NHR PerfLab Seminar Talk: A Modular Precision Ecosystem Based on a Memory Accessor

Symbolic picture for the article. The link opens the image in a large view.

The NHR PerfLab is pleased to host an online seminar talk by Hartwig Anzt (Karlsruhe Institute of Technology):

Title: Pushing the Roofline: A Modular Precision Ecosystem Based on a Memory Accessor


The roofline model not only provides a powerful tool for predicting the performance an application will exhibit and quantifying its efficiency within the hardware-specific limits, but it also allows for a descriptive presentation of how the increasing machine balance is a threatening trend to applications with constant arithmetic intensity: From the roofline model, we can easily deduce that the trend of the compute performance growing faster than the memory bandwidth will push more and more algorithms into the communication bound regime, and it will become increasingly difficult to translate the increased computational power of new supercomputers into application speedup. In this work, we present the strategy of breaking up the tight coupling between the precision format used for arithmetic operations and the precision format used for memory operations. On a high level, this idea is equivalent to compressing the data in registers before invoking memory operations. While the practical use requires algorithms to acknowledge the numerical effects coming with lossy compression, we show that a “memory accessor” able to hide the data compression behind the memory access can virtually push the bandwidth-induced roofline and enable applications capable of handling the numerical effects to achieve higher compute performance. We demonstrate the performance advantages by implementing a memory accessor based on modern C++ and generating experimental rooflines for AMD and NVIDIA GPUs. We also present that the memory accessor can bring practical benefits by  demonstrating how it reduces the execution time of preconditioned Krylov solvers.

Date and time: Tuesday, May 18, 2021,  2 p.m. – 3 p.m.

Slides: MixedPrecisionTalk.pdf

Short bio:

Dr. Hartwig Anzt is a Helmholtz-Young-Investigator Group leader at the Steinbuch Centre for Computing at the Karlsruhe Institute of Technology (KIT). He obtained his PhD in Mathematics at the Karlsruhe Institute of Technology. Afterwards, he joined Jack Dongarra’s Innovative Computing Lab at the University of Tennessee in 2013 until he started his own research group in 2017. He still contributed to the Innovative Computing Lab as a Research Consultant. Hartwig Anzt has a strong background in numerical mathematics, specializes in iterative methods and preconditioning techniques for the next generation hardware architectures. His Helmholtz group on Fixed-point methods for numerics at Exascale (FiNE) is granted funding until 2022. Hartwig Anzt has a long track record of high-quality software development. He is author of the MAGMA-sparse open source software package and managing lead of the Ginkgo numerical linear algebra library. Hartwig Anzt is a co-PI of the PEEKS project and the xSDK project inside the software technology effort of the US Exascale Computing Project (ECP). He is also the technical PI of the multiprecision effort in the xSDK project, a coordinated effort aiming at integrating low-precision functionality into high-accuracy simulation codes.