Abstract: Software-based ultrasound imaging systems send raw echo data acquired by the ultrasound transducer to a computer for processing and imaging. The computation is performed on Graphics Processing Units (GPUs) to process the enormous amount of streaming data, around 5GB/s, and generate images in real-time. Computation throughput is an important metric for such a system. In this paper, we have experimented the possibility of using mixed precision data types to speed up the computation for ultrasound imaging software on an Nvidia RTX 2060 GPU. Compared to using only single precision in the computation, the updated software can speed up imaging software by over 15% without sacrificing image quality. The benefit is from faster floating-point operations and reduced memory traffic. The improved computation capacity can also be used to improve the image quality.
Keywords: ultrasound, beamforming, GPU, parallel computing, Compute Unified Device Architecture(CUDA), single precision, half precision, mixed precision, floating point, profiling, memory bandwidth.