Abstract

Fault tolerance on various adder architectures implemented on Field Programmable Gate Arrays (FPGAs) is studied in this thesis. This involves developing error detection and correction techniques for the sparse Kogge-Stone adder and comparing it with Triple Modular Redundancy (TMR) techniques. Fault tolerance is implemented on a Kogge-Stone adder by taking advantage of the inherent redundancy in the carry tree. On a sparse Kogge-Stone adder, fault tolerance is realized by introducing additional ripple carry adders into the design. The implementation of this fault tolerance approach on the sparse Kogge-Stone adder is successfully completed and verified by introducing faults either on the ripple carry adder or in the carry tree. Two types of Xilinx FPGAs were used in this study: the Spartan 3E and Virtex 5. The fault tolerant adders were analyzed in terms of their delay and resource utilization as a function of the widths of the adders. The results of this research provide important design guidelines for the implementation of fault tolerant adders on FPGAs. The Triple Modular Redundancy-Ripple Carry Adder (TMR-RCA) is the most efficient approach for fault tolerant design on an FPGA in terms of its resources due to its simplicity and the ability to take advantage of the fast-carry chain. However, for very large bit widths, there are indications that the sparse Kogge-Stone adder offers superior performance over an RCA when implemented on an FPGA. Two fault tolerant approaches were implemented using a sparse Kogge-Stone architecture. First, a fault tolerant sparse Kogge-Stone adder is designed by taking advantage of the existing ripple carry adders in the architecture and adopting a similar approach to the TMR-RCA by inserting two additional ripple carry adders into the design. Second, a graceful degradation approach is implemented with the sparse Kogge-Stone adder. In this approach, a faulty block is permanently replaced with a spare block. As the spare block is initially used for fault checking, the fault tolerant capability of the circuit is degraded in order to continue fault-free operation. The adder delay is smaller for the graceful degradation approach by approximately 1 ns from measured results and 2 ns from the synthesis results independent of the bit widths when compared with the fault tolerant Kogge-Stone adder. However, the resource utilization is similar for both adders.

Date of publication

Spring 4-27-2012

Document Type

Thesis

Language

english

Persistent identifier

http://hdl.handle.net/10950/62

Download
COinS