Design of a Field-Programmable Dual-Precision Floating-Point Arithmetic Unit

Abstract

The growth in FPGA capacity and the inclusion of embedded arithmetic cores has enabled the use of these devices for general purpose floating-point computing. Despite their clock rate handicap with respect to contemporary general-purpose processors, these devices can be field-programmable to meet the precision requirements and operator-level parallelism of a specific computation. In this paper we describe and evaluate the performance of dual-precision, pipelined, floating-point arithmetic cores for addition, multiplication and division. Each of these arithmetic cores can be switched at run-time to perform either one double-precision operation, or with the same hardware resources, perform two single-precision operations. We also implemented quad-precision cores which can be switched to perform either one quad-precision operation or two double-precision operations. As an application of these cores, we describe and evaluate the performance potential of a custom, but flexible, vector processing units as part of a system-level architecture targeting a Xilinx Virtex-II Prom 100 FPGA device connected to multiple SRAM banks