Heidelberg 2022 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

T: Fachverband Teilchenphysik

T 15: GRID Computing

T 15.6: Vortrag

Montag, 21. März 2022, 17:30–17:45, T-H28

Optimization of performance for HEP ML applications on GPU Clusters — •Tim Voigtländer, René Caspart, Manuel Giffels, Günter Quast, Matthias Schnepf, and Roger Wolf — Karlsruhe Institute of Technology, Karlsruhe, Germany

GPU clusters are gaining increased importance also in particle physics. To use GPUs most efficiently, concepts like multi-processing on a single GPU, multi-GPU usage for suitable applications or the balance between CPU and GPU resources must be considered. In particular, GPU support for applications in Machine Learning has become quite common, and they provide a wide variety of usage scenarios. The GPU performance in relation to CPUs depends on the complexity of the network topology, on the training strategy and other hyperparameters of the problem at hand. To illustrate the possible performance gains, a number of scenarios in neural network training on a shared GPU cluster attached to the TOpAS Tier3 at KIT are discussed.