Hough Transform FPGA solution for High Energy Physics online fast tracking

In the coming years, significant upgrades are planned for ATLAS and other High Energy Physics experiments at CERN. Both the technologies and methodologies employed will undergo changes for the scheduled runs at the end of the decade. The LHC accelerator itself will also undergo multiple modifications, allowing it to achieve a peak of instantaneous luminosity up to 5–7.5 × 1034 cm−2 s−1. These enhancements will necessitate the experiments to handle a greater number of events at the conclusion of the data acquisition chain. For instance, ATLAS will be compelled to employ online tracking for its inner detector, aiming to achieve a final event rate of 10 kHz from the 1MHz originating from the Calorimeters and the Muon Spectrometer trigger discrimination. Among the architectures explored to expedite fast tracking, there is consideration of a “hardware accelerator” farm, an infrastructure made of interconnected accelerators such as GPUs and FPGAs, designed to accelerate the tracking processes. The project presented here proposes a tuned Hough Transform algorithm implementation on high-end FPGA technology, specifically designed to adapt to various tracking situations. A development platform comprising software and firmware tools has been created to study different datasets. This platform utilizes software to simulate the firmware and to perform hardware tests. AMD-Xilinx FPGAs were chosen to implement and asses the system, with specific boards such as the VC709, the VCU1525 and the Alveo U250. Strategies such as low-level design for the firmware architecture, leveraging the card’s features like PCI Express data transfer, and the > 1 million gates array available have been exploited. The system underwent testing using internally simulated events generated within the ATLAS environment. Simulated 200 pile up events were used to evaluate the algorithm effectiveness. The average processing time was estimated to be below 5 μs, with the capability to concurrently process two events per algorithm instance. Internal efficiency tests have shown conditions where track finding performance for single muon tracking exceeded 95%.