TA3 WP3 Monthly meeting
https://uni-hamburg.zoom.us/j/61672663463?pwd=ZU1TWFl4OTluczlPOE9OcW82R0dFdz09
Invited talk by Johann C. Voigt (TU Dresden):
Real-time signal reconstruction using CNNs on FPGAs for the ATLAS LAr calorimeter readout
The ATLAS LAr calorimeter is responsible for measuring the energy of photons and electrons resulting from proton-proton collisions at the LHC. The main challenge lies in the signal overlaps, as the duration of the detector response exceeds the time between LHC collisions. The upgraded readout will feature over 500 Intel Agilex-7 FPGAs, where each FPGA is responsible for reconstructing the energy deposited in 384 detector cells at a frequency of 40 MHz. To mitigate the effects of the even higher luminosity at the HL-LHC, 1-dimensional convolutional neural networks (CNNs) are under investigation as a potential replacement for the current optimal filter (OF) for the task of real-time energy reconstruction on FPGA. The deployment on FPGAs constrains the network size and motivates the use of quantization. The latency needs to be below ~150 ns to meet requirements of the ATLAS trigger. Networks are trained using simulated detector sequences with the true deposited energy as the training target. A configurable, low level implementation of the inference code of these CNNs has been developed in the VHDL hardware description language in order to integrate the networks into the existing FPGA firmware project.