NUFFT-based SAR backprojection on multiple GPUs

We report on the development of a Synthetic Aperture Radar (SAR) backprojection algorithm implemented on multiple Graphics Processing Units (GPUs) in CUDA language and using a Non-Uniform FFT (NUFFT) routine to further numerically accelerating the calculations. The performance of the approach is analyzed in terms of computational speed and scalability on the GPU ”Jazz” cluster available at the Consorzio Interuniversitario per le Applicazioni del Supercalcolo per l'Universit`a e la Ricerca (CASPUR; InterUniversity Consortium for the Application of Super-Computing for Universities and Research), Rome, Italy. The results, referring to the case of an individual node with 2-GPUs, show that the processing time scales approximately by 2 as compared to the case of a single GPU. Experimental results against the Air Force Research Laboratory (AFRL) airborne data delivered under the ”challenge problem for SAR-based Ground Moving Target Identification (GMTI) in urban environments” and collected under circular flight paths are also shown. The full processing of the data took approximately 19s on the mentioned 2-GPUs node.