Robust Design Methods for Hardware Accelerators for Iterative Algorithms in Scientific Computing

The ubiquity of embedded systems of increasing complexity in domains like scientific computing requires computation on models whose complexity has grown beyond what is economical to manage purely in software to requiring hardware acceleration - a key part of which is selecting numerical data representations (bit-width allocation). To address the shortcomings of existing techniques when applied to scientific computing dataflows, we propose a methodology for determining custom hybrid fixed/floating-point data representations for iterative scientific computing applications.