2023 CERN openlab Technical Workshop

Performance profiling infrastructure comparing CUDA and OneAPI/SYCL in the MadGraph4GPU project

Not scheduled

20m

503/1-001 - Council Chamber (CERN)

Speaker

Jorgen Teig (Norwegian University of Science and and Technology (NTNU) (NO))

Poster abstract

Detailed view of the performance profiling infrastructure profiling different GPUs and CPUs with different abstraction layers in use in the MadGraph4GPU project, with an emphasis on looking at how we build and look at the performance between CUDA/C++ and OneAPI/SYCL. We will also go in detail what we did to fully automate this infrastructure to provide the team with up to date plots and measurements of the current performance of the project on different hardware and a holistic view of the performance of the project over time as the team develops it.