RNET TECHNOLOGIES INC — Department of Defense SBIR Phase I: SB152-004

Description

The proposed product will provide extreme performance on large graph data sets by developing novel support for well known graph frameworks on edge-of-the-art GPU architectures that include innovative stacked memory. This provides an ultra-high bandwidth (over 1 TB/s) and large (16+ GB) memory near the GPU processing cores. These advancements provide an enormous opportunity to leverage GPUs for extreme scale data bound applications, such as graph analytics. The GPU implementations will be specifically developed to leverage the emerging stacked memory that provides high bandwidth DRAM. In addition, performance modeling will be developed to estimate the performance of user specified graph queries. The performance modeling will account for the query, the graph dataset, and the chosen architecture. The Phase I effort will demonstrate the feasibility to efficiently leverage the upcoming stacked memory in NVidia GPUs with over 1 TB/sec of bandwidth. The Phase I effort will evaluate the optimization opportunities for the GPU optimizations in a well known graph framework, perform an initial demonstration on existing NVidia GPU platforms, and develop a complete design for the implementation of the Phase II optimizations that target the upcoming NVidia Pascal and Volta architectures (with unified memory, stacked memory, and NVLink).