: Poster 64: 416-PFLOPS Fast Scalable Implicit Solver on Low-Ordered Unstructured Finite Elements Accelerated by 1.10-ExaFLOPS Kernel with Reformulated AI-Like Algorithm: For Equation-Based Earthquake Modeling
DescriptionWe propose herein an approach for reformulating an equation-based modeling algorithm to an algorithm similar to that of training artificial intelligence (AI) and accelerate this algorithm using high-performance accelerators to reduce the huge computational costs encountered for physics equation-based modeling in earthquake disaster mitigation. A fast scalable equation-based implicit solver on unstructured finite elements is accelerated with a Tensor Core-enabled matrix-vector product kernel. The developed kernel attains 1.10 ExaFLOPS, leading to 416 PFLOPS for the whole solver on full Summit. This corresponds to a 75-fold speedup from a previous state-of-the-art solver running on full Piz Daint. This result could lead to breakthroughs in earthquake disaster mitigation. Our new idea in the HPC algorithm design of combining equation-based modeling with AI is expected to have broad impacts in other earth science and industrial problems.