In their research, Utkarsh Utkarsh and his colleagues present a method for solving ensembles of ordinary differential equations (ODEs) and stochastic differential equations (SDEs) on GPUs. Unlike other methods, their approach does not require any code changes and achieves excellent performance, outperforming the vectorizing map approach implemented in JAX and PyTorch. The method is also vendor-agnostic, delivering performance across NVIDIA, AMD, Intel, and Apple GPUs. Additionally, the solvers implemented in this method are versatile, supporting event handling, automatic differentiation, and incorporating datasets via the GPU’s texture memory. The authors have made their software available as an open-source library.
https://arxiv.org/abs/2304.06835