Speaker
Description
The ATLAS detector is concluding the Run 3 data-taking period with an average number of proton–proton collisions per bunch crossing around 65. This number is expected to reach about 200 during the High-Luminosity LHC phase, imposing increased demands on the data acquisition and trigger systems of the experiment. The ATLAS trigger software will be required to process event rates of up to 1 MHz. A large upgrade programme is investigating the possible usage of new technologies, such as Machine Learning techniques and hardware accelerators. This work focuses on the Calorimeter Topological Clustering algorithm, which is presently used both offline for event reconstruction and online in the trigger within a CPU-based environment. This is the dominant consumer of computing resources in the calorimeter trigger, hence, was selected for practical viability and performance improvement studies. As a baseline, the current CPU-based solution is discussed using the existing software under Run 4 and HL-LHC conditions. A ready-to-use GPU-based solution is also presented with its physics and processing performance results. An FPGA-based solution is also described, with emphasis on its memory requirements that limit the algorithm performance outcome. Finally, a completely new convolutional neural network approach to jet reconstruction is presented.
| I read the instructions above | Yes |
|---|