The use of configurable computing for computational kernels in scientific simulations

In many scientific simulation codes, the bulk of the floating-point arithmetic required is done by a small number of compact computational kernels. In this paper, we explore the potential use of configurable computers to instantiate the hardware required for such kernels and, thus, improve their performance. We present algorithms and analysis for two such kernels: fast, problem-specific multipliers and the efficient evaluation of Taylor series. A novel aspect of the algorithm for Taylor series evaluation is that it takes advantage of the variable precision arithmetic available to a configurable computer. Experimental results obtained on a Xilinx field-programmable gate array (FPGA) are presented for the proposed algorithms.