Direct Sparse Linear Solver Suite for Mximal Performance FPGA/CPU Heterogeneous Supercomputing - An Enhancement to the Sca/LAPACKrc Library

An estimated 70% of supercomputing cycles worldwide are spent solving large-scale systems of linear equations. Sparse linear equations are at the core ¿ and constitute the primary bottleneck ¿ of important DOE research problems in energy fusion, accelerator simulations, and astrophysics, among other research challenges. The DOE has co-funded an ambitious research program to develop the LAPACKrc platform, which has produced the world¿s fastest sparse iterative solver, based on Field-Programmable Gate Array (FPGA) technology. This program has demonstrated that it is possible to break the memory-access bottleneck through software based non-von-Neumann algorithmic schemes. This project will develop the theoretical and implementation innovations necessary to deploy the world¿s first industrial library of FPGA-based solvers, an important functionality still missing in the LAPACKrc solver suite. Phase I will produce the foundations of the algorithmic technology and demonstrate that the technology will enable modern heterogeneous CPU/FPGA systems to solve complex problems two to three orders of magnitude faster than traditional CPUs, while preserving portability, scalability, and ease of use. Commercial Applications and other Benefits as described by the awardeeThe resulting technology should enable reductions of at least one order of magnitude in both cost and power consumption, while paving the road towards sustained petascale supercomputing for a wide class of computational problems of high relevance to DOE. In addition to DOE applications the technology should benefit industries that require advanced modeling and simulation for nuclear energy, oil and gas exploration, aerospace, weather, etc