Evaluating latency and throughput bound acceleration of FPGAs and GPUs for adaptive optics algorithms

General purpose extremely large aperture optical/infrared telescopes are instrumental in vastly advancing the astrophysical knowledge on a variety of subjects such as star formation, super-massive black holes, solar magnetic atmosphere, exoplanets and protoplanetary systems. The computational requirements for data processing and real-time control are not feasible using conventional computer and cluster architectures. This paper investigates the latency and throughput bound hardware acceleration of wavefront reconstruction and real-time control algorithms using GPUs and FPGAs. Two different correlation methods are studied and targeted on the hardware accelerators for optimal performance. The GPU-based implementations exhibit lower latency due to its superior floating point capability and supporting libraries. The FPGA-based implementation is slower and requires more fine-tuning to yield more throughput than the GPU implementation.