The Compressed Baryonic Matter (CBM) experiment at FAIR will operate at interaction rates up to 10 MHz, generating data streams averaging 500 GB/s. This necessitates efficient online reconstruction capabilities, particularly for the Silicon Tracking System (STS), which is the key detector for track reconstruction and contributes a large fraction of the expected data volume. We present a GPU-accelerated hit reconstruction chain for the STS that achieves a 128 speedup over the sequential CPU implementation. The implementation features optimized data structures reducing memory footprint, parallel algorithms for sorting, cluster finding, and hit reconstruction, and portability across GPU architectures. Our custom merge sort outperforms library implementations by 10 % while using 33 % less memory. Cluster finding employs a twophase approach with atomic operations for thread-safe connections between signal clusters. Even before GPU acceleration, algorithmic improvements provide a 3 speedup in single-threaded execution. Both NVIDIA and AMD GPUs achieve comparable performance of approximately 0.12 s on a timeframe containing 1000 Au+Au events. The reconstruction chain was successfully deployed during the May 2024 mCBM beamtime, processing data rates up to 2.4 GB/s in real-time, demonstrating its viability for CBM’s triggerless data acquisition approach.
Bluhme et al. (Wed,) studied this question.