Unlocking Tensor Core Performance for Machine Learning Interatomic Potentials

NVIDIA Tensor Core Methods Boost Scientific Simulation Performance

Researchers have demonstrated substantial performance gains for machine learning interatomic potentials (MLIPs) on NVIDIA GPUs, leveraging novel Tensor Core methods that preserve the high numerical precision critical for scientific computing. An evaluation using the Point Edge Transformer (PET) model showed that these techniques can accelerate molecular dynamics simulations without requiring model retraining or sacrificing the accuracy needed for stable results. This is a significant development for fields like material science and chemistry, where the trade-off between computational speed and simulation fidelity has long been a limiting factor.

Key Performance Metrics and Technical Details

The study benchmarked two primary methods, BF16x9 on NVIDIA Blackwell GPUs and TF32x3 on NVIDIA Ampere GPUs, against standard FP32 calculations. By decomposing FP32 operations into lower-precision components and recombining the results, these methods utilize the high throughput of Tensor Cores while keeping numerical error exceptionally low—in some cases, below the model's own non-determinism. This approach proved effective even on the PET model, which is already highly optimized.

• BF16x9 on Blackwell: Accelerated large matrix multiplications (GEMMs) by 2.2–2.5x on a 17,280-atom system. The full molecular dynamics (MD) step saw a speedup of up to 1.7x.
• TF32x3 on Ampere: Sped up major GEMMs by 1.7–1.9x and the full MD step by up to 1.5x on the same large system.
• Accuracy and Stability: The TF32x3 method successfully converged in a structural relaxation test that caused standard one-pass TF32 to fail, demonstrating its superior numerical stability for sensitive scientific calculations.

These performance improvements were achieved by modifying how matrix multiplications are executed at the library level (cuBLAS) or through custom layers (CUTLASS), leaving the core model architecture and its trained weights unchanged. This provides a direct path for researchers to apply these speedups to existing, validated models with minimal implementation overhead.

Impact on Scientific Machine Learning

The findings illustrate a practical solution to the persistent gap between the raw power of low-precision hardware and the strict accuracy requirements of many scientific workloads. For the broader AI ecosystem, this validates a powerful optimization strategy that extends beyond atomistic simulation to other precision-sensitive domains. Enabling researchers to run larger, longer, and more complex simulations accelerates the pace of discovery and allows for the exploration of phenomena previously considered computationally intractable. The techniques effectively unlock more of the hardware's potential, ensuring that advancements in GPU architecture directly translate into tangible benefits for scientific research.

The successful application of BF16x9 and TF32x3 methods demonstrates a critical pathway to unlock the full potential of NVIDIA's Tensor Cores for scientific computing, proving that performance gains can be achieved in precision-sensitive workloads without the costs of model retraining or the risks of numerical instability.