NVIDIA published an evaluation on 29 October of Tensor Core FP32 methods for machine-learning interatomic potentials (MLIPs), using the Point Edge Transformer (PET) model from EPFL's COSMO lab: a 26M-parameter PET-MAD checkpoint on 192-atom and 17,280-atom water systems, plus a 255M-parameter checkpoint on the 192-atom system. On Blackwell GPUs, cuBLAS's BF16x9 path - enabled with CUBLAS_EMULATE_SINGLE_PRECISION=1, with no model or kernel changes - accelerated large GEMMs by 2.2-2.5x on the 17,280-atom system and the full MD step by 1.7x; NVIDIA says the added error stays below the model's own run-to-run non-determinism. On Ampere, a CUTLASS TF32x3 kernel replaced the linear layers and gave 1.7-1.9x on GEMMs and 1.5x per MD step. The sharpest result is a structural relaxation of a 539-atom Cantor-alloy slab: FP32 converged in 218 steps and TF32x3 in 218-222 steps, all three replicates reaching the 1.0x10-3 eV/Angstrom force tolerance, while one-pass TF32 failed to converge in all three replicates within 500 steps.
Our takeSingle-point force error is the wrong acceptance test for a potential-energy surface, because errors are integrated over timesteps. The detail worth keeping is that one-pass TF32 looks acceptable on energies and forces yet breaks relaxation convergence, and adds visible point-to-point ripple to a CO/Pt(111) energy curve that FP32 and TF32x3 reproduce smoothly. That makes trajectory-level checks - energy conservation, relaxation step counts - the real gate, not a headline speedup. Scope is one model family, one optimizer and two GPU generations, so the ratio will move with system size.