Chinese Algorithm Can Speed ​​Up Nvidia GPUs by 800 Times

The new algorithm from scientists is expected to help improve the efficiency of Peridynamics (PD) - a non-local method used to model cracks and damage in materials. This new method can increase simulation performance by up to 800 times, thereby significantly improving simulation speed for large-scale materials.

New algorithm allows to increase the capabilities of popular GPUs from Nvidia by up to 800 times

Peridynamics is now widely used in fields such as aerospace, civil engineering, and the military to predict material failure. However, traditional Peridynamics simulations often require a lot of computational resources, making large-scale research slow and impractical.

Breakthrough algorithmic solutions

Now, Associate Professor Yang Yang and her team have solved the problem by applying Nvidia's CUDA technology to optimize algorithm design and memory management. The PD-General framework they developed achieves up to 800 times speedup on an Nvidia RTX 4070 graphics card (GPU) compared to traditional sequential methods, and is 100 times faster than OpenMP-based parallel programs. In large-scale simulations with millions of particles, the algorithm can complete 4,000 iterations in just 5 minutes. In particular, in large-scale problems, the processing speed is not only faster but also more accurate than the conventional model.

This improvement in computational efficiency allows researchers to run simulations on consumer-grade GPUs, rather than relying on expensive, high-performance computer clusters. This has important implications for industries that require detailed materials analysis, including aerospace, engineering and manufacturing, and military research.

The ability to run high-performance simulations on common GPUs also reduces dependence on outside technology, especially in the current climate of sanctions and trade restrictions. This breakthrough allows China and Russia to advance research without relying on high-end computing hardware from Western countries.

The research was published in the Chinese Journal of Computational Mechanics on January 8, 2025, and the team believes that this optimization process can be extended beyond dynamics, thereby improving GPU performance for other scientific calculations.

Source: https://thanhnien.vn/thuat-toan-trung-quoc-co-the-tang-toc-gpu-nvidia-len-800-lan-185250205231104787.htm