SMuK 2023 – wissenschaftliches Programm

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

T: Fachverband Teilchenphysik

T 33: DAQ NN/ML – GRID I

T 33.4: Vortrag

Dienstag, 21. März 2023, 17:45–18:00, HSZ/0301

Profiling of GPU-based neural network trainings — •Tim Voigtländer, Manuel Giffels, Artur Gottmann, Günter Quast, Matthias Schnepf, and Roger Wolf — Karlsruhe Institute of Technology, Karlsruhe, Germany

The training of neural networks has become a significant workload of particle physics analyses. To speed up these trainings and reduce their turnaround cycle, one or more accelerators, e.g. GPUs, are typically utilized. While the increase in computational capacity is greatly beneficial, the heterogeneous hardware also adds layers of complexity to an already opaque process. In order to improve the efficiency in the usage of the available hardware, suitable profiling to identify possible bottlenecks. In this talk, solutions to a number of commonly occurring challenges found in single- and multi-GPU neural network trainings are presented, using the DeepTau neural network training as a case-study.