Cardiff University | Prifysgol Caerdydd ORCA
Online Research @ Cardiff 
WelshClear Cookie - decide language by browser settings

Comparison of logarithmic and floating-point number systems implemented on Xilinx Virtex-II field-programmable gate arrays

Lee, Barry Roland 2004. Comparison of logarithmic and floating-point number systems implemented on Xilinx Virtex-II field-programmable gate arrays. PhD Thesis, Cardiff University.

[img] PDF - Accepted Post-Print Version
Download (22MB)


The aim of this thesis is to compare the implementation of parameterisable LNS (logarithmic number system) and floating-point high dynamic range number systems on FPGA. The Virtex/Virtex-II range of FPGAs from Xilinx, which are the most popular FPGA technology, are used to implement the designs. The study focuses on using the low level primitives of the technology in an efficient way and so initially the design issues in implementing fixed-point operators are considered. The four basic operations of addition, multiplication, division and square root are considered. Carry- free adders, ripple-carry adders, parallel multipliers and digit recurrence division and square root are discussed. The floating-point operators use the word format and exceptions as described by the IEEE std-754. A dual-path adder implementation is described in detail, as are floating-point multiplier, divider and square root components. Results and comparisons with other works are given. The efficient implementation of function evaluation methods is considered next. An overview of current FPGA methods is given and a new piecewise polynomial implementation using the Taylor series is presented and compared with other designs in the literature. In the next section the LNS word format, accuracy and exceptions are described and two new LNS addition/subtraction function approximations are described. The algorithms for performing multiplication, division and powering in the LNS domain are also described and are compared with other designs in the open literature. Parameterisable conversion algorithms to convert to/from the fixed-point domain from/to the LNS and floating-point domain are described and implementation results given. In the next chapter MATLAB bit-true software models are given that have the exact functionality as the hardware models. The interfaces of the models are given and a serial communication system to perform low speed system tests is described. A comparison of the LNS and floating-point number systems in terms of area and delay is given. Different functions implemented in LNS and floating-point arithmetic are also compared and conclusions are drawn. The results show that when the LNS is implemented with a 6-bit or less characteristic it is superior to floating-point. However, for larger characteristic lengths the floating-point system is more efficient due to the delay and exponential area increase of the LNS addition operator. The LNS is beneficial for larger characteristics than 6-bits only for specialist applications that require a high portion of division, multiplication, square root, powering operations and few additions.

Item Type: Thesis (PhD)
Status: Unpublished
Schools: Engineering
ISBN: 9781303200533
Funders: QinetiQ
Date of First Compliant Deposit: 30 March 2016
Last Modified: 10 Oct 2017 15:29

Actions (repository staff only)

Edit Item Edit Item


Downloads per month over past year

View more statistics