6–13 Jul 2022
Bologna, Italy
Europe/Rome timezone

MadFlow: automating Monte Carlo simulation on GPU for particle physics

9 Jul 2022, 11:30
15m
Room 12 (Celeste)

Room 12 (Celeste)

Parallel Talk Computing and Data handling Computing and Data handling

Speaker

Dr Juan Manuel Cruz Martinez (Università degli Studi di Milano)

Description

We present MadFlow, a python-based software for the evaluation of cross sections utilizing hardware accelerators.

The pipeline includes a first stage where the analytic expressions for matrix elements are generated by the MG5_aMC@NLO framework (taking advantage of its great flexibility) and exported in a vectorized device-agonstic format using the TensorFlow library or a device specific CUDA output.

The simulation of the event is then performed using the VegasFlow and PDFFlow frameworks for the phase space integration and interpolation of PDFs and then deployed automatically to systems with different hardware acceleration capabilities (multi-threading CPU, single-GPU and multi-GPU setups from both Nvidia and AMD). We show results for Leading Order calculations with up to 5 legs in the final state offering an speed-up of orders of magnitude over traditional CPU-based calculations.

In-person participation Yes

Primary authors

Dr Juan Manuel Cruz Martinez (Università degli Studi di Milano) Marco Rossi (Università degli Studi di Milano) Marco Zaro (Istituto Nazionale di Fisica Nucleare) Stefano Carrazza (Istituto Nazionale di Fisica Nucleare)

Presentation materials