{"id":"arxiv/2611.14041","url":"https://arxiv.org/abs/2611.14041","tags":["math.NA","physics.comp-ph"],"score":0,"title":"Accelerating matrix-free fast diagonalization and preconditioned linear solvers for higher-order finite element discretized elliptic problems on distributed GPUs","domain":"math.NA","source":"arxiv","authors":["Gourab Panigrahi","Phani Motamarri"],"reviews":0,"subject":"math.NA","summary":"The time to solution of an iterative solver depends on the number of iterations and the cost per iteration. This paper develops algorithms that reduce both for elliptic problems discretized with high-order continuous finite elements on distributed GPUs, implemented in a unified matrix-free code supporting CUDA, HIP and SYCL. A distributed direct fast-diagonalization (FD) Poisson solver provides exact solutions on tensor-product meshes with direction-wise grading and coefficients varying along one coordinate direction. Its tensor data stay distributed, moved between slab layouts by collective…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-11-10","reactions":{}}