Unlocking Tensor Core Performance for Machine Learning Interatomic Potentials
Optimizing MLIP inference using Tensor Core FP32 methods for faster, stable atomistic simulations.
1Bridging Simulation Accuracy Gaps
Atomistic simulations require a delicate balance between quantum-mechanical accuracy and computational speed. Machine learning interatomic potentials, or MLIPs, offer a solution by predicting energies and forces at reduced costs. However, these models remain sensitive to numerical precision errors, which can accumulate during long molecular dynamics trajectories and destabilize physical simulations.
2Leveraging Tensor Core Throughput
Modern NVIDIA GPUs provide significantly higher throughput for lower-precision operations like BF16 and TF32 compared to native FP32. To utilize this power without sacrificing precision, researchers use Tensor Core FP32 methods. These techniques decompose FP32 operands into lower-precision components, perform calculations on Tensor Cores, and recombine them back into FP32.
3BF16x9 on Blackwell GPUs
On Blackwell GPUs, the BF16x9 algorithm splits FP32 operands into three scaled BF16 components. This approach produces nine pairwise products accumulated in FP32. By using the cuBLAS environment variable, developers can enable this path for PET models without changing the underlying architecture, achieving 2.2–2.5x acceleration for large 17,280-atom systems.
4TF32x3 Implementation on Ampere
The TF32x3 method on Ampere GPUs represents FP32 operands using leading and residual TF32 components. By omitting the smallest residual-by-residual product, the system maintains high accuracy while accelerating matrix multiplications. This implementation via NVIDIA CUTLASS allows PET models to treat these operations as standard linear layers, ensuring seamless integration into existing workflows.
5Performance Gains in PET
Testing the Point Edge Transformer (PET) model revealed substantial performance headroom. On Blackwell GPUs, BF16x9 accelerated large GEMMs by up to 2.5x. On Ampere GPUs, TF32x3 delivered a 1.7–1.9x speedup for major linear layers. These gains were achieved without requiring model retraining or custom kernel development, proving the efficiency of the approach.
6Ensuring Numerical Fidelity
For molecular dynamics, matching single-point energy is insufficient; the potential-energy surface must remain smooth. TF32x3 maintains this smoothness, whereas one-pass TF32 introduces visible point-to-point variations. In NVE simulations of a 192-atom water box, TF32x3 demonstrated bounded energy fluctuations comparable to FP32, effectively preventing the systematic energy drift observed in lower-precision alternatives.
7Structural Relaxation Success
Structural relaxation tests on a 539-atom Cantor-alloy slab confirmed the reliability of TF32x3. The method converged in 218–222 steps, matching the FP32 baseline performance. Conversely, one-pass TF32 failed to reach the required force tolerance even after 1,000 steps, highlighting that TF32x3 preserves the necessary convergence behavior for complex atomic geometry optimizations.
8System Size and Scaling
End-to-end performance gains are highly dependent on system size. Smaller systems benefit less due to overhead from graph operations and neighbor processing. Larger systems, where GEMMs dominate the runtime, see the most significant improvements. Developers should tune kernels based on specific matrix shapes and hardware configurations to maximize these performance benefits.
9Integration Without Complexity
The primary advantage of these methods is the lack of model-code changes. By utilizing cuBLAS or CUTLASS operators, researchers can swap standard linear layers for optimized versions. This allows existing MLIPs to benefit from modern hardware acceleration while maintaining the rigorous precision standards required for scientific research and physical trajectory stability.
10Comparing Tensor Core Strategies
While one-pass TF32 offers higher per-step speedups, it sacrifices trajectory stability and convergence accuracy. TF32x3 provides a more balanced approach, delivering 1.5x per-step acceleration while maintaining FP32-level fidelity. This makes TF32x3 the preferred choice for long-duration molecular dynamics where physical consistency and energy conservation are critical for valid scientific results.
11Practical Deployment Guidelines
To implement these optimizations, researchers should consult the cuBLAS documentation for Blackwell-specific BF16x9 settings. For Ampere architectures, the NVIDIA ALCHEMI Toolkit provides integration examples for TF32x3. Users should verify their specific workload requirements, as the performance and numerical stability depend on the chosen Tensor Core strategy and the model's sensitivity to precision.
12Advancing Scientific Machine Learning
These Tensor Core methods represent a significant step in scientific machine learning, allowing researchers to push the boundaries of atomistic simulations. By unlocking hardware-level acceleration for precision-sensitive workloads, the community can study larger systems and longer timescales. This approach ensures that performance gains do not come at the cost of physical accuracy.