Performance of Linear Solvers in Tensor-Train Format on Current Multicore Architectures

Attend a 34-minute NHR PerfLab Seminar featuring Melven Röhrig-Zöllner from the German Aerospace Center (DLR). Explore the node-level performance of numerical algorithms for high-dimensional problems in compressed tensor format. Focus on two key areas: approximating large dense data through lossy compression and solving linear systems using the tensor-train/matrix-product states format. Learn about optimizations for underlying linear algebra operations, including improvements for orthogonalization and truncation steps based on a high-performance "Q-less" tall-skinny QR decomposition. Discover memory layout optimizations for faster tensor contractions and a simple generic preconditioner. Examine performance results on modern multi-core CPUs, showcasing significant speedups over reference implementations. Gain insights from Röhrig-Zöllner's background in Computational Engineering Science and his work at DLR's HPC department, focusing on numerical methods performance and scientific software development for HPC systems.