Proactive power migration to reduce maximum value and spatiotemporal non-uniformity of on-chip temperature distribution in homogeneous many-core processors

Abstract

This paper presents a proactive spatiotemporal power multiplexing method to manage the thermal field in many-core processors. We first analyze the thermal field in many core processors in deep nanometer (to 16nm nodes). We show that the thermal field in many-core can have significant spatiotemporal non-uniformity along with high maximum temperature. For better reliability and improved cooling efficiency, it is important to achieve a lower peak temperature and a more uniform thermal field under all workload or utilization conditions. We propose proactive power migration to reduce spatial and temporal temperature difference, by redistributing the heat generating locations. The effectiveness of the proposed method is demonstrated for a 256 core many-core processor in predictive 16nm nodes.