Exploiting Parallelism in the TileCal Trigger System with GPGPU

Abstract

After the 2022 upgrades, the Tile Calorimeter (TileCal) detector at ATLAS will be generating raw data at a rate of approximately 41 TB/s. The TileCal triggering system contains a degree of parallelism in its processing algorithms and thus presents an opportunity to explore the use of general-purpose computing on graphics processing units (GPGPU). Currently, research into the viability of an sROD ARM-based co-processing unit (PU) is being conducted at Wits University with especial regard to increasing the I/O throughput of the detector. Integration of GPGPU into this PU could enhance its performance by relieving the ARMs of particularly parallel computations. In addition to the PU, use of GPGPU in the front-end trigger is being investigated on the basis of the used algorithms having a similarity to image processing algorithms - where GPU can be used optimally. The use of GPUs in assistance to or in place of FPGAs can be justified by GPUs’ relative ease of programming; C/C++ like languages as opposed to assembly-like Hardware Description Languages (HDLs). This project will consider how GPUs can best be utilised as a subsystem of TileCal in terms of power and computing efficiency; and therefore cost.