Modeling and Analyzing CPU Power and Performance:
Metrics, Methods, and Abstractions
The power dissipated by current computer systems is an
increasingly pressing problem. As mobile and handheld computers
become ubiquitous in our society, extending the battery life of these
battery-powered devices becomes a major design goal. Power efficiency
also impacts the thermal design of computer systems. In desktop
systems, the extreme amounts of power consumed by high-end
microprocessors necessitate elaborate packaging and cooling techniques
that add significantly to the cost of these systems. By the next
generation of high-end processors, power dissipation is expected to be
a major limiter on the design choices available to designers, and
therefore a major limiter on the performance of next-generation CPUs.
Given the importance of the power problem, it is clear that developing
solid power modeling techniques will become increasingly crucial to attack
and understand the problem. The power modeling problem is a complex one.
Even more so than performance, the best power models are quite low-level,
relying on capacitance data from a chip that has already been fully designed.
Increasingly, however, architects need basic feedback on power consumption
as they make design decisions, much earlier in the process of architecting
a system. Therefore, modeling techniques that abstract key aspects of power
dissipation, while maintaining as much accuracy and speed as possible, are
key to addressing this issue.
This tutorial will offer an introduction to power and performance
modeling for computer design. We will discuss a range of modeling
methods, from low-level methods based on capacitance estimates
employed by CAD tools such as Synopsys PowerMill, to higher-level
methods intended mainly for early-stage architecture-level power
modeling. Finally, we will present strategies for determining and
validating abstractions that can be useful in accelerating power
modeling without dangerously degrading its fidelity . In addition to
the lecture format, we will also give demonstrations of tools,
measurements and validation techniques in action.
Who should attend?
The tutorial is primarily intended for an audience of computer
architects, performance modelers and/or simulator builders who are
interested in learning about modeling challenges and opportunities in
the arena of power dissipation modeling. We also welcome the
participation of other experienced power modelers who would like to
engage in discussion about the pros and cons of various techniques.
Prof. Margaret Martonosi is currently an associate professor in the
Princeton University Department of Electrical Engineering, where she
has been on the faculty since 1994. Martonosi's research interests
include computer architecture, the hardware/software interface, and
related modeling/simulation issues. Her current work particularly
focuses on power-efficient systems. Martonosi received her Ph.D. in
Electrical Engineering from Stanford University in 1993. She also
holds a master's degree from Stanford University and a Bachelor's
degree with distinction from Cornell University, both in Electrical
Dr. Pradip Bose is a research computer scientist at IBM T. J. Watson
Research Center, Yorktown Heights, NY. He has been with IBM since
1983. Bose currently leads a research project at Watson on
power-aware microprocessor design and modeling. His research interests
include computer architecture, modeling and validation. Bose received
his M.S. and Ph.D degrees in Electrical and Computer Engineering from
University of Illinois, Urbana-Champaign in 1981 and 1983,
respectively. He also holds a Bachelor's degree in Electronics and
Electrical Communication Engineering from Indian Institute of
Technology (IIT), Kharagpur, India.
David Brooks is currently pursuing a Ph.D degree in Electrical
Engineering at Princeton University. His research interests include
architecture-level power-performance modeling and the definition of
power-efficient, high-performan ce microarchitectures. He received
his BSEE from the University of Southern California and his MA degree
in Electrical Engineering from Princeton. By Summer, 2001, he is
planning to be finished with his PhD and is planning to take a faculty
position as of September, 2001.