FPGA-Implemented CNN Enables Real-Time Pileup Tracking

According to AI TIMES, a research team working with GSI has unveiled a technology that tracks the position and arrival time of individual particles in real time with a CNN implemented directly on an FPGA, even in pileup conditions where signals overlap in particle accelerator detectors. Introduced on October 6, the achievement centers on processing about 120 Gbit of data per second with a 300 ns latency and completing inference within the same hardware without sending data to a separate external computing device. Until now, when multiple particles passed through in a short time and pulses overlapped, conventional threshold-based methods have had difficulty separating each signal. The researchers adapted a trained model to the hardware so it could read both the shape and timing characteristics of overlapping pulses. This reduces the problem of counting fewer particles than actually passed and suggests the possibility of reconstructing an entire spill more precisely. The work also points to the potential to handle higher event rates under the same detector conditions, drawing attention to whether improvements in signal processing alone can broaden the usable range of measurement systems.