NVIDIA reports up to 1.7x faster molecular simulation steps using Tensor Cores

NVIDIA reports that alternative matrix-multiplication methods accelerated complete molecular dynamics steps in the PET machine learning model by up to 1.7x on its B300 GPU and 1.5x on A100. PET predicts atomic energies and forces for simulations of molecules and materials. The methods split single-precision calculations into lower-precision components, process them on Tensor Cores and recombine the results. Neither requires retraining or changing the model architecture. The Blackwell method can be enabled through cuBLAS; the Ampere implementation uses custom linear layers. In tested energy-conservation and structural-relaxation workflows, the Ampere method retained behavior close to the single-precision baseline. Gains depend on system size, matrix workload and kernel tuning.