Abstract

With large-scale integration and high power density in current generation microprocessors, thermal management is becoming a critical component of system design. Specifically, accurate thermal monitoring using on-die sensors is vital for system reliability and recovery. Achieving an accurate thermal profile of a system with an optimal number of sensors is integral for thermal management. This work focuses on a sensor placement mechanism and an on-chip sensor mini-network to combine temperatures from multiple sensors to determine the full thermal profile of a chip. The sensor placement mechanism proposed in this work uses non-uniform subsampling of thermal maps with k-means clustering. Using this sensing technique with cubic interpolation, an 8-core architecture thermal map was successfully recovered with an average error improvement of 90% over sensor placement via basic k-means clustering. All the simulations were run using HotSpot 5.0 modeling Alpha 21364 processor as a baseline core. The sensor mini-network using both differential encoding and distributed source coding was analyzed on a 1024-core architecture. Distributed source coding compression required fewer transmissions than differential encoding and reduced the number of transmitted bits by 36% over a sensor mini-network with no compression.

Library of Congress Subject Headings

Microprocessors--Thermal properties; Sensor networks--Design and construction

Publication Date

8-1-2010

Document Type

Thesis

Department, Program, or Center

Computer Engineering (KGCOE)

Advisor

Kudithipudi, Dhireesha

Comments

Note: imported from RIT’s Digital Media Library running on DSpace to RIT Scholar Works. Physical copy available through RIT's The Wallace Library at: TK7870.25 .D45 2010

Campus

RIT – Main Campus

Download

Share

COinS