Authors: Tsuyoshi Ichimura and Kohei Fujita (University of Tokyo, RIKEN); Kentaro Koyama (Fujitsu Ltd); Ryota Kusakabe (University of Tokyo); Kazuo Minami (RIKEN); Hikaru Inoue (Fujitsu Ltd); Seiya Nishizawa and Miwako Tsuji (RIKEN); Tatsuo Nishiki (Fujitsu Ltd); Muneo Hori (Japan Agency for Marine-Earth Science and Technology); Lalith Maddegedara (University of Tokyo, RIKEN); and Naonori Ueda (RIKEN)
Abstract: We propose an HPC-based scalable implicit low-order unstructured nonlinear finite-element solver that uses data generated during physics-based simulations for data-driven learning. Here, a cost efficient preconditioner is developed using the data-driven learning method for accelerating the iterative solver. Together with Arm scalable vector extension-aware SIMD and multi-core tuning of core sparse matrix-vector multiplication kernel on Fugaku, the developed solver achieved a 15.2-fold speedup over the conventional preconditioned conjugate gradient solver with 96.4% size-up scalability up to 1,179,648 cores of Fugaku (11.8% of FP64 peak with 8.87 PFLOPS). Furthermore, the developed solver achieved a 10.3-fold speedup over the state-of-the-art SC14 Gordon Bell Prize finalist solver on a high resolution urban model with over 11 billion degrees of freedom. Such development in merging HPC-enhanced physics-based simulations with data-driven learning is expected to enhance physics-based simulation capability and is expected to contribute to various applications such as digital twins of cities.
Best Poster Finalist (BP): no
Poster summary: PDF
Back to Poster Archive Listing