Optimizing Capacitance and Switching Activity to Reduce Dynamic Power - Computer Architecture Techniques for Power-Efficiency - page 119

Such a low accuracy for the estimator is disheartening for pipeline gating. Most of the time it
would stall correct execution. However, this holds for a single low-confidence branch.
If more than one low-confidence branch enters the pipeline then the chances of going
down the wrong path increase substantially. In fact, for N low-confidence branches and an
average estimator accuracy of P (for each), the probability of going down the wrong path (i.e.,
having at least one misprediction) becomes: 1

P ) N . Conveniently enough, evidence
shows that low-confidence predictions do tend to cluster together [ 88 ]. Pipeline gating is thus
engaged with more than one low-confidence branch in the pipeline—the actual number is called
gating threshold . This makes the coverage of the estimator (detecting many low-confidence
branches) more important than its accuracy because it is the number of low-confidence branches
in the pipeline that matters—not their accuracy. Manne et al. discuss several possible confidence
estimators for the gshare and the McFarling predictors, including

− (1 −

perfect (oracle) confidence estimation,
static (profiled) estimation allowing the customization of coverage versus accuracy,
Miss Distance Counter ( MCD ) estimator that independently keeps track of prediction
correctness,
for the McFarling predictor an estimator—called “both strong” —based on the agreement
of the saturating counters of the gshare and bimodal components, and
finally, for the gshare predictor a simple estimator based on the distance of a branch
from the last low-confidence branch.

Estimator details are not of much importance here, but rather the fact that different
estimators can be designed trading coverage and accuracy. Choosing the distance for gshare
and both-strong for McFarling and with a gating threshold of 2, a significant part of incorrect
execution is eliminated without any perceptible impact on performance.
To conclude this approach, one last question that needs to be addressed is the specific
pipeline stage to gate. The earlier the pipeline is gated, the more incorrect work is saved but also
the larger the penalty of stalling correct execution. This is not simply a function of the number of
pipeline stages before gating. The important factor here is the number of incorrect instructions
as we go deeper into the pipeline. Gating at the issue stage hardly saves any extraneous work
since very few incorrect instructions make it that deep in the pipeline. In contrast, the initial
stages of fetching, decoding, etc. can be full of incorrect-path instructions. With a gating
threshold of two or more, the chances of stalling correct execution are miniscule, so it pays to
gate as soon as possible (i.e., at the fetch stage).
Selective throttling : Subsequent work by Aragon, J. Gonzalez and A. Gonzalez followed
a different path. Instead of having a single mechanism to stall execution as in Manne et al.,