Madgraph5 GPU development

Tuesday 20 Oct 2026, 15:00 → 16:00 Europe/Zurich

513/R-068 (CERN)

Meeting minutes 2026-10-20

Quick recap

The meeting focused on progress updates for the MadGraph7 paper and related development work.
Enrico reported that the one-loop results manuscript was being prepared, with energy measurements and stability checks against quadruple precision remaining as the final missing pieces.
The group discussed the integration of the Rex library into the MadGraph repository, agreeing that a separate directory at the same level as MadSpace and AmpliCol was the best approach.
Olivier presented new results on improving folding using cubic splines and an improved virttricks method, both of which showed a reduction in negative weights.
Daniele shared benchmarking results comparing MadNIS and Vegas, noting that speedup was closely related to unweighting efficiency but showed some hardware dependency.

Next steps

Enrico

• Start energy measurements for one-loop numerators on GPU using a server in Glasgow.
• Explicitly check stability of one-loop results compared to quadruple precision reference results in infrared limits.

Frank

• Continue running benchmarks for all backends and FP types on a subset of processes.

Daniele

• Prepare a draft of the paper structure and example plots, including the speed comparison between MadSpace and MadGraph 5.
• Run end-to-end benchmarks for MadGraph5 and MadGraph7 using GridPack on 8 cores, including event generation and LHEF writing.

Zenny

• Create a pull request to clean up MadGraph7 by removing the old version of MadtRex and integrating the new version via a fork of the Rex library.
• Discuss with Theo where exactly the Rex library fork should sit in the MadGraph repository (proposed: same level as MadSpace and AmpliCol).
• Update Section 7 of the paper.
• Discuss with Daniele the implementation of helicity reweighting for MadtRex, if needed.

Olivier

• Discuss the new cubic spline folding method with Stefano on Thursday.
• Discuss with Stefano how to incorporate the negative weights paper (based on MadGraph5) into the MadGraph7 paper.

Summary

One-Loop Results Manuscript Preparation

Enrico reported on the preparation of the one-loop results manuscript, noting two missing components: energy measurements (being set up with help from a contact in Glasgow) and stability checks against quadruple precision results in infrared limits.
His colleagues plan to participate in an upcoming workshop at Argonne National Laboratory and Anthropic to finalize remaining work using Claude tokens.

Rex Library Development Transition Plans

The team discussed plans for the Rex library, with Zenny suggesting creating a fork of it within MadGraph7 while keeping it generic outside of MadGraph.
Olivier proposed following a consistent approach used for other projects like AmpliCol and MadSpace, where everything is kept in one repository for consistency but can be released independently.

MadGraph7 and MadtRex Implementation Updates

The team discussed updates to MadGraph7 and MadtRex implementation, with Zenny confirming they no longer use MadMatrix directly and can create a clean implementation.
They agreed to place REX as a separate directory at the same level as MadSpace and AmpliCol, with Zenny responsible for creating a test suite and release suite.
The discussion also touched on polarization support in MadMatrix, with Zenny offering to implement helicity reweighting if needed.

Folding Method Improvements Discussion

Olivier presented improvements to the folding method using cubic splines, which reduces the number of negative weights by approximately 80% with a small CPU cost.
The team discussed incorporating these improvements into MadGraph7, though final decisions depend on discussions with Stefano scheduled for Thursday.
The group also addressed progress on the virttricks method improvements and a paper about negative weights based on MadGraph5.

Benchmark and Validation Progress Update

Frank reported that benchmark runs are still in progress due to CPU constraints and will take a few more days to complete.
The team discussed which processes to include in the validation, proposing to use expensive processes like TTbar, DY, VBF, and either Unitary or FD Gauge.
Oliver suggested including plots for medical precisions in the FD Gauge analysis, and both agreed to potentially have a follow-up meeting to discuss this further.

NLO acceleration

Daniele also mentioned ongoing work on NLO, including comparing MadGraph5 and MadGraph5 within MadGraph7, and the process of merging pull requests.

Hardware Performance Comparison Discussion

The team discussed performance comparisons between different hardware configurations, focusing on generation time speedups and unweighting efficiency, for the work on MadNIS vs VEGAS comparison.
Daniele presented plots showing speedups relative to VEGAS, with Olivier suggesting they should compare unweighting efficiency rather than raw speedups to understand the possible hardware dependencies.
The discussion revealed that hardware differences between A100 and H100 systems were affecting the results.

Regarding MadGraph7, Daniele is planning to conduct more comprehensive comparisons using gridpack runs (with gridpacks generated using MadNIS with 200k event batches) on H100 hardware.

15:00 → 15:10 News

15:10 → 15:30 Topical discussion

Speaker: Dr All

15:30 → 15:50 Round table

15:50 → 16:00 AoB