Conclusions - Computer Architecture Techniques for Power-Efficiency - page 181

CHAPTER 6
Conclusions
While performance considerations have always played a prominent role in computer architec-
ture, power consumption has been an increasingly significant driver of the field in recent years.
Both the increase in dynamic power and even more so the exponential increase of static power
have brought sweeping changes in the way we design and build processors.
In the last two decades or more, the industry has benefited from exponential performance
improvments stemming from a combination of exponential increase in clock frequency, along
with architectural innovations. Clock frequencies today, however, have hit the power wall. They
are currently increasing at a much slower pace than ever before, and in some cases have decreased
from prior chip generations. Instead of frequency increases as the main performance driver,
other options become paramount. In particular, there is much greater focus today on chip
multiprocessors (CMPs) and on application-specific designs. Both of these have interesting
and important power implications, as we will discuss below. In the subsections that follow, we
conclude this topic with summaries of current state of the art, as well as views of promising
future techniques and trends.
6.1 DYNAMIC POWER MANAGEMENT VIA VOLTAGE AND
FREQUENCY ADJUSTMENT: STATUS AND FUTURE TRENDS
Much of the early focus on dynamic power management concerned possible voltage and fre-
quency adjustments, and a wide range of dynamic or offline policies to guide them. This work
has had high leverage both in terms of published research results as well as in terms of com-
mercial adoption. Part of the success of these techniques has stemmed from the quadratic (or
more) influence of ( V , f ) scaling on power; it has unequivocably been a high-payoff technique.
Another reason for its success stems from the fact that the mechanism of DVFS has been built
into many processors, and its existence has given researchers an excellent platform with which
to expeirment with increasingly effective policies for managing it.
Looking forward, it seems unlikely that reliance on further DVFS techniques will offer
as significant further payoffs in the future. First, much of the “low-hanging fruit” has been