SMuK 2023 – wissenschaftliches Programm

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

T: Fachverband Teilchenphysik

T 33: DAQ NN/ML – GRID I

T 33.3: Vortrag

Dienstag, 21. März 2023, 17:30–17:45, HSZ/0301

Implementation of neural networks for live reconstruction using AI processors — Klaus Desch¹, Jochen Kaminski¹, Michael Lupberger^1,2, and •Patrick Schwäbig¹ — ¹Physikalisches Institut, Universität Bonn, Deutschland — ²Helmholtz-Institut für Strahlen- und Kernphysik, Universität Bonn, Deutschland

For years, data rates generated by modern detectors and the corresponding readout electronics exceeded by far the limits of data storage space and bandwidth available in many experiments. The solution of using fast triggers to discard uninteresting and irrelevant data is a solution used to this day. Using FPGAs, ASICs or directly the readout chip, a fixed set of rules based on low level parameters is applied as a pre-selection. Only a few years ago, live track reconstruction for triggering was rarely possible but with the emergence of fast and highly parallelized processors for AI inference attempts to sufficiently accelerate tracking algorithms become viable. The Xilinx Versal AI Adaptive Compute Acceleration Platform (ACAP) is one such technology and combines FPGA and CPU resources with dedicated AI cores. Our approach is to utilize the unique combination of FPGA and AI cores to leverage neural networks for live triggering which will be relevant for future experiments and upgrades of already existing setups.

In this talk AI algorithms for track reconstruction, especially their quantized and non-quantized implementation on the Xilinx VC1902, will be shown. They will be used in an envisioned mid-size ultra-high rate fixed-target dark matter experiment (Lohengrin) at the ELSA accelerator at the University of Bonn.