Deciding What to Keep: Real-Time Reconstruction at LHCb

by Lennart Uecker (Universität Heidelberg)

Tuesday 21 Jul 2026, 15:00 → 16:00 Europe/Berlin

CP-03-123 (TU Dortmund)

At the LHC, protons collide up to 40 million times a second, producing far more data than any storage system could absorb. Every experiment must therefore throw away the overwhelming majority of its collisions within seconds, permanently and irreversibly. These decisions are made by the trigger, and the physics reach of an experiment is set by how well it works.

The LHCb experiment is specialised for studying decays of hadrons containing beauty quarks, whose signature is difficult to distinguish from the vast QCD background. LHCb has thus taken a new approach in Run 3: it has removed the hardware trigger entirely. Instead, every collision is read out and fully reconstructed in real time on a farm of GPUs that form the first trigger stage, HLT1. There is a gap in reconstruction quality between HLT1 and the slower second trigger stage, HLT2, which produces an offline-quality reconstruction and final selections.

Using the fit of charged particle tracks as a case study, I will show what it takes to perform real-time analysis at LHCb on GPUs and the compromises required to fulfil the throughput demands. I will present the design and integration of the parameterised Kalman filter, which mostly closed this quality gap to HLT2 for the track fit.

With the end of Run 3, I will review the performance and evolution of HLT1 and explore possible pathways to sustain the LHCb physics programme for the high-luminosity Run 5.