Exploiting Mathematical Structures for Scalable Algorithms in Scientific Computing and Deep Learning | Dr. Carolin Penke
About
I am a mathematician and high-performance computing researcher working at the Accelerating Devices Lab at the Jülich Supercomputing Centre at Forschungszentrum Jülich. My research interests lie at the intersection of numerical linear algebra, scientific computing, and artificial intelligence. My recent work includes contributions to the HPC.nrw competence network and algorithm development for novel computing devices.
Research Topics
- Numerical algorithms for emerging computing architectures
- Scalable exploitation of low-rank structures in deep learning
- Structure-preserving and scalable algorithms for scientific computing
Scalable Numerical Linear Algebra for Machine Learning and Scientific Computing
My research lies at the intersection of numerical linear algebra, machine learning and high performance computing. It develops numerical algorithms that treat mathematical structure, application requirements and computer architecture together, reducing computational, memory and communication costs while keeping accuracy and numerical robustness.
I exploit low-rank structure in deep learning utilizing methods from randomized numerical linear algebra. Furthermore, I’m develop numerical algorithms for emerging computing architectures, such as dataflow accelerators. Here, classic methods are reformulated and extended to minimize data movement and energy consumption. Another focus lies on structure-preserving algorithms in high performance environment, including scalable eigensolvers for skew-symmetric and Bethe–Salpeter eigenvalue problems in electronic-structure theory.
My overarching goal is mathematically well-founded, scalable methods and scientific software for demanding problems in scientific computing and AI.
More Information
ATML Accelerating Devices
The ATML Accelerating Devices invatigates GPUs and other accelerators. We support researchers using the devices and trying to get the most of the hardware; we consult, analyze and optimize. In close collaboration with vendors, we investigate performance opportunities and mitigate limiters. We bring our knowledge into a number of projects, working with researchers of various domains on different levels.
HPC.nrw
The North Rhine-Westphalian Competence Network for High Performance Computing (HPC.nrw) is a competent and thematically broad contact point for users of High Performance Computing (HPC) systems in NRW.
JuWinHPC: Jülich Women in HPC
JuWinHPC (Jülich Women in HPC) is the local chapter of the international Women in High Performance Computing (WHPC) initiative. It aims to strengthen the community of women working in HPC and to promote diversity.
The ATML Accelerating Devices invatigates GPUs and other accelerators. We support researchers using the devices and trying to get the most of the hardware; we consult, analyze and optimize. In close collaboration with vendors, we investigate performance opportunities and mitigate limiters. We bring our knowledge into a number of projects, working with researchers of various domains on different levels.
Selected Publications:
C. Penke and A. Herten. "A GPU-Accelerated Blocked Adaptive Randomized Range Finder Based on an Implicit Householder QR Decomposition". Accepted at ScalAH26: 17th Workshop on Latest Advances in Scalable Algorithms for Large-Scale Heterogeneous Systems. DOI: 10.48550/arXiv.2608.28941.
P. Benner, Y. Nakatsukasa, and C. Penke. "A Structure-Preserving Divide-and-Conquer Method for Pseudosymmetric Matrices". In: SIAM Journal on Matrix Analysis and Applications 44.3 (2023), pp. 1245–1270. DOI: 10.1137/22M1484985.
C. Penke, A. Marek, C. Vorwerk, C. Draxl, and P. Benner. "High performance solution of skew-symmetric eigenvalue problems with applications in solving the Bethe-Salpeter eigenvalue problem". In: Parallel Computing 96 (2020), p. 102639. DOI: 10.1016/j.parco.2020.102639.
All Publications in Google Scholar