论文部分内容阅读
The particles-induced soft errors are a major threat to the reliability of microprocessors. Even worse,multi-bits upsets(MBUs) are ever-increased due to the rapidly shrinking feature size of the IC on a chip. Several architecture-level mechanisms have been proposed to protect microprocessors from soft errors, such as dual and triple modular redundancies(DMR and TMR). However, most of them are inefficient to combat the growing multibits errors or cannot well balance the critical paths delay, area and power penalty. This paper proposes a novel architecture, self-recovery dual-pipeline(SRDP), to effectively provide soft error detection and recovery with low cost for general RISC structures. We focus on the following three aspects. First, an advanced DMR pipeline is devised to detect soft error, especially MBU. Second, SEU/MBU errors can be located by enhancing self-checking logic into pipelines stage registers. Third, a recovery scheme is proposed with a recovery cost of 1 or 5 clock cycles.Our evaluation of a prototype implementation exhibits that the SRDP can successfully detect particle-induced soft errors up to 100% and recovery is nearly 95%, the other 5% will inter a specific trap.
The particles-induced soft errors are a major threat to the reliability of microprocessors. Even worse, multi-bits upsets (MBUs) are ever-increased due to the rapidly shrinking feature size of the IC on a chip. been proposed to protect microprocessors from soft errors, such as dual and triple modular redundancies (DMR and TMR). However, most of them are inefficient to combat the growing multibits errors or can not well balance the critical paths delay, area and power penalty. This paper proposes a novel architecture, self-recovery dual-pipeline (SRDP), to provide provide error error detection and recovery with low cost for general RISC structures. We focus on the following three aspects. First, an advanced DMR pipeline is devised to detect Second, SEU / MBU errors can be located by enhancing self-checking logic into pipelines stage registers. Third, a recovery scheme is proposed with a recovery cost of 1 or 5 clock cycl es.Our evaluation of a prototype implementation exhibits that the SRDP can successfully detect particle-induced soft errors up to 100% and recovery is nearly 95%, the other 5% will inter a specific trap.