Enhancing Sca/LAPACKrc with Extremely-Fast Least-Squares Solvers for Heterogeneous CPU/FPGA Supercomputers

An estimated 70% of supercomputing cycles are dedicated to solving large-scale least squares problems, which are at the core of important DOE computational problems in energy fusion research, bioinformatics, geodetics, structural analysis, tomography, and astrophysics. This project will extend the already successful Sca/LAPACKrc FPGA-accelerated solver suite for linear equations, with added functionality to tackle large-scale least squares problems at unprecedented speeds. Phase I will demonstrate the feasibility of creating an FPGA-accelerated core algorithm that will increase the speed of calculation by a factor of 1,000 over the CPU-only least-squares solvers now used in supercomputing. Phase II will incorporate the core algorithm and technology into different least-squares solvers and deliver a fully functional prototype. Its use with an FPGA platform or network will be at least two orders of magnitude less expensive than a CPU cluster exhibiting the same performance. Commercial Applications and other Benefits as described by the awardee: Beyond enabling supercomputing applications that clearly meet the needs of DOE and other governmental agencies, a large market potential for this technology is expected in the commercial sector. The technology should enable in key applications that rely on large-scale least squares solutions, including bioinformatics, geodetics, structural analysis, tomography, finance, nuclear energy, and statistics.