Madgraph5 GPU development

Tuesday 10 Nov 2026, 15:00 → 16:00 Europe/Zurich

513/R-068 (CERN)

15:00 → 15:10 News

15:10 → 15:30 Topical discussion

Speaker: Dr All

Agenda

Actions

Enrico

• Start GPU energy measurements for one-loop numerators using the Glasgow server.
• Check one-loop stability against quadruple-precision reference results in infrared limits.

Frank

• Continue benchmarks across all backends and floating-point types for the selected process subset.

Daniele

• Results of MadGraph7 runs with comparison with full CPU (95 cores)
• Are we limited by IO, it would be nice to profile:
• Strategies: HDF5? RNTuple?

Zenny

• Create a PR removing the old MadtRex implementation and integrating the new version through a fork of the Rex library.
• Discuss with Theo the placement of the Rex fork in the MadGraph repository.
• Add the appropriate test and release suites for the new Rex/MadtRex integration.
• Update Section 7 of the paper.
• Discuss with Daniele whether helicity reweighting needs to be implemented for MadtRex.

Olivier

• Discuss the cubic-spline folding method with Stefano.
• Discuss with Stefano how to incorporate the MadGraph5 negative-weights paper into the MadGraph7 paper.

Open points

• Finalize the placement and integration structure of the Rex library within the MadGraph repository.
• Decide whether helicity reweighting is required for MadtRex.
• Decide how the cubic-spline folding improvements and previous negative-weights work should be incorporated into MadGraph7 and its paper.
• Finalize the process/precision configurations to use for the benchmark and validation plots.
• Understand the observed hardware dependence in the MadNIS vs VEGAS speedups and its relation to unweighting efficiency.

15:30 → 15:50 Round table

15:50 → 16:00 AoB