A Heterogeneous FPGA/GPU Architecture for Real-Time Data Analysis and Fast Feedback Systems

We propose a versatile and modular approach for a realtime data acquisition and evaluation system for monitoring and feedback control in beam diagnostic and photon science experiments. Our hybrid architecture is based on an FPGA readout card and GPUs for data processing. To increase throughput, lower latencies and reduce overall system strain, the FPGA is able to write data directly into the GPU’s memory. After real-time data analysis the GPU writes back results back to the FPGA for feedback systems or to the CPU host system for subsequent processing. The communication and scheduling processing units are handled transparently by our processing framework which users can customize and extend. Although the system is designed for real-time capability purposes, the modular approach also allows standalone usage for high-speed off-line analysis. We evaluated the performance of our solution measuring both processing times of data analysis algorithms used with beam instrumentation detectors as well as transfer times between FPGA and GPU. The latter suggests system throughputs of up to 6 GB/s with latencies down to the microsecond range, thus making it suitable for fast feedback systems.