Unlocking Tensor Core Performance for Machine Learning Interatomic Potentials
According to NVIDIA Developer. NVIDIA details how Tensor Core FP32 methods accelerate machine learning interatomic potentials on Blackwell and Ampere GPUs while preserving baseline accuracy.
- • BF16x9 on NVIDIA Blackwell GPUs accelerates large GEMMs in the Point Edge Transformer by 2.2–2.5x for a 17,280-atom system.
- • TF32x3 on NVIDIA Ampere GPUs speeds up major GEMMs by 1.7–1.9x and the full molecular dynamics step by up to 1.5x.
- • TF32x3 preserves potential-energy smoothness and energy conservation in NVE trajectories, matching FP32 behavior.
- • Structural relaxation of a 539-atom Cantor-alloy slab converges in 218–222 steps with a 1.5x per-step speedup under TF32x3.
Tensor Core Performance
Atomistic simulations help researchers understand molecular and material behavior across chemical reactions and phase transitions, though they face trade-offs between quantum-mechanical accuracy and computational cost. Machine learning interatomic potentials bridge this gap using neural networks. Tensor Core FP32 methods decompose FP32 operands into lower-precision components, evaluate cross-products on Tensor Cores, and recombine results in FP32 to boost matrix-multiplication throughput.
Benchmark Evaluation
BF16x9 on NVIDIA Blackwell GPUs accelerates large GEMMs in the Point Edge Transformer by 2.2–2.5x for a 17,280-atom system and 1.2–1.4x for a 192-atom system with error below model non-determinism. Meanwhile, TF32x3 on NVIDIA Ampere GPUs speeds up major GEMMs by 1.7–1.9x and full molecular dynamics steps by up to 1.5x while keeping force error near the FP32 baseline and preserving potential energy smoothness.
- 1. developer.nvidia.comOpen source