI Tracks Typhoons, Tames Energy Use NVIDIA Blog · 2 June 2024 Own record

   “Compared to traditional methods under these conditions, it slashes cost by 50x cost and energy use by 25x a year.”

NVIDIA's March 2024 Earth-2 announcement [E1] puts the figures on CorrDiff, a generative model that produces images at 12.5x higher resolution than current numerical models (a downscaling task). It says CorrDiff does this 1,000x faster and 3,000x more energy efficiently. Tom's Guide repeats the same figures and attributes them to NVIDIA [E2]. The claim swaps the two numbers (3,000x faster, 1,000x less energy) and presents a narrow downscaling benchmark as general weather prediction, which the criteria rate mostly false.