40. CHARACTERIZING THE VARIABILITY IN POWER CONSUMPTION IN MODERN COMPUTING PLATFORMS

Department: Computer Science & Engineering
Faculty Advisor(s): Yuvraj Agarwal | Rajesh Gupta

Primary Student
Name: Bharathan Balaji
Email: bbalaji@ucsd.edu
Phone: 858-367-9462
Grad Year: 2014

Abstract
Variability in performance and power consumption is slated to grow further with scaling down of process technologies. While this variability has been studied and modeled before, there is lack of real empirical data on its extent, as well as the factors which affect it, in current general purpose microprocessors which are already aggressively binned by manufacturers. In this paper we show using detailed power measurements that the within-die variability for a modern nehalem class processor is significant and it changes based on operating conditions. Our data shows within-die power variability for a Core i5-540M part to be a maximum of 8.8% and average of 1.7% across several applications. Furthermore, we show that die-to-die variability across 10 identically binned parts is even larger, and it changes based on applications and workload characteristics. Our measurement data shows die-to-die variations in power consumption to be a maximum of 65% and on average 15% across different applications. Finally, we discuss the implications of this observed variability in power consumption across presumed identical parts.

« Back to Posters or Search Results