Accelerating High-Order Finite Element Simulations at Extreme Scale with FP64 Tensor Cores
New technique achieves 2x performance and 83% energy gains on 10,000 GPUs for real-world tsunami forecasting.
A team of computational scientists has achieved a breakthrough in high-performance computing by directly programming NVIDIA's FP64 (double-precision) tensor cores—specialized hardware typically used for AI—to massively accelerate scientific simulations. Their work, detailed in a new arXiv paper, focuses on optimizing the widely-used MFEM (finite element method) library. By integrating FP64 tensor core operations with kernel fusion techniques, they demonstrated up to 2x performance improvements and a remarkable 83% gain in energy efficiency on NVIDIA's latest Grace Hopper GH200 and Grace Blackwell GB200 supercomputing architectures.
This is the first successful demonstration of using FP64 tensor cores for large-scale finite element applications, a cornerstone of scientific computing for fields like automotive design, electromagnetics, and natural disaster modeling. The team proved the scalability of their approach by running optimized kernels across nearly 10,000 GPUs on the Alps supercomputer, achieving near-perfect weak scaling and 90% strong scaling efficiency. The enhancements directly benefit production codes, including the application that won the prestigious 2025 Gordon Bell Prize for real-time tsunami forecasting, moving us closer to practical exascale computing for urgent global challenges.
- Achieved 2x performance gain and 83% energy efficiency improvement by programming NVIDIA FP64 tensor cores for finite element simulations.
- Demonstrated near-perfect scaling efficiency across nearly 10,000 GPUs on the Alps supercomputer, enabling exascale-class computations.
- Directly enhances real-world production codes, including the 2025 Gordon Bell Prize-winning application for real-time tsunami forecasting.
Why It Matters
Enables faster, more energy-efficient exascale simulations for critical scientific challenges like disaster forecasting and advanced engineering.