Memory-communication model for low-latency X-ray video processing on multiple cores

Abstract. Despite the speed up of PC technology over the years, real-time performance of video processing in medical X-ray procedures continues to be an issue as the size and number of concurrent data streams is increasing steadily. Since the computing evolves quicker than memory technology, there is an increasing pressure on an efficient use of the off-chip memory bandwidth. Additionally, as a multitude of video functions is carried out in parallel, the memory-bandwidth problem is further stressed. In this paper, we present an architecture study for performance prediction and optimization of medical X-ray video-processing on multiple cores. By carefully modeling the critical stages of the architecture, bottlenecks are known in detail. Model descriptions for the video-processing algorithms are inserted into the architecture model, making explicit where data and functions needs to be partitioned to obtain higher throughput. For the application under study, we propose a combined 2-level data partitioning with functional partitioning scheme that result in a bandwidth and latency reduction of 40-70% compared to straightforward implementations.