Stream Oriented Modular Architecture with Polymorphic Processing Engines

摘要

Stream computing has shown to be an effective technique to decouple communication from computation in many application domains. It provides an efficient mitigation of bandwidth restrictions, by reducing the amount of memory accesses and by maximizing the available computational resources, potentiating the parallel processing using multiple execution engines. However, it frequently implies significant development costs, since efficient stream-based architectures are usually attained through application-specific full-custom processors, often tightened to the application at hand. To circumvent this limitation, a modular stream computing architecture aiming generic and high performance applications is presented. The proposed architecture, designed for reconfigurable hardware, is composed of modular processing engines that can be customized by the end-user in many ways, such as in terms of their number, type and precision of the functional units. Furthermore, these processing engines are designed as programmable cores, allowing the execution of a wide set of applications using the same configuration. From the conducted evaluation by using a series of benchmark case studies, it was observed that the proposed architecture achieves competitive results when compared with alternative solutions.