A Study on the Performance of Distributed Training of Data-driven CFD Simulations
Sergio Iserte's team shows multi-GPU training predicts fluid dynamics 10x faster than traditional solvers.
A new study from Sergio Iserte and colleagues at Universitat Jaume I and University of Łódź demonstrates that distributed GPU training can dramatically accelerate data-driven computational fluid dynamics (CFD). The team trained a deep learning time-series forecasting model to predict future states of a fluid simulation, replacing iterative PDE solving with AI inference. They benchmarked three training strategies: CPU-only, multi-GPU on a single node, and distributed multi-GPU across multiple nodes.
The results show that with slight code adaptations, distributed GPU training reduces training time while maintaining high prediction accuracy, achieving speedups that make real-time fluid simulation feasible. The paper, published in the International Journal of High Performance Computing Applications (Vol. 37, pp. 503–515, 2023), highlights how leveraging distributed parallelism can overcome the non-negligible cost of training data-driven surrogates, opening the door to faster CFD for engineering and climate modeling.
- Compared CPU, multi-GPU (single node), and distributed GPU training for a DL-based CFD surrogate model.
- Distributed training achieved high-accuracy predictions in a fraction of the time required by traditional CFD solvers.
- Minimal code changes were needed to adapt the model for distributed GPU parallelism.
Why It Matters
Distributed GPU training makes AI-driven fluid simulation practical, cutting compute time for engineering and climate models.