An FPGA Linear Algebra Library for Maximal-Performance Petascale Supercomputing

Over 70% of supercomputing usage today is dedicated to solving large systems of linear equations. Linear equations are at the core of the most important DOE computational problems in energy fusion research, accelerator simulations, weather modeling, oil & gas exploration, chemistry, materials, and astrophysics, among others. This project will provide computational power to the solution of large dense and sparse linear equations through both direct and iterative algorithms especially tuned to achieve maximal performance in a supercomputer enhanced with FPGA (field-programmable gate array) accelerators. The result of this research will produce the core of LAPACKrc, a linear algebra library that, by the end of Phase II, will be 2 to 3 orders of magnitude faster than traditional CPU-based solvers. The library will be easy to integrate with the leadership facilities of DOE supercomputers. Its use within a supercomputing network will be 1 to 2 orders of magnitude less expensive than CPU-based technology with similar performance. The Phase I project demonstrated the world¿s first FPGA-accelerated solver to achieve speed increases >90x for large-scale dense linear equations with 64-bit accuracy. Similar levels of performance were demonstrated for a complex sparse iterative solver. Phase I also established the technical foundation upon which the LAPACKrc library will now be built. The Phase II project will produce the LAPACKrc core. Special attention will be given to industrial-quality software features, including scalability, portability, ease-of-use, and numerical robustness. Commercial Applications and other Benefits as described by the awardee: This technology should enable the highest performance for linear algebra problems in future supercomputers, improving the ability to solve large systems of linear equations in fusion energy, aerospace, automotive, bridge building, logistics, financial engineering, and linear programming.