CompassUT : study of a GPU track reconstruction for LHCb upgrades

We present a fast, data-oriented GPU tracking algorithm, CompassUT, as a potential option to cope with the expected throughput of 40Tbit/s for LHCb upgrade. We present a parallel version of the raw input decoding, optimized for SIMD architectures. We sort the hits by X and Y into group sectors while decoding, to have a fast sorting and searching of the hits. We implement the tracking by reducing the memory footprint, reducing branching to a minimum and making the algorithm data-oriented for SIMD architectures. We show the achieved throughput in a variety of consumer and server GPUs, and present the impact on both the computing and physics performance for different configurations of the algorithm.