Fast online triggering in high-energy physics experiments using GPUs

Conclusions

We investigated some issues related to the implementation of a real-time GPU-based first-level trigger system. The processing times and latencies of modern GPU devices were found to be compatible with the requirements of a high-rate first-level trigger system such as the one of the NA62 experiment. For the RICH single ring fitting problem studied, processing times around 50 ns per event were achieved, with satisfactory resolutions and without the need of processing more than few hundred events in a batch, thus resulting in maximum overall latencies (data collection, transfer and processing times) which can be around 200 us. These are well below the requirements of a modern middle-sized high-rate HEP experiment such as NA62.

More complex algorithms, required to deal with multiple rings and different trigger problems, can be more time consuming, but the approach looks quite promising. The next step is the implementation of a complete test system, involving multiple hardware GbE data sources and links, on which continuous long-term performance can be studied.

While we cannot expect the complete DAQ chain of HEP experiments to evolve into a fully commercial implementation (at least because of radiation issues in the earliest stages), a trend is firmly established towards standardized systems involving HEP-specific radiation-hard digitizing electronics and high-speed data links carrying the data to low-radiation areas where only commodity devices are used to perform all the required processing in real time. We believe the GPU-based approach described here to be a promising step for moving in such direction.