Efficient utilization of massively parallel computing resources is crucial for advancing scientific understanding through complex simulations. However, existing adaptive methods often face challenges in implementation complexity and scalability on modern parallel hardware. Here we present dynamic block activation (DBA), an acceleration framework that can be applied to a broad range of continuum simulations by strategically allocating resources on the basis of the dynamic features of the physical model. By exploiting the hierarchical structure of parallel hardware and dynamically activating and deactivating computation blocks, DBA optimizes performance while maintaining accuracy. We demonstrate DBA's effectiveness through solving representative models spanning multiple scientific fields, including materials science, biophysics and fluid dynamics, achieving 216-816 central processing unit core-equivalent speedups on a single graphics processing unit (GPU), up to fivefold acceleration compared with highly optimized GPU code and nearly perfect scalability up to 32 GPUs. By addressing common challenges, such as divergent memory access, and reducing programming burden, DBA offers a promising approach to fully leverage massively parallel systems across multiple scientific computing domains.
© 2025. The Author(s), under exclusive licence to Springer Nature America, Inc.