A Fast Scalable Implicit Solver for Nonlinear Time-Evolution Earthquake City Problem on Low-Ordered Unstructured Finite Elements with Artificial Intelligence and Transprecision Computing

To address problems that occur due to earthquake in urban areas, we propose a method that utilizes artificial intelligence (AI) and transprecision computing to accelerate a nonlinear dynamic low-order unstructured finite-element solver. The AI is used to improve the convergence of iterative solver leading to 5.56-fold reduction in arithmetic count from a standard solver, and FP16-FP21-FP32-FP64 computing is used to accelerate the sparse matrix-vector product kernel, which demonstrated 71.4% peak FP64 performance on Summit. This is 25.3 times faster than a standard solver and 3.99 times faster than the state-of-the-art SC14 Gordon Bell Finalist solver. Furthermore, the proposed solver demonstrated high scalability (88.8% on the K computer and 89.5% on Piz Daint), leading to 14.7% peak FP64 performance on 4096 nodes of Summit. The proposed approach utilizing AI and FP16 arithmetic has implications for accelerating other implicit solvers used for earthquake city simulations as well as various fields.