An FPGA Linear Algebra Library for Maximal-Performance Petascale Supercomputing

It is estimated that over 70% of supercomputing usage today is dedicated to solving large systems of linear equations, such as those at the core of DOE computational problems in fusion research, accelerator simulations, and astrophysics. This project will develop technology to provide unprecedented computational power to the solution of large, unstructured, dense and sparse linear equations. The approach involves the development of both direct and iterative algorithms, especially tuned to achieve maximal speedup in a Field Programmable Gate Array (FPGA) platform. The FPGA technology will be distributed through direct connections of single nodes to a supercomputer network infrastructure. Phase I will develop a working prototype (both software and hardware) that will be pluggable to an Opteron node of a DOE supercomputer, at minimal cost. In Phase II, a full working prototype of the software will be delivered to DOE. The results will be packaged in a reconfigurable computing software library that will be at least 1,000 times faster than traditional commodity processors. Commercial Applications and other Benefits as described by the awardee: With all major supercomputer vendors turning to hybrid architectures, in which CPUs and FPGAs co-exist, this technology should enable the highest performance for linear algebra problems in future supercomputers. Its use within a supercomputing network will be at least two orders of magnitude less expensive than CPU-based technology with similar performance. The technology should revolutionize many industries that depend heavily on the ability to solve large systems of linear equations: fusion energy, aerospace, automotive, bridge building, logistics optimization, and linear programming,