Optimizing Capacitance and Switching Activity to Reduce Dynamic Power - Computer Architecture Techniques for Power-Efficiency - page 46

many cases, such techniques inadvertently affect capacitance—at design time. Because they can
affect both factors, more broadly, their goal is to reduce the effective switched capacitance .An
orthogonal approach to such optimizations is to dynamically steer work to structures designed
to offer a range of power/performance levels.
The structure of this chapter is based on a categorization of the excess switching ac-
tivity (and by extension a categorization of the effective switched capacitance). The following
section, Section 4.1, presents this categorization (summarized in Table 4.1) and creates the
road map for the main body of the chapter. Work steering approaches conclude the chapter
(Section 4.13).

4.1 A ROAD MAP FOR EFFECTIVE
SWITCHED CAPACITANCE
Techniques to reduce effective switching capacitance have been developed for every major pro-
cessor structure—the datapath, the ALU, the caches, the instruction scheduling hardware, and
so on. In many cases, techniques for different structures share a common line of attack because
they focus on the same root cause of excess switching activity . Motivated by this observation,
the main body of this chapter unfolds based on a categorization of the excess switching activity
(presented in Section 4.1.1).
The categorization of excess switching activity also implies a corresponding categoriza-
tion of the effective switched capacitance when one takes into account the potential impact on
capacitance. Section 4.1.2 gives a short overview on how architectural techniques presented in
this chapter can affect capacitance. However, for the rest of the chapter, we will not explicitly
refer to capacitance except in a few places where it is needed.
Finally, the chapter concludes with a section on dynamic work steering .

Approaches in this class statically provide multiple units that can perform the same work but
with different power/performance characteristics. Rather than trying to optimize away excess
switching activity in a single unit, these approaches simply steer computation to the appropriate
unit according to runtime requirements for power or performance. Because any of the types of
excess switching activity can potentially be handled with dynamic work steering, we consider
this strategy to be orthogonal to the optimization techniques for excess switching activity.

(Section 4.13)

4.1.1 Excess Switching Activity
To classify excess switching activity, we ask the following question: why is it there in the
first place? We distinguish a number of causes, which lead to seven different activity types.
These types are summarized in Table 4.1, along with the corresponding causes for their
existence (column two), the typical granularity where they appear (column three), a quick