Salvaging chips with caches beyond repair

Abstract

Defect density and variabilities in values of parameters continue to grow with each new generation of nano-scale fabrication technology. In SRAMs, variabilities reduce yield and necessitate extensive interventions, such as the use of increasing numbers of spares to achieve acceptable yield. For most microprocessor chips, the number of SRAM bits is expected to grow 2× for every generation. Consequently, microprocessor chip yields will be seriously undermined if no defect-tolerance approach is used. In this paper, we show the limits of the traditional spares-based defect-tolerance approaches for SRAMs. We then propose and implement a software-based approach for improving cache yield. We demonstrate that our approach can significantly increase microprocessor chip yields (normalized with respect to chip area) compared to the traditional approaches, for upcoming fabrication technologies. In particular, we demonstrate that our approach dramatically increases effective computing capacity, measured in MIPS-per-unit-chip-area. Our approach does not require any hardware design changes and hence can be applied to improve yield of any modern microprocessor chip, incurs low performance penalty only for the chips with unrepaired defects in SRAMs, and adapts without requiring any design changes as the yield improves for a particular design and fabrication technology.