# NVIDIA's PET Benchmarks: Faster Molecular Simulation With Tensor Core FP32

New results separate matrix-multiplication speedups from the gains that survive a complete molecular dynamics step.

By Paras Madan | 2026-10-30


NVIDIA published new Point Edge Transformer benchmarks on October 29 showing up to 1.7x faster molecular dynamics steps on Blackwell using BF16x9 arithmetic. The evaluation examines whether faster matrix calculations can preserve the numerical behavior needed for atomistic simulation, rather than simply improve a kernel benchmark.

## What NVIDIA measured

The [technical report](https://developer.nvidia.com/blog/unlocking-tensor-core-performance-for-machine-learning-interatomic-potentials/) compares Tensor Core methods with FP32 baselines in PET, a machine-learning interatomic potential. Its larger water-system workload contains 17,280 atoms.

| Method and hardware | Large matrix multiplications | Complete molecular dynamics step |
| --- | --- | --- |
| BF16x9 on Blackwell | 2.2 to 2.5x faster | Up to 1.7x faster |
| TF32x3 on A100 | 1.7 to 1.9x faster | Up to 1.5x faster |

These are NVIDIA's reported results for the tested workloads. **A 2.5x matrix speedup is not a 2.5x simulation speedup.** The distinction matters when estimating how much additional scientific work a fixed compute budget can support.

The [UPET repository](https://github.com/lab-cosmo/upet) provides the PET-based models and interfaces for atomistic workflows. These models predict energies and forces; the repository recommends the PET-MAD-1.6 family for molecular dynamics. This is a scientific-model optimization result, not a general claim about chatbot inference.

## How BF16x9 keeps FP32 inputs

The [cuBLAS documentation](https://docs.nvidia.com/cuda/cublas/#floating-point-emulation) describes floating-point emulation as splitting higher-precision inputs into lower-precision components, computing partial products and recombining them. BF16x9 represents each FP32 value using three BF16 components and evaluates nine pairwise products.

The point is to use faster Tensor Core hardware while retaining the higher-precision interface. It is different from simply converting a model's weights and activations to BF16 and accepting the resulting numerical change.

cuBLAS lists BF16x9 support for compute capabilities 10.0 and 10.3 with CUDA 12.9 or later. Compatible single-precision operations can enable it with `CUBLAS_EMULATE_SINGLE_PRECISION=1`. That hardware requirement is important: the same environment variable is not a universal acceleration switch for every NVIDIA GPU.

The underlying [BF16x9 research paper](https://arxiv.org/abs/2605.16617) was submitted in May 2026. October's news is the PET application and its benchmark evidence, rather than the invention of this arithmetic method.

## Accuracy has to survive the simulation

NVIDIA's Ampere implementation replaces compatible linear layers with a custom TF32x3 layer. In the reported tests, TF32x3 preserved energy-conservation behavior close to FP32, while ordinary one-pass TF32 introduced more drift. Faster arithmetic was not equally useful in every configuration.

For a simulation team, this suggests a practical order of evaluation. Establish the original model's energy, force and trajectory behavior first, then compare the accelerated version against that baseline. Measure the complete step, including work outside matrix multiplication, and keep the model checkpoint and hardware configuration fixed.

A successful single-point prediction is only one part of that decision. Repeated updates can expose numerical differences that are difficult to see in an isolated calculation. The useful performance metric is therefore time to a scientifically acceptable result, with the acceptance criteria made explicit before comparing runtimes.

That is the promising part of this work: another performance option for an existing model. Whether it saves useful compute depends on the workload's numerical requirements, not the largest number in the benchmark table.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical teams explain research and engineering work clearly, without turning it into marketing language.

[Book a 30-minute call](https://cal.com/paras-madan/30min)