Automated Range and Precision Bit-Width Allocation for Iterative Computations

As scientific computing becomes more widespread in environments where form-factor considerations necessitate hardware acceleration, the problem of selecting numerical data representations (bit-width allocation), key to accelerator design, is faced with shortcomings in the existing techniques. To address this problem for scientific computing dataflows, we propose a methodology for determining custom hybrid fixed/floating-point data representations for iterative computations.