Fig. 15From: Scalable training of trustworthy and energy-efficient predictive graph foundation models for atomistic materials modeling: a case study with HydraGNNAverage percentage of waiting time during three parallel tasks—data loading, forward pass, and backward passBack to article page