- presentation by Anne, update on ATLAS LAr calorimeter signal processing with CNNs
- Ramesh: a lot of progress has been done; plan how to proceed; split FPGA programming and training; Keras tuner for training; HLS4ML limits the size of the network; distribute training by horovod; trainings shall be done continuously while taking/analysing; alveo board with Xilinx tools; stream junks through 100 Gb interface; inference on buffered data and then select; 3 out of 4 alveo boards were blocked but now repaired; 2 slides
- Ramesh: HLS4ML limits combination number of filters and filter size to 4096; this is to meet timing constraints on FPGA; if networks get larger need to be put on several FPGAs
- Dennis: working on alternative architecture - it is promising, based on CNN layers; issue with data - needed to be reprocessed (real cells, MC21 egam8 derivation with raw calorimeter cells, but not in HEC only EMEC; legacy sliding window drops HEC cells); Bachelor + Master student will / is working on it; HLS4ML supported FPGA types? (Versal? Ultrascale?)
- Marcel Trattner (HTW Berlin):
- Expected attendance at PUNCH General Meeting: Ramesh + 3 colleagues incl. Michael;
- Next meeting Nov 6 or 13 (depending availability)
- Another workshop in February? to be discussed at TA5 meeting
There are minutes attached to this event.
Show them.