GPU performance in Run3 ALICE online/offline reconstruction

After several software and hardware upgrades during LS2, ALICE records 50 KHz of minimum bias Pb--Pb collisions in continuous readout mode. To cope with the high data rate of 3.5 TB/s from the detectors, multiple stages of compression are employed during data taking, the last one requiring full TPC tracking. This compression chain is part of the synchronous (online) processing and largely relies on GPU computing to reduce the data rate to 170 GB/s. After the data taking, a second phase called asynchronous (offline) processing is performed, where the compressed data are read and a full global reconstruction is performed, which results in the production of data that can be analysed. When the online computing farm is not fully utilized for online processing, it is capable of running the offline reconstruction leveraging GPUs, which proved to speed up also this phase. ALICE is making significant efforts to offload more of the asynchronous reconstruction to GPUs, to speed up the overall execution time and efficiently use as much GPU computing resources as possible. This talk will focus on the performance that GPUs have provided in Run 3 for the ALICE experiment, both for online and offline reconstruction. Moreover, it will also highlight the offloading process to GPU of the track-model decoding, an asynchronous reconstruction task which is one of the software routines that have been targeted to be offloaded to GPUs during offline reconstruction.