Enhancing Precision and Bandwidth in Cloud Computing: Implementation of a Novel Floating-Point Format on FPGA

Abstract

Cloud computing is a type of Internet-based service computing that provides computing, storage and networking services to multiple users. With the increase of data size, computing capacity runs out quickly in cloud computing services. To fill the shortage of computation capacity, we propose to adopt variable precision by implementing unum (universal number), which is a number format different from IEEE Standard for Floating-Point Arithmetic - IEEE 754 floats. Compared with IEEE 754 floats, the outstanding features of unum are clearance of rounding errors, high information-per-bit and variable precision. As a candidate replacement of IEEE 754 floats, the application of unum can improve the precision in computing, decrease the bit width for high precision numbers. However, unum was only implemented in software model before due to technical complexity, in order to validate the performance on chip, we implement this arithmetic on FPGA for the first time. We also implement an unum based 16-point FFT on FPGA. We validate the design and compare the bit width in computing with IEEE 754 floats, evaluate the power dissipation on FPGA. The experimental results of comparison show that unum arithmetic can ensure correctness even in some extreme arithmetic cases in which IEEE 754 floats cannot work properly, furthermore the bit width of unum is much less than IEEE 754 floats in the same precision.