Abstract Climate exerts both short-term and long-term impacts on the ecosystem carbon assimilation. However, the main climatic drivers for the variability of gross primary productivity (GPP) remain unclear across various timescales and vegetation types. Here, we combine the state-of-the-art machine learning algorithms with a well-established explanatory method to explore the impacts of climatic factors on long-term GPP variability at global FLUXNET sites across four timescales and six plant functional types. Results show that diffuse shortwave radiation (SWdif) dominates GPP variability at the sub-daily (half-hourly to three hourly) timescales especially for the tree species, and acts as the secondary contributor after air temperature at the daily or longer timescales. Attribution analyses further showed that the main effects of SWdif are much higher than their interactive effects with other climatic factors in regulating the GPP variability. By identifying the main climatic drivers, this study improves the understanding of the climate-driven GPP variability and provides important implications for the future projection of ecosystem carbon assimilation under global climate change.