NVIDIA blog details Tensor Core speedups for ML interatomic potentials

News

NVIDIA says Tensor Core FP32 methods can accelerate PET molecular dynamics while keeping energy, force and relaxation behavior close to FP32.

NVIDIA said its Tensor Core FP32 methods can speed up machine learning interatomic potentials without retraining or changing model architecture, while keeping behavior close to an FP32 baseline in the tested workloads, according to a developer blog published Oct. 29, 2026.

Evidence and context

The post evaluates the Point Edge Transformer, or PET, a transformer-based machine learning interatomic potential developed by Lab COSMO at EPFL. NVIDIA said PET is a demanding test case because much of its runtime is already in optimized general matrix multiplications, leaving limited headroom for further optimization. The company compared BF16x9 on Blackwell GPUs and TF32x3 on Ampere GPUs, measuring kernel and end-to-end molecular dynamics performance as well as trajectory-level behavior.

For the BF16x9 path on Blackwell, NVIDIA said large PET GEMMs were accelerated by 2.2x to 2.5x on a 17,280-atom workload, while molecular-dynamics step speedups ranged from 1.2x to 1.4x on a 192-atom water box and 1.4x to 1.7x on the 17,280-atom system, depending on whether the cuBLAS strategy was “performant” or “eager.” NVIDIA said the BF16x9 error was below PET’s own level of non-determinism.

For TF32x3 on Ampere, NVIDIA said large PET GEMMs accelerated by about 1.7x to 1.9x on an A100 Tensor Core GPU. In the small 26-million-parameter checkpoint, TF32x3 sped up the molecular-dynamics step by 1.5x on the larger system, according to the post. The company also said a 255-million-parameter checkpoint reached a 1.4x speedup on the 192-atom system, with a force error of 3.9 × 10⁻⁶ eV/Å.

NVIDIA said the precision methods mattered because atomistic simulations are sensitive to small numerical differences that can accumulate during force calculations and over many timesteps. In its tests, TF32x3 followed the smooth FP32 CO/Pt(111) potential-energy curve, while one-pass TF32 showed visible point-to-point variation. In a 30 ps NVE simulation of a 192-atom water box, the company said FP32 and TF32x3 showed comparable bounded energy fluctuations, while one-pass TF32 drifted more.

The post also described a structural-relaxation test on a 539-atom Cantor-alloy slab using the FIRE2 optimizer in the NVIDIA ALCHEMI Toolkit. NVIDIA said FP32 converged in 218 steps and TF32x3 in 218 to 222 steps across three replicates, all reaching the 1.0 × 10⁻³ eV/Å force tolerance. By contrast, the post said one-pass TF32 failed to converge within 500 steps in all three runs, and still failed after the limit was doubled to 1,000 steps.

What the methods do and where they differ

According to NVIDIA, BF16x9 on Blackwell decomposes each FP32 operand into three BF16 components and recombines nine products in FP32, with cuBLAS able to enable the path through the CUBLAS_EMULATE_SINGLE_PRECISION=1 environment variable or related settings. TF32x3 on Ampere splits each FP32 operand into a leading TF32 component and a residual component, evaluates three cross-products and omits the residual-by-residual term. NVIDIA said the TF32x3 implementation used CUTLASS’s OpMultiplyAddFastF32 operator and preserved FP32 inputs and outputs at the model interface.

The company said end-to-end gains depend on system size and how much runtime is spent in GEMMs. Smaller systems benefit less because launch overhead, neighbor processing and graph operations make up a larger share of runtime, while larger systems expose more GEMM work. NVIDIA also said TF32x3’s performance depends on kernel tuning for the model’s matrix shapes and hardware, and that BF16x9 and TF32x3 should be validated against workload-specific accuracy and stability requirements.

The blog is a technical performance report rather than a product launch. It does not provide pricing, market impact or a broader adoption estimate, and its results are limited to the PET workloads and validation tests described in the post.