Algorithm Development for Reconfigurable Computing Architectures

Partial Differential Equations (PDEs) are at the core of Air Force scientific priorities in air vehicle design. As part of an ambitious research program funded by the Air Force, NASA, and DOE, Accelogic is spearheading the development of LAPACKrc, a groundbreaking family of FPGA-based linear algebra solvers able to achieve speedups larger than 1,000x with a single chip. Once completed, LAPACKrc will subsume both LAPACK and ScaLAPACK functionalities, as well as several ultra-fast sparse direct and iterative matrix solvers.     The goals of this STTR are: (1)to develop the fundamental RC algorithmic technology for LAPACKrcs high-performance sparse iterative solvers; (2)to produce these solvers and their underlying math components as portable and easy-to-use LAPACKrc utilities; and (3)to aggressively integrate the LAPACKrc technology (sparse and otherwise) into higher-level solutions that enable the acceleration of PDE problems of high priority to the Air Force.    Phase I produced the worlds fastest RC sparse iterative solver (150x speedups), demonstrating that it is possible to break the memory-access bottleneck through software-based non-von-Neumann algorithmic schemes. Phase II will continue this strong trend, using Accelogics proprietary technology and know-how to achieve 1,000x speedups through non-von-Neumann algorithms.   Phase II/III follow-on funding commitments have been secured for a total of $550,000. BENEFIT: ---Anticipated Direct Benefits: A strong emphasis of this STTR is on maximizing technology transfer value to the Air Force. To this end,  we will move aggressively into integrating all of the necessary elements of the LAPACKrc technology (including those produced by this STTR and all others already in place) to produce groundbreaking acceleration and/or enable high-priority Air Force applications. By the end of Phase II, we anticipate at least 6 success stories for which groundbreaking acceleration will be demonstrated. At least three of these successes will be attained with commercial partners. Four of the most dominant players in the CFD software industry ANSYS, Metacomp, NUMECA, and CHAM have provided letters of interest/commitment to participate in these Phase II integration efforts (letters attached). The other three success stories (and likely more) will come from Air Force programs. Four high-priority Air Force programs from WPAFB and Edwards AFB have submitted letters of interest/commitment to establish collaborations that will lead to the acceleration and/or enabling of: (1) applications of high-fidelity CFD for improved energy consumption and aerodynamic efficiency in Air Force's aircraft fleet, (2) advanced research in high temperature structural metals, (3) plasma physics, and (4) hypersonic propulsion systems; (letters attached). ---DoD Applications: As leading industry technologist and President of Metacomp Technologies, Dr. Sukumar Chakravarthy, says jointly with Dr. Dan Risha, Senior Research Scientist at AFRLs Aerospace Propulsion Division, Accelogics pioneering work on linear equations promises to set a tremendous milestone that will transform the landscape of modeling and simulation science as we know it today, thus enabling a significant leap in performance for important applications that are at the core of the Air Forces mission and of DoD missions in general.  Those applications include CFD, structural analysis, computational molecular dynamics, chemistry, computational electromagnetics, machinery, combustion, plasma physics, weather and climate modeling, and earth sciences. These fields depend heavily on large-scale linear algebra kernels, and, prior to LAPACKrc, they had not yet seen significant benefits from the typical performance leaps of reconfigurable supercomputing, due to the difficulties that the RC community has had in achieving large speedups for linear equations.  As highlighted in several of our Phase I reports,  Accelogic is the first R&D group to have demonstrated large, sustained speedups for non-toy, large-scale linear equations of relevance to the aerospace industryand particularly to DoD. ---Non-DoD Applications: In addition to the DoD applications, many other applications are highly dependent on high-performance linear algebra, and those applications will be enabled by the acceleration levels (1,000x and up) targeted by Accelogic.  They include circuit simulation, financial portfolio management optimization (interior point methods and linear programming), nuclear physics, power distribution network design, and radar signal processing.  Each one these applications represents a significant market by itselfwith many of them crossing the $1 billion threshold. The ability to solve large-scale linear equations rapidly is a clear and fundamental driver for the following activities and/or industries both inside and outside of DoD:  product design and performance simulation (aerospace, automotive, defense, boats, nuclear energy, electronic chips, power networks, radar, telecommunications, acoustics, buildings and bridges, drugs, chemistry, antennas and radio systems, machinery, combustion); economic forecasting (insurance, financial markets, budgeting, hedge funds management); physical forecasting (weather, environmental engineering, geophysics); and data analysis (defense, logistics, search engines, national security, gaming, marketing).