Pangu-Weather

Training compute

4×10²² FLOP

Parameters

256M

Published

Jul 5, 2023

Pangu-Weather is an AI model developed by Huawei (China), first published in July 2023. It works in the earth science domain, on tasks such as weather forecasting.

Training it took an estimated 4×10²² FLOP of compute (estimation method: hardware). The model has 256,000,000 parameters. It was trained on roughly 24.5T datapoints. Training ran on 192 NVIDIA V100 for about 1.5K hours. The compute alone is estimated at $51K in 2023 dollars.

Access: Open weights (non-commercial). Its weights are openly available. The reference paper has 197 citations. Epoch AI rates the confidence of this record as confident.

Full record

Organization: Huawei
Country of organization: China
Domain: Earth science
Task: Weather forecasting
Training compute: 4×10²² FLOP
Compute estimation method: Hardware
Parameters: 256,000,000
Dataset size: 24.5T
Training hardware: NVIDIA V100
Chips used: 192
Training time: 1,536 h
Training power draw: 114.6 kW
Training cost (2023 USD): $51K
Model accessibility: Open weights (non-commercial)
Open weights: Yes
Citations: 197
Epoch confidence: Confident