Unlocking Tensor Core Performance for Machine Learning Interatomic Potentials
AI summary NVIDIA tested FP32-emulating Tensor Core methods on the Point Edge Transformer, a machine learning interatomic potential built by Lab COSMO at EPFL. On Blackwell GPUs, cuBLAS's BF16x9 path made large matrix multiplications 2.2–2.5x faster for a 17,280-atom system, with errors smaller than the model's own non-determinism. On Ampere A100 hardware, TF32x3 via CUTLASS delivered 1.7–1.9x faster GEMMs and up to 1.5x faster molecular dynamics steps while keeping energy conservation close to FP32. In relaxing a 539-atom Cantor-alloy slab, TF32x3 converged in 218–222 steps, whereas single-pass TF32 never reached the force tolerance.
