Parts | Days | Selection | Search | Updates | Downloads | Help
T: Fachverband Teilchenphysik
T 66: DAQ, trigger and electronics III
T 66.8: Talk
Wednesday, March 17, 2021, 17:45–18:00, Tp
Development of an FPGA Implementation of Convolutional Neural Networks for Signal Processing for the Liquid-Argon Calorimeter at ATLAS — Anne-Sophie Berthold, Nick Fritzsche, Rainer Hentges, Arno Straessner, and •Johann Christoph Voigt — TU Dresden, Germany
With the planned Phase 2 upgrade of the ATLAS detector at LHC, the number of proton-proton collisions occuring at the same time will increase significantly. This leads to higher requirements for the data processing, since the rate of detected particles in one detector cell will increase. New machine learning solutions are under development to better reconstruct the energy deposited in the calorimeter and its timing information than the current optimal filter approach.
Here an implementation of convolutional neural networks for FPGA hardware is introduced. The network architecture is flexible and can be configured directly from the model files after network training. It is optimized regarding signal delay and resource usage. Especially the efficient use of the digital signal processors used for multiplications is crucial, since their availability is the limiting factor for network size. Respective performance and resource usage results are presented. The current status of the time division multiplexing, which is necessary to handle the high number of detector readout channels and process multiple input streams per network, is shown.