Hot spare components for performance-cost improvement in multi-core SIMT

Abstract

Adding redundant components is a well known technique for replacing defective components either before shipment or in the field, resulting yield improvement and consequently cost reduction. However, most yield improvement strategies utilize redundant components only when another component fails (i.e., cold spares). In this paper, we investigate the cost and performance implications of employing hot spares in multi-core single-instruction, multiple-thread (SIMT) processors. Hot spares are available to increase yield (and reduce costs) when the components are defective; otherwise, they can be used to improve performance in the field. Starting with a baseline architecture with six cores, and 32 lanes each, we added three hot spare cores, with two lanes each. When we make the lanes of the hot spares available to replace defective lanes in the baseline cores, we observe that expected performance per cost improved more than 2.5 and 1.7 times relative to systems integrating no redundancy and cold spares, respectively.