Acceptability of TF32 Tensor Core Precision for NNP-based Molecular Dynamics
Determine whether using TF32 reduced-precision arithmetic on NVIDIA Tensor Cores introduces excessive numerical noise that compromises the physical accuracy and stability of neural network potential evaluations in molecular dynamics simulations, and, if necessary, characterize acceptable error bounds and mitigation strategies for such computations.
References
So far, tensor cores arithmetics is based on a reduced floating point format (TF32) compared to FP32 of CUDA cores. It is unclear if the level of noise introduced is excessive for physics.
                — Machine Learning Potentials: A Roadmap Toward Next-Generation Biomolecular Simulations
                
                (2408.12625 - Fabritiis, 17 Aug 2024) in Co-evolution of models, software, and hardware