FAULT TOLERANT SIGNAL PROCESSING FOR VLSI CIRCUITS

Open Access
Author:
Radhakrishnan, Chandrasekhar
Graduate Program:
Electrical Engineering
Degree:
Doctor of Philosophy
Document Type:
Dissertation
Date of Defense:
December 14, 2010
Committee Members:
  • William Kenneth Jenkins, Dissertation Advisor/Co-Advisor
  • William Kenneth Jenkins, Committee Chair/Co-Chair
  • John F Doherty, Committee Member
  • Vijaykrishnan Narayanan, Committee Member
  • Dr Y Xie, Committee Member
Keywords:
  • Discrete-time correlation
  • discrete-time convolution
  • Fermat Number Transforms
  • Discrete Fourier Transforms
  • Transient errors
  • Soft errors
  • Hardware faults
  • Fault tolerant signal processing
  • Fault tolerant transform adaptive filtering
  • adaptive filtering
  • Residue Number Systems (RNS)
  • Modulus Replication Residue Number System (MRRNS)
  • Arithmetic Fault Tolerance
Abstract:
To achieve higher speed, higher density, and lower power dissipation CMOS VLSI circuits continue to be scaled down in terms of feature sizes and power supply voltages. Although technology scaling has enabled the design of high performance processors, susceptibility to transient and permanent errors, process variation, and high leakage power are some bottlenecks that need to be overcome for continued device scaling into sub nano-meter technologies. Digital Signal Processing (DSP) applications are numerically intensive and many of them are implemented on hand-held devices making power efficiency an important constraint in these applications. The proposed solutions must be application aware as otherwise they may not be the best in terms of efficiency. The goal of this work is to identify important DSP applications and develop algorithmic and arithmetic level solutions to overcome the problem of transient and permanent errors in scaled down devices. In this work transform domain adaptive filters operating on real-valued signals are considered first and error tolerance is introduced by exploiting redundancy in the transform domain. A comprehensive analysis of fault coverage, complexity and performance is done, and architectures to implement the developed solutions are also proposed. Residue and polynomial arithmetic techniques are then combined with algorithmic concepts to design error detection and correction for Walsh-Hadamard based transform domain adaptive filters. Finally, a class of modified transforms is developed to implement efficient convolution operation. The modified transforms are used to implement fault tolerant convolution and a detailed analysis of fault coverage, complexity and performance is done. The applications of modified transforms in reliable adaptive filtering and speech de-noising is illustrated.