Nvidia’s Tensor Core methods speed up scientific simulations without sacrificing FP32 stability

Nvidia says two Tensor Core methods can speed up machine-learning interatomic-potential simulations while keeping results close to the FP32 baseline. The practical prize is faster molecular modelling without the numerical wobble that can turn a long simulation into an expensive guess.

NVIDIA Watch analysis

What happened

In a technical post, Nvidia tested BF16x9 on Blackwell GPUs and TF32x3 on Ampere GPUs using the Point Edge Transformer, a model for predicting atomic energies and forces. The company reports that BF16x9 sped up large matrix multiplications by 2.2 to 2.5 times; on the largest molecular-dynamics workload, the full simulation step ran 1.7 times faster.

On Ampere, Nvidia’s TF32x3 method sped up major matrix multiplications by about 1.7 to 1.9 times and a large-system simulation step by 1.5 times. Nvidia says TF32x3 preserved energy conservation and smooth potential-energy behaviour in its tests. Its faster, less accurate one-pass TF32 comparison showed energy drift and failed to meet the force tolerance in the reported structural-relaxation test.

The methods use lower-precision Tensor Core operations to calculate FP32 results. Nvidia says BF16x9 on Blackwell can be enabled for compatible matrix multiplications without changing model weights or architecture. Read Nvidia’s technical post.

Why it matters

Scientific simulations need more than a plausible answer at one moment: small force errors can accumulate over many steps and distort the behaviour being modelled. The reported results suggest researchers may be able to use more of modern GPUs’ Tensor Core throughput without accepting the instability seen with a cruder low-precision shortcut.

That could make larger or longer atomistic simulations more practical. The evidence here is a test on one model and selected workloads, though, not a guarantee that every scientific model will get the same speed-up.

Our read

This is a useful case where faster computing and scientific fidelity are tested together, rather than speed being left to do all the talking. Nvidia’s strongest result is not simply a quicker matrix multiplication: it is the reported combination of a 1.5-times faster large-system step and stable energy behaviour with TF32x3. Researchers should treat that as an encouraging recipe to test on their own workloads, not a universal performance promise.

What to watch

• Whether independent teams reproduce the speed and accuracy results on other interatomic-potential models.
• How much end-to-end benefit the methods deliver in different simulation workloads.
• Whether more frameworks expose these Tensor Core paths without requiring custom kernels.

Discussion spark: For scientific AI, would you accept a smaller speed-up to preserve stable simulation behaviour, or should researchers push for the fastest method and validate each workload separately?

Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.