AI-Accelerated CFD Simulations Adapted for IPU Platform

Researchers have adapted AI-accelerated computational fluid dynamics (CFD) simulations to run on Intelligence Processing Units (IPUs), demonstrating performance improvements. The study, detailed in a paper on arXiv (2605.00462), uses custom TensorFlow from the Poplar SDK to train machine learning models on data from OpenFOAM simulations. The team targeted the IPU-POD16 platform and investigated ease of use and scalability. By employing the popdist library, they overcame a host-side data feeding bottleneck, achieving up to 34% speedup. However, using data parallelism across two IPUs did not improve performance due to communication overhead. This work contributes to the emerging field of AI for simulation, where traditional numerical methods are augmented by artificial intelligence.