Usage of GPU for online data processing – the experience of ALICE and LHCb

<!--HTML--><p>In LHC Run 3, the ALICE and LHCb experiments will not rely on hardware triggers. The luminosity will be increased significantly, so that up to 40 Tbit/s will be processed in the online farm.</p> <p>LHCb runs a full-software trigger and ALICE performs calibration and data compression, for which the full software processing enables unprecedented flexibility and possibilities.<br /> Both experiments foresee a two-stage approach, processing the data during data taking for a first data reduction and storing it to an on-site disk-buffer for a second processing when there is no beam.</p> <p>In order to cope with the higher data rates, both experiments are exploring alternative hardware technologies instead of traditional CPU compute farms.</p> <p>We discuss the framework design, implementation details, performance, and results of LHCb's full-software trigger on GPUs "Allen" and of ALICE's online and offline data compression and reconstruction on GPUs.<br /> &nbsp;</p>