Pangu-Weather makes a 24-hour global forecast in about 1.4 seconds

According to Huawei’s announcement of its Nature paper, the Pangu-Weather deep-learning model produces a global 24-hour weather forecast in about 1.4 seconds on a single GPU - roughly a 10,000-fold speed-up over traditional numerical weather prediction, which requires large supercomputers running for far longer. The model was trained on 43 years of ERA5 reanalysis data and was the first AI system Huawei said forecast more accurately than leading numerical methods across the one-hour to seven-day range.