Speaker
Description
In modern collider experiments, rare and interesting physics processes must be identified in real time while coping with enormous data rates from a large number of detector sensors. Reconstruction algorithms and trigger decisions already need to be made within microseconds, reducing the data volume by orders of magnitude with very low latency. At next generation particle accelerators such as the High Luminosity-LHC, the substantially higher luminosity will deliver orders of magnitude more data than previous runs, together with a dramatic increase in event rates and pile up interactions. This will make the requirements on these so called online algorithms drastically more stringent.
To meet these challenges, next-generation detectors will rely on more advanced readout electronics and heterogeneous computing architectures that integrate custom ASICs, FPGA-based processing units, graphics processing units and traditional central processing units. These systems will perform local data reduction and feature extraction directly at the detector frontend (FE) to reduce data transfer and latency. Currently employed conventional compression and selection algorithms, such as thresholding, clustering, or selection of region of interest, may be complemented if not replaced by machine learning (ML)-based algorithms capable of exploiting spatial and temporal correlations in the detector signals.
A promising path forward is the development of distributed Deep Neural Network architectures that span multiple levels of the data processing chain, from the frontend electronics to the backend (BE) processors. Such systems could enable more advanced low-latency object reconstruction and intelligent event filtering directly within the trigger pipeline by exploiting the larger, non-fragmented logic resources made available through a reduction in the number of intermediary abstracted objects, while maintaining compatibility with full offline reconstruction.
To this end, we introduce DRIFTS, a framework for Distributed Reconstruction in FPGAs and ASICs for Triggering Systems. Using publicly available simulations of the High Granularity Calorimeter of the Phase II CMS (Compact Muon Solenoid) detector as a case study, we present the implementation of a distributed trigger chain for HGCAL. Taking advantage of the FE ML-based encoder available within the HGCAL FE ASIC, we use the encoded information as an input for a Graph Neural Network like algorithm implemented in the BE FPGAs in a fully distributed ML chain in order to produce trigger primitives. In addition, we show how the reconstruction of trigger data can be achieved offline at every step of the chain, preserving the chain’s low latency and saving resources while enabling debugging and interpretability. Finally, we discuss the impact of quantization and of implementing the algorithms as synthesizable logic using HLS4ML.