An FPGA implementation of redundant residue number system for low-cost fast speed fault-tolerant computations
Date of Issue2018-12-19
School of Electrical and Electronic Engineering
Centre for Integrated Circuits and Systems
Shrinking of the device feature size allows high complexity systems to be designed and integrated with into a single chip but also causes potential issues on system reliability. Existing coding techniques can only detect and correct transmission and storage errors but not errors occurred in arithmetic operations. Redundant Residue Number System (RRNS) is a number representation that offers a more versatile fault-tolerant capability that allows error corrections in arithmetic operation with lower complexity than majority vote triple modular redundancy for fault-tolerant computing in two’s complement number system. The main issue for existing RRNS-based multiple residue digit error correction algorithms is that they sacrifice a lot on either hardware cost by using large error correction lookup tables or speed due to inherent iterations of the algorithm, resulting in their inefficient hardware implementations. This thesis presents an innovative hardware-efficient FPGA implementation of RRNS without iteration for multiple residue digit error detection and correction. The implementation is based on a syndrome-based table-lookup algorithm with new architectural design to address reduce the hardware cost and increase its speed. It adaptively partitions all the information channels such that each partitioned block contains no more than two information channels. The partitioned blocks handle the smaller scale error correction task independently and their outputs are combined to provide the corrected value. Each of the partitioned block is implemented by a modified double-error correction circuit by appropriately reorder the error vectors for table-lookup and applying modulo arithmetic properties to replace lookup tables (LUTs) with logic and adder-based circuits. The syndrome generation module is realized by base-extension operation. An array of binary comparators is used for modulo reduction factor computation and common reverse conversion circuitries among modulo channels are reused to lower hardware cost. Hardware cost due to large modulo reduction operation is further reduced by multi-level table lookup technique. Virtex-UltraScale xcvu190-flgc2104-2-e FPGA is used to evaluate the hardware cost and the circuit speed of the proposed implementation and direct implementations of other RRNS-based multiple error correction algorithms. It has in total of 1074240 LUT slices and 2148480 flip-flops available. FPGA synthesis results show that the proposed work has the lowest hardware cost, critical path delay and throughput compared with other implementations. Using small arbitrary moduli set given that the size of each information channel no more than 8-bit, with eight information channels and eight correctable residue digit errors, the proposed circuit can be implemented with 84148 LUT slices (7.8% of the total available LUT slices) and critical path delay of 19.3 ns. This achieves 93% of hardware saving and 52.5% of speed improvement over implementations of the other recent RRNS-based algorithm. Meanwhile, the system reliability of the proposed multiple residue digit correction scheme is high enough compared with existing fault-tolerant methods for mission-critical applications.
DRNTU::Engineering::Electrical and electronic engineering