Results: Half-precision inference of AI weather prediction models

Results behind the paper Half-precision inference of AI weather prediction models: negligible loss of skill at a fraction of the cost (Siegert, 2026). Five publicly available AI weather models (Aurora, deterministic and ensemble checkpoints; Pangu-Weather; FourCastNet; SFNO; FastNet) were run from 52 initial dates in 2024 in fp32, and with their networks in fp16, bf16 and simulated 8-bit formats while the physical state stays in fp32, and verified against ERA5. Contents (unpack into data/results/ of the code repository, https://github.com/sieste/ai-nwp-fp16, whose nwpprec.plots then redraws the paper's figures and tables): scores/: verification against ERA5 of every forecast: <model>_<tag>.csv with columns date, hour (UTC), var, lead (hours), region, rmse, bias, mae, rmse_vs_fp32 (departure from the fp32 forecast); tags fp32, fp16, bf16, bf16-whole (Aurora converted as a whole), fp16-<8-bit format>; Aurora's ensemble: aurora_ens_ensemble_<precision>.csv points/: forecast time series at named locations, per model, precision and date sensitivity/: one-at-a-time, layer-type, outlier-channel and depth experiments headroom/: largest fp16 magnitudes per module and step storms/: tracks of four named storms benchmark.json: time, memory and energy per model step on an NVIDIA GeForce RTX 5080 Not included: the saved forecast fields (about 150 GB) and the regional statistics (6 GB) behind the maps, latitude bands and spectra; the code regenerates them. Verification data: ERA5, Copernicus Climate Change Service (CC BY 4.0).