SMuK 2023 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

T: Fachverband Teilchenphysik

T 33: DAQ NN/ML – GRID I

T 33.2: Vortrag

Dienstag, 21. März 2023, 17:15–17:30, HSZ/0301

Convolutional Neural Networks on FPGAs for Processing of ATLAS Liquid Argon Calorimeter Signals — •Johann Christoph Voigt, Anne-Sophie Berthold, Nick Fritzsche, Rainer Hentges, Christian Gutsche, and Arno Straessner — Institut für Kern- und Teilchenphysik, TU Dresden, Germany

The Phase-II upgrade of the ATLAS Liquid Argon Calorimeter allows for the energy reconstruction of all ~180000 readout channels at the LHC bunch crossing frequency of 40 MHz. Further challenges arise from the increased pile-up due to the planned higher number of simultaneous proton-proton collisions.

For the digital energy reconstruction, we propose the use of Convolutional Neural Networks (CNNs) instead of the previous Optimal Filter. The networks need be able to run on an FPGA with limited resources and are therefore limited in complexity to approximately 100 weight parameters.

This talk focuses on the firmware implementation of these networks in VHDL. The implementation is optimized for DSP usage and latency. To be able to process all readout channels on the available FPGAs, time domain multiplexing is used to process multiple channels per CNN instance. This reduces the number of required instances and increases the frequency the design needs to run at. A multiplexing factor of 12 at a frequency of 480 Mhz is demonstrated for a design processing 384 detector cells. The latest FPGA resource usage estimates are presented.