Quasi-maximum accuracy floating-point computations with GPGPU for applications in digital signal processing

An idea of the use of two accumulators for improvement of the precision of floating-point computations with graphic processing units (GPUs) is presented in this paper for applications in digital signal processing. The increase of the precision of computations does not need any increase of the length of the data words. This is particularly important if hardware limits for the precision of computations exist, which is just the case for graphic processors. A history of development of the cores of graphic cards is analyzed together with the idea of general purpose computing using GPU´s (GPGPU). Special attention has been paid to efficiency and precision of computations. The so-called maximum accuracy property has been analyzed and technically realized with no additional costs in hardware and computation time. The proposed approach has been tested with illustrative frequency modulated sine waveform generators.