Managing Static (Leakage) Power - Computer Architecture Techniques for Power-Efficiency - page 157

nodes of the memory cells when restoring power is small in comparison to the hundreds
of thousands of cycles of leakage energy saved for each cache line. A small change in the
decay interval—maybe a few tens or a few hundreds of cycles—can more than make up for
this cost.
Things are different, however, when it comes to power-gating functional units. There,
the time scales are entirely different. Inactivity must be exploited in a matter of 10's of cycles
(not 100's of thousands ). In addition, there is no substantial direct energy cost, such as the decay-
induced miss in the caches, when making a mistake and switching off a needed functional unit. 13
Instead, what matters the most in functional units is the dynamic energy cost in powering them
up or down.
Hu, Buyuktosunoglu, Srinivasan, Zyuban, Jacobson, Bose [ 105 ] provide an excellent
analysis of the costs involved when power-gating functional units. They assume that a functional
unit is power-gated either with a header (connected to V dd ) or a footer transistor (connected to
ground). In their analysis, they include both the cost of switching the gating transistor and the
cost of recharging a functional unit in relation to the length of time it was discharging. Instead
of calculating the end result for a specific example, making a number of assumptions for the
parameters involved, they give analytic formulas that yield the break-even point, in cycles, for
power-gating a functional unit. To simplify the formulas, a leakage factor L is introduced, which
specifies the ratio of the average leakage power to the average switching power dissipated per
cycle by a functional unit.
Having an accurate framework for assessing power-gating at a fine grain, Hu et al. examine
two policies for power gating functional units. The first is a time-based policy, inspired by cache
decay, while the second is an event-guided policy and in particular a branch-prediction guided
policy.
Functional unit decay : The first approach works similarly to cache decay. If an idle period is
detected in a functional unit then it is switched-off. The are three timing factors that determine
the behavior of this approach: the break-even point in cycles after which there are net gains in
energy; the time it takes for the functional unit to wake up from the moment it is needed; and
the decay interval, i.e., the time it takes to decide to put the functional unit in sleep mode. The
first two are technology and functional-unit specific, while the third, the decay interval, is an
architectural knob that one can turn to tune the policy.
The first timing factor, the break-even point, varies depending on the technology and
functional unit. In particular, it varies depending on the leakage factor L ,theratioof
leakage energy to dynamic energy.
13 But there can be indirect energy costs due to power-gating mistakes which reduce performance.