Unlocking Tensor Core Performance for Machine Learning Interatomic Potentials
NVIDIA, Thursday, October 29th, 2026
NVIDIA shows that multi-pass TF32x3 and BF16x9 tensor-core modes speed up machine learning interatomic potentials without losing accuracy.
On Blackwell GPUs, BF16x9 accelerates large matrix multiplications in the Point Edge Transformer by 2.2 to 2.5x for a 17,280-atom system and 1.2 to 1.4x for a 192-atom system, with error below the model's own non-determinism.
On Ampere GPUs, TF32x3 speeds major GEMMs by 1.7 to 1.9x and the full molecular dynamics step by up to 1.5x with force error near FP32.
TF32x3 also preserves energy conservation in NVE runs, where one-pass TF32 drifts.
In a 539-atom Cantor-alloy relaxation, TF32x3 converged in 218 to 222 steps while one-pass TF32 failed after 1,000.