Lightweight Dual Modular Redundancy through Approximate Computing

  • Gabriel L. Nazar UFRGS / Technische Universität Berlin
  • Pedro H. Kopper UFRGS
  • Marcos T. Leipnitz UFRGS
  • Ben Juurlink Technische Universität Berlin


Approximate computing has been proposed to take advantage of the resilience to imprecision found in many applications domains, providing improvements in area, performance and energy consumption. In fault-tolerant systems, this paradigm can be used to reduce costs through approximate redundant modules. Additionally, further reductions can be achieved both by avoiding traditional triplication-based techniques and by embracing imprecision in the output results, especially for those application domains that can endure approximations. Using these ideas in conjunction, in this work we focus on fully exploiting approximate computing in fault-tolerant systems. With the goal of providing substantial cost reductions, we propose applying approximations on both modules of a dual modular redundancy scheme, accepting a degree of imprecision even in the absence of faults. The technique is integrated in a high-level synthesis tool to automate the generation of a number of design choices with varying costs and degrees of approximation. The gains provided by approximation in both area and performance are used to maximize the system throughput with two checking schemes that aim at different system constraints. Results show that it is possible to substantially reduce costs, such as implementing a fully duplicated system with only 31% area overhead, while introducing intermediate quality degradation.
Palavras-chave: Degradation, Costs, Filtering, Approximate computing, Redundancy, Fault tolerant systems, Tools, fault tolerance
Como Citar

Selecione um Formato
NAZAR, Gabriel L.; KOPPER, Pedro H.; LEIPNITZ, Marcos T.; JUURLINK, Ben. Lightweight Dual Modular Redundancy through Approximate Computing. In: SIMPÓSIO BRASILEIRO DE ENGENHARIA DE SISTEMAS COMPUTACIONAIS (SBESC), 11. , 2021, Evento Online. Anais [...]. Porto Alegre: Sociedade Brasileira de Computação, 2021 . p. 73-80. ISSN 2237-5430.